User manual
A short, practical guide: from installing the app to exporting your first analyzed result. Everything here works with the bundled sample datasets — no data of your own required to follow along. The full 48-analysis, 101-command reference manual ships with the application and lives at clinstat.pages.dev.
Quick start
- Install. Run the Windows installer (a single ~562 MB file) — installation finishes in about twelve seconds. On the first launch, accept the evaluation license agreement, pick your interface language, and accept the offer to set up the app's own R engine: the runtime and the 247-package library (about 1 GB) unpack from the installer with labeled phases and real byte progress. Nothing downloads unless the bundled files are missing.
- Load a sample dataset. On the welcome screen, click one of the seven bundled clinical datasets — for example the Blood Pressure Trial (n = 80). It loads instantly, and the app profiles every variable automatically.
- Run your first analysis. On the Data page, open the analysis panel and pick a test — say, a two-group t-test. Resting the pointer on a type shows a preview card explaining what it does and which inputs it needs. Assign the variables and press Run; the statistics table, assumption checks, and figures appear in seconds on the Results page.
New in 0.2.0 License key activation is built in for institutional seats (Settings → License): a signed key file activates offline, with no server and no data leaving your machine. The free evaluation tier is unchanged, and the CONSORT checklist is updated to the 2025 revision. Recent releases also brought the intent-grouped analysis picker (six groups, with a fits-only filter); competing-risks runs that co-report Gray's test and Fine-Gray regression; logistic regression with Firth's method and backward elimination; and the TRIPOD+AI reporting checklist alongside CONSORT and STROBE in the export dialog.
If you want AI interpretation as well, configure a local language model during onboarding or later under Settings → AI provider — see AI interpretation. Skipping it is fine: every statistical feature works without an AI provider.
Importing data
Choose Import Your Data on the welcome screen, Change data on the Data page, or File → Import Data with the format preselected. The format is detected from the file extension; the file is magic-byte-checked before it ever reaches R.
| Format | Extensions | Notes |
|---|---|---|
| CSV | .csv | Most robust; UTF-8 assumed |
| Excel | .xlsx, .xls | First sheet imported by default |
| Stata | .dta | Labels preserved as factor levels |
| SPSS | .sav | Value and variable labels preserved |
| SAS | .sas7bdat | Labels preserved; haven-written files accepted |
After import, the Data page shows the dataset in a fast grid with detected column types (overridable per column) and a one-card-per-variable profile: distribution, summary statistics, and missingness. Columns that look like patient identifiers are automatically excluded from analyses unless you explicitly select them.
Running analyses
ClinStat AI offers 48 analysis types across two surfaces: the 22 everyday tests (t-tests, ANOVA, chi-square, correlation, agreement and reliability, linear and logistic regression, dose-response trends) configure in the analysis panel on the Data page, and the 26 advanced types (survival, longitudinal, causal inference, predictive performance, Bayesian, imputation) configure on the Advanced analyses page. In the panel picker, the everyday types sit under six intent group headers — compare groups, relate variables, model an outcome, reliability & agreement, nonparametric & trend tests, summarize data — with a fits-only filter that narrows the list to what the current dataset supports. The Analyze and Graphs menus route into the same panels. Both selectors show a hover preview card — what the analysis does, when to use it, which inputs it needs — in your interface language.
Before a run starts, the app validates your variable choices (type compatibility, group levels) and warns about common mistakes up front. Results include the statistics table with effect sizes, an assumption-check summary, the nonparametric companion where one exists, and publication-ready figures — organized across seven tabs (Report, Data, R code, Diagnostics, Power & n, E-value, Gallery) on the Results page.
Guided flow
Prefer not to pick tests from a menu? The Guided wizard walks a structured path:
- Pick the dataset.
- Type your research question in plain language — for example, “Does the treatment reduce blood pressure compared to placebo?”
- Review the proposed analysis plan: test, variables, and why. Accept it, modify it, or rephrase the question.
- The wizard runs the analysis and shows the standard results view, with AI interpretation and export options at the end.
Each question gets exactly one proposed plan — a deliberate scope choice, so the wizard always explains a single defensible path rather than an overwhelming menu of alternatives.
Exporting results
From the export dialog you can produce:
- Word — a manuscript-ready document with a title page, a methods paragraph generated from the analysis provenance, results tables, figures, and the AI narrative if you ran it.
- Excel — a multi-sheet workbook: data, summary statistics, one row per analysis run, embedded figures, and a provenance sheet.
- Figures — PNG, or vector SVG/PDF.
- R code — a standalone, runnable
.Rfile for every analysis (see below).
Reproducibility, built in Every exported R file contains a pure analysis function plus a wrapper that reproduces the exact run you made — with the R version, package versions, random seed, and a dataset hash in its Reproducibility Appendix. Paste it into a supplement and your methods section is done.
For trials, observational studies, and prediction-model studies, CONSORT, STROBE, and TRIPOD+AI reporting checklists can be exported as editable Word documents with per-item text areas for your study.
AI interpretation
The AI layer drafts a plain-language summary of each result, a methods-section paragraph you can paste into a manuscript, and cautions — violated assumptions, small samples, effects that are statistically significant but clinically trivial.
The privacy contract The AI model you configure receives only aggregated summaries — test statistics, group means, effect sizes, variable metadata. It never receives row-level data or patient identifiers, and identifier-like columns are excluded automatically. While HIPAA mode is on (the default), endpoints on non-private cloud hosts are blocked outright.
Setting it up
ClinStat AI talks to any OpenAI-compatible endpoint, which in practice means a local model server you run yourself — LM Studio (default port 1234) and Ollama (default port 11434) are the common choices. In Settings → AI provider, create a provider profile, enter the base URL, pick a model, then test and save. An API key is optional for local servers but required up front for a new cloud profile — Discover models and Save stay unavailable until the key is entered. A key is stored in your operating system's keychain, never in a plain file. Profiles are named and switchable without retyping keys.
A smaller, fast model is the sweet spot for interactive use — it keeps interpretation quick within the 90-second request window. Very large or “reasoning” models can exceed it; use those only for occasional, publication-grade passes. And if you skip setup entirely, the app runs in standalone mode: every statistical feature works, only the narrative panel is disabled.
Troubleshooting
The app seems to start, but no window appears
A saved window position is probably off-screen — this can happen after changing monitors. Delete the window-state file in the app's data folder and relaunch; the window reappears at its default position.
The footer says “R not found”
On Windows, accept the first-run offer to set up the app's own R engine — the installer carries the runtime and the full package library, so nothing else is needed. A fresh install unpacks them on the first launch (about a minute of labeled progress); that is expected, not stuck. Otherwise run Tools → Check R Installation, or point the app at your own R 4.x with Tools → Locate Rscript.
AI interpretation fails or times out
Check that your model server is actually running, that the base URL in Settings → AI provider is right, and use Test connection. Timeouts usually mean the model is too large or a slow “reasoning” model — switch to a smaller, faster one. Everything except the narrative works while you sort it out.
License note
ClinStat AI is currently released under an evaluation license, accepted on first launch. In brief: the software is provided for evaluation use; commercial use, clinical decision-making, and redistribution are not permitted.
ClinStat AI is a statistical analysis tool. It is HIPAA-aware but not HIPAA-certified, it is not a substitute for medical judgement, IRB approval, or applicable data-protection controls, and its output must not be used for patient care.
Questions or feedback: GitHub Issues.