Skip to main content

Regression & Generalized Linear Models

Statios's model namespace covers 50+ regression and generalized linear estimation procedures, from OLS to latent-variable extensions. Methods range from standard parametric models (linear, logistic, count) to specialized estimators for endogenous/selected samples, censoring, and complex errors. All predictor-taking handlers auto-expand categorical predictors R-style (labeled indicator columns, named reference level, factors field) — no manual dummy coding required.

142 procedures · 3 with documented limitations. Every result is computed by the open Python engine and is exportable to APA tables and reports.

ProcedureWhat it doesStatusKey reference
Linear Regression (OLS)Standard ordinary least squares with diagnostics (R², F-test, VIF, Durbin-Watson, residual normality). Categorical predictors are auto-expanded R-style (labeled indicators, named reference level, factors field) — no manual dummy coding required.✅ FullAkaike (1974); Schwarz (1978)
Linear Regression — Robust SEs (HC0–HC3)OLS with heteroskedasticity-consistent standard errors; selectable HC0 (White), HC1 (Stata default), HC2, or HC3 (MacKinnon & White 1985).✅ FullAkaike (1974); Schwarz (1978)
Linear Regression — Newey-West HAC SEsOLS with Newey-West heteroskedasticity- and autocorrelation-consistent (HAC) standard errors (Bartlett kernel, set hac_lags); standard for time-series regressions.✅ FullNewey & West (1987)
Errors-in-Variables RegressionMethod-of-moments correction for classical measurement error (attenuation bias) using per-regressor reliability ratios; reports naive OLS vs corrected coefficients.✅ FullFuller (1987); Wooldridge (2010)
SIMEX (measurement-error correction)Corrects attenuation bias from a regressor measured with known additive error variance by simulation-extrapolation: adds increasing extra noise, refits OLS at each lambda (averaging over B Monte-Carlo replicates), and extrapolates the coefficient quadratically back to lambda=-1 to recover the error-free estimate.✅ FullCook & Stefanski (1994, JASA); Carroll et al. (2006)
Frisch-Waugh-Lovell Partial SlopePartials the controls out of both Y and the focal regressor and regresses the residuals through the origin; numerically equals the focal coefficient in the full OLS, with the partial correlation.✅ FullFrisch & Waugh (1933)
Omitted-Variable Bias DecompositionDecomposes omitted-variable bias for each included regressor into short-model vs long-model coefficients and the exact gamma*delta contribution of each omitted variable.✅ FullWooldridge (2010)
Log-Y Retransformation (Duan Smearing)Fits a log-scale OLS and reports Duan's smearing factor S=mean(exp(residuals)) plus naive vs bias-corrected (smeared) retransformed predictions on the original scale.✅ FullDuan (1983)
Potential Outcomes (ATE/ATT/ATU)Reports the average treatment effect (ATE), the effect on the treated (ATT) and on the untreated (ATU) in Rubin's counterfactual notation by regression adjustment under unconfoundedness: an OLS fit of the outcome on treatment, covariates and their interactions predicts each unit's pair (Y(1), Y(0)), with delta-method or bootstrap standard errors.✅ FullRubin (1974); Imbens & Wooldridge (2009)
Selection Model (MNAR missingness)Heckman (1979) two-step selection model for non-ignorable (MNAR) missingness: a probit models whether the outcome is observed, the outcome equation is corrected for selection on unobservables via the inverse Mills ratio, and rho (with its rho=0 ignorability test) quantifies the selection correlation, reported alongside naive complete-case OLS for comparison.✅ FullHeckman (1979); Little & Rubin (2019)
Pattern-Mixture Model (MNAR)Pattern-mixture model (Little 1993) for non-ignorable (MNAR) missing outcomes: partitions the sample by the outcome's missingness, fits the completers' regression, imputes the missing outcomes under a sensitivity restriction whose delta shifts the missing-group mean (delta=0 = MAR), and reports the pattern-averaged mean, fraction missing, and a sensitivity curve.✅ FullLittle (1993, JASA)
Continuous-Treatment Dose-Response (GPS)Estimates the average dose-response function of a continuous treatment with the Hirano-Imbens generalized propensity score: models the treatment given covariates as Normal, regresses the outcome on a flexible function of the treatment and the GPS, and averages to trace E[Y(t)] over a grid of doses, returning the dose-response curve, the GPS-adjusted average marginal effect, and the naive unadjusted slope for comparison.✅ FullHirano & Imbens (2004)
Abadie Kappa-Weighting (compliers)Uses Abadie's (2003) kappa weights kappa = 1 - D(1-Z)/(1-p(X)) - (1-D)Z/p(X) to estimate the complier share, the complier mean outcome, the mean covariates among compliers versus the full sample, and a kappa-weighted LATE in a binary-instrument / binary-treatment IV setting.✅ FullAbadie (2003, J. Econometrics 113:231-263)
Marginal Treatment Effect (MTE)Estimates the Heckman-Vytlacil marginal treatment effect via local IV / generalized Roy: fits the propensity score and a flexible control function in p, differentiates to trace the MTE over the unobserved-resistance dimension, integrates it with the appropriate weights to recover ATE/ATT/ATUT/LATE, and tests for essential heterogeneity.✅ FullHeckman & Vytlacil (2005), Econometrica 73:669-738
Synthetic DiDSynthetic Difference-in-Differences (Arkhangelsky et al. 2021) combines synthetic-control unit weights with difference-in-differences time weights to estimate the ATT on a balanced treated/control panel, reporting the SDID effect, a placebo/jackknife standard error, the sparse non-negative unit weights and the pre-treatment fit.✅ FullArkhangelsky, Athey, Hirshberg, Imbens & Wager (2021, AER)
Changes-in-Changes (Athey-Imbens)Nonlinear generalization of difference-in-differences (Athey & Imbens 2006) that maps the treated-pre outcome distribution through the control group's pre-to-post distributional change to identify the entire counterfactual treated distribution, reporting the ATT plus quantile treatment effects (0.25/0.5/0.75) that reveal heterogeneous effects ordinary DiD would miss.✅ FullAthey & Imbens (2006, Econometrica 74:431-497)
Doubly-Robust DiD (Sant'Anna-Zhao)Estimates the ATT in a 2x2 difference-in-differences design by combining an outcome-regression DiD with an inverse-propensity-weighted DiD into the doubly-robust estimator of Sant'Anna & Zhao (2020), which stays consistent if either the propensity-score model or the untreated-outcome model is correct, reports influence-function standard errors with the IPW and OR components and the simple DiD for contrast, and reduces exactly to the canonical mean 2x2 DiD when no covariates are supplied.✅ FullSant'Anna & Zhao (2020, J. Econometrics 219:101-122)
Honest DiD (parallel-trends sensitivity)Rambachan & Roth (2023) sensitivity analysis for difference-in-differences: estimates the event-study coefficients, then bounds the post-treatment effect under a relative-magnitudes restriction (Mbar) on how far post-treatment trend violations can exceed observed pre-trends, reporting robust confidence sets across Mbar and the breakdown Mbar at which significance is lost.✅ FullRambachan & Roth (2023, Rev. Econ. Stud.)
