StatGuide.AIAI analysis platform
Statistical analysis platform for researchers

Publication-ready statistical analysis without the consulting price tag.

Model selection, diagnostics, fallback handling, and manuscript-ready output — one workflow, at a fraction of what statistical consulting costs.

  • Describe your question and data — the system plans the right model
  • Diagnostics and fallbacks run automatically, every decision logged
  • Get manuscript-ready methods, results, and caveats
See the full walkthrough →

Best fit for project-based research

Start with On-Demand credits

Use pay-per-analysis credits when you need them. Move to a subscription only if your workload becomes recurring.

See pricing →
The problem

The typical workflow — and why it breaks.

Most researchers don't struggle because they're incapable. They struggle because the workflow is fragmented.

Without StatGuide.AI

  • Figure out which model fits your question and data
  • Run it — hit errors, warnings, or failed diagnostics
  • Look it up, fix it, rerun — repeat until something works
  • Build tables, figures, and manuscript sections manually
  • Track versions, decisions, and outputs across scattered files

Can take days or weeks. Two months later, you can't reconstruct what you did.

Scattered files. No paper trail. Start over.

With StatGuide.AI

  • Selects the right model from your question and data structure
  • Runs the full analysis in a single governed pipeline
  • Handles errors and failures; applies valid fallbacks automatically
  • Generates manuscript-ready tables, figures, and report
  • Fixed seed, pinned packages, full decision log — every time

Takes just minutes to run. Years later, every result is exactly reproducible.

Fixed seed, pinned packages, full decision log.

The workflow

Seven guided steps — from dataset to publication-ready results.

Step 1 of 7View in demo →

Catch problems early

  • Instant profiling

    1

    Upload once. The system detects types, missingness, and structural risks automatically.

    Rows

    842

    Columns

    8

    Event rate

    12%

    age1.2% missing
    baseline_score3.6% missing

  • Flag risks

    2

    Low event rate, too many predictors for your N, or missing data — flagged before analysis begins.

    warning

    With 101 events and 7 predictors (including an 8-level site variable), events per variable (EPV) may be borderline for complex models. The system will calculate the exact events-per-parameter (EPP) ratio once you confirm variable encodings in step 2.

  • Encoding proposed

    3

    The system auto-proposes how each variable will be encoded — ready for you to review in step 2.

    treatment_group

    Binary contrast: A = reference (0), B = 1

    sex

    Binary contrast: M = reference (0), F = 1

    site

    One-hot encoding: S1 = reference; 7 dummy variables

    age

    Retain continuous; centre at median (52 y)

    baseline_score

    Retain continuous; centre at mean (49.1)

    comorbidity_index

    Retain as ordered count (0–6)

Swipe to see all 3 · 1 of 3
Step 2 of 7View in demo →

Teach the AI your data

  • Context matters

    1

    Descriptions and units tell the AI what each variable means and how to interpret the model output it produces.

    survival_90dbinary

    Display name

    90-day Survival

    Unit

    —

    Description

    Whether the patient survived to 90 days post-treatment initiation
    treatment_groupcategorical

    Display name

    Treatment Group

    Unit

    —

    Description

    Assigned treatment arm at randomisation
    agecontinuous

    Display name

    Age

    Unit

    years

    Description

    Patient age at enrollment

  • Roles, not formulas

    2

    Assign each variable's role in the analysis. The system builds the correct model specification from your assignments.

    survival_90d
    Variable type
    Outcome ▾
    Outcome
    Exposure
    Covariate
    Subject ID
    Cluster
    Row ID
    treatment_group
    Variable type
    Exposure ▾
    age
    Variable type
    Covariate ▾
    site
    Variable type
    Cluster ▾

  • Encoding you control

    3

    Reference levels and category labels carry through to every coefficient table, figure, and paragraph.

    treatment_group
    Categorical ▾
    ▲

    Reference level

    A — Standard care ▾

    Level labels

    A (ref)
    Standard care
    B
    Intervention
    age
    Continuous ▾
    baseline_score
    Continuous ▾
Swipe to see all 3 · 1 of 3
Step 3 of 7View in demo →

Define your research question and hypotheses

  • Plain language

    1

    Ask your question in plain English. The AI formalises a full model plan from it.

    Research question

    Does the active intervention (Group B) improve 90-day survival compared to standard care, adjusting for age, sex, baseline clinical status, comorbidity burden, and enrollment site?

  • Hypothesis

    2

    State what you expect to find, or leave it empty for exploratory modeling.

