kondeor.ai – Manual
Everything you need to know about analysis, text analysis, reports, and your account — with examples you can try right away. Use the search on the left to jump straight to a topic, or browse chapter by chapter.
Getting Started
Registration, uploading data, and asking your first question — the basics of kondeor.ai.
Overview of the interface
Before diving in, a quick map of the controls. This shows the maximum variant (Pro plan with the Lektor/reviewer role) — depending on your plan you'll see fewer elements; the “Visible from” column shows where each one is added. Details on each function are in the linked chapter.
Header — Tools
| Element | Function | Visible from |
|---|---|---|
| Datasets | load/edit/manage saved datasets — see A4b | Basic |
| 📂 Dataset bar | above the input field: switch between several open datasets — see A4c | Pro (Free/Basic: own dataset + examples) |
| Reports | open/manage your own reports — see D1 | Pro |
| Videos | open your own video summaries — see D4 | Pro |
| Analysis plan | plan several analysis questions in advance — see B2 | Pro |
| Batch | run planned questions automatically, one after another — see B2 | Pro |
| Lektor | dedicated panel for reviewer tasks | “Lektor” role (independent of plan) |
| Courses / My courses | course catalogue or your own enrolled courses — see F1 | when enrolled in a course |
| Admin | internal administration panel | admin accounts only |
Header — Plan & Billing
| Element | Function | Visible from |
|---|---|---|
| Plan badge, credits ring | current plan resp. balance for the running 30-day window — see E1 | always |
| + Credits | buy additional credits — see E1 | always |
| Upgrade | switch plans — see E1 | Free, Basic |
| Redeem | enter an invitation code for a plan upgrade | Free, Basic |
| ✕ Cancel subscription | end the current subscription | all except Free |
Header — Account & Meta
Always visible, regardless of plan:
| Element | Function |
|---|---|
| ⛶ Fullscreen | opens kondeor.ai as a standalone page — only visible when embedded in another page — see A1 |
| Feedback | opens a pre-filled email to office@kondeor.ai to share feedback on kondeor.ai |
| Info | info panel, including the option to delete your own account — see E7 |
| AI | AI setting of your account: AI on, EU only or AI off — see A5c |
| DE/EN | language switch for the interface and answers — see E6 |
| New session | resets the chat and the uploaded dataset, without logging you out |
| Log out | ends your session immediately — see E6 |
Footer — left icon bar
| Icon (shape) | Function | Visible from |
|---|---|---|
| upward arrow into a tray | file upload (CSV/SPSS) — see A2; on the Pro plan with the choice “Create SPSS file” — see B8a | always Free: own data up to 500 rows; unlimited from Basic |
| grid/table icon (3×3 lines) | enter data or edit the dataset — see A2, A4 | always |
| small bar chart | create a chart (for reports) — see D1c | icon always visible; chart creation itself from Basis |
| three stacked chevron layers | compute/transform variables — see B4 | always |
| downward arrow into a tray | export the current dataset (SPSS, Excel with codebook or ZIP) — see B8a | always |
| checkmark in a document frame | APA statistics check against the loaded dataset — see B6 | always, once a dataset is loaded |
| question mark in a circle | method-selection guide — see A2c | always, including Free plan |
A further icon (rectangle with lines, multiple-response set syntax for SPSS) additionally appears once at least one multiple-response set has been created in the current session — otherwise it stays hidden.
Mode switch above the input field
Four modes determine how kondeor.ai treats your next message:
| Mode | Function | Visible from |
|---|---|---|
| Analysis | statistical analysis of your data — see A5 and Part B | always |
| Theory | questions about statistical concepts, independent of a specific analysis — see A6 | always |
| Power | opens the sample size calculator below the chat window — see B5 | Basic, Pro |
| Help | questions about using kondeor.ai itself — see A7 | always |
The selected mode only applies to the next message you send; kondeor.ai usually also recognises usage and theory questions automatically outside the matching mode (see A7).
Registration & Plans
You can try kondeor.ai for free without a credit card (Free plan) and upgrade to Basic or Pro at any time. During registration you agree to the Terms and Conditions and Privacy Policy — this applies even to the free plan, since it is a usage agreement as well. Both documents together, plus the Data Processing Agreement and the English version, are collected in E4.
If you already have an invitation code (e.g. for a discounted upgrade), you can enter it directly in the registration form, in the optional “Invitation code” field — redeeming it afterwards via the “Redeem” button in the header (see A0) works just as well, in case you receive the code later.
Confirming your e-mail address: After registering, kondeor.ai sends you an e-mail with a confirmation link (valid for 24 hours). You can only log in after clicking it. This also applies with an invitation code. If the e-mail does not arrive, check your spam folder or choose “Send e-mail again” on the confirmation page. If the address is not confirmed within 7 days, kondeor.ai deletes the account automatically. If you join through a course invitation, you confirm your address with the invitation link you receive by e-mail.
For prices and limits, see Part E · Plan comparison.
If kondeor.ai is embedded in another page (e.g. a course platform), a fullscreen icon (⛶) appears in the top right — it opens kondeor.ai as a standalone page in a new browser tab, outside the embedding frame.
Uploading data
Two ways to get data into kondeor.ai:
- Upload a file — CSV or SPSS (
.sav), via the upload button at the bottom left. SPSS files automatically bring variable labels, scale of measurement, and value labels; for CSV, kondeor.ai infers the scale of measurement heuristically from the values. After an SPSS upload, kondeor.ai also checks whether the stored scale levels look plausible (e.g. a variable stored as metric with only two values) and shows the note “Please check scale levels” if in doubt. Nothing is changed; you can correct the scale level in the variable view (see A4). Limit values: If a column contains censored limit values such as “>90” or “<20” (at least two, otherwise only numbers), kondeor.ai splits it automatically after the upload:<name>_wert(measured value or limit) and<name>_zens(0 = exact, 1 = right-censored, 2 = left-censored). The original column stays unchanged. The same split is available later in the transform dialog as “Split censored values”. - Enter data directly — build a small table by hand using the table button, no file needed.
If your data is a CSV and the meaning of the numbers (labels, coding) is in a separate codeplan — for example from a REDCap export — you can have a finished SPSS file created from it, see B8a.
Without your own dataset, you always have access to the practice dataset GBOO (Austrian Guest & Visitor Survey, n = 1,153) — ideal for trying out methods before working with real data.
Data quality check
Right after every upload, kondeor.ai asks whether the dataset should be checked before analysis — with three options:
- no — skip the check, start analysing right away
- yes — basic check: missing values, outliers, constant columns, negative values, case-inconsistent categories, leading/trailing whitespace, and duplicate rows. Affected variables are grouped by issue type (🔴 critical, 🟡 warning). There is also a section “Conspicuous respondents” (see below).
- strict (Pro plan and above) — runs a deeper pattern check on top of the basic one: kondeor.ai automatically detects related item groups (scales) from the variable names and checks each case for suspicious response patterns — straightlining (the same answer throughout), extreme responding (only scale endpoints), and sawtooth patterns (regular zigzagging). If a timestamp variable is detected in the dataset, response duration is factored in too (median, 10th percentile).
Conspicuous respondents (basic check, Basis plan and above)
The basic check looks not only at columns but also at individual respondents (one row = one respondent):
- Completion time (speeding): If there is a duration column (e.g. “Duration”, “Dauer”, “Bearbeitungszeit”), respondents who took less than one third of the median time are reported. The unit (seconds/minutes) does not matter. Page or question times are not used. Without a duration column the report says so.
- Missing answers per person: Reported are respondents who left at least half of the questions unanswered. Filter questions can create such gaps, so check whether they are intended.
- Near-duplicates: Reported are respondents who match an earlier respondent in at least 95 % of the compared variables. Only variables with at least three distinct values are compared (at least eight such variables are needed); people who give the same answer everywhere do not belong here.
Quality flag (ordinal) and stage 3
The button “Add quality flag to dataset” below the report creates three columns: qualitaet_flag (ordinal: 0 = no hints, 1 = slight, 2 = medium, 3 = strong), qualitaet_punkte (weighted points) and qualitaet_grund (reason, e.g. “Speeder; many missing answers”). Each type of hint counts once per person with a weight: near-duplicate 3; speeder, straightlining and many missing answers 2 each; long-string, extreme responding and sawtooth 1 each; even-odd inconsistency and multivariate outlier 0.5 each. The sum gives the level (below 2 slight, 2 to below 4 medium, 4 or more strong). Straightlining, extreme responding, sawtooth, long-string, even-odd and multivariate outlier belong to the strict check (Pro plan and above). Nothing is deleted. To exclude cases, set the subsample filter (see B3) to e.g. qualitaet_flag <= 1 (unremarkable and slight only) or == 0.
Adding attention checks and logic rules: The second button “Add attention checks and logic rules” opens a form (Basis plan and above; reopen it any time, e.g. when you notice contradictions later during analysis):
- Attention checks: choose the variable and the correct answer(s) — several are possible (e.g. “3” or “11”); known values can be clicked. Numbers are compared tolerantly (3 = 3.0). Any number of checks; each failed check counts 2 points. An empty answer is not a failure.
- Logic rules: a rule describes a contradiction (e.g. age < 16 AND employment = retired); persons it matches are counted. Conditions are built as in the subsample filter (AND/OR, comparison with another column). Each violated rule counts 1.5 points, 4 at most.
- AI suggestions (Pro plan and above, not available when AI is switched off, costs a few credits): the AI proposes possible contradictions. It only sees variable names, labels and value labels — no data. For each suggestion you see the number of hits and example rows and tick what to adopt. The evaluation itself is computed by kondeor.ai without AI.
“Preview” counts hits without changing anything. “Apply” recomputes the three columns in place (no _2 copies); an active filter is kept. If the number of rows has changed, a new data check is required.
The deep check's result shows a summary by issue type, plus how many cases are affected within each checked variable group. The complete, sortable list with row/ID is available in the results dashboard, including download as a standalone HTML report.
kondeor.ai only reports the issues — cleaning up (e.g. excluding or correcting suspicious cases) is up to you, for example via the subsample filter (see B3) or directly in the dataset editor (see A4).
Method-selection guide
Not sure which statistical method fits your question? The question-mark icon in the footer (see A0) opens its own window with two ways in:
- Guided — a question tree walks you step by step to the right method, covering the most common standard cases.
- Free filter — a facet filter across the full catalogue of 80+ methods: category, outcome variable scale, study design, and number of groups/predictors can be combined freely, with each option showing live how many methods still match. Additional constraints (covariates, clustered data, multiple outcomes at once) are available collapsed underneath.
For each method, a result card shows an explanation from either a student perspective (technical/methodological) or a business perspective (business question + what to do with it), plus a ready-made question you can insert straight into the input field.
The guide works independently of any uploaded dataset and is available on the Free plan too — it doesn't trigger any calculation or use credits itself. If it recommends a method not included in your current plan, the usual upgrade hint appears once you actually try to run it.
Related methods: One-way and two-way ANOVA, ANCOVA and linear regression describe, with grouping variables, essentially the same linear model; they differ mainly in what the output puts first (one overall test per factor, or individual coefficients against a reference group). The result card of these four methods points this out. The calculations are not identical in every respect: kondeor.ai's one-way ANOVA uses the Welch F and Games-Howell pairwise comparisons, regression uses the classical F test; with unequal group sizes the factor tests of a multi-factor analysis also depend on how the sums of squares are calculated.
See also A3 · “Why this method?” for how kondeor.ai explains the method it chose for a question you've already asked, and B7a for giving feedback on the automatic method choice.
Variable colour coding
Before running a calculation, kondeor.ai shows a confirmation card with the detected variables — the colour of each variable name shows the detected scale of measurement:
| Colour | Scale | Example |
|---|---|---|
| ● Teal | Metric | Age, income, score |
| ● Yellow | Ordinal | School grades, satisfaction scale |
| ● Purple | Categorical | Gender, place of residence, occupation |
'group' is categorical.Dependent & independent variable
For methods with a target variable (e.g. regression, t-test), the card additionally shows which variable kondeor.ai has recognised as dependent (DV) and which as independent (IV) — as their own groups above the remaining variables. This is a display for review only, not a clickable selector: if the assignment doesn't match what you meant, correct it in the chat instead (see B1) rather than clicking around on the card itself.
“Why this method?”
Below the variable list, an expandable section explains in one sentence why this particular method fits the recognised combination of variables — e.g. “Since 'gender' has exactly two categories, a logistic regression fits.” The section is expanded by default and can be collapsed if you don't need it. Meant as much for learning along the way as for double-checking.
If kondeor.ai detects two binary variables for which both a t-test (treating one variable as quasi-metric) and a chi-square test (both stay categorical) are methodologically defensible, this is made transparent right here — including the other option as an alternative, instead of silently proposing only one of the two.
Note on many predictors
If a regression method has noticeably more predictors than can be meaningfully interpreted given the available sample size (rule of thumb: at least around 10 cases per predictor), the card shows an additional note recommending you narrow the selection — e.g. by saying “only with the most important variables” (see B1).
Unclear scale of measurement
If a variable's scale of measurement can't be determined with confidence — e.g. a numeric code with few unique values that could just as well be a grouping as a genuine measurement —, a dedicated follow-up card appears instead of the usual confirmation card. For each unclear variable it shows the values actually found in the data, plus two buttons, “Treat as metric” and “Treat as ordinal/group”, each with a short pros/cons explanation. “Continue” only becomes active once a choice has been made for every affected variable — after that, your original question proceeds automatically with the choices made, no need to rephrase it.
Deciding for all variables at once: If the question concerns several variables, a row “For all:” with the same two options appears above the list. One click sets all variables together (e.g. a whole questionnaire battery on one scale); you can still change individual variables afterwards. This applies to the chat card as well as to the collected card in batch mode.
In Batch mode (see B2) the same follow-up appears as a single, consolidated card for all affected variables before the batch starts, rather than one per question.
More than two groups for a two-group method
If a method requires exactly two groups (e.g. t-test, chi-square test, Fisher's exact test, McNemar test) but the chosen variable actually has more than two distinct values in your data, kondeor.ai shows a selection card with all the values found — simply pick the two you want to compare. For methods with two categorical variables (chi-square, Fisher, McNemar), a second, identical card can follow right after for the other variable, if it also has more than two values.
Editing your dataset
Besides data entry (“New entry”), the table button also opens the “Edit dataset” tab — this lets you adjust the currently loaded dataset directly in the browser. It has three views: Data view (table), Single case (a form for one case) and Variable view.
Data view
- Click any cell to edit it
- Add rows (+ Row) or delete them via the 🗑 icon
- Add variables/columns (+ Column) or delete them — caution: deleting a variable cannot be undone
- Paste values from Excel directly into a cell with
Ctrl+V - Sort by a column (display only — the saved row order stays unchanged)
- “Show raw values” switches between value labels (e.g. “Yes”) and numeric codes (e.g. 1)
Collect changes, then save: Edited cells are highlighted but not written yet. Only “Save changes” applies them all at once — if something fails, nothing is half-applied. “Discard” resets everything. “Save as new copy” leaves the original dataset untouched and stores the edited state as a separate dataset. “Save to dataset” (for saved datasets) instead overwrites the saved dataset under the same name — so your changes are not lost the next time you open it.
Single case
The “Single case” tab shows one case at a time as a form — as known from medical data-entry systems (e.g. REDCap). This is often more comfortable than a wide table when you enter, check or correct data of individual people.
- Browse and search: previous/next, jump to a case number, or search for a value (e.g. an ID). New case adds an empty row, Delete case removes the case shown.
- A suitable input per variable: variables with value labels appear as option buttons (up to five answers) or a drop-down, metric variables as a number field, date variables as a date field. Excel/SPSS date numbers (e.g. 43544) are shown as a date (20.3.2019); the number itself stays stored.
- Sections: With “Set up sections” you group variables into your own sections (e.g. “Baseline”, “Surgery”, “Follow-up”) — tick variables (hold Shift to select a whole range), give a name, assign. Sections can be renamed or dissolved and are saved with the dataset.
- Changes in the form are collected exactly as in the data view and applied with “Save changes”; both views show the same state.
Variable view
Here you adjust variable name, label, scale of measurement and value labels. The “Variable window” button opens the variable list (name, label, scale, value labels, with search) in its own browser window, e.g. on a second screen; clicking a name copies it — handy for pasting into questions and formulas.
Defining missing values: In the “Missing values” column you enter codes that kondeor.ai should treat like empty cells — e.g. 9 or 99 for “no answer”. Separate several codes with a comma or semicolon, write ranges with a hyphen (90-99); negative codes are possible (-9, range -99--1). For SPSS files, the user-defined missing values set in SPSS are taken over automatically. The definition applies to the analyses (frequencies, statistics, tests, comparison of missing values, see B0); the cells themselves stay unchanged, nothing is deleted. For the unchanged practice dataset GBOO the usual codes are preset; as soon as you edit it, they appear in the variable view and can be adjusted there. The variable window shows them in a column of their own.
.sav), only names SPSS does not allow are adjusted automatically; the original name is kept as the variable label. Names made of letters, digits and underscores that start with a letter are the safest choice.Saving datasets permanently
After every upload, kondeor.ai offers to keep the dataset permanently in your account — independent of the 12-hour session limit, for reuse in later sessions. Click “Save dataset” in the prompt that appears; “No thanks” dismisses the offer without affecting the current session.
The “Datasets” button always takes you to your saved datasets:
- Load — brings the dataset back into the current session, as if it had just been freshly uploaded (on the Pro plan as a further place in the workspace, see A4c)
- Edit — opens it directly in the dataset editor (see A4)
- Context — add or change the questionnaire context (see below)
- Share — passes a copy of the dataset on to other people (see A4d)
- Rename and Delete
Number of datasets you can keep at once: 3 on the Basis plan, 10 on the Pro plan. If the limit is reached, “Save dataset” reports this — delete an existing dataset first to make room for a new one.
Multiple datasets in the workspace
On the Pro plan you can keep up to 10 datasets open at the same time and switch between them with one click — e.g. several studies, sites or survey waves. There is nothing to switch on: as soon as you upload a second, different file, kondeor.ai places the previous dataset in the workspace and opens a place of its own for the new file.
The dataset bar (📂 Datasets) then appears above the input field:
- Click a tile — switches to that dataset. All following questions, the data view and transformations then refer to it. Filters, transformations and your scale-level choices are kept per dataset and are back when you switch back.
- ✎ — renames the active dataset (e.g. “Wave 2024”). The name stays even if you later upload a new version of the file.
- ✕ — closes a dataset from the workspace.
- + — uploads another file.
If you upload a file that is already in the workspace (same file name, e.g. a corrected version of StudyA.csv), no new place is created; the existing one is replaced and becomes active. The same applies when loading from “My Datasets” (A4b). If 10 datasets are already open, kondeor.ai asks before the upload whether the least recently used one should be closed — nothing is removed without your confirmation.
As soon as several datasets are open, every analysis result shows the name of its dataset at the top (📂 …), so it stays clear which result belongs to which data. Buttons such as “Add to dataset” or changing the rotation of a factor analysis afterwards only work while the corresponding dataset is active — otherwise kondeor.ai points this out instead of writing values into the wrong dataset.
Examples (all plans): If you ask for an example (B0b), kondeor.ai opens a place of its own for it, 🧪 Examples — your previous dataset, including its filters, stays in the workspace, and one click on its tile takes you back. You can then also try further analyses with the example dataset. Another example replaces the content of the examples place; close it with ✕ when you no longer need it.