Sun-Abraham Event StudyInteraction-weighted (IW) event-study estimator that saturates the model with cohort-by-relative-period interactions and aggregates them by cohort shares, recovering the true dynamic treatment path (and clean pre-trends) under staggered adoption with heterogeneous effects where conventional dynamic TWFE is contaminated.✅ FullSun & Abraham (2021, J. Econometrics 225:175-199)
Callaway-Sant'Anna Staggered DiDCallaway & Sant'Anna (2021) staggered difference-in-differences estimating group-time average treatment effects ATT(g,t) via clean 2x2 DiD against not-yet-treated/never-treated comparison units, aggregated to an overall ATT, an event-study (dynamic) profile, and per-cohort effects, with a naive two-way FE benchmark to expose heterogeneous-timing bias.✅ FullCallaway & Sant'Anna (2021, J. Econometrics); Goodman-Bacon (2021)
Causal Forest (heterogeneous effects)Honest generalized random forest (Wager-Athey) that estimates the conditional average treatment effect tau(x) via split-sample honest trees with R-learner local centering, returning the ATE with a forest-based SE, the distribution of individual CATEs, and the leading covariate driving effect heterogeneity.✅ FullWager & Athey (2018), JASA; Athey, Tibshirani & Wager (2019), Annals of Statistics
Double / Debiased MLEstimates a debiased treatment effect in the partially-linear model Y = theta*D + g(X) + e by cross-fitting gradient-boosted nuisance regressions for E[Y|X] and E[D|X] and solving the orthogonal (Neyman) moment on the residuals, reporting theta with robust SE/CI alongside the biased naive single-regression estimate for contrast.✅ FullChernozhukov et al. (2018)
McCrary Density Test (RD manipulation)McCrary (2008) density-discontinuity test that fits separate triangular-kernel local-linear density estimates on each side of an RD cutoff and tests whether the log-density of the running variable jumps at the threshold, detecting sorting/manipulation that would invalidate the RD design.✅ FullMcCrary (2008, J. Econometrics)
Propensity-Score SubclassificationEstimates the ATE and ATT of a binary treatment by fitting a propensity-score logit, splitting units into quantile strata of the score, and combining the within-stratum treated-minus-control mean differences weighted by stratum size (ATE) or treated share (ATT), with per-stratum effects and a stratified standard error.✅ FullRosenbaum & Rubin (1983, Biometrika); Rosenbaum & Rubin (1984, JASA)
Mahalanobis Matching (ATT)Nearest-neighbour matching estimator of the Average Treatment effect on the Treated that matches each treated unit to its closest control(s) in Mahalanobis covariate distance (pooled-control inverse covariance), imputes the counterfactual from the matched controls, and reports the ATT with an approximate Abadie-Imbens standard error and covariate balance.✅ FullRubin (1980); Abadie & Imbens (2006)
Box-Cox Transformed RegressionEstimates the Box-Cox power-transform lambda maximizing the concentrated profile log-likelihood, then fits OLS on the transformed (strictly-positive) response.✅ FullBox & Cox (1964)
Stochastic Frontier AnalysisMaximum-likelihood Aigner-Lovell-Schmidt stochastic frontier (half-normal inefficiency) for production or cost functions, returning frontier coefficients, sigma_v/sigma_u, lambda, gamma, mean technical efficiency E[exp(-u)] and per-observation efficiency scores.✅ FullAigner, Lovell & Schmidt (1977); Meeusen & van den Broeck (1977); Jondrow et al. (1982); Battese & Coelli (1988)
Finite-Mixture Regression (latent class)Fits a finite mixture of K linear regression regimes by EM, recovering each latent class's own coefficient vector, error variance, and mixing proportion together with the log-likelihood, BIC, and posterior class-assignment summary.✅ FullQuandt & Ramsey (1978); McLachlan & Peel (2000)
Dirichlet-Process Mixture DensityBayesian nonparametric density estimation that fits an infinite (Dirichlet-process) Gaussian mixture via truncated stick-breaking, automatically inferring the effective number of mixture components and returning their means, variances, weights, and the estimated density curve.✅ FullEscobar & West (1995); Ferguson (1973); sklearn
Finite-Mixture Count (latent class)Fits a finite mixture of K latent-class Poisson regressions by EM, recovering each regime's own log-link coefficient vector and mixing proportion together with the log-likelihood, BIC/AIC, and posterior class-assignment summary to capture unobserved heterogeneity and overdispersion in count data.✅ FullWedel et al. (1993); Cameron & Trivedi (2013)
Simultaneous Quantile RegressionFits linear quantile regressions at several quantiles at once and uses a joint paired (xy) bootstrap covariance to report per-quantile coefficient tables and to test whether slopes change across the conditional distribution (e.g. tau=0.25 vs 0.75).✅ FullKoenker & Bassett (1978); Koenker (2005)
Quantile Regression ForestEstimates full conditional quantiles (Meinshausen 2006) with a random forest by reading the weighted empirical response distribution in each leaf, revealing heteroskedastic conditional intervals that widen with predictors plus an out-of-bag interval-coverage check and variable importance.✅ FullMeinshausen (2006, JMLR)
Additive Quantile Regression (quantile GAM)Fits a tau-quantile regression in which each predictor enters through a smooth cubic B-spline term, minimizing the asymmetric check-loss via linear programming to recover nonlinear conditional quantile curves.✅ FullKoenker (2011); Fasiolo et al. (2021, JASA)
Composite Quantile RegressionEstimates a single shared slope vector by pooling the check-loss across an equally-spaced grid of quantiles (with quantile-specific intercepts), giving slope estimates that match OLS under Gaussian errors and beat it in efficiency under heavy-tailed errors, with bootstrap standard errors.✅ FullZou & Yuan (2008, Ann. Statist.)
