Command reference

Every command ChakataStat can run, with its arguments — the reference for writing syntax by hand in the Syntax editor, for a .cks script run by --run, or for driving the app from the Python client or an MCP agent.

Generated from the catalog. This page is produced by dart run tool/gen_command_reference.dart from lib/tools/tool_catalog.dart, the single source of truth for the tool surface, and a test fails if the two drift — so an argument listed here is an argument the engine accepts. Do not edit it by hand.

108 commands, by family.

• Discovering the dataset
• Descriptive statistics
• Comparing means
• Bayesian tests
• Nonparametric tests
• Correlation
• Regression and modeling
• Scale, reduction and classification
• Survival analysis
• Time series
• Evidence synthesis and resampling
• Power and sample size
• Complex samples
• Tables
• Charts
• Transforms

Reading an entry

Arguments are name=value. A variable is named bare (Age), a list is bracketed ([Age, Income]), text is quoted, and a boolean is true or false. The syntax line shows the required arguments only; the table under it lists every argument, required or not.

A ⚠️ marks an argument that writes new variables into your dataset (an analysis's "save …" option). A read-only session — the AI assistant by default, --mcp --read-only, a read-only batch run — withholds and refuses those.

Discovering the dataset

list_variables

List the dataset's variables — name, type, measurement level and label — so you know what is available to analyze. Call this first when you do not yet know the variable names.

list_variables

No arguments.

Descriptive statistics

descriptives

Descriptive statistics (N, missing, mean, std. dev., min, max, median, sum) for one or more numeric variables.

descriptives variables=[Var1, Var2]

frequencies

A frequency table of each variable's distinct values (count, percent, cumulative percent). Works on numeric or text variables.

frequencies variables=[Var1, Var2]

explore

Extended descriptives — quartiles, IQR, skewness, kurtosis — for numeric variables, optionally split by a factor.

explore variables=[Var1, Var2]

crosstabs

A contingency table of one variable against another, optionally with chi-square tests of independence (Pearson, likelihood-ratio, and Fisher's exact for a 2×2 table) and association measures (phi, Cramér's V, Goodman & Kruskal's gamma and lambda; for a matched square table, Cohen's kappa and McNemar's test). Add a layer variable for a stratified table with the Cochran/Mantel-Haenszel statistics.

crosstabs row=VarName column=VarName

missing_value_analysis

Characterizes the missingness across several numeric variables: a per-variable missing summary, the distinct missingness patterns, Little's MCAR test (is the data missing completely at random?), the EM maximum-likelihood mean/SD under missingness, and optionally EM single-imputed values saved as new variables.

missing_value_analysis variables=[Var1, Var2]

multiple_imputation

Multiple imputation with Rubin pooling: generate m completed datasets by chained equations (each incomplete variable regressed on the others, redrawing its missing entries from the Bayesian predictive distribution), run an analysis on each completed dataset, and pool the estimates and standard errors by Rubin's rules (within- plus between-imputation variance, with a Barnard-Rubin df adjustment and the fraction of missing information, FMI). Reproducible for a fixed seed. The pooled analysis is a linear regression (the variables are the outcome then the predictors) or a single variable mean.

multiple_imputation variables=[Var1, Var2]

Comparing means

means

Means (and N and std. dev.) of numeric variables across the levels of a factor.

means variables=[Var1, Var2] factor=VarName

one_sample_ttest

A one-sample t test comparing each variable's mean against a test value.

one_sample_ttest variables=[Var1, Var2] test_value=0.5

independent_ttest

An independent-samples t test comparing a variable's mean between the two groups of a grouping variable.

independent_ttest variables=[Var1, Var2] grouping=VarName

paired_ttest

A paired-samples t test comparing the means of paired variables (e.g. before vs. after), tested on each case's difference.

paired_ttest pair1=[Var1, Var2] pair2=[Var1, Var2]

equivalence_ttest

An equivalence test (TOST — two one-sided t tests) of whether a mean difference lies inside declared equivalence bounds: "not significantly different" is not "equivalent", and this procedure can support equivalence where a t test only fails to reject. Covers a one-sample / paired design (against a test value or on the case differences) and two independent samples (pooled and Welch rows). Reports both one-sided tests, the TOST p (the larger one-sided p), the 1 − 2α confidence interval of the difference and an Equivalent / Not equivalent decision at α.

equivalence_ttest variables=[Var1, Var2] bound_lower=0.5 bound_upper=0.5

anova

A one-way ANOVA of numeric variables across a factor's levels (between/within sums of squares, F and significance), optionally with a Levene homogeneity test, a Welch robust test and post-hoc comparisons.

anova variables=[Var1, Var2] factor=VarName

levene

Levene's test of homogeneity of variance for numeric variables across a factor's levels (an assumption check for t tests and ANOVA).

