The Estimation of Time-Invariant Variables in Panel Analyses with Unit Fixed Effects THOMAS PLÜMPER# # AND VERA E. TROEGER + University of Konstanz, Department of Political Science and Public Administration, Box D 86; D-78457 Konstanz, Germany; Tel: +49-7531-88 2608/3081; Fax: +49-7531-88 2774; E-Mail: Thomas.Pluemper@unikonstanz.de + Max-Planck-Institute for Research into Economic Systems, Kahlaische Strasse 10, D-07745 Jena, Germany Summary: This paper analyzes the estimation of time-invariant variables in panel data models with unit-effects. We compare three procedures that have frequently been employed in comparative politics, namely pooled-OLS, random effects and the Hausman-Taylor model, to a vector decomposition procedure that allows estimating time-invariant variables in an augmented fixed effects approach. The procedure we suggest consists of three stages: the first stage runs a fixed-effects model without time-invariant variables, the second stage decomposes the uniteffects vector into a part explained by the time-invariant variables and an error term, and the third stage re-estimates the first stage by pooled-OLS including the time invariant variables plus the error term of stage 2. We use Monte Carlo simulations to demonstrate that this method works better than its alternatives in estimating typical models in comparative politics. Specifically, the unit fixed effects vector decomposition technique performs better than both pooled OLS and random effects in the estimation of time-invariant variables correlated with the unit effects and better than Hausman-Taylor in estimating the time-invariant variables correlated with the unit effects. Finally, we re-analyze recent work by Huber and Stephens (2001) as well as by Beramendi and Cusack (2004). These analyses seek to cope with the problem of time-invariant variables in panel data. John Stephens and Pablo Beramendi provided not only access to their data but also assisted our replication attempts. We thank Neal Beck, Christian Kraft, Richard Stazinski for helpful comments. The usual disclaimer applies. 2 The Estimation of Time-Invariant Variables in Panel Analyses with Unit Fixed Effects 1. Introduction Since the publication of Nathaniel Beck’s and Jonathan Katz’s influential article on pooled time-series analysis (Beck/ Katz 1995) analyzing panel data has become a standard in comparative political science. As political scientists have frequently observed, standards provide focal points and they are applied because deviations from these standards are likely to be costly. In comparative politics, referees may impose ‘sanctions’ on researchers that deviate from pooling. Beyond this obvious incentive structure, pooling also has some objective advantages. Social scientists are often interested in explaining the variance of trends across units so that neither cross-sectional nor time-series analyses provide relevant tests of the theory. Hence, pooling is warranted. At times, researchers have analyzed panel data because pooling increases the number of observations. The most important methodical advantage of panel data relates to the growing concern of applied researchers about bias resulting from omitted variables. Contrary to a cross-section or a pure time-series panel data analyses allow controlling for unit fixed-effects that – as most researchers believe – capture the systematic influences from omitted variables. This belief is not wrong but potentially misleading, since unit fixed effects do not eliminate all kinds of omitted variable bias. Time-variant omitted variables may still bias the estimates. Thus, one danger of fixed effects models is that many researchers believe that the inclusion of unit dummies precludes problems with omitted variables. Unit effects represent “all factors (…) that do not change over time” (Wooldridge 1999: 420). In other words, unit-fixed effects account for time-invariant crosssectional effects – may they be observed or unobserved. This has two obvious but notable consequences: on the one hand, unit dummies do not necessarily eliminate omitted variable bias; on the other hand, unit fixed effects models cannot estimate the coefficients of theoretically interesting time-invariant variables. For the latter reason the quest to include or exclude country dummies has stimulated a lively debate among political scientists. While some authors suggest that country dummies are needed to account for the “underlying historical fabric (…) that is not captured by any of the time and country-varying regressors” (Garrett/ Mitchell 2001: 163), others claim that unit fixed effects throw out the 3 baby with the bath water”, because political scientist are mainly interested in institutional effects and institutions do not vary much over time (Kittel/ Obinger 2002: 21). This paper discusses potential remedies for the estimation of time-invariant variables in panel-data analyses with unit effects. We compare three procedures that have frequently been used in comparative politics – namely pooled-OLS, random effects and the Hausman-Taylor formulation – to a vector decomposition procedure that allows estimating time-invariant variables in an augmented fixed effects approach. The model we advocate consists of three stages: in the first stage we run a fixed-effects model, in the second stage the unit-effects vector is decomposed into a part explained by the time-invariant variables and an error term, in the third stage we re-estimate the first stage including the time invariant variables and the error term obtained in stage 2 by pooled-OLS. We evaluate the small-sample properties of the procedure we advocate (dubbed xtfevd for fixed effect vector decomposition) to pooled OLS, random effects and Hausman-Taylor. Monte Carlo simulations demonstrate that xtfevd is on average less biased than the available alternatives. Moreover, xtfevd has better small sample properties than Hausman-Taylor. At the same time it is less biased than random effects and pooled OLS when unit effects are correlated with time-variant variables. xtfevd outperforms pooled OLS and random effects in estimating timevariant variables that are correlated with the unit effects ( ui ) and it is more adequate than Hausman-Taylor in calculating the effect of time-variant variables that are correlated with the unit effects. Thus, employing the unit fixed effects vector decomposition technique appears to be superior for the analysis of datagenerating processes typical for panel data in comparative politics. We will also identify one situation in which neither of the four models does particularly well. When the unit-effects are correlated with exogenous variables and highly skewed, all methods give biased estimates. The remainder of the paper is organized as follows. The next section briefly (re-) introduces the alternative procedures to the reader. Section 3 presents the xtfevd estimator. In section 4, we compare existing approaches by conducting Monte Carlo simulations. Section 5 re-analyzes Evelyn Huber and John D. Stephens (2001) examination of the impact of constitutional features on the dynamics of social welfare spending and Pablo Beramendi and Thomas Cusack’s (2004) analysis of the political determinants of income inequality. Section 6 concludes. 