Expectile RegressionNewey-Powell asymmetric-least-squares expectile regression, fit by iteratively reweighted least squares to estimate any conditional tau-expectile (tau=0.5 reproduces OLS exactly), with heteroskedasticity-consistent sandwich standard errors and an asymmetry interpretation.✅ FullNewey & Powell (1987); Aigner, Amemiya & Poirier (1976)
Quantile Count Regression (jittering)Estimates conditional quantiles of an overdispersed count response by the Machado-Santos Silva (2005) jittering method -- uniform noise is added to the integer counts, quantile regression is run on the log-transformed scale, and coefficients are averaged over many jitter draws to give stable, monotone-in-tau multiplicative quantile effects.✅ FullMachado & Santos Silva (2005, JASA)
Generated-Regressor Inference (Murphy-Topel)Two-step estimation that regresses a stage-2 outcome on the fitted values from a first-stage OLS, reporting both the naive stage-2 standard errors and the Murphy-Topel/Pagan-corrected standard errors that propagate the first-stage estimation uncertainty into the generated regressor.✅ FullPagan (1984, IER); Murphy & Topel (1985, JBES)
Heteroskedastic ProbitHarvey (1976) binary probit in which the latent-error standard deviation is modelled as sigma_i = exp(z_i'gamma), estimated by BFGS maximum likelihood, returning separate mean (beta) and variance (gamma) coefficient tables plus a likelihood-ratio test of homoskedasticity (H0: gamma=0).✅ FullHarvey (1976, Econometrica); Greene (2018)
Kernel Regression (Nadaraya-Watson)Nonparametric Nadaraya-Watson local-constant kernel smoother estimating the conditional mean of one outcome on one predictor (Gaussian or Epanechnikov kernel, Silverman's rule-of-thumb or user bandwidth), returning fitted values, a 100-point smooth curve, an R-squared-like fit measure, and the bandwidth used.✅ FullNadaraya (1964); Watson (1964)
Kernel Conditional DensityEstimates the full conditional density f(y|x) with a double Gaussian kernel (Nadaraya-Watson weighting in x combined with a kernel in y), returning conditional 0.1/0.5/0.9 quantile curves and the conditional mean as functions of x so heteroskedastic spread, skew and multimodality are visible where a mean regression cannot show them.✅ FullRosenblatt (1969); Hyndman, Bashtannyk & Grunwald (1996)
Single-Index Model (Ichimura)Semiparametric single-index regression y = g(x'beta) + e estimated by Ichimura's semiparametric least squares, recovering the index direction beta (identified up to scale, unit-norm normalized) by minimizing the leave-one-out kernel-regression residual sum of squares while leaving the link g(.) fully nonparametric, and reporting standardized and raw-scale index coefficients, the data-driven bandwidth, a leave-one-out pseudo-R-squared, and the estimated link curve.✅ FullIchimura (1993), J. Econometrics 58:71-120; Klein & Spady (1993), Econometrica 61:387-421.