levene variables=[Var1, Var2] factor=VarName

Bayesian tests

bayesian_ttest

A Bayesian t test — the JZS (Jeffreys-Zellner-Siow) Bayes factor and a posterior summary for a mean, in place of a classical t test. Covers a one-sample / paired design (test against a value or on the case differences) and two independent samples. Reports BF₁₀ (evidence for a difference) and BF₀₁ (evidence for the null) — so, unlike a p-value, it can support the null — a verbal evidence category, and the posterior median and credible interval of the standardized effect size δ under a Cauchy(0, r) prior.

bayesian_ttest variables=[Var1, Var2]

bayesian_correlation

A Bayesian test of correlation — the JZS / Ly Bayes factor and a posterior for Pearson's ρ — for each pair of variables. Reports BF₁₀ (evidence that the pair is correlated) and BF₀₁ (evidence for no correlation — which a p-value cannot give), a verbal evidence category, and the posterior median and credible interval of ρ under a stretched-beta prior.

bayesian_correlation variables=[Var1, Var2]

bayesian_anova

A Bayesian one-way ANOVA — the Zellner-Siow g-prior Bayes factor for the group-means model against the no-effect null, for each dependent variable across a factor's levels. Reports BF₁₀ (evidence the factor matters) and BF₀₁ (evidence it does not — which a p-value cannot give), a verbal evidence category, and η² (the variance explained).

bayesian_anova variables=[Var1, Var2] factor=VarName

bayesian_regression

A Bayesian linear regression — the Normal-Inverse-Gamma posterior for the coefficients and residual variance, plus the regression Bayes factor for the full model against the intercept-only null. Reports each coefficient's posterior mean (the OLS estimate under the diffuse prior) and credible interval, the residual-variance posterior, R², and BF₁₀ / BF₀₁ (evidence the predictors matter, or do not).

bayesian_regression dependent=VarName predictors=[Var1, Var2]

Nonparametric tests

mann_whitney

A Mann-Whitney U test: compares a numeric variable's distribution between the two groups of a grouping variable (the nonparametric alternative to the independent-samples t test). Asymptotic, tie-corrected.

mann_whitney variables=[Var1, Var2] grouping=VarName

wilcoxon

A Wilcoxon signed-rank test: compares paired variables on their ranked differences (the nonparametric alternative to the paired-samples t test).

wilcoxon pair1=[Var1, Var2] pair2=[Var1, Var2]

sign_test

A sign test: compares paired variables using only the direction (sign) of each per-case difference — the simplest nonparametric alternative to the paired-samples t test, weaker than Wilcoxon.

sign_test pair1=[Var1, Var2] pair2=[Var1, Var2]

runs_test

A Wald-Wolfowitz runs test: dichotomizes each variable at a cut point and tests whether the resulting sequence (in case order) is random — too few or too many runs signal non-randomness.

runs_test variables=[Var1, Var2]

binomial

A binomial test: dichotomizes each variable at a cut point and tests the observed proportion in group 1 against a hypothesized proportion. Exact by default (two-tailed for a 0.5 proportion, one-tailed otherwise), with an asymptotic normal-approximation method.

binomial variables=[Var1, Var2]

kruskal_wallis

A Kruskal-Wallis test: compares a numeric variable's distribution across all levels of a grouping variable (the nonparametric alternative to one-way ANOVA). Tie-corrected.

kruskal_wallis variables=[Var1, Var2] grouping=VarName

friedman

A Friedman test: compares three or more related (repeated) measures on their within-case ranks (the nonparametric alternative to a repeated-measures ANOVA).

friedman variables=[Var1, Var2]

kendall_w

Kendall's coefficient of concordance W: the agreement among several rankers (the cases) who each rank the same set of items (one variable per item), on a 0..1 scale (1 = perfect agreement). The sibling of the Friedman test, reported with its equivalent chi-square significance.

kendall_w variables=[Var1, Var2]

cochran_q

Cochran's Q test: compares the success rate of three or more related dichotomous (0/1) variables measured on the same cases (the k-sample extension of McNemar's test, and the nonparametric counterpart of a repeated-measures ANOVA for a binary outcome).

cochran_q variables=[Var1, Var2]

jonckheere

A Jonckheere-Terpstra test: compares a numeric variable across the ordered levels of a grouping variable, testing for a monotone trend (the ordered-alternative counterpart of the Kruskal-Wallis test). Tie-corrected.

jonckheere variables=[Var1, Var2] grouping=VarName

normality

Tests of normality for numeric variables: Shapiro-Wilk and Kolmogorov-Smirnov (with Lilliefors significance). A small significance suggests the variable departs from a normal distribution.

