‘The Stata Journal (2007) 7, Number 2, pp. 197-208 Estimation of nonstationary heterogeneous panels Edward F. Blackburne it Mark W. Prank Sam Houston State University Sam Houston State University Huntsville, TX ‘Huntsville, TX. blackburne@shsu.edu_ markfrank@shsu.edu Abstract. We introduce a new Stata command, xtpag, for estimating nonstation- ary heterogeneous panels in which the number of groups and nurmber of time-series observations are both large. Based on recent advances in the nonstationary panel Iiterature, xtpag provides three alternative estimators: a traditional fixed-effects estimator, the mean-group estimator of Pesaran and Smith (Istimating long-run relationships from dynamic heterogeneous panels, Journal of Econometrics 68: 70-113), and the pooled mean-group estimator of Pesaran, Shin, and Smith (Es- timating long-run relationships in dynamic heterogeneous panels, DAB: Working Papers Amalgamated Series 9721; Pooled mean group estimation of dynamic het- erogencous panels, Journal of the American Statistical Association 94: 621-634). Keywords: st0125, xtpmg, nonstationary panels, heterogeneous dynamic panels, pooled mean-group estimator, mean-group estimator, panel cointegration 1 Introduction Tn recent years, the dynamic panel-data literature has begun to focus on panels in which the number of cross-sectional observations (NV) and the number of time-series observations (T') are both large. ‘The availability of data with greater frequency is certainly a key contributor to this shift. Some cross-national and cross-state datasets, for example, are now large enough in T such that each nation (or state) can be estimated separately. The asymptotics of large NV, large T dynamic panels are different from the asymp- toties of traditional large NV, small T dynamic panels. Small T’ panel estimation usu- ally relies on fixed- or random-effects estimators, or a combination of fixed-effects es- ‘timators and instrumental-variable estimators, such as the Arellano and Bond (1991) generalized method-ofmoments estimator. These methods require pooling individual soups and allowing only the intercepts to differ across the groups. One of the cen- tral findings from the large N, large T literature, however, is that the assumption of homogeneity of slope parameters is often inappropriate. ‘This point has been made by Pesaran and Smith (1995); Im, Pesaran, and Shin (2008); Pesaran, Shin, and Smith (1997, 1999); and Phillips and Moon (2000). 1. For more discussion ofthis Itrature, oo chapter 12 in Baltagi (2001). © 2007 StataCorp LP st0125198 Estimation of nonstationary heterogeneous panels With the increase in time observations inherent in large N, large T dynamic panels, also a concern. Revent papers by Pesaran, Shin, and Smith (1997, 1999) offer two important new techniques to estimate nonstationary dynamic panels in which the parameters are heterogeneous across groups: the mean-group (M@) and pooled ‘mean-group (PMG) estimators. ‘The MG estimator (see Pesaran and Smith 1995) relies on estimating N time-series regressions and averaging the coefficients, whereas the PMG estimator (see Pesaran, Shin, and Smith 1997, 1999) relies on a combination of pooling and averaging of coefficients. In recent empirical research, the MG and PMG estimators have been applied in a variety of settings. Freeman (2000), for example, uses the estimators to evaluate state level alcohol consumption over 1961-1995. Martinez-Zarzoso and Bengochea-Morancho (2004) use them in an estimation of an environmental Kuznets curve in a panel of 22 OECD nations over 1975-1998. Frank (2005) uses the MG and PMG estimators to evaluate the long-term effect of income inequality on economic growth in a panel of U.S. states over 1945-2001, 2 The MG and PMG estimators Assume an autoregressive distributive lag (ARDL) (p,q1,--.s48) dynamic panel specifi- cation of the form FiyXinay a Hee a) wie = > Aguie-s + > it Fa where the number of groups i = 1,2,...,N; the number of periods t = 1,2,...475 Xu is a kx 1 vector of explanatory variables; 5, are the k x 1 coefficient vectors; yy are scalars; and J isthe group-specific effect. T must be large enough such that the model can be fitted for each group separately. ‘Time trends and other fixed regressors may be included, If the variables in (1) are, for example, I(1) and cointegrated, then the error term is an 1(0) process for all i. 