Maximum Score (Manski)Manski's semiparametric maximum-score estimator for binary choice, recovering the scale-normalized coefficient direction by maximizing correctly-predicted signs of x'beta under only a conditional-median error restriction, staying consistent where logit/probit fail under heteroskedasticity or non-normal errors.✅ FullManski (1975, 1985)
Translog Production/Cost FunctionEstimates a second-order flexible (transcendental logarithmic) production or cost function by OLS on log-inputs, their squares and pairwise cross-products, reporting output elasticities at the sample mean, returns-to-scale, and an F-test of the Cobb-Douglas (no-second-order-terms) restriction.✅ FullChristensen, Jorgenson & Lau (1973)
Empirical Likelihood (mean)Owen's nonparametric empirical-likelihood inference for a population mean or mean vector, returning the -2 log EL-ratio statistic with its chi-square p-value plus a distribution-free EL confidence interval that requires no normality assumption.✅ FullOwen (1988, Biometrika); Owen (2001), Empirical Likelihood
Copula Fitting (bivariate dependence)Fits Gaussian, Clayton, Gumbel and Frank copulas to two variables' rank pseudo-observations by maximum likelihood, reporting each family's parameter, log-likelihood and AIC, the AIC-selected copula, and Kendall's tau and Spearman's rho.✅ FullNelsen (2006); Joe (1997)
Gaussian-Process RegressionNonparametric Bayesian regression with an RBF + white-noise kernel that learns its length-scale and noise level by maximizing the log-marginal-likelihood, returning a smooth posterior mean plus per-point predictive uncertainty.✅ FullRasmussen & Williams (2006)
BART (Bayesian Additive Regression Trees)Flexible nonparametric regression that models the conditional mean as a regularized sum of trees fit by Bayesian backfitting MCMC, returning posterior-mean fitted values, predictive uncertainty, variable-inclusion importance, and in-sample R-squared.✅ FullChipman, George & McCulloch (2010), Ann. Appl. Stat. 4:266-298
Gaussian-Process ClassificationFits a Gaussian-process classifier (RBF kernel, Laplace approximation) to a binary 0/1 outcome, reporting in-sample accuracy/log-loss/AUC, learned kernel hyperparameters, and a predicted-probability summary to capture smooth nonlinear class boundaries a linear logit would miss.✅ FullRasmussen & Williams (2006), ch. 3
k-Nearest-Neighbours RegressionNonparametric regression that predicts each outcome as the average of its k nearest neighbours in standardized (Euclidean) predictor space, reporting an in-sample fit, a leave-one-out cross-validated RMSE and CV R-squared, and the chosen k.✅ FullAltman (1992), The American Statistician 46:175-185.
Spline Regression (B-splines)Fits a flexible nonlinear curve by building a B-spline basis of one predictor (patsy bs() with given degrees of freedom and polynomial degree) and running OLS of the response on that basis, returning the fitted curve on a 100-point grid, the data-quantile knot locations, the basis dimension, R-squared and residual MSE.✅ Fullde Boor (1978)
Generalized Additive Model (splines)Fits y = intercept + sum_j f_j(x_j) + e with each term a penalized cubic B-spline (statsmodels GLMGam, Gaussian family), selecting the smoothing penalty by GCV and reporting per-term effective degrees of freedom, deviance, pseudo-R^2, and a fitted-value summary.✅ FullHastie & Tibshirani (1990); Wood (2017)
GAM (Poisson / Logit)Generalized additive model for count (Poisson/log) or binary (Binomial/logit) outcomes that fits each predictor as a penalized cubic B-spline smooth via penalized IRLS with GCV/UBRE smoothing selection, reporting per-smooth effective degrees of freedom, deviance explained, fitted curves with confidence bands, and approximate significance.✅ FullHastie & Tibshirani (1990); Wood (2017)
Local-Linear RegressionNonparametric degree-1 local-polynomial kernel smoother that fits a kernel-weighted least-squares line at each point and takes its intercept as the conditional-mean estimate (Gaussian or Epanechnikov kernel, Silverman rule-of-thumb or user bandwidth), correcting the boundary bias of Nadaraya-Watson and returning fitted values, a 100-point smooth curve, effective degrees of freedom, and an R-squared-like fit measure.✅ FullFan (1992); Fan & Gijbels (1996)
Local Polynomial RegressionNonparametric kernel-weighted local polynomial smoother of arbitrary degree that generalizes Nadaraya-Watson (degree 0) and local-linear (degree 1) fits, returning the fitted curve, its estimated first derivative, and a smoother effective-df fit measure, with higher degrees reducing boundary bias.✅ FullFan & Gijbels (1996)
Isotonic Regression (PAVA)Nonparametric monotone least-squares fit of a response on a single predictor via the pool-adjacent-violators algorithm, returning the fitted step function (level sets), R^2, and the imposed direction.✅ FullBarlow et al. (1972); Robertson, Wright & Dykstra (1988)
Wavelet Regression (denoising)Nonparametric denoising of a response treated as a signal ordered by a predictor via discrete wavelet transform with soft-thresholding (VisuShrink universal or hybrid SureShrink), returning the denoised fitted curve, robust MAD noise estimate, retained-coefficient sparsity, and reconstruction diagnostics.✅ FullDonoho & Johnstone (1994, Biometrika)
Wavelet CoherenceComputes squared wavelet coherence R^2 in [0,1] between two time series across time and scale via a hand-rolled FFT-based Morlet continuous wavelet transform, smoothed cross-wavelet and power spectra (Torrence & Compo 1998; Grinsted et al. 2004), summarizing the dominant common period, coherence by short/medium/long scale bands, and the cone-of-influence-masked global coherence.✅ FullTorrence & Compo (1998); Grinsted, Moore & Jevrejeva (2004)