normality variables=[Var1, Var2]

Correlation

correlate

Bivariate correlations among numeric variables (a correlation matrix with significance), using Pearson, Spearman or Kendall coefficients.

correlate variables=[Var1, Var2]

partial_correlate

Partial Pearson correlations among variables, controlling for one or more other variables.

partial_correlate variables=[Var1, Var2] controls=[Var1, Var2]

polychoric_correlation

Polychoric (or tetrachoric, for binary items) correlations: the maximum-likelihood latent correlation two ordinal/binary variables would have if each were a coarsened slice of an underlying continuous normal variable. This is the correct input to factor analysis of ordinal/Likert items. Reports the latent-correlation matrix and a pairwise table of estimates with standard errors. Integer-coded ordinal variables are expected (each distinct value is a category).

polychoric_correlation variables=[Var1, Var2]

canonical_correlation

A canonical correlation analysis between two sets of numeric variables: the canonical correlations (with their squares), the Wilks-lambda / Bartlett dimension-reduction χ² tests of whether each function's correlation and those beyond it are zero, and the standardized canonical coefficients for each set.

canonical_correlation set1=[Var1, Var2] set2=[Var1, Var2]

Regression and modeling

linear_regression

A linear (OLS) regression of a numeric dependent variable on one or more predictors (coefficients with confidence intervals and standardized betas, collinearity diagnostics, R² and significance). Supports robust standard errors and forward/backward selection.

linear_regression dependent=VarName predictors=[Var1, Var2]

logistic_regression

A binary logistic regression of a 0/1 dependent variable on one or more predictors (coefficients and odds ratios with confidence intervals). Supports robust standard errors and forward/backward selection.

logistic_regression dependent=VarName predictors=[Var1, Var2]

poisson_regression

A Poisson log-linear regression of a non-negative count dependent on one or more predictors (rate ratios Exp(B) with intervals, deviance/Pearson goodness-of-fit flagging over-dispersion). Supports robust standard errors.

poisson_regression dependent=VarName predictors=[Var1, Var2]

negative_binomial_regression

A negative-binomial (NB2) log-linear regression for over-dispersed counts: like Poisson but with an estimated dispersion parameter, so standard errors are not understated when the variance exceeds the mean.

negative_binomial_regression dependent=VarName predictors=[Var1, Var2]

zero_inflated

A zero-inflated or hurdle count model for counts with excess zeros (no-visit / no-purchase / no-defect cases) that a plain Poisson or negative-binomial fit cannot absorb. Reports a count component (rate ratios Exp(B)) and a zero component (a logit model for the excess / hurdle zeros), a Vuong test against the plain count model, and AIC/BIC. The count and zero components share the same predictors.

zero_inflated dependent=VarName predictors=[Var1, Var2]

regularized_regression

Regularized linear regression — ridge, lasso or elastic net — for prediction with many or collinear predictors and (lasso) automatic variable selection. The elastic-net penalty is fitted by coordinate descent along a path of penalties (λ), with λ chosen by k-fold cross-validation. Reports the coefficients at the selected λ (on the original and standardized scales), the regularization / CV path, and a model summary. Penalized estimates carry no standard errors.

regularized_regression dependent=VarName predictors=[Var1, Var2]

robust_regression

Robust linear regression by M-estimation (Huber or Tukey bisquare): down-weights outlying residuals in the fit itself, so a few influential points cannot drag the coefficients the way they do in OLS. Reports the robust coefficients with sandwich (H1) standard errors and t-tests, the robust scale, and the most down-weighted cases. Use when a diagnostic plot shows influential outliers.

robust_regression dependent=VarName predictors=[Var1, Var2]

nonlinear_regression

Nonlinear least-squares regression of a user-specified model — a growth curve, a dose-response, a Michaelis-Menten — fitted by Levenberg-Marquardt over the named parameters. Reference predictors by name in the model; every other name is a parameter to estimate. Reports the parameter estimates with asymptotic (Gauss-Newton) standard errors, t-tests and confidence intervals, and a model summary (residual sum of squares, residual standard error, a 1−SSE/SST pseudo-R², iterations and whether the fit converged).

nonlinear_regression dependent=VarName predictors=[Var1, Var2] model="text" start="text"

iv_regression

Instrumental-variables regression by two-stage least squares (2SLS): consistent estimation when a regressor is endogenous (correlated with the error — from simultaneity, omitted confounders or measurement error). The endogenous regressors are projected onto the instruments and the outcome regressed on the projection. Reports the coefficients with corrected (2SLS) standard errors, t-tests and confidence intervals, a model summary, and the first-stage weak-instrument F for each endogenous regressor (rule of thumb F < 10 is weak). Needs at least as many excluded instruments as endogenous regressors.