4 2. The estimation of time-invariant variables Time-invariant variables can be distinguished into two broadly defined categories. The first type subsumes variables that are time-invariant by definition. Often, these variables measure geography or inheritance. Switzerland and Hungary are both landlocked countries, they are both located in Central Europe, and there is little nature and (hopefully) politics will do about it for the foreseeable future. Along similar lines, a country may or may not have a colonial heritage or a climate prone to tropical diseases. The second category covers variables that are time-invariant for the period under analysis or because of researchers’ selection of cases. For instance, constitutions in postwar OECD countries have proven to be highly durable. Switzerland is a democracy since 1291 and the US maintained a presidential system since independence day. Yet, by increasing the number of periods and/or the number of cases it would be possible to render these variables time-variant. This suggests that many variables are time-invariant because of researchers’ deliberate choice. At times, researchers even suppress time-varying information and operationalize a theoretically plausible influence as (time-invariant) dichotomous variable even though a (time-varying) continuous alternative exists. Adding observations may be preferable to searching for remedies of estimating panel data that include time-invariant variables. We suggest that researchers first consider broadening their sample of units or lengthening the time period under observation rather than employing one of the procedures we discuss here. Nevertheless, all methods enable researchers to estimate variables that are timeinvariant by nature or by researchers’ decision. One reason for suppressing time-series variance could be that variables rarely change. The identification of the effect of an almost time-invariant variable is difficult. By taking the unit means one can treat any changing variable as timeinvariant. Whether this is an appropriate research strategy clearly depends on the alternatives. From our perspective it is likely that testing the theory within a broader research design – one that includes more observations – is a superior research strategy. Most applied researchers have at one point been confronted with the appropriate estimation of time-invariant variables of theoretical interest. Even a brief look into datasets frequently used in political science reveals that time-invariant variables 5 are by no means rare. Many datasets contain variables that are time-invariant over relatively long periods. For example, data on political institutions usually includes constitutional variables that rarely and slowly change. Since no standard method exists, astonishingly diverse ways to deal with the problem of time-invariant variables have been employed in the literature. This section briefly discusses the procedures in use. Before doing so, however, a few remarks on notation are in order. We assume the following data generating process yit = α + X it β + Zi γ + u i + εit (1) where X it β stands for a vector of time-variant variables and Zi γ for a vector of time-invariant variables, u i denotes the unit effect and εit is the normal distributed error component. We further hold that u i is correlated with at least one of the time-variant variables and at least one of the time-invariant variables: p lim N →∞ 1 ' 1 x1it u i = 0; p lim x '2it u i ≠ 0 N N →∞ N (2) 1 ' 1 z1i u i = 0; p lim z '2i u i ≠ 0 N N →∞ N (3) and p lim N →∞ These assumptions are likely to be satisfied in most data analyzed in comparative politics. Having said this, we now explain how applied researchers have dealt with the estimation problems associated with this data generating process. To begin, all three procedures have in common that researchers shied away from controlling for fixed effects because the unit effect dummies and the time-invariant variables are perfectly collinear. Scholars either run random effects (often regardless of inconsistency), employ the Hausman–Taylor formulation (which is available in Stata since version 8.0), or neglect individual effects altogether and rather run a pooled OLS model. We discuss all potential remedies in turn. The first ‘solution’ to the estimation problems posed by time-invariant variables is to ignore the possibility of unit effects.1 For instance, Daron Acemoglu et al. (2002) justify not controlling for unit effects by stating the following: “Recall that 1 Oaxaca and Geisler (2003) suggest a two-stage GLS estimator which is equivalent to pooled OLS, but gives different standard errors. 6 our interest is in the historically-determined component of institutions (that is more clearly exogenous), hence not in the variations in institutions from year-toyear. As a result, this regression does not (cannot) control for a full set of country dummies.” (Acemoglu et al. 2002: 27) Doing nothing about unit effects and running pooled OLS does not offer a way to reduce omitted variable bias. As is well understood, the exclusion of exogenous variables that influence the endogenous variable does cause bias, unless the excluded variables and the included exogenous variables are orthogonal to each other. The same holds true for excluded unit fixed effects. The exclusion of a battery of unit dummies that captures ui biases estimates if ui is correlated with both, yit and xit . Therefore, it seems unlikely to obtain unbiased coefficients by pooled OLS, especially if N is small. A strategy that in many situations should work more appropriate than pooled OLS in the presence of time-invariant variables is adapting the random effects model. Random effects should perform better than pooled OLS even if the Hausman test suggests (Hausman 1978) that random effects are inconsistent and a fixed effects specification is required. When time invariant variables preclude the estimation of unit fixed effects, random effects may serve as a viable second best option. One should keep in mind, however, that when the Hausman test rejects a random effects specification, the procedure is inconsistent and very likely to be biased, because random effects impose strict exogeneity of x it and orthogonality between x it and u i :2 E ( u it | x i , u i ) = 0; t = 1,..., T E ( x i | u i ) = E(u i ) = 0; x i = ( x i1 , x i2 ,..., x iT ) (5) Moreover, random effects are based on a feasible gls estimator where the Omega matrix has a special random effects structure. Rather than depending on T(T+1)/2 unrestricted variances and covariances as it is the case in a normal GLS model, Omega only depends on the variances of u i and εit regardless of the size of T: 2 But probably most importantly, real world data rarely satisfies the conditions under which random effects estimators are consistent. “For studies in political science using TSCS data it will almost always be the case that the unobserved local factors are captured by u i and correlated with X; indeed, the main reason to abandon the standard pooled OLS is because we think such a correlation is likely to exist.” (Wilson/ Butler 2003: 8) At least under this condition, random effects models are second best options of a potentially dubious quality. We will later see that random effects models do not work significantly better than pooled OLS. 7 Ω ≡ E ( v i vi ' ) vi = u i jT + εi ; jT is the T × 1vector of ones (6) Even if the data satisfies the random effects assumptions of strict exogeneity and orthogonality between x it and u i , random effects models share the poor small sample properties of GLS. As a consequence, one should expect that the random effects procedure gives biased and inefficient estimates of the true betas in relatively small samples. To overcome the problem of random effects inconsistency, Hausman and Taylor (1981) advocated the use of instruments for the variables that are likely to be correlated with the random effects. Unfortunately, this correlation is unobservable and thus it requires some imagination to correctly specify the Hausman-Taylor model. This is rarely a trivial problem. The estimated coefficients largely vary with researchers’ decision which variables are endogenous and which variables are exogenous to the random effects. Hence, the Hausman-Taylor procedure leaves researchers with a discretionary choice that largely influences the results. As a straightforward remedy for the apparent problem of choosing a set of instruments in the presence of an unknown correlation between