For a nonlinear model with heteroskedasticity, a maximum likelihood estimator gives misleading inference and inconsistent marginal effect estimates unless I model the variance. Using a robust estimate of the variance–covariance matrix will not help me obtain correct inference.

This differs from the intuition we gain from linear regression. The estimates of the marginal effects in linear regression are consistent under heteroskedasticity and using robust standard errors yields correct inference.

If robust standard errors do not solve the problems associated with heteroskedasticity for a nonlinear model estimated using maximum likelihood, what does it mean to use robust standard errors in this context? I answer this question using simulations and illustrate the effect of heteroskedasticity in nonlinear models estimated using maximum likelihood. Read more…

I illustrate that exact matching on discrete covariates and regression adjustment (RA) with fully interacted discrete covariates perform the same nonparametric estimation. Read more…

We estimate the average treatment effect (ATE) for an exponential mean model with an endogenous treatment. We have a two-step estimation problem where the first step corresponds to the treatment model and the second to the outcome model. As shown in Using gmm to solve two-step estimation problems, this can be solved with the generalized method of moments using gmm.

This continues the series of posts where we illustrate how to obtain correct standard errors and marginal effects for models with multiple steps. In the previous posts, we used gsem and mlexp to estimate the parameters of models with separable likelihoods. In the current model, because the treatment is endogenous, the likelihood for the model is no longer separable. We demonstrate how we can use gmm to estimate the parameters in these situations. Read more…

\(\newcommand{\Eb}{{\bf E}}
\newcommand{\xb}{{\bf x}}
\newcommand{\betab}{\boldsymbol{\beta}}\)Differences in conditional probabilities and ratios of odds are two common measures of the effect of a covariate in binary-outcome models. I show how these measures differ in terms of conditional-on-covariate effects versus population-parameter effects. Read more…

\(\newcommand{\Eb}{{\bf E}}\)The change in a regression function that results from an everything-else-held-equal change in a covariate defines an effect of a covariate. I am interested in estimating and interpreting effects that are conditional on the covariates and averages of effects that vary over the individuals. I illustrate that these two types of effects answer different questions. Doctors, parents, and consultants frequently ask individuals for their covariate values to make individual-specific recommendations. Policy analysts use a population-averaged effect that accounts for the variation of the effects over the individuals. Read more…

I want to estimate, graph, and interpret the effects of nonlinear models with interactions of continuous and discrete variables. The results I am after are not trivial, but obtaining what I want using margins, marginsplot, and factor-variable notation is straightforward. Read more…

As of Stata 16, see [BAYES] bayesstats grubin and Bayesian analysis: Gelman-Rubin convergence diagnostic.

The original blog posted May 26, 2016, omitted option initrandom from the bayesmh command. The code and the text of the blog entry were updated on August 9, 2018, to reflect this.

Overview

MCMC algorithms used for simulating posterior distributions are indispensable tools in Bayesian analysis. A major consideration in MCMC simulations is that of convergence. Has the simulated Markov chain fully explored the target posterior distribution so far, or do we need longer simulations? A common approach in assessing MCMC convergence is based on running and analyzing the difference between multiple chains.

For a given Bayesian model, bayesmh is capable of producing multiple Markov chains with randomly dispersed initial values by using the initrandom option, available as of the update on 19 May 2016. In this post, I demonstrate the Gelman–Rubin diagnostic as a more formal test for convergence using multiple chains. For graphical diagnostics, see Graphical diagnostics using multiple chains in [BAYES] bayesmh for more details. To compute the Gelman–Rubin diagnostic, I use an unofficial command, grubin, which can be installed by typing the following in Stata: Read more…

Initial thoughts

Estimating causal relationships from data is one of the fundamental endeavors of researchers. Ideally, we could conduct a controlled experiment to estimate causal relations. However, conducting a controlled experiment may be infeasible. For example, education researchers cannot randomize education attainment and they must learn from observational data.

In the absence of experimental data, we construct models to capture the relevant features of the causal relationship we have an interest in, using observational data. Models are successful if the features we did not include can be ignored without affecting our ability to ascertain the causal relationship we are interested in. Sometimes, however, ignoring some features of reality results in models that yield relationships that cannot be interpreted causally. In a regression framework, depending on our discipline or our research question, we give a different name to this phenomenon: endogeneity, omitted confounders, omitted variable bias, simultaneity bias, selection bias, etc.

Below I show how we can understand many of these problems in a unified regression framework and use simulated data to illustrate how they affect estimation and inference. Read more…

Overview

In the frequentist approach to statistics, estimators are random variables because they are functions of random data. The finite-sample distributions of most of the estimators used in applied work are not known, because the estimators are complicated nonlinear functions of random data. These estimators have large-sample convergence properties that we use to approximate their behavior in finite samples.

Two key convergence properties are consistency and asymptotic normality. A consistent estimator gets arbitrarily close in probability to the true value. The distribution of an asymptotically normal estimator gets arbitrarily close to a normal distribution as the sample size increases. We use a recentered and rescaled version of this normal distribution to approximate the finite-sample distribution of our estimators.

I illustrate the meaning of consistency and asymptotic normality by Monte Carlo simulation (MCS). I use some of the Stata mechanics I discussed in Monte Carlo simulations using Stata.

Consistent estimator

A consistent estimator gets arbitrarily close in Read more…

This post was written jointly with Yulia Marchenko, Executive Director of Statistics, StataCorp.

As of update 03 Mar 2016, bayesmh provides a more convenient way of fitting distributions to the outcome variable. By design, bayesmh is a regression command, which models the mean of the outcome distribution as a function of predictors. There are cases when we do not have any predictors and want to model the outcome distribution directly. For example, we may want to fit a Poisson distribution or a binomial distribution to our outcome. This can now be done by specifying one of the four new distributions supported by bayesmh in the likelihood() option: dexponential(), dbernoulli(), dbinomial(), or dpoisson(). Previously, the suboption noglmtransform of bayesmh‘s option likelihood() was used to fit the exponential, binomial, and Poisson distributions to the outcome variable. This suboption continues to work but is now undocumented.

For examples, see Beta-binomial model, Bayesian analysis of change-point problem, and Item response theory under Remarks and examples in [BAYES] bayesmh.

We have also updated our earlier “Bayesian binary item response theory models using bayesmh” blog entry to use the new dbernoulli() specification when fitting 3PL, 4PL, and 5PL IRT models.