iv_regression dependent=VarName endogenous=[Var1, Var2] instruments=[Var1, Var2]

gamma_regression

A gamma generalized-linear model for a strictly positive, right-skewed continuous dependent (costs, durations, concentrations) with variance proportional to the mean squared. The log link (default) gives a multiplicative model (Exp(B)); the inverse link is also available. Reports an estimated dispersion, deviance and Pearson goodness-of-fit.

gamma_regression dependent=VarName predictors=[Var1, Var2]

quantile_regression

A quantile regression of a numeric dependent variable on numeric predictors: instead of the mean (OLS), it models a chosen conditional quantile (the median by default) by minimizing the asymmetric check loss. Reports the coefficients with asymptotic standard errors and a pseudo-R¹ goodness of fit. Robust to outliers and to heteroscedastic effects that differ across the response distribution.

quantile_regression dependent=VarName predictors=[Var1, Var2]

gee

A generalized estimating equations (GEE) population-averaged model for correlated / clustered data: it fits a GLM (Gaussian, binomial or Poisson) while accounting for within-subject correlation through a working correlation structure, and reports the coefficients with robust (sandwich) standard errors and the estimated correlation parameter.

gee dependent=VarName predictors=[Var1, Var2] subject=VarName

panel_regression

Panel-data regression for repeated observations of the same entities (firms, people, countries): fixed effects (the within transformation), random effects (Swamy–Arora GLS) or pooled OLS, reporting within/between/overall R², a side-by-side estimator comparison, the Hausman test of fixed vs random effects, and optionally cluster-robust (by entity) standard errors.

panel_regression dependent=VarName regressors=[Var1, Var2] entity=VarName

loglinear

A loglinear (Poisson-GLM) model of the cell counts of a multi-way contingency table formed from two or more categorical variables: the parameter estimates and a deviance / Pearson goodness-of-fit test. The main-effects (independence) model tests whether the factors are independent; the saturated model fits the table perfectly.

loglinear factors=[Var1, Var2]

multinomial_regression

A multinomial logistic regression of a categorical dependent (3+ unordered categories) on one or more predictors: baseline-category logits versus the lowest category, with relative-risk ratios.

multinomial_regression dependent=VarName predictors=[Var1, Var2]

ordinal_regression

An ordinal (proportional-odds) logistic regression of an ordered categorical dependent on one or more predictors: ordered thresholds and one slope per predictor, with odds ratios.

ordinal_regression dependent=VarName predictors=[Var1, Var2]

glm

A factorial general linear model (between-subjects ANOVA/ANCOVA) of a numeric dependent on one or more categorical factors and all their interactions, with optional continuous covariates, Type III sums of squares and F tests.

glm dependent=VarName factors=[Var1, Var2]

mixed_model

A two-level linear mixed model: a numeric dependent on optional predictors with a random intercept (and optional random slopes) for a grouping variable (fixed effects, variance components, the intraclass correlation). Supports ML or REML estimation.

mixed_model dependent=VarName group=VarName

glmm

A generalized linear mixed model: a random-intercept logistic (0/1) or Poisson (count) regression for clustered or repeated data (patients within clinics, repeated yes/no measures). Reports the fixed effects with Exp(B), the random-intercept variance and AIC/BIC, by adaptive Gauss-Hermite quadrature.

glmm dependent=VarName predictors=[Var1, Var2] group=VarName

repeated_measures_anova

A one-way repeated-measures ANOVA over several numeric measures of the same subjects (the within-subjects effect, with Mauchly's sphericity test and the Greenhouse-Geisser / Huynh-Feldt corrections).

repeated_measures_anova measures=[Var1, Var2]

manova

A one-way multivariate analysis of variance (MANOVA): several numeric dependent variables tested jointly across a grouping factor, with Pillai's trace, Wilks' lambda, the Hotelling-Lawley trace and Roy's largest root.

manova dependents=[Var1, Var2] factor=VarName

mediation

Mediation analysis: decomposes the effect of an independent variable X on an outcome Y into a direct effect and the indirect effect(s) that pass through one or more parallel mediators M. Fits the mediator model(s) M ~ X (+ covariates) for the a path, the outcome model Y ~ X + M (+ covariates) for the b paths and the direct effect c′, and the total model Y ~ X (+ covariates) for the total effect c, all by OLS. The indirect effect through each mediator is the product a·b; its confidence interval comes from a nonparametric case-resampling bootstrap (percentile or BCa), reproducible for a fixed seed. An indirect interval that excludes zero is evidence of mediation. With a moderator W the model becomes moderated mediation (PROCESS models 7 / 14 / 58): W and its product term enter the moderated path(s), and the result reports conditional indirect effects at probe values of W (each level of a categorical moderator, else mean ± 1 SD), the index of moderated mediation with a bootstrap CI (when exactly one path is moderated), conditional simple-slope effects, and an interaction plot of the simple slopes; johnson_neyman adds the regions of significance of each moderated path with a plot; a second moderator (moderator2) makes the interaction three-way; serial chains two mediators (X → M1 → M2 → Y) with the serial indirect effect. Continuous outcome. Unweighted like the bootstrap (Select Cases and Split File are honored).