📚 Practice dataset: The tile “📚 Practice dataset” always sits at the far left of the bar. One click takes you back to the practice dataset (GBOO); your own datasets and examples stay in the bar. The practice dataset is always loaded in its original state — filters or changes to it are not kept, so that courses and test questions always work with the same data. If you want to keep an edited version, save it under “My Datasets” (A4b); it gets a new name and then counts as a dataset of your own.
On the Basis and Free plans, the workspace holds your own dataset and the examples place; a new upload replaces your own dataset as before, and saved datasets can be loaded one after another via “My Datasets”. There, data in the workspace is deleted after 12 hours without use — as the session was before; on the Pro plan after 30 days.
Sharing datasets, recipes and reports
You can pass saved datasets, transform recipes and reports on to other people — e.g. colleagues, supervisors or co-authors. Sharing is free of charge (no credits).
How it works
- Dataset: in the “Datasets” panel, click “Share” next to the dataset
- Recipe: select the recipe in the transform dialog and choose “Share” (see A5)
- Report: in “My reports”, click “Share” next to the report
In the dialog you enter the recipients' e-mail address(es) (up to 5), optionally write a short message, and confirm that you are entitled to share the content. When sharing a dataset or a report you can also include recipes (optional) — e.g. the data preparation that belongs to the data or report.
What recipients get
- Their own copy in their account (under “Datasets”, in the recipe dropdown, or under “My reports”, each marked “shared by …”). They can edit and use the copy freely; your original stays unchanged, and later changes to it do not reach the copy.
- An e-mail about the share. After their next login, kondeor.ai also reports newly arrived copies.
- People without an account receive an invitation. The copy is created as soon as the person has registered with exactly this e-mail address and confirmed it.
Plan, storage and revoking
- The copy counts towards the recipient's quota (e.g. number of saved datasets). If the plan or storage does not allow it yet, the share waits and appears automatically as soon as it is possible. Anyone who does not want it can decline.
- You can revoke a share as long as the copy is unchanged. Once the person has edited the copy, revoking is no longer possible. Under “Share” you can see what you have already shared.
- If you delete your account, copies already delivered stay with the recipients; open shares lapse.
Sharing is available from the Basis plan. Anyone can use the inbox (receiving and declining copies), but whether a copy can be created depends on the recipient's plan.
Asking your first question
The start assistant
On the home page an assistant guides you in three steps — no statistical vocabulary needed:
- 1 Data: Drag a file onto the area or click it (CSV or
.sav, see A2), or try the GBOO practice dataset. Afterwards you can run the data quality check (A2b). “Enter data by hand” takes you to data entry. - 2 Question: The question builder reads your variables. “Quick start” offers four ready-made suggestions; or choose a task (Describe, Compare, Associations, Predict, Scales & items, Time series, Clinical & causal) and click the matching variables. Chips are coloured by scale of measurement as in A3, the question builds itself live, and one line shows which method kondeor.ai will run. Tasks your data cannot support are dimmed and say what is missing.
- 3 Result: “Run analysis” sends the question like a typed one. “Copy to input field” lets you edit it first, “Build a new question” starts the next one.
Check before you start. The assistant warns you before an analysis runs into nothing:
- “n=…” on a variable means it is filled in for only that many cases (e.g.
n=217), because the question was shown to a subgroup only. Variables with fewer than half of the answers are highlighted in yellow. The one-click suggestions avoid such variables. - Shared cases: below the question you see for how many cases all chosen variables have a value at the same time. If it is 0 (the questions were answered by different respondents), the combination cannot be analysed — “Run analysis” is disabled. With very few cases a warning appears.
- Group sizes: for comparisons by group you see the number of cases per group (only cases with all values). Groups with fewer than 5 cases are highlighted in yellow — tests are of little value there; consider merging groups (B4).
An active subsample filter is taken into account. If you type the question straight into the chat and there are no shared cases, you get a message with the same numbers instead of an analysis.
Below the assistant, the “Other ways” line leads to the data view and editing (A4), transforming (B4), the method-selection guide (A2c), free typing in the chat and this manual. “Don't show again” hides the assistant (a note with “Reopen” stays; later always available via Info → “Open the start assistant”); examples for individual methods are requested in the chat (B0b).
An answer with a dashboard stays at the top, together with the question above it — the view does not jump to the end of the analysis. The input field sits fixed below the chat.
Phrasing questions freely
Just phrase your analysis question in plain language. For unambiguous matching, write variable names in single quotes:
- What is the median of 'age'?Descriptive statistics for a single variable
- Does 'income' differ by 'gender'?kondeor.ai automatically suggests the appropriate test (e.g. t-test or Mann-Whitney)
- Show me a complete overview of all the data.Automatic descriptive statistics across the entire dataset
- What hypothesis could I test for 'income' and 'gender'?Instead of running the analysis right away, kondeor.ai first formulates H0/H1 and suggests the appropriate method — useful when you're not yet sure which test you actually need
R syntax
If you're used to the classic R formula notation (dependent ~ independent), kondeor.ai understands that too — handy if you already think about regressions this way:
- 'grade' ~ 'gender' + 'age' + 'attendance'Linear or logistic regression, depending on the type of 'grade' — 'gender', 'age' and 'attendance' as predictors
- 'satisfaction' ~ 'gender'Also works with just one independent variable
- 'income' ~ 'education', 'age'A comma between predictors works just as well as a plus sign
If a question already names the method, dependent variable, and all independent variables fully and unambiguously in quotes (as in the examples above), kondeor.ai skips the usual confirmation card and runs the calculation right away — the recognised configuration then appears as a short, expandable summary next to the result instead of before it. If even one variable stays unclear or unquoted, the confirmation card appears as usual.
Number format (decimal places)
You set how many decimal places results show with the “0.00” button in the header (next to the language switch). There are two options:
- APA standard (default, recommended): statistics with 2 decimals, p-values with 3; p-values and standardized statistics (r, β, R², η²) without a leading zero, very small p-values as “< .001”.
- Fixed number of decimals (2 to 6): applies to all values including p-values; the leading zero stays. The calculation itself returns at least the chosen number of places — so no places are padded with zeros that were never computed.
What the setting affects — and what it does not
- It applies to new analyses in dashboards, key-figure tiles and tables. Results already shown do not change retroactively.
- APA sentences (e.g. “t(48) = 2.31, p = .025”) always stay in APA notation, whatever is set.
- The tables in the chat text show the precision of the calculation (usually 2 to 4 places). A higher fixed number makes them more precise; a lower one does not shorten them.
- Saved reports keep the numbers as they were displayed at analysis time.
- Your browser remembers the setting; it applies again after every login.
For APA-style publications the standard remains the right choice. A fixed number of decimals is useful when you compare results with other programs or want to see very small coefficients (e.g. B = 0.0225) more precisely.
Significance level (α)
With the “α” button in the header (next to the number format) you choose from which p-value on a result counts as significant: .05 (default, recommended), .01 or .10. The setting is being rolled out step by step; where the button is not yet visible, .05 applies.
- What changes: traffic lights and markers (“Significant” / “Not significant”), assumption checks and the decisions on the dashboard. Traffic lights based on a p-value show the level used (“α = .01”); traffic lights that only rate a convention (e.g. effect-size classes) show no α and the neutral pill “Result”.
- AI texts follow the level: interpretation, chat answer, overall interpretation and project synthesis call a result “significant” only if p is below the chosen α. Example at α = .01: p = .011 counts as not significant.
- Confidence intervals follow automatically: the confidence level is always 1 − α, i.e. 95 % at α = .05, 99 % at α = .01 and 90 % at α = .10. Test and interval therefore always match; labels (“99% CI”), charts and AI texts show the level used. In the descriptive dashboard the default follows α but can still be changed there individually.
- What stays the same: effect-size classes, APA stars (*, **, ***), the Bland-Altman limits of agreement (±1.96 SD) and fit criteria such as CFI ≥ .95 are fixed conventions.
- Saved reports keep the level they were calculated with; the setting applies to new analyses.
Light and dark mode
The colour-mode button in the header switches between a light and a dark appearance. Without a choice, kondeor.ai follows your operating system setting. Your choice is stored in this browser only (not in your account) and does not affect print graphics from R, which always look the same. The change is being rolled out in stages: some areas may still show dark-mode colours.
AI setting (AI on, EU only, AI off)
You decide how much AI kondeor.ai may use for you. The “AI” button is in the header (next to the language switch) once you are logged in. The setting is stored in your account and applies on all your devices.
| Level | What happens |
|---|---|
| AI on (default) | All AI functions are available. Processing runs on EU servers first; a provider outside the EU is used only with your explicit consent (see E4). |
| EU only | AI only on EU servers or our own infrastructure. No data is ever sent to providers in the USA, not even after consent. This removes: explainer videos with an avatar (HeyGen). |
| AI off | No AI is used at all, not even on EU servers or our own infrastructure. The statistics run unchanged. |
What keeps working with “AI off”
- Uploading, data view, transforming, filtering and weighting your data.
- All statistical calculations with R, including dashboards, key figures, charts and APA sentences.
- Method detection via keywords. With very loosely worded questions kondeor.ai then chooses more cautiously or asks back; with variable names in quotation marks (
'satisfaction') and the method in the text it works reliably. - The method-selection guide, the sample size calculator and the course content.
What is not available with “AI off”
- Interpretation texts for the results (the tab shows a note; the calculations stay visible).
- Analysis plans, hypothesis and research-question checks, questionnaire draft and check.
- Qualitative text analysis, batch overall interpretation and project summaries.
- AI-assisted time-series forecasting (TiRex); where provided, a classical method (ARIMA) is used instead.
- AI grading of free-text answers in course exams — instead, you see the model answer after submitting, to compare yourself.
- Explainer videos.
Good to know
- Switching takes effect immediately for new requests. Results already shown and saved reports stay unchanged.
- You can switch back to “AI on” at any time.
- With “EU only” and “AI off” the consent prompt for providers outside the EU never appears.
- When someone opens a report you have shared, the stricter of your setting and the viewer’s setting applies to AI translations; without a login, “EU only” applies.
- If an area is blocked, you see the message “AI functions are disabled for this account.” (or a note in the respective tab).
The legal bases and the service providers involved are listed in the Privacy Policy, point 3.5.
Theoretical background
Under every analysis result you'll find a “Show theoretical background” button — kondeor.ai then explains the statistical theory relevant to the specific question you asked, not just the method in general. Ask specifically about the degrees of freedom of a crosstab, for example, and you get an explanation of degrees of freedom, not just a generic chi-square introduction.
You can also ask about theory independently of a specific analysis — just ask a question about a statistical concept, without variable names in quotes:
- What's the difference between a confidence interval and a credible interval?
- When do I use Mann-Whitney instead of a t-test?
Answers are drawn from a curated knowledge base (several statistics textbooks), including a source citation at the end of every answer.
Help mode
For questions about using kondeor.ai itself — not about statistics, not about your data — there's a dedicated mode button “Help” next to Analysis/Theory. In help mode, kondeor.ai treats your messages preferentially as usage questions, even if they don't contain a clear keyword — especially useful for follow-up questions like “and how exactly do I do that?”, which would be hard to classify on their own.
- How do I compare two groups?Explains the matching question phrasing, doesn't run a calculation
- How do I filter my data down to a subset?
- How do I save a result to a report?
- Can you analyse the answers to open-ended questions?
- How do I group values together?Help mode also covers transformations — not just app usage itself, see B4
kondeor.ai also recognises most usage questions automatically outside help mode — the mode mainly helps with ambiguous follow-up questions in an ongoing conversation, where the connection to the previous message might otherwise get lost.
Exception: if your question in help mode names something that actually matches a column in your currently loaded dataset (e.g. “How do I compute X per Y?”, where X and Y are real column names), kondeor.ai recognises this as a concrete transformation request and shows a formula preview right away instead of an explanation — usually the more useful outcome anyway.
Online office hour and appointments
In addition to kondeor.ai, we regularly offer an online office hour where you can ask questions about data analysis and about using the app. It takes place via Microsoft Teams — in your browser or in the Teams app; you do not need a Teams account of your own. Participation is voluntary; camera and microphone are optional.
- Appointments on the start pageAbove “Your start in three steps” the start page shows a slim line “Sessions” with up to three upcoming appointments: date, title and, right next to it, the main action (“Reserve seat”, “Join via Teams” or “Register”). Clicking the title unfolds the description and all buttons below. Only appointments that have actually been entered appear; depending on the appointment, it applies to everyone or only to certain plans.
- Reserve seat“Reserve seat” reserves your place for that particular appointment; with several appointments you can sign up for one, several or all of them. Only then do “Join via Teams” and “Add to calendar” (.ics file) appear. This is separate from logging in to kondeor.ai: first you log in, then you reserve your seat.
- Release seat“Release seat” withdraws your reservation at any time.
- Hide“Hide this session” (in the unfolded details) removes just that appointment; “Hide” at the end of the line hides the whole line (appointments and roadmap) and shows a note with “Show again” and “OK”. Your browser remembers this. You can always find the appointments and Teams links under Info → “Next session”; one click there brings back hidden appointments and unfolds the first one.
- Events by other organisersOccasionally we also point to events organised by someone else. These appointments have no “Reserve seat” and no Teams link from us: the button (usually “Details and registration”) takes you to the organiser's page, where you register. “Add to calendar” and “Hide” are available here too.
- Reminder e-mailPeople with a reserved seat receive an e-mail with the Teams link before the appointment.
No account yet? The free Free plan is enough to take part: register and log in first — a note about the appointment is also shown on the login and registration pages and on the website. Reserving a seat lets us plan the number of participants and send you the link.
Which data is processed and how long it is kept (30 days after the appointment) is described in the privacy policy — see E4.
From research goal to report (roadmap)
In the slim line above “Your start in three steps”, next to the appointments, you find the entry “Roadmap: step n of 7”. It opens “From research goal to report” with seven phases: the whole path of an analysis and the tools kondeor.ai offers for each step. The roadmap gives you a short explanation for every phase and buttons that open the matching tool. There is no fixed order: start wherever you are — if you already have data, go straight to step 6.
- 1 Research question and hypothesesFour sub-steps from topic to method: 1a check the research question, 1b check the hypotheses, 1c draw the hypothesis model, 1d choose the method and plan the analysis (guide, analysis plan). Each sub-step has its own button and can be used on its own, see B6a.
- 2 Plan the sampleThe sample-size calculator shows in advance how many responses you need for a meaningful analysis.
- 3 Develop and check the questionnaireQuestionnaire draft from your hypotheses, check of items, scales and order. The survey itself runs in your survey tool (e.g. LimeSurvey). See B8b.
- 4 Collect the dataUpload a file (CSV or SPSS), enter data by hand, or create an SPSS file from study data and a code plan.
- 5 Check and prepare the dataView and edit values, compute or recode variables, select subgroups, merge files, weight if needed.
- 6 AnalyseWith the start assistant (“Your start in three steps”), freely in the chat, or as a batch with several analyses.
- 7 ReportSave results as a report and create charts for your work.
The entry detects your progress automatically: without a loaded dataset it shows step 1, with data step 6. After the first answer in the chat the entry disappears from the line; the roadmap stays available via Info. You do not have to tick anything off. At the bottom of the window the roadmap points to the next office hour if one is scheduled: there you can ask questions about any step. Which buttons you see in the roadmap depends on your plan; tools your plan does not include are left out.
“Hide” at the end of the line removes the line (appointments and roadmap) from the start page. Your browser remembers this, also after your next login. You can still find the roadmap under Info → “From research goal to report” and show the line again there. In another browser, or after clearing browser data, it appears again.
Quantitative Analysis
Statistical analysis: from your first question to filters, transformations, and power analysis.
Methods at a glance
kondeor.ai covers over 100 statistical methods, organised into the following analysis families:
| Family | Example methods | Plan |
|---|---|---|
| Descriptive statistics | Mean, spread, distribution shape, normality table (Shapiro-Wilk, Anderson-Darling, Lilliefors, see B1), outlier diagnostics, Table 1 (baseline table by group, see B9g), Kaplan-Meier estimate for censored measurements (see B9g) | |
| Group comparisons | t-test, ANOVA (one-, two-, three-way, repeated measures, mixed), ANCOVA, Mann-Whitney-U, Kruskal-Wallis, Wilcoxon, Friedman, MANOVA, MANCOVA | |
| Proportion tests | Chi-square (with odds ratio, relative risk, risk difference, NNT), Fisher's exact test, McNemar, Cochran-Mantel-Haenszel (stratified 2×2 tables, see B9g), Cochran's Q | |
| Association & correlation | Pearson, Spearman, Kendall's tau, partial correlation, point-biserial correlation, Kendall's W | |
| Prediction & regression | Linear, logistic, Poisson, negative binomial, gamma; ordinal, multinomial, zero-inflated, beta, robust, quantile regression; Tobit regression for censored measurements (see B9g); regularisation (LASSO/ridge) | |
| Multivariate methods | Factor analysis (also as principal component analysis, PCA), cluster analysis, latent class analysis, discriminant analysis | |
| Causal analysis & structural models | Mediation (also moderated/serial), moderation, multilevel analysis (mixed models), SEM, CFA, growth curve models, propensity score matching, propensity score weighting (IPTW, see B9b) | |
| Equivalence and non-inferiority testing | TOST (independent groups, paired measurements, sample vs. reference value, proportions, correlation); non-inferiority (independent groups, proportions) | |
| Missing data | Multiple Imputation (t-test/correlation/regression, with selectable auxiliary variables, see B9); also replacement (mean/median/mode) and FIML (SEM/CFA); comparison of the missing rates of several variables (McNemar or Cochran's Q, see B0) | |
| Bayesian methods | Bayesian t-test/ANOVA/correlation, Bayes factor | |
| Scale analysis & reliability | Cronbach's alpha (for ordinal items also ordinal alpha and omega on polychoric correlations), McDonald's omega, intraclass correlation, interrater agreement (Cohen's/Fleiss' kappa, see B9b), item discrimination | |
| Medical statistics | Kaplan-Meier, Cox regression, competing risks (cumulative incidence, Fine-Gray, see B9g), ROC/AUC, sensitivity/specificity, Bland-Altman, Passing-Bablok regression | |
| Machine learning | Decision tree, random forest, conditional inference tree, gradient boosting, SVM, KNN, naive Bayes, neural networks | |
| Time series analysis | ARIMA, forecasts with confidence interval | |
| AI forecast (TiRex-2) | Zero-shot foundation model, complements ARIMA (limited early access) | |
| Exponential Smoothing (ETS) | Automatic error/trend/season model selection, standalone alternative to ARIMA | |
| Structural break / changepoint detection | Chow test (known point in time), Bai-Perron multiple breakpoint search, segment trends (growth/decline) | |
| Outlier / anomaly detection (time series) | STL/loess remainder + IQR-robust threshold, accounts for seasonality | |
| CUSUM test (parameter stability) | OLS-CUSUM (Brown-Durbin-Evans), no known break point required | |
| STL decomposition (trend/season/remainder) | Seasonally-adjusted series as input for structural break/CUSUM, trend/seasonal strength | |
| Meta-analysis | Random-effects model (REML, Knapp-Hartung) or, on request, fixed-effects model, prediction interval, forest & funnel plot, Egger's test, trim-and-fill — input as a study table, see B9f | |
| Data weighting | Frequency, analytical and design weights, weighted regression, design effect; a design weight can also be set permanently (chat and batch then calculate weighted) — see B3b | |
| Method comparison | e.g. t-test vs. Mann-Whitney-U — recommendation based on your data |
FFree BBasis PPro — from which plan the methods of a family are available (several marks: the family spans several plans). Basis includes everything from Free, Pro everything from Basis; Basis roughly matches what bachelor's programmes teach, Pro offers the advanced methods. Every single method with its mark: method overview by plan; a plain-language explanation of each method: method-selection guide on the assistant page (opens directly in English via the page's own ?lang=en parameter — the page chrome around the widget stays German, but the guide itself loads in English).