    Hypothesis

    Group B improves 90-day survival over standard care (Group A).

    Include direction of effect if known. Leave empty for exploratory modeling.

  • Interactions

    3

    Describe in plain language if one variable's effect depends on another. The AI formalises it.

    Suspected interactions

    I suspect the treatment benefit may differ depending on how unwell patients were at baseline — sicker patients at entry might respond differently to the intervention.

Swipe to see all 3 · 1 of 3
Step 4 of 7View in demo →

Statistical rigor, without the PhD.

  • The complete plan

    1

    The AI proposes the full model plan — family, formula, diagnostics, and fallbacks — from everything you've defined.

    Model selected

    Logistic regression (binomial family, logit link)

    Binary outcome (0/1) with ~12% event rate — logistic regression is the canonical choice.
    Interaction term treatment_group × baseline_score included per protocol hypothesis (does benefit differ by baseline severity?).

  • Fallbacks planned

    2

    Convergence issues, violated assumptions, influential outliers — the plan defines the fallback before any data is fit.

    Separation detected → refit with Firth-penalised logistic regression (logistf / statsmodels penalized).
    Max VIF > 10 → flag, suspend interaction term, re-evaluate collinear predictor.
    EPP = 7.8 — advisory flag raised (below recommended threshold of 10). Analysis proceeds with caution; interpret coefficient estimates carefully.

  • Verify, then run

    3

    Check that the model type, outcome, and covariates match your intent. Regenerate any time your question or variables change.

    LockedSeed 42 · machine-readable plan
    {
  "schema_version": "statguide-plan/1.0",
  "created_at": "2024-03-14T09:12:05Z",
  "seed": 42,
  "dataset": {
    "file": "trial_cohort_v3.csv",
    "n": 842
  },
  "outcome": {
    "variable": "survival_90d",
    "family": "binomial",
    "link": "logit"
  },
  "formula": "survival_90d ~ treatment_group + age + sex + baseline_score + site + comorbidity_index + treatment_group:baseline_score",
  "missing_data": {
    "strategy": "complete_case",
    "sensitivity": "multiple_imputation_m20"
  },
  "diagnostics": [
    "separation",
    "vif",
    "hosmer_lemeshow",
    "roc_auc",
    "calibration",
    "influence"
  ],
  "safeguards": {
    "separation": "firth_penalized",
    "vif_threshold": 10,
    "epv_minimum": 5
  },
  "outputs": [
    "coef_table",
    "roc_curve",
    "calibration_plot",
    "forest_plot",
    "report_html"
  ]
}
Swipe to see all 3 · 1 of 3
Step 5 of 7View in demo →

Automated execution — assumptions checked, fallbacks applied

  • No copy-paste between tools

    1

    Design matrix, fitting, diagnostics, figures, and report text run in a single audited session.

    Preparing design matrix

    Fitting logistic regression model

    Running diagnostics

    Generating figures

    Drafting manuscript text

  • Fallbacks applied for you

    2

    Model assumptions violated? The system applies a valid fallback or flags the issue immediately.

    Diagnostic pass · fallback ready

    SeparationCLEARfallback: Firth
    VIF (max)2.48threshold 10
    EPP7.8advisory < 10
    HL p-value0.39PASS

  • Reproducible by design

    3

    Random seed and package versions are fixed at run time. Run again months later: same numbers.

    09:12:08

    Plan validated. Proceeding with complete-case analysis (n = 820).

    09:12:09

    Design matrix constructed: 820 × 14 (1 outcome, 13 predictors incl. dummies).

    09:12:11

    Separation check passed. Proceeding with standard MLE.

    09:12:13

    Model converged in 7 iterations. Log-likelihood: −312.7.

    09:12:15

    VIF calculated. Max VIF: 2.48 (treatment_group:baseline_score). PASS.

Swipe to see all 3 · 1 of 3
Step 6 of 7View in demo →

Results and diagnostics — the full picture

  • Trust indicators

    1

    A confidence badge (Green / Yellow / Red) summarises model quality in plain language.

    Moderate confidence

    Model fit is adequate (AUC 0.781, HL p = 0.39, Nagelkerke R² 0.241). Two caveats flagged: EPP below recommended threshold (7.8) and 2.6% missing data on key predictors. Primary conclusions are robust; a multiple imputation sensitivity run is recommended before finalizing.

    AUC

    0.781

    HL p

    0.392

    R²

    0.241

    n

    820

  • No silent failures

    2

    VIF, Cook's D, calibration, and HL test shown alongside coefficients — nothing hidden.