mediation dependent=VarName independent=VarName mediators=[Var1, Var2]

Scale, reduction and classification

reliability

A reliability analysis (Cronbach's alpha) of a set of numeric items measured on the same scale: the alpha coefficient, item statistics and item-total statistics (corrected item-total correlation and the alpha if each item were deleted), so you can see how internally consistent the items are and which weaken the scale.

reliability items=[Var1, Var2]

intraclass_correlation

An intraclass correlation analysis of a set of numeric rater columns (one column per rater, one row per subject): the six Shrout & Fleiss forms (single and average measures, for the one-way random, two-way random absolute-agreement and two-way mixed consistency models), each with its F test and confidence interval. Measures the reliability / agreement of ratings.

intraclass_correlation raters=[Var1, Var2]

bland_altman

A Bland-Altman method-comparison (agreement) analysis of two numeric measurement columns: the bias (mean difference), the standard deviation of the differences and the limits of agreement (bias ± 1.96 × SD), each with a t-based confidence interval; a proportional-bias regression of the differences on the pair means; and the mean-vs-difference plot with the bias and limit lines. Differences are method1 − method2, over the cases complete on both.

bland_altman method1=VarName method2=VarName

irt

An item response theory (IRT) analysis of binary items: Rasch or two-parameter logistic (2PL) item parameters — discrimination and difficulty with standard errors — estimated by marginal maximum likelihood (Bock-Aitkin EM, latent trait N(0,1)), with the model fit block and item/test information curves showing where on the trait the test measures well.

irt items=[Var1, Var2]

factor_analysis

A factor analysis / principal component analysis that reduces several correlated numeric variables to a few underlying components: communalities, the variance each component explains, and the loading of every variable on each component (optionally Varimax-rotated for a cleaner structure).

factor_analysis variables=[Var1, Var2]

cfa

Confirmatory factor analysis (v1): tests a hypothesized measurement structure — which observed indicators load on which named latent factors — by maximum likelihood, unlike exploratory factor_analysis which finds a structure. Reports the model chi-square with the conventional fit indices (CFI, TLI, RMSEA with its 90% CI, SRMR), unstandardized and completely standardized loadings with expected-information standard errors, factor correlations, and factor and residual variances. Each factor's first indicator sets its scale (loading fixed at 1). Identification problems, non-convergence and Heywood cases are reported clearly, never as silent numbers. Optionally reports modification indices (the expected chi-square drop from freeing each fixed cross-loading or residual covariance, with the expected parameter change). With a grouping variable it becomes a multi-group CFA with a mean structure and tests measurement invariance: the configural, metric (equal loadings), scalar (equal intercepts, latent means free after the first group) and strict (equal residual variances) models up to the requested level, with the model-to-model chi-square, CFI and RMSEA differences, then every parameter per group. For ordinal (Likert) indicators the WLSMV estimator fits polychoric correlations by diagonally weighted least squares with the mean-and-variance-adjusted test, robust standard errors and thresholds. The covariance is weighted under Weight Cases; Select Cases and Split File are honored.

cfa model="text"

kmeans_cluster

A k-means cluster analysis that partitions the cases into a fixed number of groups by their numeric variables: the final cluster centers, the number of cases per cluster, and a descriptive ANOVA of how each variable differs across clusters.

kmeans_cluster variables=[Var1, Var2]

hierarchical_cluster

An agglomerative hierarchical cluster analysis that repeatedly merges the closest cases/clusters: an agglomeration schedule and, when a cluster count is given, each case's cluster membership. Suited to small datasets (capped at 200 cases by default; raise "max_cases" to lift it).

hierarchical_cluster variables=[Var1, Var2]

twostep_cluster

A two-step cluster analysis that automatically chooses the number of clusters by Schwarz's BIC (or uses a fixed count): the BIC for each candidate count, the final cluster centers and the number of cases per cluster. Suited to finding a natural grouping when the cluster count is unknown.

twostep_cluster variables=[Var1, Var2]

discriminant_analysis

A linear discriminant analysis that finds the combinations of numeric predictors best separating known groups: the canonical functions (eigenvalues and canonical correlations), the Wilks' lambda test of group separation, and Fisher's classification function coefficients.