Descriptive multi-variable comparison: When several variables are named at once (in quotes or as a range, e.g. 'mpreis' to 'mshop'), kondeor.ai automatically recognises a comparison question from phrasings such as “highest”/“lowest” or “the highest agreement”, and instead of separate figures for each variable returns a sorted comparison table plus a bar chart (including a 95% confidence interval as error bars, when the requested statistic is the mean). You can ask for the mean, median, standard deviation, width of the confidence interval, or the lower/upper quartile — kondeor.ai works out which statistic you mean from the wording of the question (without any AI call for explicit terms like “standard deviation”, via a brief AI-assisted lookup for content-based phrasing like “satisfaction”).
- 'mpreis' to 'mshop': where is satisfaction highest?Comparison across eight variables, sorted by mean, with bar chart
Pairwise significance matrix: As a natural follow-up to such a comparison, you can also ask which of the compared variables differ from each other in a statistically significant way — for phrasings like “which ones differ significantly?” or “is there a significant difference?”, kondeor.ai runs paired t-tests for every pair of variables and displays the result as a half-matrix (Holm-adjusted p-value per cell, highlighted where p < .05). Since each pair only uses respondents who answered both variables involved, the sample size can differ between cells — it is shown per cell.
- Which of the values from 'wkult' to 'wfam' differ significantly?Pairwise comparison matrix with Holm correction
Descriptive statistics: one dashboard per variable. Each variable gets its own dashboard, organised into location, spread, shape and frequency table. What is shown depends on the scale of measurement (the one you set in the confirmation or in the variable view): metric shows everything (mean, spread, distribution shape, frequency table up to 20 values), ordinal shows median, quartiles, interquartile range and the frequency table with value labels (mean, standard deviation and shape can be expanded), nominal shows the mode and the frequency table. For metric variables you can also expand: further location measures (trimmed and winsorised means, quadratic mean, sum), percentiles (P5, deciles, P95) and the coefficient of variation. The coefficient of variation only makes sense on a ratio scale with a true zero; for a mean ≤ 0 it is not shown, and for rating scales only on request.
Choosing the confidence level: In the descriptive dashboard you can change the level of the confidence interval (90, 95, 99 % or a custom value between 80 and 99.9 %). The default is the level 1 − α (see A5d), i.e. 95 % at α = .05; here it can be changed independently. The statistics are recalculated for the same cases. For ordinal and nominal variables the intervals of the proportions are called “credible intervals” (Bayesian); they can also be expanded on request.
Frequencies of several variables as a table: If you ask about several categorical variables at once (e.g. “How many respondents visited the place, how many the museum?”), a comparison table appears before the individual analyses: one row per variable, one column per category, each cell as “n (x %)”. The percentages refer to the valid cases of each variable (“n valid”), and the columns follow the order of the value labels. A “—” means the category does not occur for this variable. The table appears for up to eight categories.
Statistics of several variables as a table: If you ask for mean, median or mode of several variables, a table with one row per variable appears instead of running text. A “—” means that the statistic is not provided for the variable's scale of measurement (e.g. no mean for ordinal variables).
Comparing missing values (from Basis): With a question like the following you see, per variable, how many values are missing and how many are valid, and whether the missing rates differ.
- Does the number of missing values differ between 'fbort' and 'urlart'?Table “missing n (%) / valid n (%)” plus a McNemar test on the missing indicators of the same cases
- Compare the missing values of 'q1', 'q2' and 'q3'.For three or more variables Cochran's Q with pairwise exact follow-up tests (Holm correction)
At the top is the table with the shares per variable and the number of cases with at least one missing value, below it the test including its direction (“variable A is missing more often than variable B”). Empty cells and the missing codes defined in the variable view count as missing (see A4). For the question to be recognised it needs the word “missing”, a comparison word (e.g. “differ”, “compare”, “more often”) and at least two variables in quotes; recognition currently reacts to German wording, so in English prefer phrases containing “missing” plus “differ” or “compare”. The test says nothing about why values are missing or whether they are missing at random; imputation and tests for missing at random are separate procedures (see B9).
Scale recognition & comparable studies (demo)
After upload, kondeor.ai automatically checks whether variable names match known, standardised questionnaire instruments (e.g. based on an abbreviation contained in the name, such as IPSS or FSFI). A match appears as its own chat message right after upload, with the recognised variables and a source reference for the instrument.
For some analysis methods, kondeor.ai additionally adds a brief note on which other studies known from the literature used the same statistical methodology — as context for orientation, not as an evaluation of your own result.
Both hints are deliberately labelled “🔬 Demo”: the underlying knowledge base is continuously expanded with further fields and instruments, but doesn't yet cover every topic.
Trying out example analyses
You don't need your own data to get to know kondeor.ai: just ask in chat for an example of a specific method (e.g. “Show me an example of a factor analysis”) — kondeor.ai automatically loads a practice dataset and runs the analysis live.
| Area | Sample question |
|---|---|
| Descriptive Statistics | “Show me an example of a descriptive statistic” |
| Proportion Tests | “Show me an example of a chi-square test” |
| Group Comparisons | “Show me an example of a t-test” |
| Association & Correlation | “Show me an example of a correlation” |
| Prediction & Regression | “Show me an example of a linear regression” |
| Multivariate Methods | “Show me an example of a factor analysis” |
| Causal Analysis & Structural Models | “Show me an example of a mediation analysis” |
| Scale Analysis & Reliability | “Show me an example of Cronbach's alpha” |
| Bayesian Methods | “Show me an example of a Bayesian test” |
| Medical Statistics | “Show me an example of a Kaplan-Meier analysis” |
| Machine Learning | “Show me an example of a random forest” |
| Time Series Analysis | “Show me an example of a time series analysis” |
| Meta-Analysis | “Show me an example of a meta-analysis” |
| Data Weighting | “Show me an example of an analytical weighting” |
You can also just ask “Which example analyses can you show me?” — kondeor.ai will then show you exactly this list in chat. Name any other method directly to check whether an example exists for it too.
The example opens as a dataset of its own, 🧪 Examples, in the dataset bar above the input field (A4c) — an own dataset you have already loaded is not replaced but stays alongside. You can then also try further analyses with the example data; the results are marked with 🧪 and are based on synthetic, not real data.
The three examples suggested on the homepage (t-test, correlation, linear regression) run free of charge. All other example analyses in this list are computed live and billed with credits like a regular analysis.
Follow-up questions & changing topics
After an analysis, you can directly dig deeper into the results — just keep going, kondeor.ai automatically refers back to the last analysis:
- What hypotheses can be derived from these results?
- Which group deviates the most?
- Are the assumptions met?After a linear regression: checks the model assumptions, see B1a
To switch to a new topic (same data), simply ask the new question directly — if there's a misunderstanding, an explicit hint helps:
- I mean a new question – please ignore the last analysis.
Correcting the last analysis
If the last suggested or already-run analysis isn't quite right, you don't need to rephrase the whole question — a short, targeted correction in the chat is enough. This works both for a still-open confirmation card and for an already-finished result:
- Change 'age' as dependent to 'gender'.Swaps the target variable — the previous DV automatically becomes available as a possible independent variable. If the previous method no longer fits the new variable type, kondeor.ai switches automatically (see “Why this method?”, A3)
- Without 'age'.Removes one or more variables from the analysis, the rest stays as is
- Also with 'gender'.Adds an extra variable without removing the existing ones — can be combined with “without” too
- Only with 'age' and 'gender'.Replaces the previous selection entirely with a new, closed list
- 'income' instead of 'age'.Replaces one specific variable with another
If a correction doesn't clearly relate to a question you've already asked (e.g. if you asked about something else entirely in between), kondeor.ai generally recognises that on its own and treats your message as a new, standalone question instead.
Distribution check for t-test and ANOVA (robustness check)
Before a group comparison (t-test/Mann-Whitney U, one-way ANOVA/Kruskal-Wallis), kondeor.ai checks the distribution shape in each group separately — skewness, excess kurtosis and Shapiro-Wilk; the groups are not pooled. The result determines how the result is presented:
- Unremarkable: You get the main result as usual. Below it is a “Robustness check” line: kondeor.ai has also calculated the distribution-free test in the background and reports whether the conclusion is the same. If the other test would reach a different conclusion, the method comparison is shown instead.
- Remarkable or strongly remarkable: Instead of calculating straight away, kondeor.ai shows a method comparison with a recommendation and a group table (sample size, skewness, excess kurtosis, Shapiro-Wilk per group); you choose which test is run.
For the independent t-test and ANOVA the result also contains a normality line per group (Shapiro-Wilk), supplemented by a description of the histogram and Q-Q plot in words (number of peaks, share beyond ±3 standard deviations, Q-Q pattern, deviation of the ends). The same is provided for the one-sample t-test, Mann-Whitney U, Kruskal-Wallis and the residuals of linear regression. All numbers come from R. On the Free plan, in batch mode and for weighted analyses the method choice works as before, without these stages.
Normality table (Shapiro-Wilk, Anderson-Darling, Lilliefors)
Ask directly about the distribution of one or more variables — kondeor.ai shows a table with three normality tests side by side: Shapiro-Wilk, Anderson-Darling and Lilliefors (Kolmogorov-Smirnov with estimated parameters). Each variable gets one row with sample size, mean, standard deviation, skewness, excess kurtosis, the three p-values and the number of tests that reject normality; a Q-Q plot is included.
- Normality test for 'score'.One row with all three tests
- Check 'score' and 'duration' for normality, by group 'gender'.One row per variable and group (exactly one grouping variable)
After a t-test, one-sample t-test, ANOVA, ANCOVA, Pearson or partial correlation, a follow-up chip below the result checks the normality assumption of the variables involved with exactly this table; for the t-test and one-way ANOVA optionally per group. For linear regression the chip checks the regression assumptions (see B1a). Paired and repeated-measures procedures have no chip: the assumption applies there to the differences or residuals, not to the single variables.
Read the tests together with the Q-Q plot, skewness and excess kurtosis, not in isolation: with large samples all three tests become significant for trivial deviations, with small samples they also miss clear ones. Missing and user-defined missing values are excluded. The normality table also runs in batch mode.
Regression: checking the assumptions
After a linear regression you can have the model assumptions checked directly. Just ask a follow-up question (see B1) — kondeor.ai automatically refers back to the regression model you ran last:
- Are the assumptions met?
- Residual analysis
The result is a dashboard with a traffic light (“Model assumptions met” or “Assumption violation(s) detected”) and the individual checks. The numbering follows the Gauss-Markov assumptions:
| Check | Method | How to read it |
|---|---|---|
| 1. Linearity | Ramsey RESET: adds the fitted values squared and cubed as extra terms and tests whether they explain additional variance | p ≥ .05 supports a sufficiently linear model |
| 2. Random sample | not testable from the data (depends on the sampling design) | shown as “not testable” |
| 3. Multicollinearity | Variance Inflation Factor (VIF) per predictor; not applicable with a single predictor | largest VIF below 5 is considered unproblematic |
| 4. Exogeneity | not testable from the data (theory: no omitted variables, no reverse causality) | shown as “not testable” |
| 5. Homoscedasticity | Breusch-Pagan test (Koenker variant, against the fitted values) | p ≥ .05 supports constant residual variance |
| 6. Independence of residuals | Durbin-Watson test, see below | p ≥ .05 argues against autocorrelation |
| Extra: normality | Shapiro-Wilk test of the residuals (skipped above 5,000 cases), plus skewness and excess kurtosis | not a Gauss-Markov assumption itself, only important for exact t-/F-tests in small samples; feeds into the traffic light only at n ≤ 30, above that information only (ⓘ), see below |
| Extra: randomness of the order | Runs test (Wald-Wolfowitz around the median) on the residuals in the order of the data rows; plus a plot of residuals against order | p ≥ .05 supports a random sequence of signs; too few runs point to blocks (e.g. a trend over time); does not feed into the traffic light |
| Extra: influential cases | Cook's D: number of cases above the common threshold 4/n | does not feed into the traffic light, but worth a look |
The check needs at least 10 complete cases (valid values on the dependent and all independent variables). The result of the linear regression itself also shows a response plot (observed against predicted values, diagonal = ideal prediction, R² below): the closer the points lie to the diagonal, the better the model fits. The runs test and the order plot assume that the data rows have a meaningful order (e.g. chronological); for arbitrarily sorted data the result is not meaningful.
While a weighting is active, these classic tests do not run; instead kondeor.ai explains what matters for a weighted regression (residual plot, influential cases).
Normality of residuals in larger samples
Estimating the coefficients does not require normally distributed residuals; normality matters mainly for exact t- and F-tests in small samples. In larger samples the central limit theorem usually carries the inference. At the same time, formal normality tests such as Shapiro-Wilk become significant for even trivial deviations when n is large. kondeor.ai therefore works as follows:
- n ≤ 30: A significant deviation counts as a violation (✗) and turns the traffic light.
- n > 30: The row stays visible but does not trigger a violation (ⓘ deviation). Judge the size of the deviation with the Q-Q plot and with skewness and excess kurtosis (both 0 for a normal distribution): if the points in the Q-Q plot stay close to the line and skewness and kurtosis are moderate, the deviation is usually unimportant for inference; strong skewness or single outliers deserve a closer look.
The limit of 30 is a rule of thumb about the central limit theorem; depending on the distribution, “large enough” can require considerably more cases.
Durbin-Watson: order and p-value
The Durbin-Watson test checks whether consecutive residuals are related. The statistic lies between 0 and 4: values around 2 argue against autocorrelation, values clearly below point to positive and values clearly above to negative autocorrelation.
- The order of the records matters. The test compares neighbouring rows. The value therefore depends on the order in which the cases appear in the dataset — if you re-sort, it changes (statistic and p-value). The test is only meaningful if the order carries meaning, such as time series, measurement series or a fixed order of data collection. For survey data with an arbitrary row order the result says little.
- The p-value is computed analytically. kondeor.ai uses the exact test from the R package
lmtest(two-sided). The same data in the same order always give the same p-value. Programs that determine the p-value by simulation (resampling) can show a slightly different p-value for the same statistic, and it may change from run to run there — for example jamovi (as of October 2026). - Comparing with other programs. The statistic itself is identical for the same data in the same order. If it differs, first check that both programs use the same cases in the same order (missing values, filters, sorting).
Analysis plan & batch mode
For structured work (e.g. a prepared survey analysis) you can save multiple analysis questions as an analysis plan and run them all at once in batch mode — both available from the Pro plan.
Generating an analysis plan automatically
Via the “Analysis Plan” toolbar button, describe your research question in your own words — kondeor.ai proposes a complete plan from it, each individual question already in the correct kondeor.ai format:
- Context questions (only if needed): Before the plan is created, kondeor.ai may ask one to three questions that concern the plan as a whole — e.g. whether the study is more exploratory or hypothesis-driven. Not every research question needs this; often the plan is generated right away.
- The plan itself: Each proposed analysis appears as its own card (title, question phrasing, H0/H1, a short justification for why this particular method fits the research question, and a confidence estimate). If kondeor.ai is unsure about one specific analysis — e.g. which of two plausible variables is meant —, a clickable choice appears directly on that card; the analysis is only finalised after your decision, the rest of the plan is unaffected.
- Available, but not in the batch: Text analysis of open answers and multiple imputation do not run in the batch (follow-up questions and background processing, or a button under the result). If one of them fits the research question, the plan lists it in a separate block “Available in kondeor.ai, but not part of the batch” — for text analysis with a ready-made question to copy into the input field, for multiple imputation with a pointer to the button. Methods kondeor.ai cannot do appear separately under “Methodologically useful here, but not (yet) available in kondeor.ai”.
Use “⚡ Load into batch” to move the finished plan directly into batch mode, or save it as a file (.md) via “Save analysis plan”.
Further analyses from comparable studies (optional). If kondeor.ai holds comparable studies on your topic (only the case for selected topics), an offer with the number of those studies appears below the plan. “Show suggestions” (the price is shown on the button) loads a separate, collapsible section “Further analyses from comparable studies”: methods that were used there and that your plan does not yet cover. You are charged only if suggestions actually appear; if nothing suitable comes up, it costs nothing. The plan itself still starts from your data and your research question; the suggestions are an addition you can take or leave.
- The hypotheses of the suggestions are worded neutrally and do not anticipate any findings from the studies.
- Each suggestion shows the number of studies, how the method was used there, and the sources (authors, year, DOI link).
- “Add to the plan” inserts the analysis as supplementary, that is, as an exploratory analysis.
The studies are a small, selected collection and not a complete literature review. They show what is common, not what is correct.
Carry over to hypotheses. “Carry over to hypotheses” (in the analysis plan and the batch window, from Pro) moves the H0/H1 of the analyses to step ② of the hypotheses dialog (B6a). You choose H1 (default) or H0 for all rows and can switch or deselect single rows. Because plan and batch come from the data, kondeor.ai ticks “Data already exist”: the results count as exploratory.
Results are exploratory: A plan that kondeor.ai derives from your data yields discoveries, not assumptions fixed in advance (what that means is explained in B6a). A collapsible note above the plan points this out; the same thought appears briefly below the batch's overall interpretation and is appended to the report when you save it. The note also appears for follow-up analyses and for hypotheses with “Data already exist” ticked.
Batch mode
The batch window holds one analysis question per line — either taken from the analysis plan or typed manually:
- Compare 'age' between 'gender' = 0 and 'gender' = 1
Compare 'satisfaction' between 'gender' = 0 and 'gender' = 1
Correlation between 'age' and 'satisfaction'Several questions in the batch field, one per line — kondeor.ai works through them one by one
kondeor.ai collects the individual results and, on request, generates an AI-assisted overall interpretation across all analyses — with four fixed sections (key findings, non-significant results, overall picture, methodological notes) plus a separate hypothesis summary: for every analysis with H0/H1, it states whether the hypothesis is confirmed, not confirmed, or ambiguous — including catching the case “significant, but in the wrong direction”, which a plain p-value check would otherwise wave through as “confirmed”. If the batch contains more than one analysis with a p-value, the table also shows a Holm-Bonferroni-adjusted column — it accounts for the fact that running several tests at once increases the chance of a purely coincidental significant result. If the adjusted result diverges from the single test (only possible in one direction: significant alone, no longer significant after correction), this is flagged as its own category — both in the table and in the hypothesis summary.
After the batch run, “🔬 Suggest follow-up analyses” becomes available: kondeor.ai looks at the actual results (not just the original research question) and proposes sensible follow-up analyses from them — in the same card format as the original plan (title, question phrasing, H0/H1, justification, confidence), including a clickable choice where it's unsure. The suggestions appear directly in the batch window below the results (if the window is closed, as a card in the chat). “Load into batch” takes them over into the input field — there you can delete lines before you start. During the current session kondeor.ai remembers which analyses were started in a batch and which suggestions it has already shown, and does not suggest them again when you call it once more; if nothing genuinely new remains, it says so. Not covered are analyses you ran individually in the chat, and anything from before a page reload.