    VIF · multicollinearity check

    treatment_group
    1.42
    age
    1.18
    sex
    1.09
    baseline_score
    2.31
    comorbidity_index
    1.07
    site
    1.22
    treatment_group:baseline_score
    2.48

  • Ready-to-cite bundle

    3

    Coefficients, fit statistics, and diagnostic plots export as one citation-ready package.

    ROC · AUC 0.781

    0.00.20.40.60.81.00.00.20.40.60.81.0AUC = 0.781False Positive RateTrue Positive Rate
Swipe to see all 3 · 1 of 3
Step 7 of 7View in demo →

Submission-ready methods, results, and limitations

  • Two output modes

    1

    Switch between peer-review-ready text and an explanatory version.

    Methods

    AcademicStudent

    Binary logistic regression (binomial family, logit link) was applied to model 90-day survival probability. The primary exposure was treatment group (A = standard care, B = active intervention). Covariates included age (continuous, centered at median 52 years), sex (binary), baseline clinical score (continuous, centered at mean 49.1), enrollment site (8-level categorical, dummy-coded with S1 as reference), comorbidity index (Charlson, ordered count), and a pre-specified treatment × baseline_score interaction term. Complete-case analysis was conducted (n = 820 after exclusion of 22 records with missing age or baseline_score values). Model diagnostics comprised complete-separation detection, variance inflation factors (VIF), Hosmer–Lemeshow goodness-of-fit test (8 groups), bootstrapped ROC/AUC (1 000 replicates, seed 42), calibration plot, and Cook's distance influence analysis; no violations were identified (maximum VIF = 2.48; maximum Cook's D = 0.047). All analyses were conducted in Python 3.11.8 (statsmodels 0.14.1; scikit-learn 1.4.1; numpy 1.26.4).

  • Ready to copy

    2

    Paragraphs ready to paste directly into your manuscript.

    Results

    Copy

    The complete-case model (n = 820; 101 events, EPP = 7.8) demonstrated adequate overall fit (AUC 0.781, 95% CI 0.737–0.826; Hosmer–Lemeshow χ²(8) = 8.43, p = 0.392; Nagelkerke R² = 0.241). The active intervention (Group B) was associated with 85% higher odds of 90-day survival compared to standard care (OR 1.85, 95% CI 1.31–2.62, p < 0.001). Among continuous predictors, higher baseline clinical score independently increased survival odds by 4.2% per point (OR 1.042, 95% CI 1.021–1.063, p < 0.001), while older age (OR 0.973 per year, 95% CI 0.958–0.988, p < 0.001) and greater comorbidity burden (OR 0.734 per CCI unit, 95% CI 0.658–0.819, p < 0.001) were associated with reductions of 2.7% and 26.6% per unit, respectively. Sex was not significantly associated with outcome (OR 0.811, 95% CI 0.570–1.154, p = 0.244). The pre-specified treatment × baseline score interaction was non-significant (OR 1.018, 95% CI 0.996–1.041, p = 0.107), providing no evidence of effect modification by baseline score.

  • Caveats included

    3

    Limitations reviewers expect, written before you're asked.

    Limitations

    Complete-case analysis excluded 22 participants (2.6%) with missing baseline data; a multiple imputation sensitivity analysis (m = 20) is recommended to assess the impact on primary estimates.

    The events-per-parameter ratio (EPP = 7.8) falls below the commonly recommended minimum of 10; coefficient estimates may be unstable, and replication in an independent cohort with greater event volume is advised.

    Site was modeled as a fixed effect; a mixed-effects model with site as a random intercept may better account for between-site clustering and improve generalizability.

    The study is observational in design; unmeasured confounders cannot be excluded and causal inference is limited.

    Generalizability is restricted to clinical settings resembling the 8 enrolled hospital sites.

Swipe to see all 3 · 1 of 3

In one run, you generated

Everything you need. Nothing you have to reassemble later.

  • Model plan
  • Diagnostics
  • Valid fallback (if needed)
  • Publication-ready results text
  • Reproducible bundle (report + figures + script)
  • Stored analysis you can revisit anytime
See plans

How it compares

Not another stats tool. A governed workflow system.

Most tools help you run models. StatGuide.AI helps you run them correctly and keep everything together.

CapabilityExcel / SPSSR / PythonChatGPTStatGuide.AI
No coding required✔❌❌✔
Model planning guidance❌❌Partial✔
Automatic diagnostics❌Manual❌✔
Valid fallback methods❌ManualSuggestedAutomatic
Manuscript-ready output❌❌Partial✔
Reproducible bundle (code + seed + versions)❌Manual❌✔
End-to-end in one place❌❌❌✔

Ever tried to rerun analysis from 3 years ago and couldn't reconstruct what you did?