the right-hand side variables and the random effects, Hausman and Taylor suggest using the exogenous variables that vary over time and are not correlated with the individualspecific part of the error term u i (x1) to instrument the variables correlated with u i (x2 and z2 of eqs. 2 and 3). Deviations from the mean of x1 are used to produce unbiased estimates for the time varying variables (x2) and the mean of x1 is used as an instrument for the time-invariant variable (z2). While Hausman and Taylor simply assume that x1 and z1 are uncorrelated with u, the applied researcher faces the problem of distinguishing endogenous from exogenous right-hand side variables. Moreover, it must not be the case that x1 and z1 are good instruments for x2 and z2, respectively. All three procedures used in the presence of time-invariant variables and unit effects suffer from omitted variable bias (pooled OLS), are likely to be inconsistent and biased (random effects), or have poor small sample properties 8 and leave the researchers with discrete choices that are hard to justify (HausmanTaylor).3 3. Fixed effects vector decomposition In this section we suggest an alternative procedure for the estimation of timeinvariant variables in the presence of unit effects that maintains the small sample properties and the unbiasedness of the fixed effects model. The logic of our specification is straightforward: Unit fixed effects are a vector of the mean effect of omitted variables, including the effect of time-invariant variables. Thus, in a first stage we obtain the unit fixed effects vector by estimating a fe-model that excludes the time-invariant variables. In a second stage, the vector can be decomposed into a part explained by the time-invariant variables and an errorterm.4 In stage 3, this error-term accounts for the unobserved unit fixed effects and, therefore, captures the potential of omitted variable bias. This enables us to re-run stage 1 by pooled OLS. Since we include only one variable (the error term of the second stage) to account for all remaining unobservable individual effects in the third stage regression, we have to adjust the degrees of freedom. In brief, the fixed effect vector decomposition technique carries out the following three steps: (1) estimation of the unit fixed effects by the baseline panel fixed effects model excluding the time-invariant right hand side variables; (2) regression of the fixed effects vector on the time invariant explanatory variables of the original model (by OLS); (3) estimation of a pooled OLS model by including all explanatory timevariant variables, the time-invariant variables and the unexplained part of the fixed effects vector. This stage is required to control for multicollinearity and to adjust the degrees of freedom.5 3 4 5 The poor quality of the instrument suggested has frequently raised some concerns and alternative proposals. See Breusch, Mizon and Schmidt’s (1989) proposal of feasible instruments and Amemiya and MaCurdy (1986). Having said this, we hasten to admit that there is at least one potential source of bias: As in all pure cross-sectional models, we cannot control for omitted variable bias in the estimation of the time-invariant variables in the second stage. An additional source of bias is the distribution of the fixed effects. If unit-fixed effects are not normally distributed, OLS estimates may be biased if N is small . Upon request we provide a STATA program (ado-file) that executes all three steps and adjusts the variance-covariance matrix. Several options, like ar1 errorcorrection and robust VC-matrix are allowed. 9 This procedure adopts the robustness of the fixed effects model and allows for correlation between explanatory variables and the unobserved individual effects. The effects of time varying factors are consistently analyzed and remain unbiased. Since fixed effects are robust with respect to potential correlation between right hand side variables and the individual specific effects the exogeneity of explanatory variables is not required. In econometric terms, the fevd technique works as follows. Recall the datagenerating process of eqs. 1-3. The within estimator de-means the data and removes the individual effects u i : yit − yi = ( x it − x i ) ß + εit − εi ≡ yit = x itβ + εit (7) Thus, the fixed effects are û i = yi − x iβˆ FE (8) In the second stage we regress the û i on the z-variables. û i = ω + z i γ + ηi (9) where ϖ is the intercept of the stage 2 equation and ηi is the error. Note that we get a biased estimate of γ if we exclude variables that are simultaneously correlated with the unit-effects û i and the time-invariant variables z i . As one can see, ηi is the part of u i that is not explained by the time invariant z-matrix. In the third stage, we re-run the full model without the unit effects but including the decomposed unit fixed effect vectors including ηi obtained in stage 2. This stage is estimated by pooled OLS (or Prais-Winston in the presence of serial correlation). yit = α + x it β + z i γ + ηi + εit (10) ηi is no longer correlated with any of the z i , but by including the error term of stage 2 we are able to account for individual specific effect that can not be observed. The coefficient of ηi is either equal to 1.0 or at least close to 1.0 (by accounting for serial correlation or panel heteroscedasticity) in stage 3. Stage 3 is necessary for basically two reasons. On the one hand we must adjust the degrees of freedom by u i -1 in calculating the variance-covariance matrix of β and γ . Not correcting the degrees of freedom would lead to potentially severe underestimation of standard errors and overconfidence in the results. In adjusting 10 the standard errors we explicitly control for the specific characteristics of the three step approach.6 On the other hand, stage 3 also accounts for the potential multicollinearity between the time-variant variables and the time-invariant variables. Estimating stage 3 by pooled OLS further requires that heteroscedasticity and serial correlation must be eliminated beforehand. We suggest running a robust Sandwich-estimator or/and model the dynamics by an MA1 process (PraisWinston transformation of the original data).7 At least in theory this method has three obvious advantages: a) the fixed effects vector decomposition does not require prior knowledge of correlation between the explanatory variables and the unit specific effects, b) the estimator relies on the robustness of the within-transformation and does not need to meet the orthogonality assumptions of random effects, and c) xtfevd maintains the consistency and efficiency of OLS. 4. Monte Carlo Experiments Given the properties of the unit effects decomposition technique there are good reasons to believe that this estimator is superior to pooled OLS, random effects and Hausman-Taylor if the empirical model includes time-invariant variables in panel data. We now report a series of Monte Carlo experiments which aim at assessing the performance of the available procedures in fairly small samples which are typical for comparative politics. We will demonstrate that xtfevd does better than its alternatives in situations common to applied political research. Before discussing the results of the Monte Carlo experiments, however, we briefly describe the design of the experiments. Design of the Experiments All experiments use simulated data, which are generated to discriminate between the various estimators, while at the same time mimic some properties of panel data. Specifically, the data generating process underlying our simulations is as follows: yit = α + β1x1it + β2 x2it + β3 x3it + β 4 z1i + β5 z2i + β6 z3i + u i + ε it 6 7 (11) Upon request we provide a STATA program (ado-file) called xtfevd that executes all three steps and adjusts the variance-covariance matrix. The STATA ado allows for an ar1 error-correction and robust VC-matrix. 