Methods with a confirmation card in the batch: equivalence and non-inferiority tests, SEM/CFA, serial and moderated mediation and growth curve models run without a card in the batch. If all required details are in the question, it is calculated directly and the output states the recognised assignment as “not confirmed”. If something is missing (e.g. margin, direction, test group or event), the question is not run and marked with a message — kondeor.ai never guesses these details. In the analysis plan, kondeor.ai only proposes these methods if the hypothesis supplies the details; otherwise it picks the matching group comparison. You can then start the analysis individually in the chat. Tobit regression follows the same principle: it runs in the batch if the question names the variable with the censored values and the predictors and the censoring can be read from the data (value label or text such as “>90”); you define the censoring itself in the single chat on the card (see B9g). The Kaplan-Meier estimate for censored measurements follows the same rule: the question names the variable (a group variable is optional) and the censoring can be read from the data (value label, text such as “>90” or a column “<name>_zens”).
Saving as a report: After the overall interpretation there are three ways. “Save summary to report” (right below the summary in the chat) saves only the text of the summary as its own entry. “Save as report” (in the batch dialog) creates a report with the overview table and the overall interpretation as text. “Save entire batch analysis” — in the batch dialog and also right below the summary in the chat — additionally saves every single analysis as its own entry — in the report you can then move them by drag & drop, delete them and add a note (H0/H1 are taken over as the note). The overview table and the overall interpretation can be saved as their own entries too. The overall interpretation can also be set directly as the report’s final interpretation (no further AI call): for a new report, or a report without a final text, this is preselected and the interpretation is then not additionally saved as its own entry; if the report already has a final interpretation (e.g. with a follow-up batch), the option is off and the overall interpretation is added to the report as an entry — you can change both in the dialog. If you later choose “Generate with Claude” for the final interpretation in the report, the synthesis builds on the saved batch interpretations and on any existing final text (Holm correction and methodological notes are carried over) — worthwhile mainly when a report combines several batches and individual analyses; after a single, completely saved batch you do not need it. The target is a new or an existing report. Saving all analyses costs a flat price, regardless of the number of analyses (shown in the dialog); if you only need certain analyses, save them individually — that stays free. For “Suggest follow-up analyses”, the report of the first batch is preselected by default when saving; you can change it. Results without a dashboard (e.g. plain text answers) are not saved as an entry and appear only in the overview table.
'opmeth' — this prevents mix-ups with similarly named variables.Subset filters
You can restrict an analysis to a subset of your data without changing the dataset itself — the filter is set per session and can be viewed or removed at any time. Supported:
- Restrict the analysis to 'age' > 40 AND 'gender' = 1Multiple conditions, combined with AND
- Restrict the analysis to 'holiday_type' = 1 OR 'holiday_type' = 3Multiple conditions, combined with OR
- Restrict the analysis to cases without a missing value in 'satisfaction'Missing/not missing as its own condition, no value needed at all
- Restrict the analysis to cases where 'income' is greater than 'expenses'Cross-column comparison instead of a fixed value
In the data view, filtered-out rows are subtly highlighted in red instead of simply disappearing — so you always keep track of exactly what the filter excludes.
The filter applies to all analyses — including equivalence and non-inferiority tests, SEM/CFA, mediation, growth curve models, multiple imputation and the APA check against the dataset. When a filter is active, the output states the row count of the subsample (“n of N rows”), so a forgotten filter is noticed immediately.
Directly in the question, in plain language: For a restriction that should only apply to a single question — e.g. within a batch hypothesis or a subsample hypothesis generated by the hypothesis model — a plain-language phrase directly in the question text is enough, no quotes needed:
- Only among women: satisfaction with the museum shop affects overall satisfaction with the museum visit.“only among/for”, “restricted to”, “where…”, “among those…” trigger the detection
- Where age is over 30, satisfaction is higher.Also works for comparison/numeric restrictions, as long as a signal phrase (“where”, “among those”, etc.) is present
This restriction applies only to that one question — unlike the session-wide filter described above, it does not carry over to later questions. If kondeor.ai can't confidently match the phrase to a variable, it asks (a card with a text field in chat, a summary card before the run in batch) instead of silently continuing without a filter.
'holiday_type' = 1 (above) is not a subset filter but a permanent dataset transformation (needs its own confirmation, see B4) — for “only among women”-style restrictions scoped to a single question, you don't need this syntax.Analysis per group (group mode)
Group mode runs every analysis separately for each group of a variable — e.g. the same regression for every branch, country or treatment group. This corresponds to “Split File” in SPSS. The dataset itself stays unchanged.
Set it once, then work as usual
- Open Transform and pick the grouping variable in the “Analysis per group” section. kondeor.ai counts the groups and shows how many are large enough. You can exclude small groups (minimum size adjustable) and individual groups.
- From then on the footer shows a note with ✕ — while it is there, the mode applies. Click ✕ to end it.
- Ask your questions in the chat as usual or start a batch. Every question is run per group.
An explicit phrase in the question (“per branch”, “separately by gender”) takes precedence over the configured mode and also works without it.
Confirmation and cost
In the chat, every analysis question first shows a confirmation card with the number of groups and the credit cost. There you can also choose “Run once, pooled for all” without ending the mode. In a batch, a prompt before the start shows “questions × groups = runs” and the minimum cost. The balance is checked before each group; if it runs out, the run stops and delivers what is finished.
Result
- A comparison table per question with one row per group (Holm correction per question across several tests) and an AI summary across all groups.
- After the first successful group the method is pinned, so all groups use the same method and variables. A note names it.
- Skipped groups (too small, errors) are listed with the reason.
In recipes
Like a subset filter, group mode can be saved as part of a Transform recipe (see B4). Loading the recipe sets it up again; if the variable is missing in the current dataset, this is flagged visibly.
Weighting (design weight)
A weighting adjusts the sample to the population (post-stratification, scaling up): a case with a weight above 1 counts more in the analysis, a case with a weight below 1 counts less — the mean of the weights is usually around 1. You set the weighting variable once, and from then on chat and batch calculate weighted. The dataset itself stays unchanged.
Set it once, then work as usual
- Open Transform and pick the weight variable in the “Weighting (design weight)” section. Numeric variables with only positive values are offered.
- “Turn on” activates the weighting. The footer shows a note with ✕ — while it is there, the weighting applies. Click ✕ (or “Turn off” in the dialog) to end it.
- Ask your questions in the chat as usual or start a batch. If you switch the dataset, the weighting ends.
Which methods calculate weighted
| Method | Calculated weighted |
|---|---|
| Descriptive statistics | mean, standard deviation, quartiles and frequencies with design-based standard errors and confidence intervals |
| Crosstab / chi-square | weighted table, Rao-Scott test (F-test), Cramér's V |
| t-test, one-way ANOVA | design-based tests; group means, difference with confidence interval, Cohen's d or η² |
| Mann-Whitney, Kruskal-Wallis | design-based rank tests |
| Linear regression | weighted least squares; coefficients with standard error, 95% confidence interval and beta, Wald F-test of the whole model, weighted R²; predicted values and residuals can be added to the dataset |
What changes — and what does not
- Estimates (means, proportions, coefficients) are weighted.
- Standard errors, confidence intervals and tests take into account that unequal weights cost information: the sample is “effectively” smaller than the actual number of cases. The result shows the effective sample size and the design effect.
- The number of cases stays the real one — weights do not multiply cases. Rows without a valid weight (empty or not positive) are excluded and reported in the result.
Why p-values can be higher than elsewhere
Programs that treat weights as known frequencies or measurement precisions calculate as if all cases were equally informative — standard errors then come out too small and p-values too low. kondeor.ai treats the weights as part of the sampling design: unequal weights cost information (the sample is effectively smaller than the actual number of cases), and this shows up in wider standard errors, larger p-values and wider confidence intervals. The estimates themselves agree; only the uncertainty is reported more honestly.
Regression: assumptions
If you ask “Are the assumptions met?” while a weighting is active, kondeor.ai does not run the classic assumption tests: on unweighted residuals they would check a different model than the one you interpret. You get an explanation instead of what matters with design-based standard errors: linearity and influential cases (both visible in the residual plot of the regression dashboard), no perfect multicollinearity, and independence of observations as a question of the sampling design. No longer a prerequisite are normality and equal variance of the residuals — design-based standard errors, confidence intervals and tests are robust against both. If you want the classic tests, turn the weighting off for that analysis.
Charts, recipes, cost
- Charts: The chart dialog proposes the active weighting automatically for bar, line, pie, mean and histogram charts (see D1c). Box plots and scatter plots do not support weights; they are drawn unweighted, and a note says so explicitly.
- Recipes: Like a subset filter and group mode, the weighting can be saved as part of a Transform recipe (see B4). Loading the recipe preselects the variable in the “Weighting” section; you turn it on with “Turn on”. If the variable is missing in the current dataset, a note is shown.
- Cost: A weighted analysis costs the price of the respective method, possibly plus a surcharge for weighting. You see the exact values in the price overview (E1).
Transforming variables
Compute new variables, build categories, or recode values — either via the transform dialog or directly in plain language:
- Log-transform 'income' and build 3 groups from it.Multiple chained operations in a single request are supported
- Rank 'test_score', separately within each 'class'.Rank cases — smallest value = rank 1, ties get the average rank
- Split the censored values in 'gfr' into value and indicator.Values such as “>90” become a value column and an indicator column — preparation for Tobit regression, see B9g
Simple calculations work the same way, e.g. “Calculate the sum of 'Q1' and 'Q2'”, “Compute the mean of 'item1', 'item2' and 'item3'”, “Add up 'a' + 'b'” or “Difference between 'post' and 'pre'”. Write variable names containing dots or spaces exactly, in single quotes.
In the Transform dialog, “Known values” chips help: they show the values actually present in a column so you can click them, instead of guessing whether to enter the numeric code or the label in filters and recodes.
kondeor.ai always shows the result as a preview first — nothing is applied automatically. Only clicking “Apply” actually writes the new variable into your working dataset.
Saving transformations as a recipe: Especially with more complex formulas or several chained steps, it's worth not re-typing them every time. The transform dialog offers “💾 Save as recipe” for this — a recipe is simply the currently assembled operation (or chain of operations) under a name you choose. Saved recipes can be loaded again anytime from the dropdown in the same dialog, edited before running, and either updated under the same name or saved as a new recipe. Recipes are deliberately stored across projects — like a personal formula library that can also be applied to other datasets. If a loaded recipe references a variable that doesn't exist in the current dataset (e.g. because it's named differently), this is flagged directly in the card instead of being silently dropped. Since version 1.16.53 a recipe can also contain a subset filter and group mode (B3a) as well as the weighting (B3b, since version 1.16.54) — so a complete data preparation including its restriction can be restored in one click.
“Start fresh” tidies up the dialog: all cards and the loaded recipe are discarded, so you can create a new recipe instead of accidentally overwriting the one you opened last. A recipe can also be passed on with “Share” (see A4d).
Five further operations are available in the transform dialog:
- Count occurrences — e.g. “how many of 5 items were answered yes”: for each row, counts how many freely definable conditions (one per variable) are met
- Replace missing values — mean, median, or mode, optionally grouped (e.g. mean income per region); the original variable with its missing values is left unchanged, and an optional second column shows which cells were replaced. For more than a handful of missing cells, Multiple Imputation is preferable — see B9
- Rank cases — smallest value = rank 1, ties (equal values) get the average rank, optionally computed separately within groups
- Split censored values — splits limit values such as “>90” into a value and an indicator column, preparation for Tobit regression (see below and B9g)
- Restructure Wide ↔ Long — for repeated-measures data: turns several time-point columns (e.g.
anxiety_t1,anxiety_t2,anxiety_t3) into one row per person and time point, or vice versa. Needed, among other things, to prepare data for a multilevel (mixed models) analysis — which expects one row per observation.
Split censored values: If a column holds a limit value — as a value label on a numeric code or as text such as “>90” or “<15” — kondeor.ai splits it into two new columns: the value (in censored rows the limit itself, otherwise the measurement) and the indicator (0 = exactly measured, 1 = right-censored, 2 = left-censored). You start the operation with the “≷ Split censored values” button in the transform dialog — the card shows right below the chosen source variable which censoring was recognised — or give it as a question in the chat, like any plain-language transformation. The feedback names the recognised limits with their frequency. Only entries with >, ≥, < or ≤ and a number are recognised; rows that cannot be assigned stay empty and are reported. The limit comes from the label or text, never from the numeric code. Both new columns can be used in the Tobit card in the chat as “separate column” (see B9g); in the batch Tobit still reads the censoring from the value labels or text of the original column — there is no indicator column there.
Sample size calculator (power analysis)
The target icon at the bottom left next to the input field opens an online calculator for the required sample size — it appears in the chat window, usable immediately, no AI request needed; your current mode and dataset stay unchanged. Three separate tabs, depending on your question:
- Power-based — for group comparisons and association tests: how many cases do I need to detect an assumed effect with a given statistical power? Covers t-test (independent, paired, one-sample), Mann-Whitney U and Wilcoxon tests, one-way and two-way ANOVA, repeated-measures ANOVA, Kruskal-Wallis and Friedman tests, correlation, linear and logistic regression, chi-square, proportions, survival (log-rank test), and mediation. For rank tests (Mann-Whitney, Wilcoxon, Kruskal-Wallis, Friedman) the sample is increased by a cautious allowance for the efficiency loss — the result is also sufficient for non-normal data. For the log-rank test you additionally get the number of required events, not only the total number of people.
- Precision-based — for classic upfront planning of a survey, with no hypothesis test involved: how many cases do I need to estimate a proportion or mean within a given margin of error? Optionally with a correction for small or known populations.
- Equivalence / non-inferiority — for planning a TOST or non-inferiority test (see B9b): how many cases do I need to show that two procedures are equivalent, or that a new one is not worse than the reference? Available for metric outcomes (independent groups, paired measurements, one sample against a reference value) and for proportions (e.g. complication rates). You enter the margin and, for non-inferiority, the direction (is a higher value better or worse?) yourself; they are never pre-filled or guessed. You can also set the true difference or the expected proportions, the allocation ratio of the groups and the one-sided significance level (preset to α/2 for non-inferiority). For proportions kondeor.ai uses the exact power of the same Miettinen-Nurminen score test used in the analysis; with very large groups (more than about 500 per group) the calculation can take somewhat longer. The result shows n per group, total n and the achieved power.
“I am not sure — find the method with the guide”: Don't know the test procedure? The button in the power tab opens the method selection guide (A2c). Answer the questions about your project and pick the matching method; “→ Plan sample size for this method” takes you back to the calculator with the corresponding procedure already selected. The assignment is rule-based, no AI request is made. Where the calculator can only use a closely related procedure (e.g. for ANCOVA, partial correlation, Fisher's test or Cox regression), a note helps you interpret the number — e.g. “for Fisher's test, plan about 10% more people per group”. The assignments were checked by simulation (actual power at the suggested n, target 80%).
For more complex designs (e.g. three-way ANOVA, mixed designs, sensitivity analyses), we recommend G*Power.
Post-hoc statistical power in results: Beyond upfront planning, kondeor.ai automatically shows statistical power for many test results — calculated from the actually observed effect size and the available sample size. When power is low or borderline, an additional methodological note appears: this figure is closely tied to the p-value and is systematically low for non-significant results — it does not reliably indicate whether a larger sample would have reached significance. A critical effect size is also shown: the smallest effect that the test could have reliably detected at the given sample size and significance level with 80% power — information that is independent of the p-value. Covered: t-test (independent, paired, one-sample), Mann-Whitney U, Wilcoxon, Kruskal-Wallis, ANOVA, correlation (including partial and point-biserial), chi-square test, chi-square goodness-of-fit test, binomial test, linear regression, and mediation and moderation.
Effect sizes & hover hints on the dashboard: Many results show several effect sizes side by side rather than just one — e.g. Cohen's d, Hedges' g and Glass' Δ for the t-test (independent, paired, and one-sample, each appropriate to a different degree depending on the design), η², ω² and Cohen's f for ANOVA, ε² for Kruskal-Wallis (including both Dunn and DSCF post-hoc comparisons — DSCF uses a correction tailored to the studentized range distribution instead of Dunn's generic Holm correction, often considered more powerful for all-pairs comparisons), plus odds ratio, relative risk, contingency coefficient, and gamma for the chi-square test. For Kendall's Tau, kondeor.ai automatically distinguishes between Tau-b (standard case) and Stuart's Tau-c (rectangular contingency tables, i.e. differing numbers of row/column categories) — the appropriate variant is selected without any action needed. Where a figure needs explanation, the term is shown with a dotted underline on the dashboard — hovering over it reveals a short methodological note (e.g. when one effect size is preferable to another). This marking currently only appears on desktop: tooltips cannot technically be triggered by tapping on touch devices, so the explanation is unavailable there, even though the figure itself is still shown.
Bootstrap confidence interval (ANOVA): If you mention the word “bootstrap” in your question, kondeor.ai additionally computes a bootstrap-based confidence interval for η² and ω² in one-way ANOVA — unlike the analytical confidence interval shown by default, this does not assume normality within groups. It does not run automatically, as the procedure is noticeably more computationally intensive than the analytical variant.
One- vs. two-sided hypothesis for contingency tables: For Fisher's exact test (2×2) and the proportion-difference calculation, the dashboard offers a “Hypothesis” control (Group1 ≠/>/< Group2) — choosing a directional hypothesis changes the computed p-value accordingly. For larger contingency tables (beyond 2×2), the test remains two-sided for statistical reasons, and a transparency note to that effect appears on the dashboard.
Reading crosstabs: The corner cell of the crosstab lists “Rows: <variable>” and “Columns: <variable>” on separate lines, so it is clear which variable forms the rows and which the columns, even when both have identical categories.
Style of the AI texts: Interpretations on the dashboard and overall interpretations in reports are factual and without direct address (“The groups differ …”); finding before interpretation, effect size before significance. Learning plans in courses still address you informally.
APA check
For editors and anyone who wants to double-check published statistics: paste a reported statistic directly into the analysis question field — kondeor.ai recognises the pattern automatically and checks internal consistency.
- t(38) = -0.71, p = .481kondeor.ai recomputes the p-value from the statistic and degrees of freedom and reports whether it's consistent, a rounding discrepancy, or a decision error
Supported: t-test, F (ANOVA), Chi², correlation r, z, Mann-Whitney U, Kruskal-Wallis H, Wilcoxon. By default a pure internal consistency check of the reported values against each other, with no dataset needed. If a dataset is also loaded, you can have individual statistics recomputed directly against the actual data.
Check research question and hypotheses, draw the hypothesis model
The icon in the toolbar below the chat window opens a dialog with five steps: ① Research question → ② Hypotheses → ③ Draw model → ④ Questionnaire → ⑤ To the batch. Each step works on its own: if you already have hypotheses, jump straight to ②; if you already have a finished model, straight to ③. The feature is available from the Pro plan.
① Check the research question. Paste your research question (it does not have to be phrased as a question; a research goal or topic works too). kondeor.ai checks it against the requirements from the paper “Research question and hypothesis model”: one sentence, no hidden sub-questions, subject, population, geographic and time frame, readable, free of contradictions, empirically testable, a clear focus, and feasible. A free rule check catches simple formal flaws instantly, the AI covers the rest. Each criterion appears with a traffic light (fine, unclear, not fine, note) and a short reason, with a suggested wording where useful.
② Check the hypotheses. Paste your hypotheses, one per line (up to 30). Each hypothesis is checked on its own (if-then form, exactly one statement, generally worded, operationalisable, direction, link to the research question) and all together (contradictions, overlaps, coverage of the research question, H0/H1). Numbering is only flagged from H2 on, because “H1” is reserved for the alternative hypothesis; an H0/H1 pair is fine. Lines starting with “+ Variable” count as pure description, not as hypotheses.