Partially Linear (Robinson)Robinson (1988) sqrt(n)-consistent semiparametric partially-linear regression y = x'beta + g(z) + e, estimating the linear slopes beta by the double-residual method (Nadaraya-Watson partialling-out of y and each x on z, then residual-on-residual OLS) while leaving g(z) fully nonparametric.✅ FullRobinson (1988, Econometrica 56:931-954)
Overdispersion Test (Cameron-Trivedi)Cameron-Trivedi (1990) regression-based test of the Poisson equidispersion assumption: fits a Poisson GLM, then runs auxiliary through-origin OLS of the studentised squared-residual statistic on g(mu) for g(mu)=mu (NB1) and g(mu)=mu^2 (NB2), reporting per-variant alpha-hat, a one-sided t-test of H0 alpha=0, and an overdispersion verdict.✅ FullCameron & Trivedi (1990, J. Econometrics 46:347-364)
Conway-Maxwell-Poisson RegressionCOM-Poisson GLM (log link on lambda) with a jointly-estimated dispersion parameter nu that flexibly handles under- (nu>1) and over-dispersed (nu<1) counts, fit by maximum likelihood with a stable truncated normalizing constant and a Wald test of nu=1 against the Poisson.✅ FullShmueli et al. (2005); Sellers & Shmueli (2010)
Endogenous Count (control function)Poisson count regression with an endogenous regressor corrected by two-stage residual inclusion (control function): a first-stage OLS residual on instruments is added to the count model, yielding the endogeneity-corrected effect, a control-function residual coefficient that doubles as an endogeneity test, and bootstrap standard errors.✅ FullMullahy (1997); Terza, Basu & Rathouz (2008)
System OLS (multi-equation)Estimates a system of linear equations equation-by-equation by OLS (intercept added), reporting per-equation coefficient tables (coef/se/t/p/CI), R2 and n, plus the cross-equation residual covariance Sigma (symmetric, PSD) and residual correlation on the common listwise-complete sample - the SUR/FGLS baseline.✅ FullZellner (1962); Greene (2018)
Joint Confidence RegionFits OLS and reports the joint F test that a chosen set of coefficients is simultaneously zero, the exact confidence-ellipse geometry (centre, covariance submatrix, eigen-axes and half-axis lengths) for exactly two coefficients, and Bonferroni simultaneous intervals, contrasting the tilted exact region with the conservative axis-aligned box.✅ FullGreene (2018); Scheffe (1959)
General GLS (specified Omega)Feasible generalized least squares (Aitken) under a user-specified error covariance: AR(1) Toeplitz errors, groupwise heteroskedasticity (one variance per group level), or power heteroskedasticity (var proportional to |fitted|^delta), reporting the estimated structure parameter and an efficient coefficient table.✅ FullAitken (1936); Greene (2018)
Complementary Log-Log RegressionBinary regression with the asymmetric complementary log-log link P(y=1)=1-exp(-exp(x'beta)) for rare-event data and grouped/discrete-time proportional-hazards (survival) models, returning a coef/se/z/p/CI table with log-likelihood and AIC/BIC.✅ FullMcCullagh & Nelder (1989)
GEE (Population-Averaged)Generalized estimating equations fit a population-averaged (marginal) regression for clustered or longitudinal data, modelling within-cluster dependence through a working correlation (independence / exchangeable / AR-1) while reporting robust sandwich standard errors valid even under correlation misspecification.✅ FullLiang & Zeger (1986); Zeger & Liang (1986)
LIML (Limited-Information ML)Limited-information maximum likelihood IV estimator that is less biased than 2SLS under weak or many instruments, reporting the k-class kappa eigenvalue, robust coefficient table with confidence intervals, first-stage instrument-strength diagnostics, and an overidentification test.✅ FullAnderson & Rubin (1949); Davidson & MacKinnon (2004)
Two-Sample IV (TS2SLS)Two-sample two-stage least squares: estimates a single endogenous regressor's structural coefficient by fitting the first stage in one half of the data and carrying the fitted endogenous value into the reduced-form/structural equation in the other half (Inoue-Solon TS2SLS; Angrist-Krueger TSIV design), reporting the coefficient with a robust SE, first-stage strength, and sample sizes.✅ FullAngrist & Krueger (1992, JASA); Inoue & Solon (2010, ReStat)
Semiparametric IV (sieve)Newey-Powell / Ai-Chen series (sieve) two-stage instrumental-variables estimation of an unknown dose-response g(endog) approximated by a polynomial sieve and instrumented by a sieve basis of the instruments, returning the estimated g(.) curve on a grid, the average derivative, HC1-robust standard errors, first-stage relevance, and a naive sieve-OLS comparison to expose endogeneity bias.✅ FullNewey & Powell (2003); Ai & Chen (2003)
Lewbel Heteroskedasticity IVEstimates a regression with an endogenous regressor when no external instruments exist, by constructing Lewbel (2012) generated instruments from heteroskedasticity in the first-stage residual and running heteroskedasticity-robust 2SLS, reporting the first-stage F and an OLS-vs-Lewbel comparison.✅ FullLewbel (2012, JBES)
Generalized Empirical Likelihood (GEL)Information-theoretic estimation of a linear IV / moment-condition model via the GEL saddle-point (empirical likelihood, exponential tilting, or continuous-updating), returning coefficients with sandwich SEs, an LR-type overidentification test, and nonnegative implied probabilities; reproduces 2SLS exactly when just-identified.✅ FullSmith (1997); Newey & Smith (2004)
Empirical-Likelihood RegressionEstimates linear-regression coefficients on the OLS estimating-equation moments E[x(y - x'beta)]=0 and tests/intervals them via Owen's empirical likelihood, giving distribution-free, generally-asymmetric EL confidence intervals plus chi-square-calibrated EL-ratio p-values whose point estimates coincide with OLS.✅ FullOwen (1991); Qin & Lawless (1994)