11 where the x-variables are time varying and the z-variables are time-invariant, both groups are drawn from a normal distribution. Variables x3 and z3 are correlated with the unit specific effect ui .8 Variables x1, x2, z1 and z2 are uncorrelated with ui . We draw ui from a normal distribution in the first series of experiments and from a gamma distribution in the second series of experiments. This allows to test the sensitivity of the various procedures to unit effects that are not normally distributed. In real data, unit fixed effects are unlikely to be normally distributed. The idiosyncratic error εit is white noise and repeatedly drawn from a normal distribution, and the R-squared is fixed at 50 percent for all experiments.9 Experiments are conducted with various combinations of N and T to reflect the datasets that are typically analyzed in comparative politics. We use all permutations of N=20, 30, 50, 80 and T=20, 40, 60, 100. For each combination of T and N we conduct 1000 experiments. Finally, we hold the coefficients of the true model constant throughout all experiments at the following values: α = 1, β1 = 0.5, β2 = 2, β3 = −1.5, β4 = −0.4, β5 = 1.8, β6 = 4.5 . Analysis of simulated data We are interested in the bias and the efficiency of the estimators. We define bias as the average deviation of the coefficients from the true coefficients, () K ( b βˆ = ∑ βˆ − β true k =1 ) K, (12) where k = {1, 2,..., K} is the number of simulations. For the purpose of this paper we define efficiency as the standard deviation of the betas ( ) ( βˆ − β ) σ βˆ = true 2 . (13) () To begin, we consider the bias b βˆ of the first series of experiments where ui is drawn from a normal distribution. table 1 about here 8 9 We have experimented with various correlations, but since our results are robust with respect to various correlation coefficients 0.5 ≤ ρ ( ui , x3i ; ui , z3i ) ≤ 0.95 , we do not report the results from varying correlation coefficients here. All reported results are therefore based on a correlation of app. 0.5. For each permutation we generated 1000 replicates of εit and yit . Code was written in STATA and is available upon request. 12 All procedures have difficulties in calculating the effect of time-invariant variables that are correlated with the unit effects. The average deviations from the true betas are smallest for the fixed effects vector decomposition technique and largest for Hausman-Taylor. In fact, xtfevd is approximately 50 percent less biased than the second best estimator, pooled OLS and random effects. Hausman-Taylor is not only the most biased, but it also performs poorly with respect to the time-invariant variables. It is the only specification that is likely to report biased coefficients of time-invariant variables uncorrelated with the unit effects. Since this result may come as a surprise it is important to note that we correctly specified the endogenous variables. Bias surges if researchers make a incorrect decision about exogeneity. OLS and random effects models fail to give an unbiased coefficients of the timevariant variable that is correlated with the unit fixed effects (x3). This outcome is expected because – having assumed unit effects correlated with x3 – we clearly observe omitted variable bias here. Neither pooled OLS nor random effects models capture this effect unless the number of observations approaches infinity and random effects converge to the fixed effects estimator. Figures 1.1-1.8 present the Kernel density distribution of the coefficients for x3 and z3 when N=30 and T=20. We have chosen this combination of N and T for illustration purposes, since this represents a fairly common sample size in comparative politics. The figures highlight the difficulties in obtaining the effect of time-variant and time-invariant variables correlated with the unit effects. Figure 1.1 -1.8 about here Pooled OLS and random effects produce biased results for both x3 and z3. Both, the fixed effects vector decomposition technique and Hausman-Taylor correctly estimate the coefficient of the time-variant variable that co-varies with the unit effects. Both models vary with respect to the estimation of z3, the time-invariant variable that is correlated with the unit effects. The poor performance of Hausman-Taylor in obtaining the coefficient of z3 stems from its small sample inefficiency. Hausman-Taylor apparently requires very large samples to become 13 reliable in respect to time-invariant variables correlated with the unit effects. Thus, while Hausman-Taylor and xtfevd are less biased than pooled OLS and random effects, xtfevd has by far more favorable small sample properties than Hausman-Taylor. On average, the fixed effect vector decomposition technique behaves best. It is almost as good as pooled OLS and random effects models in estimating the coefficients of time-invariant variables and it provides less biased results than both alternatives in calculating the effect of time-variant variables that are uncorrelated with the unit fixed effects. This finding is additionally supported by table 2 that reports the variance of the estimates. table 2 about here Table 2 displays the standard deviation of the estimated betas from the true betas, σ βˆ . Pooled OLS and random effects give a higher variance of betas than () Hausman-Taylor and xtfevd when calculating coefficients of time-invariant variables. Since both methods also generate more biased coefficients for timeinvariant variables correlated with the unit effects, pooled OLS and random effects should be avoided in this case. More favorable results occur for both estimators when we consider time-invariant variables. This is clearly not the domain of the Hausman-Taylor model, which has by far the largest variance of all four procedures. In sum, xtfevd is least biased and most efficient – at least when we analyze data that mirrors the data generating process underlying our Monte Carlo experiments. Sample size and bias The preceding subsection has shown that on average and when analyzing wellbehaved data xtfevd performs better than its competitors. In this subsection we examine whether the estimators’ bias and efficiency vary with the sample size. We study the performance of the procedures for all permutations of N = {20,30,50,80} and T = {20, 40, 60,100} . We focus on the results of the ‘problematic’ variables (x3 and z3) and do not further consider the variables uncorrelated with the unit effects. In addition, we contrast the results of a model in which unit effects are drawn from a normal distribution to the results of a model 14 where unit effects are drawn from a gamma distribution. Table 3 reports the bias of estimators with respect to both the distribution of unit effects and the sample size. table 3 about here The Hausman-Taylor and xtfevd procedures are superior in estimating the effects of time-variant variables correlated with the unit effects. This holds true for all combinations of N and T. Both OLS and random effects produce biased coefficients when the time-variant variable is correlated with the unit fixed effects. The estimates of xtfevd and Hausman-Taylor remain unbiased even if the unit effects are gamma distributed. The impact of skewed unit effects on the bias of coefficients is large for random effects and pooled OLS. On average, b βˆ of () pooled OLS and random effects almost triple when unit effects are skewed. In our view, this constellation is likely to exist in real data. There is little reason to believe that unit effects are normally distributed and it is likely that the unit effects are correlated with the time-variant variables. Estimating the empirical model by either pooled OLS or random effects cannot be recommended. Pooled OLS and random effects work better when we consider time-invariant variables correlated with the unit fixed effects. Table 4 shows that no procedure is likely