discriminant_analysis group=VarName predictors=[Var1, Var2]

decision_tree

A decision / classification tree: a recursive partition of the predictor space into interpretable if-then rules for a categorical target (a classification tree) or a continuous target (a regression tree). CART splits maximize the impurity decrease (Gini / entropy / variance); CHAID splits by a χ²/F significance test. Reports the tree structure, variable importances and a classification (or fit) table.

decision_tree target=VarName predictors=[Var1, Var2]

latent_class

A latent class analysis of two or more categorical indicators: an EM-estimated mixture model finding unobserved subgroups (classes), with class sizes, class-conditional item-response probabilities, a fit block (log-likelihood, AIC/BIC/aBIC, entropy, G², chi-square), average-posterior classification diagnostics, and an optional class-enumeration comparison table. Multiple seeded random starts; classes are reported in descending-size order.

latent_class indicators=[Var1, Var2]

roc_curve

A receiver-operating-characteristic (ROC) analysis of a continuous test score against a binary state variable: the area under the curve (AUC) with its standard error, confidence interval and test against chance (0.5), the curve coordinates (sensitivity vs 1 − specificity at each cut-point), and the cut-point maximizing Youden's J. Pairs with binary logistic regression — fit it, save the predicted probabilities, then run ROC on that column.

roc_curve score=VarName state=VarName

propensity_score

Propensity-score matching / weighting (v1) for observational data where treatment was not randomized. Fits a logistic model of the binary treatment on the covariates (the fitted probability is the propensity score), then either matches each treated case to its nearest control(s) — greedy 1:k without replacement on the logit of the score, optional caliper, estimand ATT — or computes inverse-probability-of-treatment (IPTW) weights (estimand ATE or ATT). Reports the score model, a match/weight summary, and covariate balance as standardized mean differences before and after (|SMD| ≤ 0.1 is the common benchmark). The result is written back as columns: matching saves a 1/0 filter variable (propensity_matched) for Select Cases, weighting a weight variable (propensity_weight) for Weight Cases, so any downstream analysis can run on the matched or weighted sample. Computed unweighted by design (a pre-existing Weight Cases variable only selects cases); Select Cases is honored; Split File cannot be combined with the saved columns. No optimal/genetic/full matching and no doubly-robust estimators in v1.

propensity_score treatment=VarName covariates=[Var1, Var2]

correspondence_analysis

A correspondence analysis of the two-way contingency table of two categorical variables: the principal inertias (singular values) of each dimension with their share of the total inertia (= χ²/N), and the row and column principal coordinates for plotting the categories in a low-dimensional map.

correspondence_analysis row=VarName column=VarName

multidimensional_scaling

A classical (metric) multidimensional scaling of the cases from the Euclidean distances between them over a set of numeric variables: the eigenvalue and share of variance of each recovered dimension, and the case coordinates in that low-dimensional space.

multidimensional_scaling variables=[Var1, Var2]

Survival analysis

kaplan_meier

A Kaplan-Meier (product-limit) estimate of a survival curve from a time-to-event variable that may be right-censored: the survival table (survival probability and Greenwood standard error at each time), the case summary with restricted-mean and median survival, and — when a grouping factor is given — a log-rank (Mantel-Cox) test comparing the groups.

kaplan_meier time=VarName status=VarName

cox_regression

A Cox proportional-hazards regression of the event hazard on numeric covariates (right-censored survival data): coefficients with standard errors and Wald tests, the hazard ratios Exp(B) with confidence intervals, and an omnibus likelihood-ratio test of the model. Ties are handled by Breslow's method.

cox_regression time=VarName status=VarName covariates=[Var1, Var2]

parametric_survival

A parametric accelerated-failure-time (AFT) survival model on right-censored time-to-event data: the exponential, Weibull or log-normal model fitted by maximum likelihood, with the coefficients on the log-time scale (their Exp() is the time ratio / acceleration factor), the scale parameter, and the model log-likelihood and AIC. Covariates are optional (an intercept-only model fits the chosen distribution to the times).

parametric_survival time=VarName status=VarName

competing_risks

A competing-risks cumulative-incidence analysis: with a status variable that codes several mutually exclusive event types (plus censoring), the cause-specific cumulative incidence function for each cause over time (each increment is the all-cause survival times the cause hazard), and the final cumulative incidence per cause. The CIFs sum to 1 − overall survival.

competing_risks time=VarName status=VarName

Time series

autocorrelation

The autocorrelation (ACF) and partial autocorrelation (PACF) functions of a numeric series taken in case order: the correlations at each lag with their standard errors and the cumulative Box-Ljung Q test, optionally after regular and/or seasonal differencing. Useful for identifying ARIMA orders and checking for serial dependence.