Continuously-Updated GMM (CUE)Continuously-updated GMM for a linear IV model that re-estimates the optimal weight matrix as a function of the parameters inside the objective and minimises jointly, returning CUE coefficients with robust standard errors, z/p-values, the Hansen J overidentification test, and (when just-identified) an exact 2SLS match.✅ FullHansen, Heaton & Yaron (1996, JBES)
Optimal-Instrument GMMEfficient GMM for a linear IV model that builds Chamberlain (1987) optimal instruments A(z)=E[x|z]/Var(e|z) from a flexible projection of the regressors on the instruments plus a fitted conditional-variance model, reporting beta with efficient-GMM standard errors, first-stage fit, Hansen's J overidentification test, and an efficiency comparison versus plain 2SLS.✅ FullChamberlain (1987); Newey (1990); Hansen (1982)
Nonlinear GMM (exponential moments)Two-step efficient GMM for the exponential / multiplicative-error model y = exp(x'beta)eta with instruments, using Mullahy's multiplicative moment E[z(y*exp(-x'beta)-1)]=0 to identify possibly-endogenous regressors (a nonlinear analogue of 2SLS) and reporting the coefficient table plus Hansen's J overidentification test.✅ FullHansen (1982); Mullahy (1997)
Method of Simulated MomentsMethod of Simulated Moments / indirect inference for a univariate AR(1)-plus-measurement-noise model whose likelihood is intractable: parameters (phi, sigma_eta, sigma_e) are chosen so simulated auxiliary statistics (variance and first two autocovariances) match the data's, minimizing a scale-free weighted distance with common random numbers and reporting estimates, sandwich SEs with the (1+1/S) simulation-noise inflation, the objective value, and the moment-match table.✅ FullMcFadden (1989); Pakes & Pollard (1989); Gourieroux, Monfort & Renault (1993)
Minimum Distance (CMD)Classical minimum-distance (minimum-chi-square) estimation that recovers a structural IV slope from the OLS reduced-form coefficients by minimizing (pi_hat - h(theta))' W (pi_hat - h(theta)) with the efficient weight W = inv(Var(pi_hat)), reporting theta_hat, CMD standard errors, the overidentification chi-square, and (in the just-identified case) exact agreement with 2SLS.✅ FullRothenberg (1973); Chamberlain (1982); Wooldridge (2010)
IV Quantile Regression (Chernozhukov-Hansen)Estimates the structural effect of an endogenous regressor on a chosen conditional quantile of the outcome using one or more instruments, via the Chernozhukov-Hansen inverse-quantile-regression grid search that drives the instrument coefficient to zero, correcting the endogeneity bias that contaminates ordinary quantile regression.✅ FullChernozhukov & Hansen (2005, Econometrica; 2006, J. Econometrics)
Oaxaca-Blinder DecompositionDecomposes the mean outcome gap between two groups from separate within-group OLS fits into the threefold endowments/coefficients/interaction components and the twofold pooled-reference (Neumark) explained vs unexplained split.✅ FullOaxaca (1973); Blinder (1973); Jann (2008)
Prais-Winsten AR(1) RegressionIterative AR(1) FGLS that quasi-differences the data and retains a scaled first observation, correcting OLS for first-order serial correlation; reports rho and before/after Durbin-Watson.✅ FullPrais & Winsten (1954)
Linear Probability ModelOLS of a binary 0/1 outcome with HC1 robust SEs; coefficients are marginal effects on P(y=1), and the share of out-of-[0,1] fitted values is flagged.✅ FullWooldridge (2010)
Generalized Linear Model (family x link)Unified GLM: choose the exponential family (gaussian, poisson, binomial, gamma, inverse-Gaussian) and link (identity, log, logit, probit, inverse, sqrt). Reports coefficients, deviance, AIC/BIC.✅ FullNelder & Wedderburn (1972)
Linear Combination of Coefficients (lincom)Tests an arbitrary linear combination g'β of OLS coefficients with its standard error, t, p-value, and 95% CI (Stata-style lincom).✅ FullWooldridge (2010)
Delta-Method (nonlinear combination)Delta-method standard error for a sum / difference / ratio / product of two OLS coefficients, with t, p-value, and 95% CI.✅ FullWooldridge (2010)
Restriction Test (F / Wald / LR)General joint-exclusion test: F, Wald, and likelihood-ratio statistics for dropping a subset of regressors (unrestricted vs restricted OLS).✅ FullWooldridge (2010)
Binary Logistic RegressionBinary outcome (0/1) modeled as probability; reports odds ratios, McFadden R², classification accuracy.✅ FullNelder & Wedderburn (1972); McFadden (1974)
Probit RegressionBinary choice via normal CDF link; includes marginal effects and classification accuracy.✅ FullSeabold & Perktold (2010)
Multivariate ProbitJointly models 3+ correlated binary outcomes by maximum simulated likelihood with the GHK simulator, returning per-equation probit coefficients plus the full cross-equation error correlation matrix.✅ FullCappellari & Jenkins (2003); Train (2009)
Bivariate ProbitJointly estimates two binary-outcome probit equations with correlated normal errors by maximizing the bivariate-normal log-likelihood, reporting both coefficient tables, the error correlation rho with a Wald rho=0 test that detects whether the outcomes can be modeled as two independent probits, the log-likelihood and n.✅ FullAshford & Sowden (1970); Greene (2018)
Ordered LogitOrdinal outcome; reports coefficients, threshold parameters, and McFadden pseudo-R².✅ FullSeabold & Perktold (2010)
Generalized Ordered LogitFits an ordinal-outcome model that relaxes the proportional-odds (parallel-lines) assumption by letting each cumulative split have its own ML-estimated coefficient vector, reporting per-threshold coefficient tables, a Brant-style Wald test (with an LR cross-check) of parallel lines, log-likelihood, AIC/BIC, and an interpretation.✅ FullWilliams (2006); Fu (1998)
