Skip to main content

Post-Estimation & Diagnostics

Statios provides 31 post-estimation and regression diagnostic procedures spanning heteroskedasticity tests, serial correlation diagnostics, influence measures, marginal effects, contrasts, and model comparison tools.

47 procedures · 1 with documented limitations. Every result is computed by the open Python engine and is exportable to APA tables and reports.

ProcedureWhat it doesStatusKey reference
Normality tests (comprehensive suite)Shapiro-Wilk, Kolmogorov-Smirnov, D'Agostino-Pearson, Anderson-Darling; includes skewness/kurtosis and Q-Q plot data.✅ FullShapiro & Wilk (1965); D'Agostino & Pearson (1973)
Homogeneity of variance testLevene and Bartlett tests for equality of variances across groups; concludes on Levene (robust).✅ FullLevene (1960); Bartlett (1937)
Outlier detection (multiple methods)IQR, Z-score (|z|>3), and modified Z-score (MAD-based) methods with index identification.✅ FullVirtanen et al. (2020)
Multicollinearity checkVariance Inflation Factor (VIF) analysis with concern classification (low/moderate/high).✅ FullVirtanen et al. (2020)
Automatic test suggestionRecommends appropriate statistical tests based on data characteristics.✅ FullVirtanen et al. (2020)
Breusch-Pagan test for heteroskedasticityLM and F variants test if residual variance depends on predictors; requires n > k+2.✅ FullBreusch & Pagan (1979)
White test for heteroskedasticityGeneral heteroskedasticity test not assuming specific form; more flexible than Breusch-Pagan.✅ FullWhite (1980)
Goldfeld-Quandt heteroskedasticity testSorts by first predictor, splits at the midpoint, drops a block of observations starting at the midpoint, and F-tests the lower vs upper group residual variances.✅ FullGoldfeld & Quandt (1965)
Breusch-Godfrey serial correlation testLM and F tests for autocorrelation up to specified lags in OLS residuals.✅ FullBreusch (1978); Godfrey (1978)
Durbin-Watson autocorrelation statisticDW near 2 indicates no autocorrelation; <1.5 positive, >2.5 negative autocorrelation.✅ FullDurbin & Watson (1950)
Ljung-Box autocorrelation testTests joint significance of autocorrelations at specified lags; generic for series/residuals.✅ FullLjung & Box (1978)
Heteroskedasticity-robust LM testWooldridge robust LM test that a set of regressors are jointly zero, valid under heteroskedasticity of unknown form (n − SSR auxiliary regression on residualized products).✅ FullWooldridge (2010)
AR(1) serial-correlation testWooldridge's t-test for first-order serial correlation: regress OLS residuals on their own lag (plus the regressors) and test ρ = 0.✅ FullWooldridge (2010)
Durbin's h TestDurbin's h test for first-order serial correlation in OLS models that include a lagged dependent variable, where the Durbin-Watson statistic is biased toward 2.✅ FullDurbin (1970)
Moulton Factor (clustered SE inflation)Moulton (1986) standard-error inflation factor: estimates the residual intra-class correlation across groups and reports naive OLS vs Moulton-adjusted standard errors.✅ FullMoulton (1986)
Vuong Test (non-nested)Vuong's (1989) likelihood-ratio test for two non-nested models (OLS/logit/Poisson), comparing per-observation log-likelihoods to decide which specification is closer to the true data-generating process, with AIC/BIC complexity-corrected variants.✅ FullVuong (1989, Econometrica)
Cox Test (non-nested)Pesaran-Cox test discriminating between two non-nested linear regression models via the difference in log-likelihoods adjusted by its expectation under each model, reported in both directions as standard-normal statistics with a preferred-model verdict.✅ FullCox (1961); Pesaran (1974, Rev. Econ. Stud.)
Encompassing Test (non-nested)Mizon-Richard (1986) encompassing test that asks, by artificial nesting, whether one non-nested linear model explains everything its rival adds: each direction adds the rival's non-overlapping regressors to the base model and F-tests their joint significance (insignificant => the base model encompasses the rival), returning F/df/p, per-direction verdicts, and a preferred model.✅ FullMizon & Richard (1986, Econometrica)
J Test (non-nested models)Davidson-MacKinnon J test that discriminates between two non-nested linear specifications by adding each model's OLS fitted values to the other and testing their t-statistic in both directions, returning per-direction rejection verdicts and a preferred model.✅ FullDavidson & MacKinnon (1981)
Information-Matrix Test (White)White's (1982) omnibus information-matrix specification test for OLS: IM = n*R^2 of the auxiliary regression of squared residuals on the unique squares and cross-products of the regressors (chi-square), detecting general misspecification — wrong functional form, heteroskedasticity, or non-normality — not just heteroskedasticity.✅ FullWhite (1982)