autocorrelation variables=[Var1, Var2]

seasonal_decompose

Classical seasonal decomposition of a numeric series taken in case order: splits it into trend, seasonal and remainder components at a given period, additively (y = trend + seasonal + remainder) or multiplicatively (y = trend × seasonal × remainder).

seasonal_decompose variable=VarName period=3

cross_correlation

The cross-correlation function (CCF) of two numeric series taken in case order: the correlation between the first series at time t and the second at time t+k across a range of lags k, with white-noise standard-error bands. A positive lag means the second series leads the first.

cross_correlation series1=VarName series2=VarName

var_model

A vector autoregression (VAR) of two or more numeric series taken in case order: each series regressed on p lags of all of them by per-equation OLS, with lag-order selection criteria (AIC, HQ, BIC, FPE), residual covariance and correlation matrices, and pairwise Granger causality F tests — does one series's history improve the prediction of another?

var_model variables=[Var1, Var2]

arima

A Box-Jenkins ARIMA(p,d,q)(P,D,Q)ₛ model of one numeric series taken in case order — fitted by conditional least squares or exact maximum likelihood, optionally with the orders chosen automatically by an information criterion: the AR and MA parameter estimates (including any seasonal terms) with standard errors and t tests, the model-fit statistics (residual variance, log-likelihood, AIC and BIC), and a Box-Ljung white-noise test of the residuals.

arima variable=VarName

stationarity

Unit-root / stationarity tests of one numeric series taken in case order — the step before choosing ARIMA differencing. Runs the Augmented Dickey-Fuller test (null: a unit root, i.e. non-stationary) and the KPSS test (null: stationarity), reporting each statistic with a p-value and a 5% conclusion. The two are complementary.

stationarity variable=VarName

exponential_smoothing

Exponential smoothing of one numeric series taken in case order: simple (level only), Holt's linear trend (level + trend) or additive Holt-Winters (level + trend + season). The smoothing parameters are fitted by minimizing the one-step error; reports the estimates, a fit summary, and an optional forecast with prediction intervals.

exponential_smoothing variable=VarName

Evidence synthesis and resampling

meta_analysis

Meta-analysis (v1): combines per-study effect sizes — one row per study, an effect column and its standard-error column, already on a common scale (log odds ratio, standardized mean difference, …) — into fixed-effect and random-effects pooled estimates (the τ² estimator is DerSimonian-Laird or REML) with Wald confidence intervals, heterogeneity statistics (Cochran's Q, I², τ²), a forest plot (per-study effects with CIs and the pooled diamond) and a funnel plot with Egger's regression test for small-study asymmetry. Every included row must have both an effect and a positive standard error; a missing cell is an error, not a silent drop. The dataset's Weight Cases setting does not apply (the inverse-variance weights are the statistic's own); Select Cases and Split File are honored.

meta_analysis effect=VarName se=VarName

bootstrap_ci

A nonparametric case-resampling bootstrap confidence interval for a chosen statistic: it draws B resamples of the cases with replacement, recomputes the statistic on each, and reports the observed value, the bootstrap bias and standard error, and a percentile or BCa (bias-corrected and accelerated) interval. Reproducible for a fixed seed. The statistic is one of the mean, median, standard deviation, variance or coefficient of variation of one variable; the Pearson or Spearman correlation of two variables; or the difference in means of a variable across a two-level grouping factor.

bootstrap_ci variable=VarName

Power and sample size

power_means

Statistical power, required sample size, or minimum detectable effect for a t test (one-sample, paired or two independent samples), from Cohen's d. Takes no dataset — all inputs are numbers.

power_means

power_anova

Power, required total sample size, or minimum detectable effect for a one-way ANOVA, from Cohen's f. Takes no dataset.

power_anova groups=0.5

power_correlation

Power, required sample size, or minimum detectable correlation for a test of Pearson's r against zero (Fisher z approximation). Takes no dataset.

power_correlation

power_proportions

Power or required per-group sample size for a test of two independent proportions (Cohen's h). Takes no dataset.

power_proportions p1=0.5 p2=0.5

power_regression

Power, required total sample size, or minimum detectable effect for an R² test in multiple linear regression, from Cohen's f². Takes no dataset.

power_regression predictors=0.5

Complex samples

survey_design

Declare (or clear) the dataset-level complex survey design: the sampling-weight variable, and optionally the PSU (cluster) and stratum variables. While a design is declared, the covered procedures (explore, frequencies, two-way crosstabs, means, linear and logistic regression) report design-based estimates with Taylor-linearized standard errors, confidence intervals on the design df, and design effects; every other analysis states that it ran ignoring the design. Mutually exclusive with Weight Cases. Returns the design summary (strata, PSUs, design df). Declaring a design changes every subsequent analysis, so a read-only profile blocks it.