'A principal feature of cointegrated variables is their re- sponsiveness to any deviation from long-run equilibrium. This feature implies an error correction model in which the short-run dynamics of the variables in the system are influenced by the deviation from equilibrium. ‘Thus it is common to reparameterize (1) into the error correction equation 6X) + SapAwier + TE HGAKey tm tee — Q) Aver = dali Maegtt dim J = where $; = —(1 — DF xj), = Dfo bis/(1— Dy Aum), Ap 1,2,...5p—1, and J =—Djgi8im FQ gh ‘The parameter ¢; is the error-correcting speed of adjustment term. If 4, 0, then there would be no evidence for a long-run relationship. This parameter is expected toE. F. Blackburne and M. W. Frank 199 be significantly negative under the prior assumption that the variables show a return to.a long-run equilibrium. Of particular importance is the veetor 6, which contains the Jong-run relationships between the variables. ‘The recent literature on dynamic heterogeneous panel estimation in which both N and T are large suggests several approaches to the estimation of (2). On one extreme, a fixed-effects (FE) estimation approach could be used in which the time-series data for each group are pooled and only the intercepts are allowed to differ across groups. If the slope coefficients are in fact not identical, however, then the FE approach produces, inconsistent and potentially misleading results, On the other extreme, the model could. be fitted separately for each group, and a simple arithmetic average of the coefficients, could be calculated. This is the MG estimator proposed by Pesaran and Smith (1995). With this estimator, the intercepts, slope coefficients, and error variances are all allowed to differ across groups. More recently, Pesaran, Shin, and Smith (1997, 1999) have proposed a PMG estima tor that combines both pooling and averaging. ‘This intermediate estimator allows the intercept, short-run coefficients, and error variances to differ across the groups (as would the MG estimator) but constrains the long-run coefficients to be equal across groups (as, would the FE estimator). Since (2) is nonlinear in the parameters, Pesaran, Shin, and Smith (1999) develop a maximum likelihood method to estimate the parameters Expressing the likelihood as the produet of each cross-section’s likelihood and taking the log yields 1& ze FL arta 8(OF HAD A6(0} (3) na? nO, for i = Ay... 4.M, where (8) = yieo1 —Xi0iy Hi = Ir ~ Wi(W{W,) WW, Tris am identity matrix of order T, and Wy = (Avie-ts--- Bvie-pt1» 4Xi, AXie-,--y AXig-o4t) -xtpng uses Stata’s powerful ml framework to implement the PMG estimator. Specif- ically, we take advantage of the undocumented hold option of ml to maximize the likelihood via “back-substitution”.? Beginning with an initial estimate of the long-run coefficient vector, 8, the short-run coefficients and the group-specific speed of adjustment terms can be estimated by regressions of Ay, on (&,,W,). These conditional estimates are in turn used to update the estimate of 8. The process is iterated until convergence is achieved. ‘The parameter estimates from iterated conditional likelihood maximization are as- ymptotically identical to those from full-information maximum likelihood. But the ‘estimated covariance matrix is not. However, since the distribution of the PMG param- ters is known, we can recover the full covariance matrix for all estimated parameters. As shown in Pesaran, Shin, and Smith (1999), the covariance matrix can be estimated by the inverse of 2. Although (@) looks benign, the model is dificult to program directly as a Stata ml program, Looking forward to (8), readers will note that the model cannot be readily adapted to nl's familiar “theta” notation200 Estimation of nonstationary heterogeneous panels nBvxyév -3.xiWs mBiy XW ° Sy 0 eo Ba ue wt ‘The Mo parameters are simply the unweighted means of the individual coefficients, For example, the MG estimate of the error correction coefficient, ¢ is d=NI DH « with the variance 1 aay Le 6 ‘The mean and variance of other short-run coefficients are similarly estimated. 