Suggest hypotheses (optional). If you have a research question but no hypotheses yet, kondeor.ai suggests up to five fitting hypotheses at the click of a button. The basis is your research question only, not your dataset and not the analysis plan, so the order “hypotheses first, data second” is preserved. The suggestions are starting points from general subject knowledge. You tick what is added to the hypotheses field (nothing happens automatically) and then support each hypothesis with literature. If “Data already exist” is ticked, the HARKing note appears as well.
Hints only, nothing is overwritten. kondeor.ai never changes your text on its own. You accept suggestions with a click, or ignore them.
What does “exploratory” mean? A confirmatory analysis tests an assumption that was fixed beforehand. An exploratory analysis looks into the data and searches for what stands out. Both are legitimate, but they carry different weight: if you try many relationships, you will almost always find one that is “significant”, even when it is only chance. An exploratory finding is therefore a pointer for future studies, not yet proof. Say so in the report (“exploratory”) and word it with care. With many tests, a correction for multiple testing helps as well.
③ Draw the model. After the hypothesis check you can have the model drawn directly from the check result, with no further AI call. Alternatively, paste research question and/or hypotheses as text in step ③ as before, and kondeor.ai translates them into a hypothesis model following the Kondeor convention (boxes = variables/constructs, arrows = hypotheses, including null/alternative hypotheses, one-/two-sided direction, and multivariate relationships). The result is freely arranged, central variables are highlighted, and kondeor.ai usually recognises thematically related independent variables on its own and shows them grouped by colour. Lines starting with “+ Variable” appear as description boxes and never enter the batch analysis.
Overriding construct and category classification explicitly: If kondeor.ai doesn't classify a variable as a construct (shown in italics) the way you intended, or groups it into a different colour category than you meant, you can override this directly in the text — as an addition right after the relevant variable/phrase:
- Age influences how strongly people are interested in culture (Variable) [Attitude].“(Variable)” forces the classification as a directly measured variable, “[Attitude]” still groups it under the “Attitude” colour category — the two brackets act independently
- Eating habits (Construct) influence overall satisfaction.“(Construct)” forces the italic construct rendering, even if the term might look directly measurable at first glance
Without these additions, kondeor.ai decides automatically based on content — the round brackets (Construct)/(Variable) and the square category bracket […] are deliberately separate and can be freely combined or used on their own.
④ Questionnaire. This step shows the hypotheses you entered in ② or drew in ③ and hands them to the questionnaire dialog: “Draft from these hypotheses” opens the “Draft from hypotheses” tab with research question and hypotheses already filled in, “Check an existing questionnaire” opens the check tab with the hypotheses (up to five), to which you then only assign method and variables. The step itself calculates nothing and costs nothing; you see prices in the questionnaire dialog before each action. More in B8b.
⑤ To the batch. The hypotheses from the model can be taken over as questions into the batch analysis, see B6.
The other way round. If you already have an analysis plan or batch, “Carry over to hypotheses” moves their H0/H1 to step ②, from where they go to the model and the questionnaire. Your text in ② is added to, not overwritten; duplicate lines are skipped.
Cost: Drawing the model from the check result is free; the other AI steps cost credits (price shown on the button each time, see E1); generating the model from free text costs more for many hypotheses. Credits are only charged once a usable AI answer exists; the rule check is always free. The result stays available as editable JSON and can be downloaded as an SVG graphic.
Limits: The check is purely text-based and judges your wording, not your dataset. It does not (yet) check whether the variables exist in the dataset and are scaled suitably. The AI may name the same variable differently across hypotheses, so check this in the model.
Writing results back into the dataset
For several methods (factor analysis, cluster analysis, regression [linear, logistic, multinomial], discriminant analysis, latent class analysis, propensity score matching and weighting) you can write the computed values — e.g. factor scores, cluster membership, residuals, or predicted probability/class — into the working dataset as new column(s) with one click. For factor analysis/cluster, this happens using the parameters currently set in the dashboard (e.g. after changing the rotation or number of clusters); for the other methods, directly from the analysis result.
Afterwards, download the extended dataset via the toolbar button “Export dataset” — as an SPSS file (.sav), as an Excel file with a codebook, or as a ZIP with CSV files, see B8a.
👍/👎 on the method choice
The results dashboard shows two small thumbs icons next to the suggested method. Clicking one gives anonymous feedback on whether the automatically suggested method fit your question — this helps us keep improving automatic method selection. You can change your feedback at any time by simply clicking the other thumb instead.
Before rating, it's worth checking “Why this method?” right on the confirmation card (see A3) — it already explains why this particular method was suggested, often helpful for putting your own assessment in context.
See Privacy Policy, point 3.4, for details on how this feedback is processed anonymously.
Merging files
Combine two datasets into one — e.g. append a second wave of data collection, or bring in additional variables via a shared ID. You find the feature in the “Transform variables” dialog (the icon with three stacked chevron layers), top right, as the button “Merge files” (from Pro). A dedicated dialog opens:
- Add cases — the second file is appended as additional rows (e.g. a further data-collection wave with the same variables).
- Add variables — the second file is joined to the current dataset via an ID column on both sides (e.g. additional follow-up variables for cases you already have).
kondeor.ai automatically matches column names between both files — even with different spelling or language (e.g. “geschlecht” and “gender” are recognised as the same). Nothing is decided automatically when uncertain: a mapping table shows every suggestion with a dropdown to review and override before anything is written.
For “add variables”, the ID column is matched tolerantly as well: spellings like P07, p7, and 007 are recognised as the same person (case, leading zeros, and special characters don't matter). Cases that only appear in the second file are added as a new row instead of being discarded — the resulting missing columns from your original dataset stay blank.
Create an SPSS file from study data and codeplan
Study data often comes as a plain CSV file (e.g. an export from REDCap), while the meaning of the numbers — variable labels, value labels such as “1 = yes, 2 = no”, units, scale level — is written down in a codeplan. This tool combines both into a .sav file that you can open directly in SPSS or upload here. You open it via the upload button (arrow up): on the Pro plan, kondeor.ai asks there whether you want to upload a file or create an SPSS file. The dialog works independently of any loaded dataset, and your current dataset stays untouched.
Steps
- 1. Data file — a CSV file with the raw data (encoding and separator are detected automatically). Codes are stored as numbers in the cells, e.g.
1and2. - 2. Codeplan — as CSV, Word file (
.docx) or SPSS syntax (.sps), see below. - 3. Check — kondeor.ai compares both files and shows a check report with all findings and a table of all variables (label, type, scale level, number of value labels). This step is free.
- 4. Decide — you resolve anything unclear directly in the table (see “Decisions in the report”). Every change re-runs the check immediately.
- 5. Create — one click on “Create SPSS file” downloads the
.savfile. With “Load as dataset” you take it over directly as your working dataset — the same flow as a normal upload, including the question about the data check.
Which codeplans are read?
| Format | What is recognised |
|---|---|
| CSV | One row per variable. Columns are recognised by their header (e.g. Variable/Variable name, Definition/Description/Label, Unit, Coding/Values/Choices, Missing values, Scale level/Usage, Note). If the header is not recognised, kondeor.ai asks which column contains what. |
| Word, table | The same logic for tables in a .docx file; several tables and header rows repeated across pages are no problem. |
| Word, running text | One paragraph per variable, e.g. “geschlecht – Gender. Values: 1 for male, 2 for female (categorical).” This is read by meaning and is less reliable than a table — kondeor.ai points this out in the report. Please check labels and scale levels especially carefully then. |
SPSS syntax (.sps) | Variable labels and value labels from VARIABLE LABELS and VALUE LABELS, e.g. from the SPSS export of REDCap. The syntax contains no scale level; it is derived from the value labels (nominal with value labels, otherwise scale) and can be changed in the report. MISSING VALUES are detected but not applied. |
Value labels are read in several notations, e.g. 1 = yes; 2 = no, 1=pT1a, 1, yes | 2, no, one value per line, “1 for yes, 2 for no” and range forms such as 1–4 = ASA I–IV. Anything that cannot be read as value labels (e.g. a formula description) is treated as a description and not applied as a label.
An entry can also contain a rule for missing values, e.g. “NA bei ischaemie_check = 1”. It is checked against the data; violated rules appear in the report.
What the check report flags
- Values that have no label in the codeplan (e.g. a code
5although only 1–4 are defined) — they are kept as numbers. - Text in mostly numeric variables (e.g. “n.b.” or “>90”).
- Violations of missing-value rules, duplicate ID values and values outside a specified range.
- Codeplan entries without a matching column in the data. Columns without a codeplan entry are recognisable in the table by their name being used as the label; their scale level is determined from the data.
- Dates in columns named after “date” (or described that way in the codeplan) — ISO format,
DD.MM.YYYYand Excel serial numbers are stored as dates; unreadable values are reported and set to missing — variable names that are invalid or reserved in SPSS, and duplicate value labels.
Decisions in the report
- Deviating variable names — if a column is called e.g.
rel_nephrometry_scorebut the codeplan saysrenal_nephrometry_score, kondeor.ai only suggests the assignment. It is applied only after you click “Accept”; alternatively you pick the codeplan entry yourself. - Mixed values such as “>90” — for such columns you choose: keep as text (default), take the bound as a number (>90 becomes 90), or treat the text values as missing. A column in SPSS cannot hold numbers and text at the same time; your choice decides whether you can calculate with it later. “Bound as number” biases means systematically — choose it only if that is acceptable for your analysis.
- Scale level — for numeric variables you change it via dropdown (nominal, ordinal, scale). An asterisk * marks a level derived from the codeplan (e.g. from “ordinal/categorical”).
- Omit — individual variables can be excluded from the export, e.g. a case number. If the codeplan says “do not analyse”, the field is highlighted.
- REDCap form status — fields such as
pat_info_completecan be labelled 0 = Incomplete, 1 = Unverified, 2 = Complete.
The reverse direction: export a dataset with a codebook
It also works the other way round: with the toolbar button “Export dataset” you choose the format. The SPSS export is available from the Basis plan without credits; Excel and ZIP are — like “Create SPSS file” — Pro features and cost credits (the price is shown on the “Export” button in the dialog; credits are only deducted once the file is finished, and nothing is charged if it fails):
- SPSS (.sav) — with variable and value labels. Variables with value labels (e.g. gender) are stored in the file as a numeric code (
1,2) with their labels, not as text; only variables whose cells are neither a code nor a known label (e.g. free text) remain text. The scale level (nominal, ordinal, scale) is included. - Excel (.xlsx) with codebook — two sheets: “Data” and “Codebook” (variable label, unit, value labels, scale level).
- ZIP: CSV + codebook — two text files (UTF-8, comma-separated), e.g. for archiving, data repositories, R or Python.
For Excel and ZIP you decide how the values appear in the data: as codes (recommended, e.g. 1 and 2, matching the codebook “1 = yes; 2 = no”) or as labels (e.g. “yes” and “no”, easier to read but not uniquely traceable without the codebook). The codebook uses the same notation that “Create SPSS file” reads: you can load the ZIP files back in there later and get the same labels and scale levels. The tool does not read Excel files — save them as CSV for that.
- The codebook is only as good as the metadata in the dataset. For a plain CSV file without labels, scale levels come from automatic detection; you can correct them in the variable view beforehand.
- The unit appears in the codebook only if it is part of the variable label in square brackets (e.g. “Age [years]”).
- If a filter is active, it is included as an additional column
subset_filter_aktiv(0 = no, 1 = yes), as with the SPSS export. - If a variable with value labels contains text that matches no label, it is kept unchanged; kondeor.ai points this out after the export and notes it in the codebook's note column.
- If a value label contains a semicolon, a vertical bar or a line break, it is adjusted in the codebook (otherwise it could no longer be read unambiguously); the data stays unchanged.
Costs, data protection and limits
- Costs: Checking and the check report are free. Creating the file costs credits (so does exporting as Excel or ZIP, see above); the price is shown on the “Create SPSS file” button. Credits are only deducted once the file is finished — if creation fails, nothing is charged.
- Data protection: The files are processed only for the respective step. They are not stored and not sent to an AI model; reading and checking follow fixed rules, and the statistics file is written in R. For personal or health data, your institution's rules still apply (e.g. pseudonymisation before uploading).
- Limits: The data file may be at most 8 MB, the codeplan at most 3 MB. Excel files (
.xlsx) are not read — please save as CSV. Date and time fields in SPSS syntax format are taken over as text. If a text cell contains more than 255 characters, the scale level (nominal/ordinal) cannot be set in the file; the dialog then says so, and you can set it in the SPSS variable view if needed.
Check a questionnaire (LimeSurvey)
Before you send a questionnaire into the field, kondeor.ai checks it for typical planning errors — and, if you like, compares it with the hypotheses and methods of your study. The check (steps 1–3) uses fixed rules; no AI is queried. Only if you choose the AI item check or the draft from hypotheses are question texts or hypotheses sent to an AI model (see below). You open it via the upload button (arrow up, choice “Check a questionnaire”), via “More ways” in the start assistant, or via the info window.
.lss, tested with LimeSurvey 7). Further programs will follow step by step.You reach the questionnaire via the “Research question · Hypotheses · Hypothesis model” icon (step ④, where your hypotheses are taken along) and via the roadmap (phase 3).
Steps
- 1. Read the questionnaire — in LimeSurvey save the survey structure as
.lssunder “Export” and select it here. kondeor.ai shows how many groups and questions were read and names elements that could not be imported (e.g. unsupported question types). This step is free. - 2. Describe the study (optional) — hypotheses with the planned method and the questions involved. Only with this can the check tell whether the response formats suit the analysis. You also set the upper limit for completion time (default 8 minutes).
- 3. Run the check — free of charge you see the number of errors, warnings and notes and the estimated completion time.
- 4. Show findings — the full list with location and suggested improvement. Only this step costs credits (price shown on the button; charged only after a successful check). The findings can be saved as CSV.
Draft from hypotheses (AI item suggestions)
On the “Draft from hypotheses” tab you write your hypotheses or research question as plain text. kondeor.ai derives the constructs and their roles (independent, dependent, mediator, moderator, control), assigns matching scales from the library and proposes items of its own for all remaining constructs. The result is a questionnaire with an introduction, one item group per construct and the necessary personal details at the end, a LimeSurvey file (.lss) to import, a codebook (CSV) for “Create SPSS file” and an automatic check with the same rules as above. You see the price first; it is charged only once the draft exists. Your hypothesis text is sent to an AI model for this.
AI item check (step 4)
Optionally the AI checks the wording of your questions for typical errors: leading questions, double-barreled items (two matters in one item), double negatives, vague terms (“regularly”, “often”), absolute terms (“always”, “never”), mismatched response scales and jargon. Each finding comes with a suggested better wording. Library scales are not assessed; their validated wording stays untouched. You see for free how many questions will be checked, the price and your balance; you are charged only if the AI evaluation worked.
What is checked
| Area | Examples |
|---|---|
| Structure & response formats | duplicate or statistics-unfriendly question codes, unlabelled answer steps, gaps or overlaps in classes, scales with inconsistent response format, very long radio-button lists, multiple choice without “none of these”, overlong matrices and rankings |
| Fit with the analysis | Is the level of measurement sufficient for the planned method (e.g. comparing means with a nominal answer)? Is a hypothesis variable measured by no question? Are there questions in no hypothesis, constructs with a single item, or age collected in classes? |
| Order, filters & duration | filter conditions that refer to later or unknown questions or are never met; demographics as the first question; blocks without introduction; open questions too early; dependent variable directly before the independent one; estimated completion time against your upper limit |
| Scales from the library | If you use a known scale (e.g. satisfaction with life, self-esteem, Big Five short scales), kondeor.ai recognises it automatically — by the wording of the items or by an abbreviation in the question code (e.g. swls1) — and checks wording, completeness, validated language version and licence against the library. Since LimeSurvey does not store reverse coding, kondeor.ai lists the items that must be reverse-coded in the analysis. A note also shows in which studies the instrument was used and how it was analysed there (pilot feature, only for stored instruments). |
| Data protection | Questions on personal or specially protected information, missing note on data protection or consent, mandatory questions on sensitive topics, content of the welcome text. This is a memory aid, not legal advice. |
| Multilingual | Questions or answer options missing in one of the questionnaire's languages |
Each finding is rated as error (should be fixed before fieldwork), warning (probably a problem) or note (for your information).
Good to know
- Data protection: The questionnaire is processed only for the check and is not sent to an AI model. The duration is an estimate based on question type, not a measurement.
- Costs: Reading and counting are free, the list of findings costs credits. The function is available from the Pro plan.
- Limits: The wording of questions is checked only by the optional AI item check (step 4); its assessments are suggestions. Max. 4 MB and 2000 questions per file.
Upload your own studies (“My studies”)
If you work with a study that kondeor.ai does not know yet, you can upload it as a PDF. kondeor.ai extracts study details, the scales used and the analysis steps, and then uses them in your literature suggestions, in scale recognition and in the questionnaire check. The “My studies” button is in the header next to “Invoices” and is visible from the Pro plan.
How it works
- 1. Choose a PDF — read the notes, tick the confirmation, select the PDF (up to 5 MB, text PDF; scanned pages without text are not recognised).
- 2. Check the preview — kondeor.ai shows title, authors, year, journal and DOI, the publication type, the study design and whether the study is peer reviewed. This classification is an estimate from the PDF text; you can correct everything. You can deselect analysis steps and scales.
- 3. Choose a topic and save — either an existing topic (recommended) or an own category that only you can see.
What is stored
- Stored are only the extracted details: study data, scales with abbreviations and short own descriptions of the analysis steps. The PDF itself, the full text and the results (effects, p-values) are not stored.
- Data protection: The text of the PDF (without the reference list) is passed to an AI model for extraction, as described in the privacy policy; the upload is not available while AI is switched off.
- Cost: Extraction costs credits (the price is shown before the upload). It is charged only once a usable result exists. If the study is already available, no price applies.
- Limits: Up to 60 own studies per account.
Private or for everyone
Your studies are private at first: only you see them, and they feed only into your own suggestions. With “Propose for all” you can submit a study for inclusion in the general collection. kondeor.ai reviews the proposal; only after approval do all users see the extracted details (with no link to your account). Until then you can withdraw the proposal. In the “My studies” list you see the status of each study and can delete studies that are not approved at any time.
Structural Equation Models, Multiple Imputation & Growth Curves
Three advanced extensions that all work on the same principle: kondeor.ai proposes a configuration based on your question, shows it in a confirmation card, and only runs the calculation after you click “Apply”/“Run” — nothing happens automatically in the background.
Confirmatory Factor Analysis (CFA) & Structural Equation Models (SEM)
- Confirmatory factor analysis with items item1, item2, item3 for the factor Satisfaction.
- Structural equation model: factor Satisfaction from item1-item3, factor Loyalty from item4-item5, Satisfaction affects Loyalty.Combined measurement and structural models are supported
The confirmation card shows the proposed factor/path structure in plain language and the raw lavaan syntax in an editable text field — you can correct the syntax directly or add your own covariances/second-order factors instead of relying on the automatic assignment. A checkbox lets you choose between listwise (complete cases only, default) and FIML (Full Information Maximum Likelihood — uses all available values, no row is dropped; methodologically superior to listwise deletion when data are missing at random).
Multiple Imputation
Unlike the simple “replace missing values” option in the transform dialog (see B4 — a single replacement value per missing cell), Multiple Imputation generates five complete, plausibly imputed versions of the dataset, runs the analysis on each one separately, and combines the five results using Rubin's rules into a single result that correctly accounts for imputation uncertainty — the methodological standard for missing data whenever more than a handful of cells are affected.