to produce unbiased coefficients for time-invariant variables correlated with the unit effects, but Hausman-Taylor performs worse than the other three alternatives. table 4 about here Bias decreases as the number of cross-sections increases but applied researchers should nevertheless keep in mind that the coefficients of time-invariant variables correlated with the unit effects are likely to be upward biased. In many cases this knowledge will prevent applied researchers from making careless inferences. The random effects specification works best for time-constant variables correlated with individual effects and this holds especially true when the unit effects are skewed. If the unit effects are drawn from a normal distribution the difference 15 between random effects models and xtfevd becomes small and lies within the sampling variance. Most noteworthy, the Hausman-Taylor procedure is worse than any alternative including pooled OLS. The bias is large regardless of the distribution of unit effects. This outcome is probably due to a combination of poor small sample properties and an insufficient quality of instruments. It appears that using Hausman-Taylor has some similarity to organizing a coefficient lottery. To sum up, all procedures give unbiased estimates of time-variant and timeinvariant variables that are uncorrelated with the unit-effects. In most real data analyzed in comparative politics, the unit effects are likely to be correlated with the explanatory variables. While xtfevd and Hausman-Taylor work much better than random effects and pooled OLS when time-variant variables are correlated with the unit effects, Hausman-Taylor is unreliable in cases where time-invariant variables are correlated with unit effects. This possibility should not be dismissed. Fortunately, xtfevd and random effects generate relatively unbiased coefficients for time-invariant variables. Under normal conditions, the fixed effects decomposition works in no respect worse than any of its alternatives. Thus, we recommend that applied researchers employ this procedure. Only if a) time-invariant variables are correlated with the unit effects and b) the unit effects are skewed, researchers should avoid using xtfevd.10 Under these circumstances the random effects model seems to be the least biased choice. One should keep in mind, however, that random effects obtains biased coefficients for time-variant variables correlated with unit-effects. Sample size and efficiency Small samples are prone to large variation in the estimated betas. This holds especially true for the effect of time-invariant variables, because the number of observation is T times smaller than the number of observations of time-variant variables. Consequently, efficiency is a crucial issue. The calculation of efficiency is based on the same experiments as the preceding discussion of bias. Table 5 presents the estimators’ efficiency with respect to the time-variant variable co-varying with the unit effects. 10 The output of xtfevd.ado provides information on the correlation between the unit effects and the endogenous variables and on the skewness of the unit effects. 16 table 5 about here Hausman-Taylor and xtfevd are more efficient in estimating the effect of timevariant variables correlated with the unit effects. The difference to random effects and pooled OLS increases with the number of cross-sections. The larger N, the more efficient Hausman-Taylor and xtfevd estimate time-variant variables. When the unit effects are skewed, the efficiency of all procedures declines. The relative advantage of Hausman-Taylor and xtfevd over random effects and pooled OLS remains intact. With the exception of Hausman-Taylor, efficiency is generally smaller for timeinvariant variables. The efficiency of Hausman-Taylor apparently depends on large samples. table 6 about here As table 6 demonstrates, no procedure can adequately cope with the coincidence of complications such as time-invariant variables correlated with skewed unit effects. Applied researchers should keep this in mind when deriving inferences from datasets where both unfortunate conditions are combined. Under any other circumstance, researchers should find results produced by xtfevd acceptable. In the absence of fully reliable solutions, conservatism in making inferences is warranted. 5. Re-analyses This section reports the results from replicating three contemporary analyses that use pooled OLS in the presence of time-invariate variables: Evelyn Huber and John D. Stephens study of the ‘Development and Crisis of the Welfare State’ and Pablo Beramendi and Thomas Cusack (2004) analysis of the determinants of income inequality. The aim of this section is to demonstrate that coefficients may, but do not need to, vary if researchers apply more appropriate estimators for the problem at hand. Re-estimation of Huber/ Stephens 2001 Huber and Stephens aim at explaining the long-term patterns of welfare state development in respect to the short- and long-term dynamics as well as the cross- 17 sectional variance.11 They take issue with the de facto Beck-Katz standard of estimating panel data by OLS and including the lagged dependent variable (to eliminate autocorrelation) plus unit fixed effects. They justify their deviation from the standard procedure by claiming that the “analysis of short term change can lead to quite misleading conclusions about long-term change.” (HS 2001: 4) Specifically, analyses of short-term changes “are unable to explain the increasing divergence of welfare states over time.” (HS 2001: 9) Huber and Stephens argue that the long-term development of the welfare state has been influenced by constitutionally created veto points. These veto points tend to make welfare state expansion (but also welfare state retrenchment) more difficult for the government. Political systems with more checks and balances should maintain lower levels of government spending in general and also less social security contributions in particular. It is well understood among political scientists that the constitutional factors do not vary much over time. Therefore, analyses of constitutional effects typically have to deal with the problem of time-invariant variables in one way or another. Huber and Stephens opt for aggregating constitutional features into a ‘constitutional structure’ variable that counts various veto points. Specifically, their measure is a simple additive index including categorical data for federalism (0,1,2), presidentialism (0,1), bicameralism (0,1,2) and the use of referenda (0,1). This constitutionalism variable is ordinally scaled and ranks from 0 to 6. Since it is almost time-invariant, they estimate the model by pooled OLS. This research strategy has at least one obvious drawback: It is impossible to identify the effect of each single constitutional variable. However, the impact of the various constitutional features on government spending is likely to be unequal across the various veto points. Implicitly, Huber and Stephens assume that the effect of federalism and bicameralism on government spending and social security transfers is exactly twice as large as the effects of presidentialism and referenda. For that reason, we have not only re-analyzed the Huber-Stephens specification (table 3.3. model 1) but also run a specification that uses the four original veto points variables rather than the composite index.12 11 12 This part has profited from communication with Neal Beck. We left Huber and Stephen’s estimation approach unchanged. This might be problematic since the Im, Peseran, Shin test (2002) does not allow to reject the hypothesis of a unit root. Keep in mind, however, that unit root test performs 18 Using STATA 8.0 and the data set that Stephens has posted on the Luxemburg Income Study (LIS) webpage,13 we have not been able to exactly replicate the results reported in the book by