3 The xtpmg command 3.1. Syntax xtpmg varlist [if] [in] [, 1x(varlist) ec(string) replace constraints(string) noconstant cluster(varname) level(#) technique(algorithm spec) difficult full model] 3.2 Opt Lr(varlist) specifies the variables to be included when calculating the long-run cointe- grating vector. ns ‘2ec(string) is used to specify the name of the newly created error-correction term; default replace overwrites the error-correction variable, if it exists. constraints (string) specifies the constraints to be applied to the model. This option is currently used only with option png.E. F. Blackburne and M. W. Frank 201 noconstant suppresses the constant term. ‘This option cannot be used with option dfe, cluster (varname) specifies that the observations are independent across groups (clus- ters), but. not necessarily within groups. varname specifies to which group each observation belongs, e.g., cluster (personid) in data with repeated observations on individuals. cluster() affects the estimated standard errors and variance covariance matrix of the estimators (VCE), but not the estimated coefficients; see (v] 20.14 Obtaining robust variance estimates. LeveL(#) sets the confidence level; default is Level (95). technique algorithm.spec) specifies the ml maximization technique. algorithm.spec is algorithm [# [algorithm [#]] ...], where algorithm is {nr|btgs|afp}. ‘The buh algorithm is not compatible with xtpag. technique() can be used only with option pag. 4iff{cult will ue a different stepping algorithm in nonconcave regions ofthe likelihood. full specifies that: all V cross-section regression results be listed. Only the averaged coefficients are listed by default. ‘model is the type of estimator to be fitted and is one of the following: png is the default and specifies the PMG estimator. ‘This model constrains the long- Fun coefficient vector to be equal across panels while allowing for group-specific short-run and adjustment coefficients. ng specifies the MG estimator. ‘This model fits parameters as averages of the N individual group regressions. fe specifies the dynamic fixed-effects estimator. 4 Empirical example: OECD consumption 4.1 Data , We illustrate the use of xtpmg with annual aggregate consumption data for 24 Organisa- tion for Economic Co-operation and Development (OECD) nations. These data are taken from Pesaran, Shin, and Smith (1997, 1999) and encompass the years 1960-1993.° ‘The 1993 annual observation for Belgium is not included in the estimation sample, leaving ‘an estimation period of 1962-1992 for Belgium and 1962-1998 for the other 23 OECD countries. xtpmg requires that the data be tsset before estimation. use jasa2 ‘eset id yoar ‘panel variable: 44 (unbalanced) year, 1960 to 1998, 3. The original data and GAUSS code are available on Pesaran's web site: ttpi//wwwcon.cam.2e:uk/faculty/pesaran202 Estimation of nonstationary heterogeneous panels Assume the long-run consumption function Cat = Bou + Oremne + Oremee + os + ie (6) where the number of nations i = 1,2,...,N; the number of periods ¢ = 1,2,....7; cw is the log of real per capita consumption; yx is the log of real per capita income; and ‘mis the inflation rate. If the variables are I(1) and cointegrated, then the error term is 1(0) for all 7. The ARDL(1,1,1) dynamic panel specification of (6) is exe = Srowse + Suravse—1 + Goose + Saremi—1 + Ach e—aH + Ge a) ‘The error correction reparameterization of (7) is ese = bs (C421 ~ 804 — where $5 = —(1— As), Boi = 7H; Bie = SMHS, and By, = Sani fans, The error-correction speed of adjustment parameter, gi, and the long-run coeffi- cients, 61, and O24, are of primary interest. With the inclusion of #o;, a nonzero mean of the cointegrating relationship is allowed. One