Sequential / Continuation-Ratio LogitModels an ordinal outcome as a sequence of binary continuation-ratio logits -- one per ordered transition, fit on the at-risk subset {Y>=r} -- yielding stage-specific coefficients for outcomes whose categories are stages literally passed through in order.✅ FullFienberg (1980); Tutz (1991)
Ordered ProbitOrdinal outcome via normal CDF; threshold parameters and pseudo-R² included.✅ FullSeabold & Perktold (2010)
Multinomial LogitMulti-category outcome; reports coefficients and relative-risk ratios per category relative to base.✅ FullSeabold & Perktold (2010)
Choice-Based Sampling (WESML)Corrects a conditional/multinomial logit for endogenous (choice-based) stratified sampling using the Manski-Lerman WESML estimator, reweighting each observation by population-share / sample-share of its chosen alternative to recover consistent population coefficients with robust sandwich standard errors, alongside a side-by-side comparison to the biased unweighted fit.✅ FullManski & Lerman (1977)
Multinomial Probit (GHK)Estimates a random-utility multinomial probit on long-format choice data by maximum simulated likelihood with the Geweke-Hajivassiliou-Keane (GHK) recursive simulator, recovering attribute coefficients and the identified correlated-error covariance while relaxing the IIA restriction of the conditional logit.✅ FullGeweke (1991); Hajivassiliou & McFadden (1998); Keane (1994)
Nested LogitTwo-level nested logit for discrete choice estimated by FIML, relaxing IIA within nests via per-nest dissimilarity (lambda) parameters with a 0<lambda<=1 RUM-consistency check and a likelihood-ratio test against the conditional logit.✅ FullMcFadden (1978); Train (2009)
Mixed (Random-Parameters) LogitRandom-parameters (mixed) logit estimated by simulated maximum likelihood on long-format discrete-choice data, drawing Normal mixing coefficients via Halton draws to recover preference means and the std devs that quantify unobserved heterogeneity.✅ FullTrain (2009); Revelt & Train (1998)
Maximum Simulated LikelihoodFits a random-coefficient (mixed) binary-choice probit by Maximum Simulated Likelihood, approximating the intractable integral over unobserved slope heterogeneity with shifted-Halton low-discrepancy draws and reporting the mean coefficients, the heterogeneity standard deviation, the simulated log-likelihood, the number of draws, and a likelihood-ratio comparison to the pooled (no-heterogeneity) probit.✅ FullTrain (2009); Gourieroux & Monfort (1996)
Conditional Logit (McFadden)Estimates McFadden's choice-specific conditional logit on long-format discrete-choice data, recovering how alternative-varying attributes drive the probability that each alternative is chosen within a choice occasion, with a full coefficient table, log-likelihood, and McFadden pseudo-R2.✅ FullMcFadden (1974)
Poisson RegressionCount data (non-negative integer); reports incidence-rate ratios (IRRs) and dispersion diagnostics; optional exposure/offset term for rate (frequency) models.✅ FullSeabold & Perktold (2010)
Negative Binomial (NB1 / NB2)Overdispersed count data; MLE-estimated dispersion (alpha). Selectable NB2 (quadratic variance, default) or NB1 (linear variance). Includes likelihood-ratio test of Poisson vs NB.✅ FullHilbe (2011); Cameron & Trivedi (2013)
Zero-Inflated Poisson (ZIP)Count data with excess zeros via mixing; separate count and zero-inflation equations with Vuong test.✅ FullSeabold & Perktold (2010)
Zero-Inflated Negative Binomial (ZINB)Overdispersed count data with excess zeros; count + inflation + alpha parameter estimated jointly.✅ FullSeabold & Perktold (2010)
Hurdle ModelTwo-part: logit for zero vs nonzero + truncated Poisson for positive counts.✅ FullSeabold & Perktold (2010)
Tobit Regression (Censored)Outcome censored at known bound (lower/upper); MLE estimation with sigma parameter.✅ FullSeabold & Perktold (2010)
Censored LAD (Powell CLAD)Powell's (1984) Censored Least Absolute Deviations estimator for left-censored (Tobit-type) outcomes, fit by an iterated median-regression linear-programming scheme with bootstrap standard errors, consistent under conditional-median restrictions and robust to heteroskedastic or heavy-tailed errors where Tobit is biased.✅ FullPowell (1984, J. Econometrics)
Truncated RegressionSample truncated (observations above threshold retained); MLE with sigma.✅ FullSeabold & Perktold (2010)
Heckman Selection Model (Two-Step)Endogenous sample selection via probit + IMR augmentation; reports rho correlation coefficient.✅ FullSeabold & Perktold (2010)
Endogenous Switching (Roy model)Estimates a two-regime Roy switching regression with regime-specific outcome equations and a probit selection equation, correcting each regime for selection on unobservables via its own inverse Mills ratio and reporting the rho_0/rho_1 error-correlation terms.✅ FullRoy (1951); Lee (1978); Maddala (1983)
Beta RegressionOutcome in (0,1); precision parameter (phi) and mean-model coefficients estimated jointly.✅ FullSeabold & Perktold (2010)
Fractional-response quasi-MLE (Papke-Wooldridge)Outcome in [0,1]; Papke-Wooldridge (1996) Bernoulli quasi-MLE — GLM with binomial family + logit link; includes marginal effects.✅ FullPapke & Wooldridge (1996); Nelder & Wedderburn (1972)
Quantile RegressionConditional quantile (median or specified tau); pseudo-R² and quantile-specific inference.✅ FullSeabold & Perktold (2010)
Ridge Regression (L2)L2 regularized OLS on standardized predictors; alpha controls shrinkage intensity.✅ FullSeabold & Perktold (2010)