Huber and Stephens (p. 72-73). However, we arrive at exactly the same numbers that John Stephens kindly provided in assisting our replication attempts and – most importantly – these coefficients are Sala-iMartin-robust (signs of the coefficients do not differ) to the results reported by Huber and Stephens.14 We re-analyze their specification based upon the updated dataset and employing STATA instead of SHAZAM, which was used by Huber and Stephens.15 Huber and Stephens regress government spending in 19 OECD countries on a battery of political, social and economic variables. Their sample covers the years 1960 to 1994. However, for reasons beyond our knowledge, the model they report analyzes only the years from 1960 to 1986. Table 7 presents the STATA output that John Stephens kindly provided upon request and our re-analyzes using the four procedures we found relevant in the presence of time-invariant variables. table 7 about here The coefficients reported by Huber and Stephens are fairly but not completely robust. Our results deviate from those reported by Huber and Stephens in respect to the effect of the population share of individuals older than 65 years, of authoritarian legacy and of inflation. First, we find that the population share of individuals above 65 has no significant effect on social security transfers. Second, once we split Huber and Stephens’ composite constitutional constraints index into its original components, we find that authoritarian legacy significantly reduces social security transfers. Third, we also obtain a significant impact of inflation on social security transfers if we control for unit effects. Finally, the composite index 13 14 15 poorly, and that failure to reject the unit root hypothesis does not imply that data is non-stationary. We nevertheless add an appendix in which we estimate a model in differences. http://www.lisproject.org/publications/welfaredata/welfareaccess.htm. As a matter of fact, the STATA results of the updated dataset exhibit more significant relations between the exogenous and the endogenous variable and provide larger support for Huber and Stephens theoretical claims than the models reported in their book. We kindly thank John Stephens for his support with the data and with the replication analysis. 19 of Huber and Stephens conceals more than it reveals. The effect of constitutional constraints on social security benefits is not homogenous. While federalism and presidentialism reduces social security benefits as suggested by Huber and Stephens theoretical claims, neither bicameralism nor referenda (a Switzerland dummy) appear to be significantly related to social transfers. Probably to the satisfaction of the authors, we find no difference to any of the coefficients Huber and Stephens are theoretically interested in. Our analyses, therefore, provide additional support for the theories more broadly discussed by Huber and Stephens.16 However, we find significance for our assumption that the aggregation of constitutional rules into a summary variable is not useful. To be sure, Huber and Stephens correctly argue that veto points limit the observed rise of the welfare state. But different constitutional settings do not have identical effects. Bicameralism exerts the strongest influence on government spending, followed by presidentialism. Neither federalism nor referenda are significantly related to government spending if we rely upon fixed effect vector decomposition. Re-estimation of Beramendi/ Cusack 2004 Pablo Beramendi and Thomas Cusack’s (unpublished) paper studies the determinants of income inequality in OECD countries. Beramendi and Cusack distinguish between three types of income, wages, total income from economic activity and disposable income. Thus, their analyses allow isolating the redistributive efforts of governments. We re-estimate model 11 of their discussion paper version, which examines the determinants of disposable income, broadly defined as market income plus government transfers minus taxes. The model controls for pre-tax inequality and also contains union density and ‘left government inheritance’. Model 11 includes one time-invariant variable, a dummy that controls for coordinated market economies. The theoretical justification for the inclusion of this variable is provided by Michael Wallerstein and Miriam Golden (2000) as well as by Peter Hall and David Soskice (2001). In the view of these authors, coordinated market economies are characterized by centralized wage negotiations between unions and employer associations and possibly governments. Centralized bargaining tends to moderate the unions’ demands for higher nominal wages. In exchange for wage moderation the government provides social security systems to 16 See appendix A for a regressions in differences. 20 workers. As a consequence, coordinated market economies are associated with more redistribution and less after tax and transfers income inequality. Beramendi and Cusack justify running a pooled OLS model without eliminating serial correlation of errors and controlling for unit effects in the following way: “In the context of data sets in which the variance is dominated by the ‘between units’ (as opposed to the “within units”) component and where there are time invariant independent variables, the inclusion of both FE and a LDV frequently does more harm than good. Let us briefly explain why. Consider first the inclusion of fixed effects in any specification with a time invariant independent variable, such as the institutional terms in the wage and disposable income inequality equations or the wage inequality measure itself in the market income equation. In such cases, the specific value each country takes in the independent variable is going to be highly collinear with that country’s FE. As a result, the inclusion of FE yields substantively uninteresting and generally misleading results in which none of the variables of interest are statistically significant and yet the adjusted rsquared increases dramatically due to multicollinearity.” (Beramendi/ Cusack 2004: 29) We do not agree, but that’s another issue (see Plümper et al. 2004). Table 8 reports our re-analysis of the Beramendi-Cusack model. Note that we simply do not have a sufficient number of time-invariant variables to be able to estimate the Beramendi-Cusack model by the Hausman-Taylor formulation. Hausman-Taylor requires at least as many exogenous time-invariant variables as there are endogenous time-invariant variables.17 table 8 about here In passing, we may say that – quite to our surprise – the information on the disposable income inequality used by Beramendi and Cusack appears to be stationary. It should be emphasized that the results reported by Beramendi and Cusack are only mildly influenced by the switch from Panel OLS to both random effects and fixed effects vector decomposition. While the former outcome is not surprising given the similarity between Panel-OLS and random effects, the latter result may come as a surprise. However, the time-variant explanatory variables included in the Beramendi-Cusack model are uncorrelated with the unit effects. The union density score shows the highest correlation with u i , but the correlation 17 Needless to say that this is another disadvantage of the Hausman-Taylor procedure. 