would expect ¢ to be negative if the bles exhibit a return to long-run equilibrium. Most aggregate consumption theories Indicate that the long-run income elasticity, ys, should be equal to one. ‘The inflation effect, 0x, is generally thought: to be negative je — Oaemin) + BrsiAie + Sarid + ee @ 4.2 PMG estimation ‘The first example estimates the PMG estimator for model (8). In this context, the PMG model allows for heterogeneous short-run dynamics and common long-run income and Inflation elasticities. Often only the long-run parameters are of interest. The default results of the pmg option include the long-run parameter estimates and the averaged short-run parameter estimates. The PMG standard ervors match the GAUSS output provided by Pesaran, Shin, and Smith (1909). ‘The standard errors in table | of Pesaran, Shin, and Smith (1090) are, however, differnt from those reported here and ffom the original GAUSS program.E, F. Blackburne and M. W. Frank 203, 4-pi d.y 4 yoar>=1962, 1r(..c pi) ec(ec) replace pag ‘og likelshood = 2270.2017 (not concave) + apg 4, eration °: Hteration 1: log likelihood = 2319.1636 Teeration 2: log likelihood = 2322.9801 Iteration 3: log likelihood = 226.7588 Kteration 4: log likelihood = 227.0742 Hteration 6: log ixelihood = 2827.0749 6 Teeration 6: log likelihood = 227.0748 Pooled Nean Group Regression (Estimate results saved as pag) Panel Variable (4): ia Munber of be = TOT ‘Tine Variable (): year Number of groups = FM Obs per group: nin at es 92.0 m= 32 Log Likelihood = 2327.075 Goof. Std. Err. = Pola! ‘(08% Conf. Interval] pi | -.4ssese .osevaa2 0.000 .s770388 -. 2546487 y | 19084836 “o0s8i5 0-000 ‘savaiat —“oztaaa1 a ec | ~-t996761.oszi6ss -6.21 0.000 ~.26202e7 ~. 1368275 0192588 027529 -0.66 0.511 0726522861347 nt | .sasoa5s osrazae 5.69 0.000 .2iaae7a.asoaaae cons | "1584507 | “oat6949 7-12 0.000 ‘1119808 “1969707 In the output, the estimated long-run inflation elasticity is significantly negative, as expected. Also, the estimated income elasticity is significantly positive. ‘Theoretically, the income elasticity is equal to one. This hypothesis is easily tested: = vest (oc}y=t CD beady =1 eMC 1) = 121.18 Prob > chi? = 0.0000 ‘The corresponding x? value of 121.2 leads to rejection of the null hypothesis of unit, income elasticity. ‘The ful option estimates and saves an N+ 1 multiple-equation model. The first equation (labeled per option ec) presents the normalized cointegrating vector.° The remaining NV equations list the group-specific short-run coefficients. ‘5 The vector fas been normalized such that the coefficient on the fist term in the eointegrating vector is 1. Accordingly, the normalize term is omitted from the estimation output.204 Estimation of nonstationary heterogeneous panels xtmng d.c 4.pi day ££ yoar>ei962, 1e(Z-c ply) ec(ec) full pag Heeration 0; log likelihood = 270.3017 (not concave) Teeration 1: log likelihood = 2319.1636 Iteration 2: log likelihood = 2372.9001 Iteration 3: log likelihood = 2526.7546 Hteration 4: log likelihood = 289710742 Teeration 5: log likelihood = 227.0748 Teeration 6: log likelihood = 2877.0748 Pooled Mean Group Rogression (Gstinate results saved as PHC) Panel Variable (2): 1d umber of obs = TT Tine Variable (0): year Munber of groups = 2 Obs per group: min = a mes 82.0 nar 32 Log Likelshood = 2527-075 Coot. Std. Ber. 2 Poizi (96% Conf. Interval] pi | -.asse428 .0se7a92 8.21 0,000 -.9770388 3546467 'y | 9064336 ‘00g6815 104.18 0,000 .g874isi __-s2t4aat sant ~.097e81S 0280694 1.67 0.115 fossosri 0092742 mates 0866013 0.015 3813648 0415314 y ptr | 8198067 .055876 0.000 4009018 6290217, aeons | 10336383 0147912 0023 “oogeasi “0526285, (output omitted) 34.198 ec | ~.a9ve608 .oaer77a 0.000 6718605 -.5259608 pt pi. | .oratosa -orarsas 0.817 0502575134464 y pt. | 0300857 10016 0.705 =.10546 2015515 eon | ‘2743629 0690309, 01000, 1607979 ‘3979098, Since each group has its own estimated equation, we can, for example, predict vari- ables intuitively. predict deli sf ddee4it, og(ia.110) (783 asseing values generated) Similarly, cross-equation restrictions