What follows is a longish controversy (two critiques, a reply and two rejoinders) over the quality of the estimates and associated SEs provided by Plümper and Troeger's (2007) “fixed-effect vector decomposition” (FEVD) procedure; Plümper and Troeger (PT) will refer to that article and not any persons. My role is to lay out some issues that separate the authors rather than to adjudicate between them. As with many controversies, a bit of heat is generated along with some light. Readers care a bit less than the authors about what was said when, but they do care a lot about what appropriate method to use when a panel data model has both unit-specific intercepts and variables that are invariant over a unit. Thus, I also take it upon myself to discuss some things that I gleaned from this controversy; this discussion has a bit less heat than what follows, but of course readers should judge the evidence for themselves.

This paper analyzes the properties of the fixed-effects vector decomposition estimator, an emerging and popular technique for estimating time-invariant variables in panel data models with group effects. This estimator was initially motivated on heuristic grounds, and advocated on the strength of favorable Monte Carlo results, but with no formal analysis. We show that the three-stage procedure of this decomposition is equivalent to a standard instrumental variables approach, for a specific set of instruments. The instrumental variables representation facilitates the present formal analysis that finds: (1) The estimator reproduces exactly classical fixed-effects estimates for time-varying variables. (2) The standard errors recommended for this estimator are too small for both time-varying and time-invariant variables. (3) The estimator is inconsistent when the time-invariant variables are endogenous. (4) The reported sampling properties in the original Monte Carlo evidence do not account for presence of a group effect. (5) The decomposition estimator has higher risk than existing shrinkage approaches, unless the endogeneity problem is known to be small or no relevant instruments exist.

Plümper and Troeger (2007) propose a three-step procedure for the estimation of a fixed effects (FE) model that, it is claimed, “provides the most reliable estimates under a wide variety of specifications common to real world data.” Their fixed effects vector decomposition (FEVD) estimator is startlingly simple, involving three simple steps, each requiring nothing more than ordinary least squares (OLS). Large gains in efficiency are claimed for cases of time-invariant and slowly time-varying regressors. A subsequent literature has compared the estimator to other estimators of FE models, including the estimator of Hausman and Taylor (1981) also (apparently) with impressive gains in efficiency. The article also claims to provide an efficient estimator for parameters on time-invariant variables (TIVs) in the FE model. None of the claims are correct. The FEVD estimator simply reproduces (identically) the linear FE (dummy variable) estimator then substitutes an inappropriate covariance matrix for the correct one. The consistency result follows from the fact that OLS in the FE model is consistent. The “efficiency” gains are illusory. The claim that the estimator provides an estimator for the coefficients on TIVs in an FE model is also incorrect. That part of the parameter vector remains unidentified. The “estimator” relies upon a strong assumption that turns the FE model into a type of random effects model.

This article reinforces our 2007 Political Analysis publication in demonstrating that the fixed-effects vector decomposition (FEVD) procedure outperforms any other estimator in estimating models that suffer from the simultaneous presence of time-varying variables correlated with unobserved unit effects and time-invariant variables. We compare the finite-sample properties of FEVD not only to the Hausman-Taylor estimator but also to the pretest estimator and the shrinkage estimator suggested by Breusch, Ward, Nguyen and Kompas (BWNK), and Greene in this symposium. Moreover, we correct the discussion of Greene and BWNK of FEVD's asymptotic and finite-sample properties.

Fixed effects vector decomposition (FEVD) is simply an instrumental variables (IV) estimator with a particular choice of instruments and a special case of the well-known Hausman-Taylor IV procedure. Plümper and Troeger (PT) now acknowledge this point and disown the three-stage procedure that previously defined FEVD. Their old recipe for SEs, which has regrettably been used in dozens of published research papers, produces dramatic overconfidence in the estimates. Again PT concede the point and now adopt the standard IV formula for SEs. Knowing that FEVD is an application of IV also has the benefit of focusing attention on the choice of instruments. Now it seems PT claim that the FEVD instruments are always the best choice, on the grounds that one cannot know whether any potential instrument is correlated with the unit effect. One could just as readily make the same specious claim about other estimators, such as ordinary least squares, and support it with similar Monte Carlo assumptions and evidence.

Contingency tables are among the most basic and useful techniques available for analyzing categorical data, but they produce highly imprecise estimates in small samples or for population subgroups that arise following repeated stratification. I demonstrate that preprocessing an observed set of categorical variables using a latent class model can greatly improve the quality of table-based inferences. As a density estimator, the latent class model closely approximates the underlying joint distribution of the variables of interest, which enables reliable estimation of conditional probabilities and marginal effects, even among subgroups containing fewer than 40 observations. Though here focused on applications to public opinion, the procedure has a wide range of potential uses. I illustrate the benefits of the latent class model—based approach for greatly improved accuracy in estimating and forecasting vote preferences within small demographic subgroups using survey data from the 2004 and 2008 U.S. presidential election campaigns.

In this paper, I introduce changepoint models for binary and ordered time series data based on Chib's hidden Markov model. The extension of the changepoint model to a binary probit model is straightforward in a Bayesian setting. However, detecting parameter breaks from ordered regression models is difficult because ordered time series data often have clustering along the break points. To address this issue, I propose an estimation method that uses the linear regression likelihood function for the sampling of hidden states of the ordinal probit changepoint model. The marginal likelihood method is used to detect the number of hidden regimes. I evaluate the performance of the introduced methods using simulated data and apply the ordinal probit changepoint model to the study of Eichengreen, Watson, and Grossman on violations of the “rules of the game” of the gold standard by the Bank of England during the interwar period.

If ignored, noncompliance with a treatment or nonresponse on outcome measures can bias estimates of treatment effects in a randomized experiment. To identify and estimate causal treatment effects in the case where compliance and response depend on unobservables, we propose the parametric generalized endogenous treatment (GET) model. GET incorporates behavioral responses within an experiment to measure each subject's latent compliance type and identifies causal effects via principal stratification. Using simulation methods and an application to field experimental data, we show GET has a dramatically lower mean squared error for treatment effect estimates than existing approaches to principal stratification that impute, rather than measure, compliance type. In addition, we show that GET allows one to relax and test the instrumental variable exclusion restriction assumption, to test for the presence of treatment effect heterogeneity across a range of compliance types, and to test for treatment ignorability when treatment and control samples are balanced on observable covariates.

Although methodologists have provided us ample notice of both the problem of nonproportional hazards (NPHs) and the means of correcting for them, less attention has been paid to the postestimation interpretation. The suggested inclusion of time interactions in our models is more than a statistical fix: These corrections alter the substantive meaning and interpretation of results. Framing the issue as a specific case of multiplicative-interaction modeling, I provide detailed discussion of the problem of NPHs and present several appropriate means of interpreting both the substantive impact and the significance of variables whose effects may change over time.