Harvey Multiplicative HeteroskedasticityHarvey's (1976) LM test for multiplicative heteroskedasticity, regressing the log of squared OLS residuals on chosen variables to estimate the variance model Var(e)=exp(z'alpha) and test whether the variance depends on them.✅ FullHarvey (1976)
LM / Score Test (omitted regressors)Classic Lagrange-multiplier (score) test for omitted regressors: regresses the residuals of the restricted OLS (dependent on kept) on the full regressor set and reports LM = n*R-squared of that auxiliary regression as a chi-square(q) test that the tested coefficients are jointly zero, with the equivalent F-form and a reject/not verdict.✅ FullBreusch & Pagan (1980); Engle (1984)
Multivariate Normality (Doornik-Hansen)Tests whether a set of variables (or VAR residuals) is jointly multivariate normal via Mardia's multivariate skewness and kurtosis statistics and the Doornik-Hansen omnibus chi2(2p) test, returning each statistic with its df and p-value, per-variable skew/kurtosis, and an overall normal/non-normal verdict.✅ FullMardia (1970); Doornik & Hansen (2008)
Ramsey RESET specification testTests for functional form misspecification by adding polynomial terms of fitted values.✅ FullRamsey (1969)
Hosmer-Lemeshow goodness-of-fit testGroups logit predictions by deciles, compares observed vs expected event counts via chi-square.✅ FullHosmer et al. (1980)
Variance Inflation Factor (VIF) analysisVIF per predictor with concern thresholds (VIF>5 moderate, >10 high); includes mean VIF.✅ FullMarquardt (1970)
Cook's distance influence measureIdentifies observations with disproportionate influence on regression; threshold default 4/n.✅ FullCook (1977)
Leverage (hat values) analysisDetects high-leverage observations (unusual X values); threshold 2*(k+1)/n.✅ FullBelsley et al. (1980)
DFBETAS parameter-specific influenceMeasures change in each coefficient when one observation removed; threshold 2/sqrt(n).✅ FullBelsley et al. (1980)
Marginal effects (logit/probit)Average Marginal Effects (AME) or Marginal Effects at Means (MEM) for binary choice models.✅ FullSeabold & Perktold (2010)
Predicted probabilities with CILogit/probit predictions on observed data or at specified covariate values with 95% CI.✅ FullSeabold & Perktold (2010)
Post-hoc multiple comparisonsTukey, Bonferroni, Holm step-down, Scheffe, Games-Howell with p-value adjustments.✅ FullTukey (1949); Games & Howell (1976)
Effect sizes across designsCohen's d, Glass's delta, Hedges' g, eta-squared, partial eta-squared, omega-squared, Cohen's f.✅ FullCohen (1988)
Planned contrasts (pairwise/polynomial)Pairwise, polynomial (Legendre), Helmert, deviation contrasts with t-tests and CI.✅ FullSeabold & Perktold (2010)
Prediction intervals (OLS)Confidence and prediction intervals for fitted values; wider than CI due to residual variance.✅ FullSeabold & Perktold (2010)
Coefficient table (with VIF/robust SE)Formatted OLS results with standardized betas, VIF, HC3 robust SE option; significance stars.✅ FullSeabold & Perktold (2010)
Model comparison (R-sq, AIC, BIC)Side-by-side OLS models with delta AIC/BIC relative to best; RMSE and F-stat per model.✅ FullSeabold & Perktold (2010)
Marginal effects at specified valuesConditional marginal effects (OLS/logit/probit/Poisson) at user-supplied covariate values.✅ FullSeabold & Perktold (2010)
Margins contrast (difference in effects)Compare marginal effects of one predictor across two or more levels; delta-method SE.✅ FullSeabold & Perktold (2010)
Conditional marginal effects (interaction)Marginal effect of focal variable at 10th, 25th, 50th, 75th, 90th percentiles of moderator.✅ FullSeabold & Perktold (2010)
Nested contrasts (hierarchical)Two-sample Welch t-tests on specified pairs within a grouping variable.✅ FullSeabold & Perktold (2010)
Elasticities (full/partial/arc)dy/dxx/y (full), dy/dxx (partial), or arc elasticity across specified predictor values.✅ FullSeabold & Perktold (2010)
Semi-elasticities (dy/dx/y and dy/dx*x)Proportional change in y and linear change in y per unit change in x.✅ FullSeabold & Perktold (2010)
Predicted means over groups (Stata-style)Marginal cell means with 95% CI; pairwise comparisons across factor levels.✅ FullSeabold & Perktold (2010)
Adjusted predictions (holding covariates constant)Linear predictor at specified covariate values (mean/median/min/max); 95% CI included.✅ FullSeabold & Perktold (2010)
Probability profile (logit/probit curve)Binary outcome probability as function of one predictor; other predictors at mean.✅ FullSeabold & Perktold (2010)
Treatment effect contrasts (ATE/ATT/ATC)Average treatment effect via regression-based g-computation; balanced or flexible interaction.⚠️ LimitedSeabold & Perktold (2010)

Notes & limitations

  • Treatment effect contrasts (ATE/ATT/ATC) — Uses OLS-based potential outcomes, not designed for instrumental variables.