are easily applied. + test [id_ttt]oceLid_112100~0 CD Gatitiec - Gasizlec = 0 (2 Ualiitlec = 0 chi 2) = 2.58 prob > chi? = 0.2614E. F. Blackburne and M. W. Frank 205 4.3. MG estimation ‘The MG estimates are the unweighted mean of the NV individual regression coefficients. xtpag with the ng option loops through all panels in the sample to estimate the param- eters of (8) xtpag d.c d.pt diy if year>=1962, Ir(1.c pi y) ecCec) replace ag Mean Group Eetinetion: Error Correction Fors (Getinate rosules saved at mg) Coot. Std. Err. = Pola! (98% Conf. Interval] pi | ~.3529005 1168025, 5. seteaet -.1239800 y | 9161304 loa72673 33.1 ‘eeweoia lorasrra ae ec | ~-206s473 0201599 -10.16 0.000 -.ses4se7 | -. 2472940 pt pi, | -.02s9682 0294774 -0.86 0.990 0831989 .0s24108 7337888 0489602 4.78 0.000 .sa7e4e2 996004 2082185 ‘1088365 1.91 0.056 ~.005297 14217390 The MG estimates are presented as a two-equation model; the normalized cointe- ‘grating vector and the short-run dynamic coefficients. In comparing the PMG and MG estimators, we note that the estimated long-run income and inflation elasticities are statistically significant. and properly signed in both models. However, the PMG esti mate of the inflation elasticity is larger in magnitude than the estimate from the MG model (—.47 and —.85, respectively). ‘The opposite is true for the estimated long-run income elasticity (.90 and .92, respectively). The speed of adjustment estimates from each model imply significantly different short-run dynamics (compare ¢ = —.20 from $= 31 from Me). . Recall that the PMG estimator constrains the long-run elasticities to be equal across all panels. This “pooling” across countries yields efficient and consistent estimates when the restrictions are true, Often, however, tle hypothesis of slope liomogeneity is rejected empirically, If the true model is heterogeneous, the PMG estimates are inconsistent; the MG estimates are consistent in either case. The test of difference in these models is performed with the familiar Hausman test.” . Actually since (8) ie nonlinear inthe parameters, xtpag estimates the reduced-form regressions for each group, dcr = deen + Baye + Bare + dye +7aAze, and then apples Stata’s mlcon command to recover the underlying parameter estimates. 7, Stata’s hausnan test offers a signanore option. This option fores the variance-covariance matt: from the efficient model (PMG here) to be used in calculating the vst statistic. This is what is prevented here. See Baum, Schaffer, and Stillman (2008) for more details.206 Estimation of nonstationary heterogeneous panels ‘heusnan ag pag, signanore — cootticienta — © ® oe) sgre(diag (V2) me pisterence SE. pi | 3520006 -.4658408 178042 326218 y | -s18isea ——“ooseaa6 ‘0137008 ater = consievent under Ho and Ha; obtained from xtpa B= inconsistent under Ha, efficient under Ho; obtained fron xtpag Hor ditference in coefficients not systematic, eb822) = (b-B)? [OV B-V_B) CD14 5 41.06 Protech = 0.5087 ‘The calculated Hausman statistic is 1.06 and is distributed x2(2). Here we conclude ‘that the PMG estimator, the efficient estimator under the null hypothesis, is preferred. 4.4 Dynamic FE ‘The dynamic PE estimator, like the PMG estimator, restricts the coefficients of the coin- tegrating vector to be equal across all panels. The FB model further restricts the speed of adjustment coefficient and the short-run coefficients to be equal. xtpag with the dfe option fits the model in (8) while allowing panel-specific intercepts.