Technical details (for your methods section): R package mice (van Buuren & Groothuis-Oudshoorn, 2011), 5 imputations with 5 iterations each, predictive mean matching for numeric and logistic or multinomial regression for categorical variables, fixed random seed – the same analysis therefore always yields the same result. Five imputations are the classic standard and suffice when few values are missing; with a high share of missing cases (roughly above 20%), more recent literature recommends more imputations, roughly as many as the percentage of incomplete cases.
Available for t-test, Pearson correlation, and linear regression: a “Re-run with Multiple Imputation” button appears on the dashboard after any of these three analyses — the existing variables are picked up automatically, no need to re-describe the analysis. If the analysis has no missing values at all, kondeor.ai reports this briefly instead of running a pointless calculation. Each run costs credits; the plan mark of the heading shows from which plan multiple imputation is available.
Auxiliary variables & comparison with complete cases: So that the imputation does more than fill values in from the group, kondeor.ai automatically adds up to five further numeric variables of the dataset to the imputation model — those related to the incomplete variable or to its missingness (e.g. the baseline value in a follow-up measurement). They only serve the imputation and do not enter the analysis; the upper limit depends on the number of observed values (about one auxiliary variable per ten). The dashboard shows the chosen auxiliary variables, the share of missing values and, next to it, the result of the analysis on the complete cases only — so you can see whether the imputation changes the result (a simple sensitivity analysis for missing values).
Choosing auxiliary variables yourself: Under the result, “Change auxiliary variables of the imputation model” opens a list of the numeric variables that qualify (with the share of observed values). You can tick up to eight variables and impute again with “Recalculate with this selection”, or use “Without auxiliary variables” to impute from the analysis variables only on purpose. Each run costs credits, like the first one, and appears as a new answer — so you can compare selections. The dashboard and the interpretation note that the selection is yours.
Growth Curve Models
- Growth curve model for 'mood' over time, grouped by 'person_id'.
- Wachstumskurvenmodell für stimmung über die Zeit, gruppiert nach person_id.
- Quadratic growth curve model for 'score' over time, grouped by 'person_id', controlling for 'stress' over time.Curved trend + time-varying covariate in one request
Models how a measurement develops across repeated occasions — in two ways at once: as a multilevel model (random intercept + random slope per person) and as a latent growth model (structural equation model with intercept and slope factors). The confirmation card shows dropdowns for person ID, time, value, method, and growth form, plus a multi-select for covariates — all populated from your column names and editable before running. The dataset must be in long format (one row per person and occasion) — if it's wide (e.g. mood_t1, mood_t2, mood_t3), reshape it first using “Restructure wide ↔ long” (see B4).
The result shows the estimates from both models (if both were requested) as well as a trajectory chart: thin individual trajectories in the background plus a bold mean curve with a 95% confidence band — the core of the feature, not an afterthought.
Growth form: linear (default) or quadratic (an additional curvature term for a trend that flattens out or accelerates). For the multilevel model, curvature is estimated as a population-average effect (no individual differences in curvature per person — that would be a third random-effects level and would rarely converge reliably at typical sample sizes). For the latent model, a third factor is added instead.
Time-varying covariates: up to 5 additional variables that were themselves measured over time (e.g. “controlling for stress over time”) — not to be confused with a time-constant characteristic like sex. In the multilevel model they enter as additional fixed effects; in the latent model, their effect is estimated freely at each occasion (no assumption that the effect stays constant over time).
Further advanced methods
Further Pro methods: equivalence and non-inferiority tests, latent class analysis and extended mediation are operated like SEM/CFA via a confirmation card with editable variable dropdowns; for propensity score matching you change the settings directly in the dashboard after the calculation. The chapter ends with interrater agreement (kappa), which you start with an ordinary question naming the rater columns.
Equivalence and non-inferiority testing (TOST)
Statistically tests whether two values are practically equivalent — unlike an ordinary test, which only tests for a difference and doesn't let you interpret a non-significant result as “equal”. TOST (“Two One-Sided Tests”) flips the logic: the null hypothesis is the difference, and what gets confirmed is equivalence.
- Equivalence test for 'score' between 'group' = 1 and 'group' = 2, equivalence bounds ±5.Two independent groups (design “independent”)
- Equivalence test for 'before' and 'after', equivalence bounds ±0.3 in Cohen's d.Paired measurements (design “paired”), bounds either in raw units or standardised as Cohen's d
Also available: design “one_sample” (one sample against a fixed reference value), design “proportions” (two independent proportions, e.g. approval rates of two groups), and design “correlation” (a correlation against a reference value, via Fisher-z transformation). You need to set the equivalence bounds (the range around zero within which a difference still counts as negligible) yourself and justify them on subject-matter grounds: for means kondeor.ai suggests ±0.3 Cohen's d as a flagged suggestion you can change in the card if you give none; for proportions and correlations there is no suggestion, you enter the bounds yourself. For proportions (e.g. complication yes/no) you also explicitly choose in the card which value counts as the event and which group is group 1; the calculation uses a score test by Miettinen and Nurminen, which remains reliable for rare events. The significance level α (per one-sided test, 0.05 is customary) can be changed.
Non-inferiority test (one-sided): checks whether a new method (test group) is not worse than a reference by more than a pre-specified margin Δ — typical in medicine for new surgical procedures or therapies. Available for two independent groups with a metric outcome and for two proportions (e.g. complication yes/no).
- Non-inferiority test for 'complication' by group 'procedure', test group 'new', event 'yes', margin 0.05, higher is worse.Two proportions: the margin is a difference in proportions (0.05 = 5 percentage points)
- Non-inferiority test for 'pain' by group 'procedure', test group 'new', margin 1 point, higher is worse.Metric outcome (one-sided Welch t-test), margin in raw units or as Cohen's d
In the confirmation card you explicitly set: the margin, the direction (is a higher value better or worse? — no preselection, nothing is guessed), the test group (the new method) and, for proportions, the event. The one-sided significance level is pre-set to 0.025 (the lower limit of a 95% confidence interval). Superiority is only reported when non-inferiority has been shown. In non-randomised data the test does not replace adjustment for confounders (see propensity score matching).
Latent class analysis (LCA)
Identifies unobserved subgroups (classes) in the data based on response patterns across several categorical variables — e.g. to derive respondent types from several yes/no or agreement questions, without defining the groups yourself in advance.
- Latent class analysis with 'item1', 'item2', 'item3', 'item4'.At least two categorical variables required
The number of classes is chosen automatically based on BIC (Bayesian Information Criterion) across several tested solutions — no need to specify K manually. The result shows a model comparison table (AIC/BIC/log-likelihood for each tested number of classes) plus the relative size of each class found. Each case's class membership can be committed straight into the dataset just like with factor/cluster analysis (see B7).
Moderated & serial mediation
Builds on simple mediation (X acts on Y through a mediator M, available from the Basis plan), but covers two more complex models:
- Moderated mediation — the indirect effect (X → M → Y) itself depends on a fourth variable W, either at the first stage (W affects how strongly X influences M) or the second stage (W affects how strongly M influences Y).
- Serial mediation — two mediators in sequence (X → M1 → M2 → Y) instead of just one.
- Moderated mediation: 'stress' acts on 'turnover intention' through 'exhaustion', moderated by 'support'.
- Serial mediation: 'training' acts on 'performance' through 'competence' and 'confidence'.
Both methods need at least 20 usable cases (complete values on all variables involved).
Propensity score matching
Makes two groups comparable that did not arise at random — for example patients of two surgical techniques, or customers with and without programme participation who already differ in age, severity or similar. For each case, the probability of belonging to the treatment group is estimated from the covariates (propensity score); each treated case is then paired with a comparison case as similar as possible, and the effect is estimated only between these pairs (ATT). The treatment group is automatically the smaller of the two groups.
- Propensity score matching: effect of 'surgery_method' on 'ischemia_time', covariates 'age', 'bmi', 'asa'.Metric outcome: mean difference in the matched comparison; for a yes/no outcome, odds ratio and risk difference
- … covariates 'age', 'bmi', 1:2, caliper 0.1, with replacement.Settings can be given directly in the question
In the dashboard you can then change the ratio (1:1 to 1:5), the caliper (0.1 / 0.2 / 0.25 standard deviations of the score, or none) and replacement — the result is recalculated immediately, without further credits. The balance table (“Groups” tab) shows the standardised mean difference of each covariate before and after matching; below 0.1 counts as well balanced. “Add matched pairs to dataset” writes psm_gematcht, psm_paar, psm_score and psm_gewicht to the working dataset — for further analyses in the matched sample, filter on psm_gematcht = 1.
Factor analysis and principal component analysis (PCA)
For a factor analysis kondeor.ai uses maximum-likelihood extraction by default. If your question contains “PCA”, “principal component” or the German “Hauptkomponentenanalyse”, a principal component analysis is run instead. The difference: PCA decomposes the total variance of the variables and suits condensing many variables into a few summary scores; factor analysis models only the shared variance and suits the search for underlying constructs.
- Principal component analysis with 'item1', 'item2', 'item3', 'item4', 'item5'.Component loadings, variance explained per component, parallel analysis and MAP test as a hint for the number of components
- PCA with 'item1' to 'item6', varimax rotation.Rotation and number of components can be changed afterwards in the dashboard without further credits
In the dashboard the columns are called “C1, C2 …” (components) instead of “F1, F2 …”, and the row “Method” names the chosen method. “Save component scores to dataset” writes the values as komponente_1_score etc. into the working dataset (see B7).
Passing-Bablok regression (method comparison)
Checks whether two measurement methods are interchangeable — typical in laboratory medicine when a new method is compared with a reference method. Unlike ordinary regression, Passing-Bablok regression is robust against outliers and assumes neither normality nor an error-free x value. It complements the Bland-Altman analysis: Bland-Altman shows how far the methods differ in individual cases, Passing-Bablok whether they differ systematically.
- Passing-Bablok regression: 'method_a' and 'method_b'.Two metric measurements of the same samples, one sample per row
The result shows slope and intercept with confidence intervals and an assessment of agreement: if the confidence interval of the slope contains 1 and that of the intercept contains 0, the methods agree; if only the intercept deviates there is a constant difference, if only the slope deviates a proportional one (or both). In addition there is a Cusum test for linearity (p < .05 argues against a linear relationship, in which case Passing-Bablok regression is not applicable) and Spearman's rank correlation. The scatter plot shows the regression line and the identity line y = x.
At least 6 pairs and, per method, at least 3 distinct values are required; from 3,000 pairs kondeor.ai uses the large-sample approximation. With small samples the confidence interval is wide — at least 40, preferably 100 samples are usually recommended.
Interrater agreement (Cohen's & Fleiss' kappa)
Checks how well two or more raters (e.g. coders or diagnosticians) assign the same cases to categories — beyond chance. This is the standard measure when several people code findings, interviews or open answers.
- Interrater agreement (Cohen's kappa): 'rater1', 'rater2'.Two raters: Cohen's kappa with a 95% confidence interval; for numerically coded, ordered categories (e.g. 1–5) additionally a weighted kappa (quadratic; linear if the question says “linear weighted”)
- Fleiss kappa of the raters 'rater1', 'rater2', 'rater3'.From three raters: Fleiss' kappa with a test against zero and Light's kappa (mean of all rater pairs); the pairwise kappas are in the “Groups” tab
Each rater column is one rater, each row is the same case; name all rater columns in the question in single quotes. Cases not rated by all raters are excluded. The rating follows Landis & Koch (.61 and above “substantial”, above .80 “almost perfect”).
Kappa depends on how evenly the categories are used: with very uneven frequencies it can be low despite high raw agreement. For metric ratings use the intraclass correlation (ICC), for rankings Kendall's W; if a rater column has more than 15 different values, kondeor.ai points you to ICC. There is no scale-level follow-up question here — kappa works on the category codes. Interrater agreement also runs in batch mode.
Propensity score weighting (IPTW)
The alternative to matching: here too, two non-randomly formed groups are made comparable, but no case is dropped. As in matching, the propensity score is estimated from the covariates; each case is then weighted by the inverse of the probability of its actual group, so that both groups look alike in the covariates. This helps above all with small samples, where many cases would find no partner in matching.
- Propensity score weighting (IPTW): effect of 'surgery_method' on 'ischemia_time', covariates 'age', 'bmi', 'asa'.Default: effect in the whole sample (ATE) with stabilised weights; outcome metric (mean difference) or yes/no (odds ratio and risk difference)
- … with ATT, unstabilised, trim weights at 5%.ATT = effect in the treatment group; “trim” caps extreme weights (without a number at the 1st and 99th percentile)
The result shows the adjusted effect with a 95% confidence interval (robust standard errors), the effective sample size and the balance table (“Groups” tab: standardised mean difference per covariate before and after weighting; below 0.1 counts as well balanced). Notes appear for weak overlap of the groups, very large weights or remaining imbalance. Settings are part of the question; there is no dropdown to change them as with matching. With “Add weights to dataset” you write each case's weight (iptw_gewicht) and the propensity score (iptw_score) into the working dataset as new columns — for your own weighted analyses (then with robust standard errors, because the score was estimated); cases outside the analysis stay empty (see B7).
Time series analysis
Covers five related methods for time-ordered data — each reachable via its own keywords in the question:
- ARIMA forecast for 'revenue' over the next 6 months.Forecast with confidence interval
- Is there a structural break in 'visitor numbers'?Chow test if the timing is known, otherwise an automatic Bai-Perron multi-search for the most likely break point
- Are there unusual points in 'daily revenue'?Outlier detection, accounts for seasonality (STL/Loess residual + IQR-robust threshold)
- CUSUM test for 'production volume'.Tests parameter stability over time, no known break point required beforehand
- STL decomposition of 'visitor numbers'.Trend/seasonal/remainder components, also provides the seasonally adjusted series as input for the other four methods
AI forecast (TiRex-2 foundation model)
As a complement to the classical ARIMA analysis, an optional AI-powered zero-shot forecast is available — a pre-trained foundation model (TiRex-2) that works without any training on your own dataset. With the matching keyword in your question, both methods run one after the other, so you can compare them directly:
- AI forecast for 'revenue' over the next 6 months.Runs alongside the classical ARIMA analysis, both results are shown one below the other
Exponential Smoothing (ETS)
A standalone alternative to the classical ARIMA analysis — automatic selection of the best error/trend/season component combination (AICc-based). Unlike the TiRex comparison feature, this runs only exponential smoothing, not ARIMA as well:
- Exponential smoothing for 'revenue' over the next 6 months.“Holt-Winters” or “ETS” are also recognised as phrasing
Meta-analysis
Combines the results of several studies into one overall effect — e.g. for a systematic review, a literature overview in a thesis, or to pool your own sub-studies. Unlike every other method, you do not upload raw data for this, but a study table: one row per study with its effect size.
1. Prepare the study table
Create a CSV, Excel or SPSS file in which each row is one study. kondeor.ai recognises the format from the column names (case does not matter). One of the following combinations must be present:
| Your studies report … | Required columns | Pooled as |
|---|---|---|
| Standardised mean difference + standard error | d (or g, hedges_g, cohens_d, smd) + se | d or g |
| Standardised mean difference + group sizes | d/g + n1 + n2 | d or g |
| Correlation + sample size | r (or rho, cor) + n | r |
| Odds ratio with 95% confidence interval | or (or odds_ratio) + ci_lower + ci_upper | OR |
| Log odds ratio + standard error | lor (or log_or) + se | OR |
Optional but recommended: a column study (or label, author) with a short name such as “Miller 2021” — it then appears in the forest plot and the study table instead of “Study 1, 2, 3 …”. Example of a table with correlations:
| study | r | n |
|---|---|---|
| Miller 2021 | 0.32 | 120 |
| Smith 2022 | 0.18 | 85 |
| Brown 2023 | 0.41 | 210 |
n1/n2 or r with n as a row into the study table “Meta-analysis (g)” or “Meta-analysis (r)”, which kondeor.ai creates in the workspace for this purpose. You can adjust the study name. For the t-test, the dialog shows which group is subtracted from which (e.g. “women − men”) — reverse the direction if needed so that all studies point the same way; it is kept in the richtung column for checking. To analyse, switch to the study table and ask as described below. The table is an ordinary dataset: you can correct it in the data view, add values from the literature and save it under “My Datasets”. Paired t-tests and Spearman/Kendall correlations deliberately do not offer the button, because their effect sizes have a different sampling distribution.2. Upload the file and ask
Upload the study table like any other dataset (see A2) and ask a question such as:
- Run a meta-analysis across all studies.
- Combine the studies into a pooled effect and check for publication bias.Phrasings with “forest plot” or “pooled effect” are also recognised as a meta-analysis
To try it without your own file: “Show me an example of a meta-analysis” (see B0b).
3. Reading the result
- Pooled effect with 95% confidence interval and p-value — estimated with a random-effects model (REML), which allows the true effect to differ between studies. The test uses the Knapp-Hartung adjustment (t-distribution with k − 1 degrees of freedom), which gives more reliable confidence intervals with few studies; the interval is never narrower than without the adjustment.
- 95% prediction interval (from 3 studies) — the range in which the true effect of a new, comparable study is expected. It accounts for the variation between studies and is much wider than the confidence interval when studies are heterogeneous. If it includes the no-effect value, the effect may be absent or even reversed in individual settings.
- Heterogeneity — Cochran's Q test, I² (the share of variation due to real differences rather than chance) and τ². With high heterogeneity the overall effect is of limited meaning; it is then worth looking at how the studies differ.
- Forest plot (Charts) — each study with its estimate and confidence interval; the size of the square shows its weight, the diamond at the bottom the pooled effect.
- Funnel plot and Egger's test — check whether small studies with striking results are over-represented (a sign of publication bias). Trim-and-fill additionally estimates how the overall effect would change if such presumably missing studies were added.
- Study table (“Groups” tab) with effect, standard error, weight and confidence interval of each study.
Fixed-effects model
The default is the random-effects model. You can explicitly request a fixed-effects model by naming it in your question, e.g. “meta-analysis with fixed effect” or “fixed-effects model”. It assumes one common true effect for all studies and is only appropriate if the studies are practically identical in design and population — e.g. several sites or survey waves with the same protocol. The result then applies only to exactly these studies. With heterogeneous studies (see I² and the Q test) its confidence interval is too narrow; there is no prediction interval for this model.
Not currently covered: moderator analyses (meta-regression) and several dependent effects from the same study (multilevel meta-analysis). If a study reports several effects (e.g. two outcomes), include only one of them per analysis. Effect sizes from repeated-measures designs (dz) have a different sampling variance than those from independent groups — enter their standard error directly (column se) instead of n1/n2.
Clinical analyses: Table 1, Cochran-Mantel-Haenszel, risk measures, competing risks, Tobit regression, Kaplan-Meier for censored measurements
Six analyses mainly needed in medicine and the lab. Table 1, the Cochran-Mantel-Haenszel test, competing risks, Tobit regression and the Kaplan-Meier estimate for censored measurements carry the plan mark in their heading (on the Free plan a notice names the required plan). Those with a confirmation card show it before calculating whenever something is not clear from the question.
Table 1 (baseline table)
Describes the baseline characteristics of a study by group — e.g. age, sex and pre-existing conditions per treatment group — with a p-value per characteristic. You only name the grouping variable; all other columns go into the table.