21 is only 0.37. All other variables are even less correlated with the unit fixed effects. In any case, we find a smaller coefficient for the time-invariant variable – the coordinated market economy dummy – but the interpretation of Beramendi and Cusack remains intact. The negligible influence of controlling for unit fixed effects may easily stem from the comparably low correlation between the unit effects and the exogenous variables. As we know from OLS, excluded variables cause little bias if they are orthogonal to the included exogenous variables. Their results are also robust to the inclusion of additional or the exclusion of present regressors. Discussion of Re-Analyses We draw two conclusions from our re-analyzes of two interesting and important studies in comparative politics and comparative political economy. Controlling for unit effects is essential in comparative politics. Political scientists are unlikely to formulate theories and design empirical models that completely explain the variance across nation-states. Even if we analyze a relatively homogenous group of countries, for instance OECD countries, we cannot expect to capture all crosssectional variance by the exogenous variables. Neither should social scienstists aim at formulating a complete model, we believe. And this directly leads to our second conclusion: Large empirical models (as the one suggested by Huber and Stephens) are likely to suffer from collinearity and endogeneity. From our perspective, it is more promising to estimate smaller empirical specifications that more closely represent the theoretical claims. In this case, however, researchers almost must control for unit-effects and should run estimators that do not require orthogonality between the unit effects and the substantive explanatory variables. This is the domain of the unit effects vector decomposition technique. 6. Conclusion The results of the Monte Carlo experiments suggest that xtfevd is the least biased estimator when time-variant and time-invariant variables are correlated with the unit effects. This constellation appears to be common to social science data. The unit fixed effects vector decomposition technique produces the least biased and most efficient coefficients under a wide variety of data generating processes. The main advantages of xtfevd come from its desirable small sample properties and from its unbiasedness in estimating the coefficients of time-variant variables that are correlated with the unit-effects. The efficiency of random effects models 22 and especially of the Hausman-Taylor procedure depend on large samples. The three-stage vector decomposition technique shares the small sample properties with pooled OLS without inheriting pooled OLS’s potential for bias. In particular, when unit effects are uncorrelated with the time-variant variables all four estimators are unbiased. When unit effects are uncorrelated with timeinvariant variables, pooled OLS, random effects and fixed effects vector decomposition give unbiased, Hausman-Taylor biased estimates. When unit effects are correlated with the time-variant variables, pooled OLS and random effects model perform poorly, fixed effect vector decomposition and HausmanTaylor are unbiased, but Hausman-Taylor is less efficient. Finally, when unit effects are correlated with time-invariant variables, all procedures are about equally biased, but Hausman-Taylor is by far least efficient. In cases where both time-variant and time-invariant variables are correlated with the unit effects, unit fixed effects decomposition technique clearly outperforms its competitors. However, one final clarification is in order: The vector decomposition technique does slightly worse than the random effects model when the following conditions are simultaneously given: a) the time-variant variables are uncorrelated with the unit-effects, b) the time-invariant variables are correlated with the unit effects, and c) the distribution of the unit effects is extremely skewed. In this case, the vector decomposition technique suffers from biased estimates of the second stage. Literature Acemoglu, Daron/ Johnson, Simon/ Robinson, James, Thaicharoen, Yunyong (2002): Institutional Causes, Macroeconomic Symptoms: Volatility, Crises and Growth, nber working paper 9124. Amemiya, Takeshi/ MaCurdy, Thomas E. (1986): Instrumental-Variable Estimation of an Error-Components Model, Econometrica 54: 869-881. Beck, Nathaniel/ Katz, Jonathan (1995): What to do (and not to do) with TimeSeries Cross-Section Data, American Political Science Review 89: 634-647. Beramendi, Pablo/ Cusack, Thomas (2004): Diverse Disparities: The Politics and Economics of Wage, Market and Disposable Income Inequalities, Wissenschaftszentrum Berlin dp. Breusch, Trevor S./ Mizon, Grayham E./ Schmidt, Peter (1989): Efficient Estimation using Panel Data, Econometrica 57, 695-700. Garrett, Geoffrey/ Mitchell, Deborah (2001): Globalization, Government Spending and Taxation in the OECD, European Journal of Political Research 39: 145-177. 23 Hall, Peter/ Soskice, David (2001): Introduction, in: Hall/ Soscise (eds.): Varieties of Capitalism, Oxford University Press, Oxford 2001, PAGES. Hausman, Jerry A. (1978): Specification Tests in Econometrics, Econometrica 46, 1251-1271. Hausman, Jerry A./ Taylor, William E. (1981): Panel Data and Unobservable Individual Effects, Econometrica 49: 6, 1377-1398. Huber, Evelyne/ Stephens, John D. (2001): Development and Crisis of the Welfare State. Parties and Policies in Global Markets, University of Chicago Press, Chicago. Im, Kyung So/ Peseran, M. Hashem/ Shin, Yongcheol (2002): Testing for Unit Roots in Heterogeneous Panels, unp. Manuscript, Cambridge University. Kittel, Bernhard/ Winner, Hannes (2003): How reliable is Pooled Analysis in Political Economy. The Globalization – Welfare State Nexus revisited, MaxPlanck Institut für Gesellschaftsforschung discussion paper 3/2003, Köln. Leamer, Edward (1983): Let’s take the Cons out of Econometrics, American Economic Review 73, 31-43. Leamer, Edward (1985): Sensitivity Analysis would help, American Economic Review 75, 308-313. Oaxaca, Ronald L./ Geisler, Iris (2003): Fixed Effects Models with TimeInvariant Variables. A Theoretical Note, Economics Letters 80, 373-377. Plümper, Thomas/ Troeger, Vera/ Manow, Philip (2004): Panel Data Analysis in Comparative Politics. Linking Method to Theory, European Journal of Political Research forthcoming. Sala-I-Martin, Xavier (1997): I Just Ran two Million Regressions, American Economic Review 87, 178-183. Wallerstein, Michael/ Golden, Miriam (2000): Postwar Wage Setting in Nordic Countries, in: Iversen, Torben/ Pontusson, Jonas/ Soskise, David (eds.): Unions, Employers and Central Banks, Cambridge University Press, Cambridge. Wilson, Sven E./ Butler, Daniel M. (2003): Too Good to be True? The Promise and Peril of Panel Data in Political Science, unp. Manuscript, Brigham Young University, 2003. Wooldridge, Jeffrey M. (2002): Econometric Analysis of Cross Section and Panel Data, MIT Press, Cambridge. 24 Appendix A: Huber-Stephens type of model in differences lagged change in government consumption level variables military spending per capita income federalism presidentialism referenda authoritarian legacy voter turnover openness fdi outflows inflation unemployment old female labor force participation cumrelcb left left * female difference variables military spending per capita income voter turnover openness fdi outflows strikes inflation unemployment old female labor force participation cumrelcb left left * female η intercept Im Pesaran Shin t-bar F R² within F (stage2) R² (stage 2) corr ( u i , Xβ ) N Estimated by xtfevd coeff 0.0188 std. err. 0.0334 -0.0090 0.0000 0.0240 0.3970 0.0978 0.0730 0.0061 -0.0061 0.0305 -0.0106 -0.0165 -0.0092 -0.0233 0.0151 0.0135 -0.0006 0.0269 0.0000 0.0378 0.1030 0.1359 0.0597 0.0046 0.0020 0.0403 0.0213 0.0095 0.0182 0.0069 0.0069 0.0055 0.0004 0.5837 -0.0011 0.1976 -0.0063 -0.0429 0.0118 0.2271 -0.0161 0.8571 0.2663 0.0433 0.1194 0.0017 1.0000 0.7065 -3.830 27.25 0.6917 56.84 0.4216 -0.574 0.0642 0.0001 0.0838 0.0041 0.0528 0.0101 0.1192 0.0229 1.0175 0.1131 0.1023 0.0475 0.0058 0.2208 0.4412 384 **** *** * *** ** ** **** **** ** * ** ** **** 25 pooled OLS ( ) b ( βˆ ) b ( βˆ ) b ( βˆ ) b ( βˆ ) b ( βˆ ) u normal distributed xtfevd xtreg, re xthtaylor b βˆ1 0.002 -0.005 0.013 -0.002 2 0.006 0.000 0.002 -0.002 3 0.295 -0.001 0.277 0.006 4 -0.014 -0.002 0.019 0.051 5 -0.016 -0.032 -0.015 0.166 6 