® An allowance for intragroup correlation in the calculation of standard errors is made with the cluster option. sepag dic dips dy Sf yoar>=1962, Lr(2-e pi y) ec(ee) replace dfe cluster(id) ‘Standard errors adjusted with cluster ia) option Dynanie Fixed Effects Regression: Estimated Error Correction Form (stinate reault ‘2 DFE) Coot. Sta. Err. = -Poizi (96% Conf. Interval] ps | -.296943 102606 2.60 0.009 =.4672509 ~.055435 'y | 9120574 .0468008 18.49 0.000 8208795 1.008785 = 1794145 o1s4564 —-4.13 0.000 2645015 -.0942578 020826 0925622 -0.86 0.388 0919094 .0367382 seri94¢ 070878 5.380.000 exe .sn01089 1257694 080540 1.56 0.118 -.0321025 2836204 ‘3 Por the PE modal, xtpag is simply a wrapper for Stata's xtrog, fo command (see [XT] xtreg) ‘Tho underlying parameters of (8) are calculated with nicom and stored as BC. The reduoed-form model, as estimated by xtreg, £0, is stored as DFE.E. P. Blackburne and M. W. Frank 207 Alll coefficients from the dynamic FE model are properly signed and, in fact, similar to the PMG and MG estimates. As discussed in Baltagi, Griffin, and Xiong (2000), PE models are subject to a simultaneous equation bias from the endogeneity between the error term and the lagged dependent variable. ‘The Hausman test can be easily performed to measure the extent of this endogeneity. Deuenen mg DFE, sgnanore Costticients — ©) @) oe) sqrelding (v2 me DFE ister SE. pi | 3520005 260543. 0885608 25.20672 y | -st8isea 9120574 ‘0080771 6.028402 = consistent under Ho and Hla; obtained from xtpRe B= inconsistent under lis, efficient under Ho; obtained from x°pOE st: Ho: difference in cosfficionts not systenatic ebi22) = (BB)? [OV B-V_B)-C-AD}(b-B) 0.00 1.0600 Prov>ens2. Results indicate that the simultaneous equation bias is minimal for these data and, for this example, we conclude that the FE model is preferred over the MG model. 5 Conclusion ‘This paper follows recent advances offered by Pesaran and Smith (1995) and Pesaran, Shin, and Smith (1997; 1999) in the estimation of nonstationary heterogeneous panels with large V and large T. We offer a new Stata command, xtpmg, that estimates three alternative models: a traditional dynamic FE estimator that relies on pooling of cross-sections, an MG estimator that relies on averaging of cross-sections, and a PMG estimator that relies on a combination of pooling and averaging of coefficients. 6 References Arellano, M., and $. Bond. 1991. Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. Review of Economic Studies 58: 277-297. Baltagi, B. H. 2001. Econometric Analysis of Panel Data, 2nd ed. New York: Wiley. Baltagi, B. H., J. M. Griffin, and W. Xiong. 2000. To pool or not to pook: Homogeneous versus heterogeneous estimators applied to cigarette demand. Review of Economies and Statistics 82: 117-126. Baum, C. F., M. B. Schaffer, and S. Stillman, 2003. Instrumental variables and GMM: Estimation and testing. Stata Journal 3: 1-31208 Estimation of nonstationary heterogeneous panels Frank, M. W. 2005. Income inequality and economic growth in the U.S. A panel cointegration approach, Sam Houston State University Working Paper 05-03. Freeman, D. G. 2000. Alternative panel estimates of alcohol demand, taxation, and the business cycle. Southern Economic Journal 67: 325-344, Im, K. 8., M. H, Pesaran, and Y. Shin, 2003. Testing for unit roots in heterogeneous panels. Journal of Econometrics 115: 53-T4. Martinez-Zarzoso, I., and A. Bengochea-Morancho. 2004. Pooled mean group estimation of an environmental kuznets curve for CO. Economics Letters 82: 121-126. Pesaran, M. H., Y. Shin, and R. P. Smith. 1997. Estimating long-run relationships in ‘dynamic heterogeneous panels. DAE Working Papers Amalgamated Series 9721. . 1999. Pooled mean group estimation of dynamic heterogeneous panels. Journal of the American Statistical Association 94: 621-634. Pesaran, M. H., and R. P. Smith. 1995. Estimating long-run relationships from dynamic heterogeneous panels. Journal of Econometrics 68: 79-113. Phillips, P. C. B., and H. R. Moon. 2000. Nonstationary panel data analysis: An ‘overview of some recent developments. Econometric Reviews 19: 263-286. ‘About the authors Edward P. Blackburne IIIs an associate professor in the Department of Economies and Interna- tional Business, Sam Houston State University. His current research interests are benchmarking racial profiling and modeling asymmetric time series. ‘Mark W. Frank is an associate professor in the Department of Economics and International Business, Sam Houston State University. His current research interests are measuring effects of income inequality on economic growth and the economics of alcohol advertising.