StatGuide.AI makes reproducibility the default, not the afterthought.

See the full walkthrough →Join early access
Get started
Early access cohort — 40% founder discount

Choose the payment model that matches your workflow

Most early users start with On-Demand credits for thesis and project work. Subscriptions are available for labs and recurring analysis volume.

Best for project-based research

On-Demand

$8/ analysis

No subscription — buy credits when you need them

No monthly minimum
  • Full feature access per analysis
  • Full diagnostics and publication-ready report
  • Academic + student report modes
  • Export bundle (HTML + figures + script)
  • Credits never expire
  • Email support

One-time credit packs, no subscription

Project

$69/ month

Focused work, thesis projects, and small studies

20 analyses/month
  • 20 analyses per month
  • Full diagnostics and publication-ready report
  • Academic + student report modes
  • Export bundle (HTML + figures + script)
  • Email support

Cancel anytime

Lab

$179/ month

Ongoing research, multiple models, full iteration

70 analyses/month
  • 70 analyses per month
  • Full diagnostics and publication-ready report
  • Academic + student report modes
  • Export bundle (HTML + figures + script)
  • Priority email support

Cancel anytime

Questions & concerns

Common questions

Everything you need to know before running your first analysis.

Is my data secure?

Yes — files are stored in a private, user-scoped cloud bucket and auto-deleted after 90 days.

Your file is uploaded directly to a private Google Cloud Storage bucket over HTTPS via a short-lived signed URL (valid for 15 minutes). Files are stored at a path scoped to your unique user ID — no other user can access your data. All endpoints that access your uploaded datasets and analysis artifacts require an authenticated session; raw files are never returned to the browser. Data is automatically deleted after 90 days, and you can delete it manually at any time from your dashboard.
Will my data be used to train AI models?

No. Raw data never leaves the analysis environment and is never used for model training.

StatGuide.AI uses AI in two narrow, well-defined ways: (1) generating a structured analysis plan from a column-level summary of your dataset, and (2) writing methods and results prose from already-computed statistics. In both cases the model receives structured JSON summaries (for example, column distributions, warnings, and computed estimates) — not your raw row-level data. Raw data is never sent to the AI provider and none of your data is used for model training.
Are the results correct?

All computation uses peer-validated Python libraries, not AI. The AI only selects models and writes prose.

All numeric computation is performed by peer-validated Python libraries (statsmodels, scipy, numpy) — not the AI. The AI only selects the model family and writes prose from the computed numbers. Every analysis includes automated assumption checks: multicollinearity (VIF), goodness-of-fit (Hosmer-Lemeshow), separation detection, and events-per-variable (EPV). When a check fails, the system applies pre-validated fallback methods (e.g., Firth penalized regression for complete separation) and logs the decision explicitly. The analysis plan is locked before execution — preventing post-hoc adjustments.
Can I reproduce the analysis later?

Yes — every run stores its seed, package versions, and a script you can rerun independently.

Reproducibility is a core design guarantee, not an afterthought. Every run stores the fixed random seed, pinned Python package versions, and a machine-readable analysis plan. Your reproducible bundle includes a generated Python script and a requirements.txt file you can run independently to reproduce every number in the output — months or years later.
What data can I upload?

CSV and Excel (.xlsx), up to 50 MB, 2M rows, 200 columns.

StatGuide.AI accepts CSV and Excel (.xlsx) files up to 50 MB, with up to 2 million rows and 200 columns. You must confirm that your dataset does not contain PHI (Protected Health Information) or HIPAA-protected data before uploading. StatGuide.AI is designed for de-identified research datasets. Personally identifiable information should be removed before upload.
What models are supported?

GLMs, mixed models, GEEs, zero-inflated and hurdle models, beta regression, and more.

StatGuide.AI currently supports: binary logistic regression, Firth penalized logistic regression (sparse outcomes / complete separation), linear regression, Poisson regression, negative binomial regression (overdispersed counts), zero-inflated Poisson and zero-inflated negative binomial (excess zeros), Hurdle models (two-part count models), beta regression (proportion outcomes bounded 0–1), generalized linear mixed models (GLMMs with random intercepts and slopes), and generalized estimating equations (GEEs for correlated/clustered data). The system automatically selects the appropriate model family and applies fallbacks based on your data — no manual configuration required.
View pricing