0.335 0.415 0.356 0.550 0.668 0.455 0.682 0.777 sum b ( β k ) Table 1: Mean Deviation of estimated betas from true betas ( ) σ ( βˆ ) σ ( βˆ ) σ ( βˆ ) σ ( βˆ ) σ ( βˆ ) σ βˆ1 OLS u normal distributed xtreg, re xtfevd xthtaylor 0.276 0.279 0.280 0.279 2 0.083 0.082 0.083 0.082 3 0.324 0.132 0.320 0.135 4 0.217 0.206 0.230 0.427 5 0.255 0.328 0.289 0.524 6 0.388 0.474 0.400 1.300 Table 2: Efficiency of the estimated betas 26 OLS N=20 N=30 N=50 N=80 xtfevd N=20 N=30 N=50 N=80 xtre N=20 N=30 N=50 N=80 xthtaylor N=20 N=30 N=50 N=80 T=20 0.334 0.264 0.259 0.283 u normal distributed T=40 T=60 0.244 0.358 0.309 0.289 0.240 0.269 0.391 0.316 T=100 0.191 0.204 0.336 0.429 T=20 0.883 0.790 0.989 0.985 u gamma distributed T=40 T=60 0.722 0.653 1.052 0.879 0.967 0.874 1.012 0.958 T=100 1.065 0.860 0.943 0.855 -0.006 -0.005 0.003 -0.006 0.002 -0.006 -0.001 -0.001 -0.002 0.005 -0.005 -0.002 0.004 0.001 0.006 0.000 0.025 0.001 0.004 0.013 -0.022 -0.005 0.003 -0.004 0.000 -0.002 -0.002 -0.004 -0.009 -0.003 0.000 -0.005 0.170 0.297 0.475 0.330 0.253 0.135 0.308 0.312 0.180 0.144 0.337 0.205 0.176 0.399 0.372 0.339 0.703 0.897 0.732 1.009 0.902 0.949 0.932 0.894 0.477 0.896 0.842 0.921 0.493 0.918 0.798 0.903 0.012 0.014 0.006 0.005 -0.003 0.003 0.003 0.005 0.009 0.013 0.007 0.006 0.003 0.011 0.000 0.003 0.013 0.005 0.006 -0.003 0.056 -0.017 0.071 -0.001 0.017 0.008 0.005 0.008 -0.005 -0.002 0.000 -0.002 Table 3: Sample Size and Average Bias for x3 (time-variant, correlated with ui ) 27 OLS N=20 N=30 N=50 N=80 xtfevd N=20 N=30 N=50 N=80 xtre N=20 N=30 N=50 N=80 xthtaylor N=20 N=30 N=50 N=80 T=20 0.456 0.233 0.338 0.381 u normal distributed T=40 T=60 0.359 0.671 0.492 0.243 0.227 0.171 0.348 0.458 0.538 0.413 0.266 0.473 0.604 0.473 0.252 0.378 0.328 0.327 0.227 0.334 2.629 -0.036 0.320 -1.328 u gamma distributed T=40 T=60 0.835 0.799 0.646 0.520 0.634 0.522 0.508 0.323 T=100 0.346 0.264 0.395 0.273 T=20 0.580 0.651 0.536 0.371 0.501 0.442 0.387 0.363 0.508 0.457 0.254 0.335 0.711 0.973 0.798 1.217 0.319 0.704 1.228 1.384 1.041 1.265 1.351 1.305 1.199 1.218 1.043 0.981 0.555 0.397 0.352 0.246 0.510 0.502 0.462 0.425 0.278 0.187 0.264 0.309 1.078 0.604 0.973 0.545 0.603 0.515 0.601 0.667 0.786 0.884 0.781 0.405 0.996 0.341 0.638 0.478 -0.722 0.671 0.679 1.499 0.403 0.538 0.656 1.489 0.480 0.434 0.927 0.170 -4.626 2.467 1.625 0.666 1.416 2.424 1.760 1.654 5.175 0.045 -0.599 0.121 2.254 -0.061 0.560 0.792 Table 4: Sample Size and Average Bias for z3 (time-invariant, correlated with ui ) T=100 0.362 0.527 0.561 0.681 28 OLS N=20 N=30 N=50 N=80 xtfevd N=20 N=30 N=50 N=80 xtre N=20 N=30 N=50 N=80 xthtaylor N=20 N=30 N=50 N=80 T=20 0.403 0.322 0.299 0.304 u normal distributed T=40 T=60 0.290 0.379 0.333 0.305 0.257 0.279 0.397 0.321 0.242 0.213 0.154 0.131 0.171 0.141 0.111 0.083 0.297 0.361 0.497 0.349 0.263 0.206 0.157 0.120 u gamma distributed T=40 T=60 0.779 0.697 1.070 0.896 0.979 0.883 1.017 0.962 T=100 0.215 0.218 0.340 0.430 T=20 0.963 0.855 1.014 1.000 0.140 0.109 0.087 0.063 0.104 0.086 0.062 0.039 0.524 0.409 0.331 0.257 0.355 0.282 0.214 0.174 0.292 0.237 0.179 0.140 0.220 0.179 0.145 0.103 0.297 0.187 0.324 0.322 0.228 0.181 0.348 0.213 0.203 0.410 0.378 0.342 0.824 0.943 0.775 1.024 0.944 0.972 0.946 0.903 0.536 0.913 0.854 0.926 0.532 0.930 0.806 0.907 0.176 0.150 0.106 0.081 0.145 0.117 0.078 0.064 0.107 0.084 0.061 0.043 0.539 0.402 0.318 0.261 0.358 0.308 0.233 0.178 0.305 0.278 0.190 0.141 0.226 0.175 0.140 0.102 Table 5: Sample Size and Efficiency for x3 T=100 1.075 0.868 0.947 0.858 29 OLS N=20 N=30 N=50 N=80 xtfevd N=20 N=30 N=50 N=80 xtre N=20 N=30 N=50 N=80 xthtaylor N=20 N=30 N=50 N=80 T=20 0.340 0.319 0.376 0.402 u normal distributed T=40 T=60 0.385 0.687 0.510 0.276 0.256 0.207 0.365 0.467 0.614 0.651 0.317 0.491 0.630 0.502 0.278 0.393 0.410 0.383 0.273 0.366 3.013 0.798 1.016 2.434 u gamma distributed T=40 T=60 0.909 0.873 0.729 0.556 0.672 0.557 0.537 0.354 T=100 0.370 0.289 0.405 0.286 T=20 0.733 0.798 0.608 0.431 0.530 0.464 0.402 0.377 0.528 0.468 0.272 0.444 0.852 1.077 0.859 1.246 0.486 0.745 1.256 1.398 1.079 1.294 1.364 1.315 1.238 1.233 1.053 0.988 0.595 0.419 0.378 0.269 0.538 0.518 0.473 0.437 0.306 0.222 0.280 0.320 1.235 0.712 1.042 0.590 0.717 0.592 0.640 0.688 0.827 0.924 0.805 0.429 1.021 0.385 0.658 0.494 0.868 0.770 0.778 1.781 0.542 0.606 0.701 1.546 0.586 0.626 0.950 0.280 15.006 2.594 1.967 0.896 3.455 3.505 2.114 2.232 5.777 0.619 0.912 1.165 2.302 0.606 0.824 1.141 Table 6: Sample Size and Efficiency for z3 T=100 0.437 0.552 0.581 0.691 30 depvar: ssbencor procedure: Left Cumrelcb Constr97 Results provided by John Stephens xtpcse 0.1699 (0.0655) *** 0.3785 (0.0629) *** -1.1481 (0.2772) *** xtfevd 0.1456 (0.0634) ** 0.2743 (0.0656) **** -0.9119 (0.2775) ** Fed Pres Strbic Referen Flfp Left*fem Vturn Old Strks Authleg Rgdpc 0.1125 (0.0441) *** 0.0181 (0.0047) *** -0.0722 (0.0475) 0.6483 (0.1801) *** -0.0799 (0.0229) *** -0.6825 (0.5308) 0.0005 0.2382 (0.0590) **** 0.0214 (0.0046) **** -0.0543 (0.0518) 0.0631 (0.2625) -0.0433 (0.0210) ** -0.7689 (0.5665) 0.0004 xtfevd 0.1491 (0.0718) ** 0.2995 (0.0817) **** xtreg, re 0.1182 (0.0279) *** 0.4790 (0.0350) *** xthtaylor 0.0615 (0.0534) 0.3397 (0.0502) *** -1.2356 (0.6451) * -3.8533 (1.6905) ** 0.2268 (0.5469) -2.4191 (2.2729) 0.2646 (0.0835) *** 0.0228 (0.0056) **** -0.0829 (0.0632) 0.1878 (0.3056) -0.0406 (0.0230) * -1.8056 (0.9127) ** 0.0003 -1.9215 (0.2162) *** -1.1071 (0.5911) * -0.8709 (0.2508) *** -10.7307 (0.7432) *** 0.1647 (0.0332) *** 0.0277 (0.0025) *** -0.1260 (0.0262) *** 0.5983 (0.1128) *** -0.1790 (0.0102) *** -1.9427 (0.3437) *** 0.0009 -0.8315 (1.5134) -4.4428 (3.0768) -3.1422 (0.7120) *** -1.9580 (5.1658) 0.3095 (0.0536) *** 0.0253 (0.0033) *** 0.0150 (0.0540) 0.2885 (0.4792) -0.1201 (0.0170) *** -2.1188 (1.4868) 0.0009 31 Cpi Unemp Mil Fdiout Openness η Cons rho N R² Chi² norm ( u i ) Jarque& Bera norm ( ηi ) (0.0001) *** 0.2558 (0.2675) 0.4901 (0.0651) *** 0.1393 (0.1548) -0.1806 (0.1451) 0.0028 (0.0140) 1.8956 (5.2914) 0.8564 416 0.530 806.28 (0.0001) *** 0.7350 (0.2360) *** 0.4597 (0.0585) **** -0.0441 (0.1883) -0.1179 (0.1522) -0.0123 (0.0124) 0.9096 (0.1576) **** 0.5493 (5.8231) 0.8785 416 0.5017 (0.0001) ** 0.7308 (0.2535) *** 0.4416 (0.0609) **** -0.0729 (0.2133) -0.1283 (0.1532) -0.0099 (0.0130) 0.8542 (0.1918) **** 3.2458 (6.5881) 0.8941 416 0.4549 22.29 **** 22.29 **** 11.28 ** 11.28 ** Jarque & Bera F 27.12 **** 19.23 **** Table 7: Replication Results for Huber/ Stephens 2001, p. 72-73, model 1. (0.0001) *** -0.5539 (0.1224) *** 0.5617 (0.0533) *** -0.5496 (0.1554) *** 0.0582 (0.2128) -0.0040 (0.0087) (0.0001) *** -0.1141 (0.1730) 0.6015 (0.0557) *** -0.2448 (0.1780) 0.0671 (0.2173) -0.0476 (0.0132) *** 11.1813 (2.7884) *** 0.0000 416 0.920 4545.44 -2.6110 (7.4181) 0.8806 416 3219.32 32 Market Inc. Inequality Union Density Coordinated Mark. Econ Left Govern. Inheritance η Intercept rho N R² F norm ( u i ) Jarque& Bera norm ( ηi ) Pooled OLS (Beram. and Cusack) 0.3265 (0.0694) **** -0.0009 (0.0001) **** -0.0420 (0.0053) **** -0.0188 (0.0054) *** Random Effects 0.2428 (0.0323) **** 0.2333 (0.0342) **** 41 0.901 78.67 **** 41 0.900 0.3465 (0.0725) **** -0.0009 (0.0002) **** -0.0407 (0.0085) **** -0.0182 (0.0069) *** fixed effect vector decomposition 0.3108 (0.0547) **** -0.0016 (0.0002) **** -0.0285 (0.0054) **** -0.0205 (0.0050) **** 0.9967 (0.2361) **** 0.2751 (0.0266) **** 0.1183 41 0.917 89.50 **** 7.25 ** 1.59 Jarque & Bera Chi² 145.22 **** Table 8: Replication Results for Beramendi/ Cusack 2004 (table 11) 33 Kernel density graphs for x3 (N30T20) x3 z3 0 0 .5 .5 1 1 Density 1.5 Density 1.5 2 2 pooled OLS -2 -1.5 b3 -1 -.5 -2 -1.5 b3 -1 -.5 -2 -1.5 b3 -1 -.5 -2 -1.5 b3 -1 -.5 4 4.5 b6 5 5.5 4 4.5 b6 5 5.5 2 Density 1.5 1 .5 0 0 .5 1 Density 1.5 2 xtfevd -2.5 1 .5 0 0 .5 1 Density 1.5 Density 1.5 2 xtre 4 4.5 5 b6 5.5 6 6.5 0 0.1.2.3.4 .5 1 Density 1.5 Density 2 xthtaylor -2.5 2 4 b6 6 8 10