Changes your data. This is a transform, not an analysis; a read-only session cannot run it.

survey_design

Tables

custom_tables

A general / custom (pivot) table: one or more nested categorical row variables against zero or more nested categorical column variables, each cell holding one or more statistics — case counts (optionally as row, column or total percentages, so a cell can show a count and a percentage together) or summaries of a numeric measure variable (mean, sum, median, minimum, maximum, standard deviation or valid N). Generalizes crosstabs (counts) and means-by-group (a mean over a measure), and can run pairwise column-proportion z tests. The optional "stack_rows" / "stack_columns" lists add further independent variables on that axis, each its own block computed over its own cases — stacked down the row axis or laid side by side across the column axis (stack one axis at a time, not both).

custom_tables row=VarName

multiple_response

A multiple-response frequency table over a set of related variables. In the default "dichotomy" mode the variables are yes/no items sharing a counted value (e.g. several "select all that apply" items), one row each. In "category" mode they share a common code frame (e.g. first / second / third mention), with one row per distinct category pooled across the set. Reports each row's response count, its percent of all responses, and its percent of the valid cases.

multiple_response variables=[Var1, Var2]

Charts

histogram

A histogram of one numeric variable.

histogram variable=VarName

scatter

A scatterplot of two numeric variables, optionally with a LOESS smoother curve over the points.

scatter x=VarName y=VarName

boxplot

A boxplot of one numeric variable, optionally one box per level of a category.

boxplot variable=VarName

violin

A violin plot of one numeric variable, optionally one violin per level of a category: the kernel-density silhouette of the distribution with a quartile bar inside — shows the shape (skew, multiple modes) that a boxplot's five-number summary hides. Gaussian kernel, Silverman bandwidth; widths share one density scale across the violins.

violin variable=VarName

line_chart

A line chart of one or more numeric Y variables against an X variable (case order if omitted), optionally split into one line per level of a category.

line_chart y=[Var1, Var2]

bar_chart

A bar chart of a category variable: the count of cases per level (default), or a summary statistic (mean, sum or median) of a scale variable per level, optionally clustered by a second category. Value labels drive the category axis.

bar_chart category=VarName

error_bar

An error-bar (means) plot: the mean of a scale variable per level of a category, each with an error bar — a confidence interval of the mean (default), or ±SE / ±SD. With a second category it becomes an interaction/profile plot, one connected point-run per level. Value labels drive the category axis.

error_bar variable=VarName category=VarName

qq_plot

A Q-Q plot of one numeric variable against a theoretical distribution (normal by default), with a reference line fit through the quartiles — the standard visual normality check.

qq_plot variable=VarName

scatter_matrix

A scatterplot matrix (pairs plot): every pair of the chosen numeric variables drawn in one grid — the standard first look at a multivariate dataset, showing at a glance which pairs are related and which are not. The diagonal names the variable; each pair appears in both orientations.

scatter_matrix variables=[Var1, Var2]

ecdf

An empirical cumulative distribution (ECDF): the proportion of cases at or below each value, as a step curve. Answers distributional questions a histogram can obscure, since it needs no bin choice — and comparing groups is a matter of reading two curves apart.

ecdf variable=VarName

dot_plot

A Cleveland dot plot: one dot per category, placed at its value. Use it instead of a bar chart whenever the values do not start from zero — a bar's length only means something from a zero baseline, while a dot marks the value itself. Sorted by value by default.

dot_plot category=VarName

bubble_chart

A bubble chart: a scatterplot whose marks carry a third numeric variable as their area. Bubble area is proportional to the value, never the radius — mapping the value to the radius would exaggerate large values by the square.

bubble_chart x=VarName y=VarName size=VarName

Transforms

sort_cases

Sort the dataset's cases by a variable (stable, missing values last). Mutates the dataset.

Changes your data. This is a transform, not an analysis; a read-only session cannot run it.

sort_cases variable=VarName

compute

Compute a numeric variable from an arithmetic/logical expression over the other variables (e.g. "weight / (height * height)"). Creates the target variable, or overwrites it if it already exists. Mutates the dataset.

Changes your data. This is a transform, not an analysis; a read-only session cannot run it.

compute target="text" expression="text"

rank_cases

Rank cases by one or more numeric variables into new ordinal variables (one "R"+name variable per source). Mutates the dataset.

Changes your data. This is a transform, not an analysis; a read-only session cannot run it.

rank_cases variables=[Var1, Var2]

bin_variable

Group a numeric variable into a new ordinal variable by equal-width groups, equal-count (quantile) groups, or your own cutpoints. Mutates the dataset.

Changes your data. This is a transform, not an analysis; a read-only session cannot run it.

bin_variable variable=VarName into="text"