- Table 1 by 'procedure'.All characteristics by the groups of the variable “procedure” (2 to 6 groups)
- Baseline table by 'procedure', without 'patient_id'.Explicitly exclude individual columns
Metric characteristics appear as mean (SD) or — if a group is not normally distributed and small — as median [Q1; Q3], categorical characteristics as n (%) within the group. The statistics engine chooses the test: Welch t-test or Welch ANOVA, alternatively Mann-Whitney U or Kruskal-Wallis; chi-square or Fisher's exact test for small expected counts. For two groups the standardised mean difference (SMD) is added.
If the grouping variable is not in the question, the confirmation card asks for it — with no preselection. There you can exclude further columns. IDs, free text, constant and empty columns are skipped automatically and listed in the result with the reason; nothing disappears silently.
Cochran-Mantel-Haenszel test (stratified 2×2 tables)
Tests the association of two binary characteristics (e.g. procedure and complication yes/no) across several strata — such as centres or age groups — adjusting for the stratification.
- Cochran-Mantel-Haenszel test for 'complication' by 'procedure', stratified by 'centre'.Exposure, outcome and stratification variable are in the question
- Cochran-Mantel-Haenszel test for 'complication' by 'procedure', exposed group 'new', event 'yes', stratified by 'centre'.With all details the analysis runs immediately, also in the batch
The confirmation card explicitly asks for: which group is the exposed (test) group, which value counts as the event and which variable forms the strata — all without preselection, nothing is guessed. The odds ratio refers to the exposed group versus the other group.
The result shows the CMH chi-square test (with continuity correction), the common Mantel-Haenszel odds ratio with confidence interval and, for comparison, the crude odds ratio without stratification — a clear difference between the two points to confounding by the stratification variable. The Breslow-Day test (Tarone-corrected) checks whether the odds ratios are homogeneous across strata; only then is a common odds ratio a sensible summary (otherwise a warning appears). The values per stratum are also shown in the dashboard.
Requirements: exposure and outcome have exactly two levels, the stratification variable at least two (at most 30), and at least 10 complete cases. An active subset filter (B3) is respected.
Competing risks: cumulative incidence, Gray test, Fine-Gray
Competing events arise when, besides the event of interest, another event can occur that prevents the first — e.g. a relapse where patients may die from another cause beforehand. Kaplan-Meier treats these cases as censored and therefore overstates the risk. kondeor.ai instead calculates the cumulative incidence per event.
- Cumulative incidence of relapse with a competing event: time 'months', status 'outcome', by 'treatment'.Time and status variable are in the question, you set the status values in the card
- Competing risks with Fine-Gray regression for 'treatment' and 'age', time 'months', status 'outcome', event '1', censored '0'.With all details the analysis runs immediately, also in the batch
The status variable needs at least three values (censored, event of interest, competing event). The confirmation card explicitly asks which value is the event of interest and which means censored — with no preselection. All other status values count as the competing event, and the card shows you this live. Optionally you choose a group variable, covariates (up to 8) and the times at which to report.
The result shows the cumulative incidence of the event and of the competing event per group with a 95% confidence interval at the chosen times, plus the curve of the cumulative incidence with a band. With groups the Gray test is added (does the cumulative incidence of an event differ between the groups?), and with a group or covariates also the Fine-Gray regression with subdistribution hazard ratios (SHR) and the reference group.
Requirements: at least 20 complete cases, at least 5 events of interest and at least one competing event; at most 5 groups. An active subset filter (B3) is respected. If the status variable has only two values, Kaplan-Meier is the right method.
Tobit regression: measurements with limit values (censored values)
Some measurements are only known as a limit value — e.g. a kidney function (GFR) that is reported only as “>90” above 90, or a lab value “<0.5” below the detection limit. If the limit is treated like a measured value, the estimate is biased. Tobit regression instead uses only the information that the true value lies above (right-censored) or below (left-censored) the limit. Different limits in one column (e.g. >70 and >90) are possible, as is censoring on both sides.
- Tobit regression for 'gfr' with 'age' and 'sex'.Value variable and predictors are in the question, you define the censoring in the card
- Tobit regression for 'gfr' with 'age', log scale.Model on the log scale (for right-skewed values); effects then appear as the factor exp(b)
You specify the censoring in the chat. There is no separate dialog for it: the confirmation card in the chat first shows the censoring rules kondeor.ai recognised in the data — either a value label on a numeric code (e.g. code 91 with the label “>90”) or text in the cell (“>90”, “<15”). The limit is read from the label or text, never from the numeric code (code 91 with the label “>90” gives the limit 90). You can change or delete every rule and add your own; “>” or “≥” means right-censored, “<” or “≤” left-censored. Alternatively you choose a separate column with a censoring indicator and define which indicator value means right- or left-censored. If nothing is recognised the list stays empty — nothing is guessed.
If the values are in a CSV column as text such as “>90”, you can split them beforehand — in the transform dialog or in the chat (see B4, “Split censored values”) and choose the new indicator column in the card.
The result shows the coefficients for the latent (uncensored) value with 95% confidence interval and p, next to them for comparison the naive OLS estimate in which the limits count as measured values (deviations from 10% are flagged), plus σ, the likelihood-ratio test against the model without predictors, AIC/BIC and an AIC comparison normal vs. lognormal that recommends the log scale when it fits clearly better.
In the batch: Tobit runs there without a card if the question names the value variable and the predictors and the censoring can be read from the data (value label or text such as “>90”). The output states the recognised censoring as “not confirmed”. If something is missing — or the censoring cannot be recognised — the question is not run and marked with a message. A separate indicator column exists only in the single chat. In the analysis plan kondeor.ai only proposes Tobit if the value labels of the dependent variable show a limit value.
Requirements: at least 20 complete cases, at least 5 exactly measured values and at least one censored value; up to 8 predictors (categorical ones with at most 12 categories). An active subset filter (B3) is respected. Not to be confused with censoring in survival analysis — Kaplan-Meier and Cox regression are the right methods for that. If you only want to describe the distribution (median, quartiles, group comparison), use the Kaplan-Meier estimate for censored measurements (below).
Kaplan-Meier estimate for censored measurements (median, quartiles, group comparison)
For measurements with limit values (e.g. “>90” or “<0.5”, see Tobit above) there is also a distribution estimate without predictors: the measurement plays the role of “time”, a limit value is a censored value. You get median and quartiles with confidence intervals without treating the limits as measured values — optionally separately by group including a group comparison. Tobit regression is the right choice when you want to estimate the influence of predictors on the measurement.
- Kaplan-Meier-type estimate for the censored measurements 'gfr'.Median and quartiles of the distribution with a curve
- Kaplan-Meier-type estimate for the censored measurements 'gfr' by 'sex'.Plus the comparison of the groups (2 to 8 groups)
You define the censoring as for Tobit on the confirmation card in the chat (rules recognised from value labels or text in the cell, or a separate indicator column); there you also optionally choose the group variable. Right-censored values (“>”), left-censored values (“<”) and both at once are possible: with only one direction of censoring kondeor.ai computes the classic Kaplan-Meier estimate (mirrored on the axis for left-censored values), with mixed censoring a Turnbull procedure for interval data.
The result shows per group n, share of censored values, median with 95% confidence interval and quartiles, next to them for comparison the naive values in which the limits count as measured values, and the curve P(Y ≤ y) with a band. With groups the comparison follows: with one-sided censoring the log-rank test (plus Peto-Peto) with Holm-corrected pairwise tests from three groups, with mixed censoring a Gehan test with a permutation p-value, including the effect (δ = P(group 1 > group 2) − P(group 1 < group 2)) and the share of pairs that cannot be decided.
In the batch: The estimate runs there without a card if the question names the value variable and the censoring can be read from the data (value label, text such as “>90” or a column “<name>_zens”); a group variable is optional. The output states the recognised censoring as “not confirmed”. Otherwise the question is not run and marked with a message. In the analysis plan it is only proposed if the dependent variable is censored.
Requirements: at least 10 complete cases, of which at least 3 exactly measured values; groups with fewer than 5 cases are left out and named in the result; at most 8 groups. An active subset filter (B3) is respected. Not to be confused with censoring in survival analysis — Kaplan-Meier and Cox regression are the right methods for that.
Risk difference and NNT/NNH in the chi-square dashboard
For a 2×2 cross-table the chi-square test now also shows, next to odds ratio and relative risk, the risk difference (row 1 minus row 2, with a score confidence interval) and the NNT/NNH (number needed to treat or to harm = 1 / |risk difference|). The risk difference is also defined when a cell is zero. The confidence interval of the NNT only appears if that of the risk difference does not include 0 — otherwise it says “not finite” instead of invented limits.
R code to reproduce your analysis
For the most common methods, you can have kondeor.ai show you the matching, ready-to-run R code — to double-check kondeor.ai's own results, or to pick up R along the way. Just ask for it as part of your question:
- Compare 'score' by 'group'. Also show me the R code.Works for any supported method, not just the t-test
The code includes the matching read-in line with your dataset's actual filename (read.csv() or haven::read_sav(), depending on the format) — load the same dataset locally in RStudio as well, and the code runs there unchanged.
Supported: t-test, ANOVA, correlation, linear regression, chi-square, descriptive statistics, Mann-Whitney U, Wilcoxon test, Kruskal-Wallis test.
Qualitative Text Analysis
Systematically analyse open-ended answers — including the key limits you should know about.
Methods overview
Name the method directly, or make a general text-analysis request — kondeor.ai will then ask for column selection and method:
- Run a Grounded Theory text analysis.
- Perform a qualitative content analysis using Mayring.
| Code | Method | Short description |
|---|---|---|
| A | Inductive category formation | Categories emerge directly from the material |
| B | Deductive category formation | You provide the categories, kondeor.ai assigns the texts to them |
| C | Sentiment analysis | Rated as positive / neutral / negative (5-point scale) |
| D | Themes & keywords | Most frequent themes and terms |
| E | Content analysis (Mayring) | Systematic category formation with anchor examples and coding rules |
| F | Grounded Theory | Theory generation through open, axial, and selective coding |
These features currently require the Pro plan. Details on all methods: method-selection guide on the assistant page (opens directly in English).
Analysing multiple text columns together
Yes, kondeor.ai can do this automatically — manually copying multiple open-ended questions into one column is not necessary for the analysis itself. If multiple free-text columns are detected in the dataset, kondeor.ai actively asks before starting the analysis:
- ALLAnalyse all detected free-text columns together (cross-column analysis)
- ONLY oe2_improvementAnalyse only one specific column
Alternatively, you can name the desired column(s) directly in your first question — kondeor.ai recognises them automatically without asking. If there's only a single free-text column in the dataset, it's selected automatically.
With “ALL”, texts from all selected columns are shuffled together for the analysis (not processed column by column) — this way the AI doesn't pick up artificial patterns purely from column order. Interactive review and category export (see below) have recently started supporting multiple jointly analysed text columns too — manual merging is no longer needed for any part of the workflow.
Limits: case count in text analyses
For cost and time reasons, text analyses have an upper limit on the number of text units processed for the main analysis. The dual-model reliability check no longer has a fixed cap — for larger case counts it now runs automatically in the background (see below).
| Analysis part | Cap | Details |
|---|---|---|
| Main analysis (prose/categories) | 500 | … per column. With “ALL” across e.g. 3 columns, that's up to 1,500 text units in total. |
| Dual-model reliability check (Kappa/Jaccard) | no cap | From roughly 250 text units in total, the check runs as a background process and then processes the entire dataset, not just a subset. |
For inductive methods (A, D, E, F), the category system is discovered on a representative sample for very large case counts (currently around 220 text units) and then applied to all responses — category systems typically saturate well before the full case count is reached. For sentiment (C) and predefined categories (B), this step doesn't apply, since no new category system needs to be discovered there.
Dual-model reliability check
For sentiment (C), inductive category formation (A), Mayring (E), Grounded Theory (F), and deductive category formation (B), a second, independent classification by a second AI model (Mistral) runs in the background alongside the main analysis. Comparing the two models gives you:
- Cohen's Kappa (for sentiment) or Jaccard similarity (for multi-label category assignment) as an agreement measure
- A list of the divergent cases where the two models disagreed
This works like a second, independent coder's perspective — similar to an inter-rater reliability check with two human coders, just automated.
For Grounded Theory (F), there's an additional core category comparison (selective coding): both models independently propose a core category, and kondeor.ai shows the similarity between the two descriptions as a percentage, based on an embedding comparison. This is deliberately not a formal reliability statistic like Kappa or Jaccard — just an indicator of how closely the two proposed theoretical foci align. It runs independently of the rest of the category matching, even if that isn't possible for some reason.
For larger case counts (roughly 250 text units in total or more): the check now runs automatically in the background while you can already continue working with the main analysis. A brief status line in the chat shows progress; the result (Kappa/Jaccard, plus the divergent-cases card if applicable, see below) appears automatically as a new message once the check is done — usually within a minute or a few minutes, depending on case count. A “⏹ Cancel” control right next to the status line lets you stop the background run at any time.
Interactive review & category export
After the reliability check, you can review divergent cases directly in an interactive card and decide on each one — cases where both models agreed are adopted automatically. At the end, you add the result to the dataset as new column(s) with a single click — no more download detour needed (see next chapter).
This recently started working even when you analysed multiple text columns together (selecting “ALL” or naming several columns): if a divergent case comes from one of several analysed columns, the card shows a small column tag on each case so it's clear which original question the answer came from. With only one analysed column, the card looks exactly as before, with no extra tag.
Export with multiple text columns: kondeor.ai creates a separate column per category and source column (e.g. kat_satisfaction__question1 and kat_satisfaction__question2) instead of blending both answers into one column — after all, a respondent might be categorised differently in question 1 than in question 2. With only one analysed column, the column naming stays exactly as before (kat_satisfaction, no suffix).
Missing answers are correctly exported as a true missing value (NA), not as “0” — an empty cell in question 2 doesn't mean the respondent is “not in this category”, it simply means they didn't answer that question. At kondeor.ai, “0” always means: the answer was reviewed and is confirmed not to be in this category.
Writing categories back into the dataset
Clicking “Add to dataset” writes the new kat_* columns directly into this session's working dataset — no more downloading and re-uploading. The new columns are immediately usable in any follow-up analysis (e.g. a cross-tabulation right afterwards) and are automatically included the next time you use “Export dataset”.
If you want to categorise several open-ended questions one after another, simply run the next text analysis and add it too — all kat_* columns accumulate in the same working dataset (alternatively: analyse several columns together in one pass, see above — then a single “Add to dataset” click is enough).
The new category columns then behave like any other nominal variable — e.g. usable for cross-tabulations or group comparisons.
Multiple response set: when you add the categories, kondeor.ai automatically bundles the kat_* columns from one categorisation into a so-called multiple response set (SPSS's equivalent: “Define Multiple Response Sets”) — with several open-ended questions in the same chat, one set is created per question. For SPSS users, an additional toolbar button then appears offering a small syntax file (.sps) that creates the set once you run it in SPSS.
Technical note: SPSS files can't receive a multiple response set directly during export (a limitation of the underlying library, independent of kondeor.ai) — the syntax file needs to be run once in SPSS. To anchor the set permanently in the .sav file, re-save the file in SPSS afterwards.
Frequency analysis directly in kondeor.ai: instead of going through SPSS, just ask in the chat — e.g. “Analyse multiple responses” or “Frequency of multiple responses”. kondeor.ai returns a table with count, % of responses, and % of cases per category (SPSS convention: “cases” refers to everyone with at least one valid answer in the set, not the overall N). If several sets exist in the same chat, you can name the relevant open-ended question in your request (e.g. “frequency for satisfaction”); without naming one, all existing sets are shown. This is a plain count with no AI call involved — it doesn't cost any credits.
The frequency tables — weighted ones included — have a Visualisation tab with a bar chart in the result dashboard.
Reports & Export
Collecting, designing, and sharing your results.
Saving a result to a report
Under every result (quantitative or qualitative) you'll find the “Save to report” button. This lets you:
- Add the analysis to an existing report, or create a new report
- When creating a new report, optionally choose your own brand profile (see D1b) — without one, the report appears in the standard kondeor.ai design
- Optionally add a title for this specific analysis and an analyst note (a hint for the client)
A report collects multiple analyses over time this way — ideal for having a complete write-up in one place at the end of a project.
Besides analyses, charts you create yourself can also be placed in a report — see D1c. How to structure a report with headings, texts and page breaks is described in D1d.
Reports (individual and combined) are currently available from the Basis plan and up.
Custom brand profiles for reports
Instead of the standard kondeor.ai design, reports and combined reports can use a custom brand profile — with your own logo, your own header/footer text, and a light or dark colour scheme for the header and footer. The rest of the report content (tables, charts) is deliberately kept consistent.
- A brand profile is created once and can then be applied to any number of your own reports and combined reports — you can create as many profiles as you like (e.g. one per client)
- A later change to the profile (e.g. a new logo) takes effect immediately on all reports that use it
- Without a logo, a cleanly typeset wordmark using your entered header text is shown instead
- The “powered by kondeor.ai” note in the footer can optionally be hidden for a one-time additional fee
Manage profiles via “🎨 Brand profiles” next to “My Reports”. Creating a brand profile, as well as hiding the “powered by” note, costs credits (a one-time cost per profile, reusable without limit afterwards) — you'll see the exact amount directly in the app when creating one.
Creating charts
Classic reports need more than analysis results — they also need charts such as bar, pie or line charts. kondeor.ai creates them straight from the active dataset — using the currently active subsample and, if you wish, weighting — and places them in a report. You open the chart dialog with the small bar-chart icon in the footer icon bar (“Create chart”) or by writing a message in the chat (see below).
Chart types
Bar (vertical or horizontal), stacked bars, stacked bars at 100%, pie, donut, line, mean with confidence interval, histogram, box plot and scatter plot.
How a chart is made
- From dataset: choose the chart type, then the variables (category or x-axis; for means and distributions also the numeric variable; optionally a grouping and a weight) and the measure: frequency, percent (with a selectable percent base), mean, median, sum or standard deviation. Clicking “Calculate chart” has R compute the values; the preview shows the chart immediately. The numbers in the chart therefore come from the same calculation as every other result — the browser only draws.
- Manual entry: type values into a small table or paste them from Excel (first row = names of the data series, first column = categories). Available for bar, stacked bar, pie, donut and line charts. Manual entry calculates nothing and costs nothing.
- Appearance (free): colours, axis titles, legend, grid lines, showing values, decimals, sorting, and swapping axes can be changed as often as you like without recalculating.
Subsample and weighting
If a subsample is active in the dataset, the chart uses only those cases; the dialog names the subsample, and it is stored with the chart so that a later recalculation in the report uses the same cases. If you choose a weight variable, frequencies are sums of weights, means are weighted and confidence intervals take the effective sample size into account; in the report the chart is marked as weighted. If a global weighting is active, it is proposed automatically for bar, line, pie, mean and histogram charts (deselect with “— none —”); for box plots and scatter plots, which do not support weights, a note states explicitly that the chart is drawn unweighted.
Requesting a chart in the chat
Instead of the dialog, a message is enough, for example “Create a bar chart of the frequency by gender” or “Draw a histogram of the score”. kondeor.ai then proposes a chart and shows a confirmation card; only “Create chart” opens the dialog, where variables, chart type, colours and labels can still be changed. Calculation and charging happen in the dialog only. As soon as the question is really about an analysis (for example it asks about significance or a group difference), it runs as an analysis as usual, even if the word “chart” appears in it.
Print-ready (R)
For the scatter plot, box plot and histogram there is also a print-ready version drawn by R (“Draw with R”): the scatter plot with regression line, confidence band, r and n; the box plot optionally with individual points, mean and case numbers; the histogram optionally with a density or normal curve. Format (PNG, SVG or PDF), size in centimetres and resolution can be set, and a title in the image is optional. The print-ready version can be downloaded or placed in a report as an image; once the chart has been calculated it costs nothing extra.