LASSO Regression (L1)L1 regularized OLS; automatic feature selection via coefficient sparsity. Standardized fit.✅ FullSeabold & Perktold (2010)
Elastic Net (L1+L2)Combined L1/L2 penalty on standardized predictors; l1_ratio balances penalties.✅ FullSeabold & Perktold (2010)
Elastic Net (cross-validated)Elastic-net penalized linear regression that standardizes predictors and uses K-fold cross-validation to pick the penalty strength lambda at minimum CV-MSE and at the 1-SE rule, reporting the CV-MSE curve, original-scale selected coefficients, the number selected, and R^2.✅ FullZou & Hastie (2005, JRSS-B); Friedman, Hastie & Tibshirani (2010, J. Stat. Softw. / glmnet)
Robust Regression (Huber M-estimator)M-estimator resistant to outliers; asymptotic-normal Wald inference (z-stat); no t-distribution df.✅ FullSeabold & Perktold (2010)
Stepwise SelectionGreedy forward/backward/both; adds/removes by AIC or BIC until no improvement.✅ FullSeabold & Perktold (2010)
Weighted Least Squares (WLS)User-supplied or estimated weights (FGLS); corrects heteroscedasticity.✅ FullSeabold & Perktold (2010)
Feasible Generalized Least Squares (FGLS)Cochrane-Orcutt iterative GLSAR; corrects serial correlation in errors.✅ FullSeabold & Perktold (2010)
Nonlinear Least Squares (NLS)User-provided formula and starting values; scipy.optimize curve_fit minimizes residual sum of squares.✅ FullSeabold & Perktold (2010)
Seemingly Unrelated Regressions (SUR)Multiple equations fitted jointly accounting for cross-equation correlation; Breusch-Pagan test included.✅ FullSeabold & Perktold (2010)
Three-Stage Least Squares (3SLS)System of equations with endogenous right-hand-side variables; IV + SUR combined.✅ FullSeabold & Perktold (2010)
FIML (simultaneous system)Full-Information Maximum Likelihood joint estimation of a two-equation linear simultaneous-equations system, maximizing the system Gaussian log-likelihood concentrated over the residual covariance (with the ln|det(B)| Jacobian) to recover structural coefficients with FIML standard errors, the residual covariance/correlation, the system log-likelihood, and a side-by-side comparison against equation-by-equation 2SLS.✅ FullKoopmans, Rubin & Leipnik (1950); Hausman (1983); Greene (2018)
Generalized Method of Moments (GMM)IV-GMM with moment conditions; J-test of overidentifying restrictions.✅ FullSeabold & Perktold (2010)
Panel Fixed EffectsWithin-entity estimator; removes time-invariant heterogeneity. Clustered SEs on entity. Optional two-way (entity + time) effects.✅ FullSeabold & Perktold (2010)
Panel Random EffectsGLS random-effects estimator; assumes random effect uncorrelated with predictors.✅ FullSeabold & Perktold (2010)
Linear Mixed Model (Random Intercepts/Slopes)REML by default; fixed-effect Wald-z inference. AIC/BIC reported only under ML.✅ FullLaird & Ware (1982); Harville (1977)
Intraclass Correlation (ICC)ICC(1) from one-way random-intercept model; variance-partition coefficient for clustering.✅ FullShrout & Fleiss (1979); McGraw & Wong (1996)
Difference-in-Differences (DiD)Treatment×Post interaction; cluster-robust or HC1 SEs depending on clustering.✅ FullCard & Krueger (1994); Bertrand et al. (2004)
Regression Discontinuity (Sharp RDD)Local-linear at cutoff with triangular or uniform kernel. Ad-hoc or user-supplied bandwidth.⚠️ LimitedSeabold & Perktold (2010)
RD Robust (CCT bias-corrected)Sharp regression-discontinuity treatment effect via local-polynomial (triangular-kernel) fits on each side of the cutoff within an MSE-optimal bandwidth, reporting the conventional point estimate, the Calonico-Cattaneo-Titiunik (2014) bias-corrected estimate, conventional and robust standard errors, confidence intervals, and effective sample sizes.✅ FullCalonico, Cattaneo & Titiunik (2014); Imbens & Kalyanaraman (2012)
Propensity Score Matching (PSM)Logit/probit propensity score; 1:1/1:k matching with replacement or caliper.✅ FullRosenbaum & Rubin (1983)
Extended Regression Model — Linear (ERM-Linear)Combines endogeneity (IV-2SLS), sample selection (Heckman), treatment effects in one model.✅ FullSeabold & Perktold (2010)
Extended Regression Model — Probit (ERM-Probit)Endogeneity (control function) + selection (bivariate-probit IMR) for binary outcome.⚠️ LimitedSeabold & Perktold (2010)
Extended Regression Model — Ordered (ERM-Ordered)Ordered probit with sample selection correction via IMR augmentation.⚠️ LimitedSeabold & Perktold (2010)
Extended Regression Model — Interval (ERM-Interval)Outcome observed as bounds [lower, upper]; handles left/right/interval censoring via MLE.✅ FullSeabold & Perktold (2010)
ERM DiagnosticsSpecification tests for endogeneity, selection bias, and treatment effect significance within ERM.✅ FullSeabold & Perktold (2010)
LOESS RegressionLocal polynomial regression; non-parametric smoothing via weighted least squares.✅ FullSeabold & Perktold (2010)
Gamma GLM (severity)Gamma generalized linear model with log link for strictly-positive responses (e.g. claim severity/cost); supports an exposure/offset term. Dispersion via Pearson, matching R glm(family=Gamma(link="log")).✅ FullNelder & Wedderburn (1972)
Tweedie GLM (pure premium)Tweedie compound Poisson-Gamma GLM (log link) for non-negative data with a mass at zero — single-model pure-premium pricing; configurable variance power p in (1,2).✅ FullNelder & Wedderburn (1972); Jorgensen (1987)

Notes & limitations

  • Regression Discontinuity (Sharp RDD) — Bandwidth selection is ad-hoc (1.5*SD rule), not data-driven MSE-optimal.
  • Extended Regression Model — Probit (ERM-Probit) — IMR lambda SE not corrected for generated-regressor problem (Murphy-Topel correction not implemented).
  • Extended Regression Model — Ordered (ERM-Ordered) — Lambda SE uncorrected; no p-values reported due to generated-regressor issue.