Adding it to a report — and changing it later
- “Save to report” asks, as with the other building blocks, for a heading and a note that appear above and below the chart in the report.
- The chart stays editable: under “My Reports”, “Edit chart” reopens the dialog. Appearance, colours and labels can always be changed. To use a different variable, the chart's dataset must be loaded (“Load dataset” in the dialog); the chart is then recalculated.
- Charts from analysis dashboards (for example a box plot under a t-test) can be placed in a report via “Add to report (as image)” in the export window of that chart. They are stored as an image and cannot be changed afterwards.
Chart creation is available from the Basis plan.
Structuring a report
A report consists of entries: analyses, charts, texts, videos. Under “My reports” you can build it up like a document — with headings, your own texts and page breaks.
Inserting blocks
- At the top of the report list there are “Block at start” and “Block at end”; every entry also has “+ Block after”. This lets you insert at any position.
- Heading: as a chapter (level 1) or sub-chapter (level 2).
- Text: your own paragraph or section in Markdown (headings with #, lists with -, **bold**, tables), with a preview. It appears in the report as you write it.
- Methods section: kondeor.ai assembles a “Methods” section from the information in your report — dataset, sample size, subsample, and the procedures used in the order they appear. Nothing is calculated and no AI writes it; it only lists existing information. What kondeor.ai cannot know (survey, period, significance level, handling of missing values) appears as [placeholders in square brackets]. The section is ordinary text and can be extended and edited freely.
- Page break: when printing or saving as PDF, the following content starts on a new page.
- Blocks can be edited, moved by drag & drop and deleted. kondeor.ai asks before deleting; when a heading is deleted, the entries below it stay in the report.
Chapters and table of contents
Chapters follow from the order: a chapter heading (level 1) covers all entries up to the next one. The numbering (1, 1.1, 2 …) and the table of contents in the report’s left sidebar are created automatically and adjust when you move things. Dragging a chapter heading moves the whole chapter; “Collapse chapter” only hides the entries in the list temporarily (the view is not saved). Blocks do not shift the numbers of the analyses.
Charts when saving a batch analysis
In the “Save entire batch analysis” window, kondeor.ai proposes a chart for suitable analyses: for group comparisons means with confidence intervals (medians for Mann-Whitney and Kruskal-Wallis), for chi-square and Fisher 100 % stacked bars, for correlation and simple regression a scatter plot, for a descriptive analysis a histogram or bars. The proposals follow fixed rules; no AI is involved. You tick the ones you want (nothing is ticked by default); the price line updates as you go. Each chosen chart is saved as its own entry directly behind its analysis. There is no proposal for analyses with their own subsample, with more than two variables, or for paired procedures.
Organising a report with AI
For reports with at least three entries, the AI proposes an outline at the push of a button (“✨ Organise report (AI)”): chapter and sub-chapter headings and a sensible order of the entries. Contents are not changed; charts stay behind their analysis, and the position of the final interpretation stays at the end.
- You first see the proposal as a numbered preview; only “Apply” changes the report. Entries the AI did not place are listed under “Further entries” — nothing gets lost.
- Existing headings and page breaks are replaced by the new outline. Right after applying, “Undo” restores everything (as long as nothing else was changed in the report).
- Cost: credits are charged only for the proposal, and only if it succeeds. Applying and undoing are free. You see the exact number in the window before you start.
Inserting and editing blocks is part of reports (Basis plan and above); organising with AI is available from the Basis plan.
Viewing & exporting the individual report
A saved report (with all the analyses collected in it) can be viewed as a whole and exported as an HTML file — including interactive charts. For a print version or PDF, the report itself has a “🖨 Print / Save as PDF” button that uses your browser's print function with a layout specifically optimised for it (A4, clean page breaks).
A report can also be shared via a link — recipients then see a public, readable view without needing their own kondeor.ai account. If you would rather pass the report on to other kondeor.ai users as an editable copy, use “Share” (see A4d).
For publication: In addition to the overall interpretation, kondeor.ai can turn it into a press text (trade press or general media) or a social media text (Pro, credits per text). The figures in the text are checked against the report; quotes, names and contact details remain placeholders for you to fill in, and for exploratory results the text says so. “Add to report” inserts the text as a chapter at the end of the report.
Combining several reports into a combined report
Several already published individual reports can be combined into a combined report — e.g. to bundle multiple survey waves or sub-studies of a project under one shared link. Each included report appears in it as its own, renameable tab.
- Create a new combined report via “🗄️ Combined reports” (next to “My Reports”) and give it a name
- Add any number of your own published individual reports as tabs — only published reports are offered, so a shared combined report never links to a dead tab
- Reorder tabs via drag & drop and rename them individually
- The combined report gets its own, separate share link — the individual reports remain separately accessible via their own links as well
Just like the individual report, a shared “⬇ Save as HTML file” and “🖨 Print / Save as PDF” button covers all tabs at once.
If an included individual report is later set back to draft, its tab disappears from the published combined report automatically (instead of showing a broken link) — it reappears once the individual report is published again.
The management view also shows how often the combined report was opened in total, and how often each individual tab was actually clicked (not just loaded) — a simple counter, no detailed history of individual views.
Video summary
Instead of (or in addition to) text, kondeor.ai can also generate a result summary as a short presenter video — for example for a presentation in a course or in front of management. A “🎬” button appears in several places:
- After a theory answer — summarises the essence of the explained concept
- After an analysis — answers the question that led to the analysis directly
- After an analysis and its related theory request — combines both into one video
- On a project report or batch report — synthesises several analyses into one summary
Every video is designed for around 3 minutes by default — two dropdown fields next to the “🎬” button let you choose duration (30 sec. to 3 min. in five steps) and format (16:9 landscape or 9:16 portrait, e.g. for Instagram) individually. Since a presenter avatar handles formulas, exact statistical values, and Greek letters poorly, kondeor.ai rephrases the content specifically for this — results are described qualitatively (e.g. “a small, statistically significant difference”) rather than read out as a number.
Before the actual video is generated, kondeor.ai shows you the generated presenter script in an editable text field — you can adjust the wording before video creation starts.
A finished video can be downloaded with one click, or saved to your personal “Videos” folder (visible only to you). Important: the video link is only valid for about 7 days via the provider — download the video if you want to keep it permanently; saving the link alone is not enough.
In a saved project report, you can additionally add a YouTube link (your own upload, or a link to a displayed kondeor.ai video), which is then embedded directly in the report.
Video usage draws on your current credit allowance (see Plan comparison) — a video costs noticeably more than a single analysis, since it's generated via an external provider.
Account, Plans & Legal
What each plan offers, how payment and cancellation work, and where to find the legal documents.
Plan comparison
| Free | Basis | Pro | |
|---|---|---|---|
| Upload your own data | – | ✓ | ✓ |
| Standard methods at bachelor's level (t-test, ANOVA, regression, factor analysis, mediation, moderation and more, see B0) | – | ✓ | ✓ |
| Advanced methods (e.g. MANOVA, ordinal regression, Tobit regression, multilevel analysis, SEM, multiple imputation, see B0) | – | – | ✓ |
| Create reports (individual/combined) | – | ✓ | ✓ |
| Practice dataset & theory questions | ✓ | ✓ | ✓ |
| Loading additional features … | |||
Every plan includes a monthly credit allowance (rolling 30-day window from your first login, no rollover into the next cycle). Every analysis consumes credits — how many depends on the method: a simple analysis (e.g. t-test, correlation) costs little, more elaborate procedures (e.g. regression, factor analysis) cost more, and computationally intensive methods with cross-validation or a broad model search (e.g. random forest, regularization, time series analysis) or a qualitative text analysis cost the most. An individual report (creating it, adding analyses, viewing, exporting) costs nothing; a combined report (see D1b) and creating a custom brand profile, on the other hand, cost credits — you'll see the exact numbers directly in the app when creating one. Re-displaying already-computed results also costs nothing. Creating a chart from the dataset (see D1c) also costs credits, but only for a new data request — changes to colours, labels and appearance are free. The AI proposal for the outline of a report (see D1d) also costs credits; applying it does not. The credits actually charged are shown directly with each analysis result; your current balance is shown by the icon in the top header, and the exact allowances and prices per plan are always available, up to date, under “Upgrade plan” in the app. What a single function costs is shown where you trigger it (on the button or in the confirmation); in Help mode (see A7) you can also ask “What does … cost?”. The manual deliberately names no credit figures, because prices can change.
One exception to the “no rollover” rule: if you switch between plans (upgrade or downgrade), any allowance still remaining at that moment isn't lost — it carries over into the new 30-day cycle that begins with the switch. From that new cycle onward, the normal rule applies again: anything unused expires at the next cycle's end, just like any regular monthly allowance. Purchased credits (see E5) are unaffected by this and remain valid until your plan ends regardless.
Invoices & payment
Payments are processed via Mollie (SEPA, credit card, and others); invoices are generated automatically and available for download in your account. Businesses with a valid EU VAT ID from abroad automatically benefit from the reverse-charge procedure (0% VAT, verified via VIES) — private individuals and Austrian businesses pay the standard rate of 20%.
Cancellation vs. withdrawal
These are two legally distinct things:
| What happens | |
|---|---|
| Cancellation | Your subscription ends at the end of the already-paid period. Access remains fully active until then; there is no refund for the current period. |
| Withdrawal | Within 14 days of concluding the contract, you as a consumer may withdraw from it — with a refund. If you consent to immediate performance during checkout, you lose this right early (see the checkbox during the order process). |
For details, see § 12 (Termination) and § 13 (Right of Withdrawal) of the Terms and Conditions.
Both are different from deleting your account (E7): cancelling ends your subscription but leaves your account and your data in place.
Terms, privacy & data processing
The complete legal documents (Terms and Conditions, Privacy Policy, Data Processing Agreement) are collected on a dedicated page, each available in German and English:
If the European servers are briefly unreachable, we ask you whether the task may instead be carried out via a provider outside the EU — this only happens with your explicit consent. If you want no AI at all, or no providers outside the EU, set this in your account — see A5c.
Your uploaded data set is temporarily stored server-side, in encrypted form, at our hosting partner Mittwald (Germany) during the session, and is automatically deleted after 12 hours of inactivity (see G3). Datasets placed in the workspace on the Pro plan are deleted after 30 days without use (A4c). See Privacy Policy, points 3.3 and 3.6, for details.
AI-generated text (e.g. interpretations, reports, video scripts) may contain a watermark, imperceptible to humans, used for machine-readable origin verification — content and readability are unaffected. See Privacy Policy, point 3.8, for details.
The statistical calculations are based on R and freely available open-source packages (CRAN). All packages are free to use; there is no legal citation requirement. Overview, grouped by license type:
| License | Packages |
|---|---|
| Loading package list … | |
Buying additional credits
If your monthly credit allowance ever runs short, you can buy additional credits via the “+ Credits” button in the header — regardless of your plan, including the free Free plan — one flexible package whose quantity you choose yourself; you see size and price in the purchase dialog.
The key difference from your monthly allowance: purchased credits don't expire at the end of the cycle — they remain valid regardless of cancellation or a plan change, and only expire if you delete your account. Every analysis draws on your regular monthly allowance first; only once that's used up do purchased credits come into play. Which analysis methods you can use, though, always depends on your current plan, regardless of your credit balance.
Your billing address is pre-filled from your most recent invoice when you buy — you can override it at any time, for example with a new VAT ID or a changed address. Just like a plan purchase, every credit purchase generates a regular invoice.
Password, signing out & language
Forgotten or changing your password: the “Forgot / change password?” link in the sign-in dialog takes you to a field for your email address. You'll then receive an email containing a link that stays valid for one hour; use it to set a new password (at least 8 characters). This is also the route to take if you simply want to change your password deliberately — there is no separate form for that.
For security reasons you always get the same confirmation, whether or not an account exists for the address you entered. That way the form can't be used to find out who is registered with kondeor.ai. If no email arrives, please check your spam folder and the spelling of the address.
Signing out: the “Sign out” button at the top right of the header ends your session immediately. On top of that you are signed out automatically after 12 hours (see G3 · Session expired).
Switching language: the “DE”/“EN” button in the header switches between German and English — both the interface and the language your questions are answered in. Your choice is stored in your browser and reused on your next visit. On a first visit the language follows your browser setting. If kondeor.ai is embedded in or linked from another page, the language can also be set directly via the address suffix ?lang=de or ?lang=en.
This manual is available in both languages too — the switch for it sits at the top left of the sidebar.
Deleting your account
You can delete your account yourself at any time, no questions asked: open “Info”, then click “Delete account” at the bottom left of the window. You'll be asked for your password once more — that protects you from someone else removing your account at an unattended screen.
What gets deleted: your uploaded datasets, saved analyses and projects, your question history, your credit history, plus course enrolments and learning progress. This happens immediately and cannot be undone — there is no recovery window. So export anything you want to keep beforehand (see D3).
What is kept: your invoices. Austrian law requires them to be retained for seven years (§ 132 BAO), which expressly takes precedence over the right to erasure under Art. 17 GDPR (Art. 17(3)(b)). They are stored for that purpose only and used for nothing else.
You'll receive a confirmation by email afterwards. You're welcome to register again with the same address at any time — it starts as a completely new account, so old content does not come back.
Self-service deletion is blocked in two cases: while a subscription is still running, please cancel it first (see E3) and delete the account once the paid period has ended — that way deletion never triggers a payment. And if your account owns courses with enrolled participants, please contact office@kondeor.ai so the courses can be handed over first.
Courses (E-Learning)
Structured learning units with AI-guided learning steps and exams — separate from statistical analysis.
Course catalog & enrolment
The “Courses” button shows all public, active courses — with the number of units, access duration (in months from enrolment), and creator. Free courses can be enrolled in with one click; paid courses lead to a payment via Mollie (one-off payment, no subscription).
Courses you're already enrolled in or have purchased are marked accordingly in the catalog and lead straight to “My Courses” instead of another payment.
My Courses: units & progress
“My Courses” shows all courses you're enrolled in, each broken down into units. Every unit has a status:
- Locked — not yet reachable, unlocks once the previous unit is passed
- Open — ready to learn
- Passed / Failed — after completing the exam, including the score achieved; if failed, another attempt is possible
Units unlock linearly — the first unit is open immediately after enrolment, every further unit only after passing the exam of the previous one.
Whether units unlock in order is set by the instructors per course (“Enforce order”). With the order switched off, all units are accessible immediately.
Learning steps
Within an open unit, kondeor.ai first creates a structured learning plan (3–5 steps, tailored to the unit and an optional hint from the course instructor) and then guides you through it step by step. Each step provides a detailed explanation (roughly 200–350 words) with everyday examples, citing the course materials where possible — the same curated knowledge base used in normal theory mode (see A6). Relevant videos can be shown alongside each step.
Exam per unit
After the learning steps comes an exam — the questions can come from up to three sources at once, depending on how the instructor set up the unit: the instructor's own question pool, a curated question database (matched to the chapter), and/or AI-generated questions. Multiple-choice, true/false, and free-text questions are supported — free-text answers are graded by AI, including brief feedback on why an answer was right or wrong.
The minimum score needed to pass is set per unit by the instructor. If you fail, the unit stays open for another attempt; if you pass, the next unit unlocks automatically.
Statistical methods in a course (Free plan exception)
If you are enrolled in at least one active course, you can use all Basis-level methods (t-test, ANOVA, regression and more, see E1) with the unchanged GBOO practice dataset — even if you are on the Free plan yourself. The exception applies only to the practice dataset; as soon as you upload your own data or edit the practice dataset (change a cell, transform, filter), the regular Free-plan limits apply again for the rest of the session.
A brief note about this appears once in the chat, the first time the exception actually unlocks a method that would otherwise be locked on the Free plan.
The exception ends automatically once your course enrolment expires — after that, your regular plan applies again (Free, Basis, or Pro), with whatever credit allowance you had already used by then. To use Basis-level methods beyond the course, or with your own data, an upgrade to the Basis plan is required.
Board in the learning unit
Besides Learn, Ask kondeor.ai and Test questions, every learning unit has a Board tab — a shared pinboard for the course. The board also works with the “AI off” setting.
The tab only appears for task units or if the instructor has ticked “Show board” in the unit editor.
For participants: Write a card (up to 280 characters, send with Ctrl+Enter). It first appears in the “New / unsorted” column; everyone else sees it within a few seconds. You can edit or delete your own cards. Up to 30 own cards per learning unit are possible.
For instructors: Create, rename and reorder columns (template: Examples / Questions / Unclear). Move cards by drag & drop, or on a phone via “Move to”; cards can be hidden or deleted. In the toolbar you set per course whether names are shown as first name/initials or not at all, and whether each person gets their own colour.
Display name in course: Above the card list you can set your own name to appear on your cards. It takes precedence over your account name; if left empty, your first name or “Participant N” is used.
Note: even with names hidden, a card is internally linked to your account. If the unit or course is deleted, its board is deleted too. Please do not enter personal data of third parties.
Task unit
Besides units with learning steps and an exam, instructors can create a task unit. It consists of a task description (longer text, links are possible; file upload is not provided). There are no test questions and no passing threshold.
For participants: Read the task and enter your answer in the Board tab (see F6). Your first own card completes the unit and unlocks the next one. If the task does not require a board entry, finish it with the “Done” button.
For instructors: The “+ Task” button next to “+ Unit” creates a task unit directly. In the unit editor choose the type “Task” and enter the task description (up to 4,000 characters). In the instructor panel, individual courses can be collapsed and expanded to keep an overview when you have several courses.
Progress is recorded as “Passed” without a score. Instructors and admins can also complete the unit this way to test the unlocking.
Gender-sensitive wording in the task: Instructors can use placeholders of the form {weiblich|männlich} (feminine|masculine form of a German word) in the task description. Which form appears depends on the optional w/m/d entry in the participant upload; it is used for this purpose only and is not displayed or passed on anywhere else. Using “Preview as” above the task, instructors and admins can see how the text looks for w, m and d/unknown.
When Something Goes Wrong
The most common pitfalls — and how to get around them.
Question gets misunderstood
Rephrasing with clear context usually helps:
- I mean a new question – please ignore the last analysis.
- Refer to the regression analysis just performed.
If it's specifically about using kondeor.ai (not statistics or your data), switching to help mode (mode button next to Analysis/Theory) often helps too — see A7.
Tip: if a question was misunderstood, try rephrasing it rather than resubmitting it word for word — an identical repeat often triggers the same misunderstanding again.
Free-text column not recognised as text
If an open-ended question is instead treated as a frequency count (each answer appears as its own “category”), this is usually because the request wasn't clearly recognised as a text analysis. Explicitly mention one of the following keywords in your question:
- text analysis, qualitative, content analysis, categories, code, sentiment, open-ended answer, free text, Grounded Theory
Example: instead of “What should be improved?” try “Run a text analysis: What should be improved according to the open-ended answers?”
Session expired
After 12 hours of inactivity, you're automatically logged out for security reasons — this also automatically deletes any uploaded data and unsaved changes. Just upload a dataset again to start a new session; signing in again is possible immediately. Datasets in the workspace (Pro, A4c) and under “My Datasets” (A4b) are kept.
During the session, your data set is stored exclusively server-side, in encrypted form, at our hosting partner Mittwald (Germany) — not with an external provider outside the EU.
Further help
Can't find what you're looking for? Write to us directly: office@kondeor.ai