PANEL DATA Data sets that combine time series and cross sections are called longtitudinal or panel data sets. Panel data sets are more orientated towards cross section analyses – they are wide but typically short (in terms of observations over time). Heterogeneity across units is central to the issue of analysing panel data. The basic framework is a regression of the form: Yit = Xitβ + Ziπ + εit (1) X has k columns and does not include a constant term. The heterogeneity or individual effect is Ziπ where Z contains a constant term and a set of individual or group specific variables. Such as gender, location, etc. We will consider two cases: Fixed Effects Zi is unobserved, but correlated with Xit then OLS estimators of β are biased. However, in this case where i = Ziπ embodies all the observable effects and specifies an estimable equation. This takes I to be a group specific constant term. Random Effects if the unobserved heterogeneity however formulated can be assumed to be uncorrelated with Xit then : Yit = Xitβ + E[Ziπ] + { Ziπ - E[Ziπ] } + εit = Xitβ + + ui + εit (2) (3) This random effects approach specifies that ui is a group specific random element which although random is constant for that group throughout the time period. FIXED EFFECTS This assumes that differences across units of observation can be captured by differences in the constant term. Each i is estimated: Yi = Xiβ + ii + εi (4) Where Yi is a Tx1 column of observations on m individual (group) i over T time periods. Hence the total sample size is mT. I is the unit vector {1,1,….,1} an identity matrix. Collecting together groups over time we get: Y = Xβ + D + ε (5) The model is usually referred to as the least squares dummy variable model (LSDV). This is a classical regression model and no new methodology or tests are needed to analyse it. In effect we simply regress Y on X plus a dummy variable for each group. Of course if the number of groups is very large then this presents computational problems. To tackle this we write the regression as: The Within and between Groups estimators We can formulate a pooled regression in three ways: (i) The original formulation: Yit = Xitβ + + εit (6) (ii) Deviations from Group Means Yit - MYi= (Xit - MXi)β + εit - Mεi (7) Where MYi denotes the mean of the observation for the i’th group across the T observations, etc. NOTE: The R2 related to (7) is known as the within groups R2. (iii) In terms of the group means: MYi= MXiβ + + Mεi (8) NOTE: The R2 related to (8) is known as the between groups R2. All three are classical regression models and in principal could be estimated by OLS. Because we have a large number of dummy variables to put in we can proceed by estimating (7), obtaining estimates of . Now rearrange (8) and assume Mεi= 0: = MYi - MXiβ (9) That is we can get estimates of by including subtracting from the mean of Y for each group the means of X multiplied by : Yit - MYi= (Xit - MXi)β + εit - Mεi (10) Unbalanced Panels and Fixed Effects. If we have missing data we have what we call an unbalanced panel. The required modifications are relatively simple. With a balanced panel the sample size is n = mT. With an unbalanced panel it is Ti. Hence instead of calculating group means on the basis of a sample size of n we have to have a specific sample size Ti for each group. RANDOM EFFECTS The fixed effects model allows the unobserved individual effects to be correlated with the included variables. The differences between units are then modelled as shifts in the constant term. If the individual effects are uncorrelated with the regressors then this is appropriate. The gain to this approach is that it substantially reduces the number of parameters to be estimated. The cost is the possibility of inconsistent estimates should the assumption be inappropriate. For random effects we reformulate the basic model: Yit = Xitβ + ( + ui) + εit (11) There is now a single constant term which is the mean of the unobserved heterogeneity, E(Ziπ). ui is the random heterogeneity specific to the i’th observation and is constant throughout time. For example in an analysis of firms it is the factors which we cannot measure which are still specific to that firm. We define: ηit = ui + εit and ηi = [ηi1, ηi2,……….., ηiT,]' This gives us what is called an “error components model”. For this: E[ηit2 |X] = σε2 + σu2 E[ηit ηis |X] = σu2 t ≠s E[ηit ηjs |X] = 0 for all t and s, i ≠ j Then: Σ = E[ηi ηi' |X] = σε2 + σu2 σu2 ………………… σu2 σu 2 σε2 + σu2……………. σu2 ……………………………………. σu 2 σu2……………. σε2 + σu2 Since observations I and j are independent the disturbance covariance matrix for the full nT observations is: Ω= Σ 0 0 …………….. 0 0 Σ 0………………0 ……………………… = I x Σ Kronecker multiplication 0 0 ………………. Σ We then estimate the model using GLS and the standard formula: β = (X ' Ω-1X)-1X' Ω-1Y (12) It can be shown that the GLS estimator is, like the OLS estimator, a matrix weighted average of the within and between units estimators. The inefficiency of OLS (i.e. fixed effects) follows from an inefficient weighting of the two sets of estimates. Fixed effects places too much emphasis on the between units variation. As usual we have the problem that we do not know the variances and covariances which comprise Σ. We begin by estimating (7): Yit - MYi= (Xit - MXi)β + εit - Mεi (7) Using this we can get an estimate of 2ε. It remains to estimate σu2. Return to the original model in (2): Yit = Xitβ + + ui + εit (3) This is a classical least squares model in which OLS is consistent and unbiased although efficient. Therefore the probability limit of the residuals from this regression equal: 2ε + 2u Estimate this, subtract 2ε and wehave our estimate of 2u. Testing for Random effects Breusch and Pagan (1980) have devised a Lagrange multiplier test for the random effects model based on the OLS residuals. For: H0: σu2 = 0 H1: σu2 ≠ 0 We have LM=nT/2(T-1)[(Σ(Tei.)2/ Σ Σeit2)-1] Where ei.= Mεi (from 8) This is distributed as Χ2; if it exceeds the critical value we conclude OLS is inappropriate and random effects is preferable. Note in Greene this is: LM = nT/[2(T-1)] {[T2ei’ei/eit’eit]-1}2 (slightly changed his terminology) where ei is the vector of unit (e.g. firm or country) specific mean errors and eit is the vector of total residuals from the least squares regression. This appears to be the same as above with one exception, the addition of a squared term: LM=nT/2(T-1)[(Σ(Tei.)2/ Σ Σeit2)-1]2 This has been checked and is correct. Hausman’s test for the Random Effects Model The specification test devised by Hausman is used to test for whether the random effects are independent of the right hand side variables. This is a general test to compare any two estimators. The test is based on the assumption that under the hypothesis of no correlation between the right hand side variables and the random effects both fixed effects and random effects are consistent estimators of (10) but fixed effects is inefficient (This is the assumption with random effects). Whereas under the alternative assumption (i.e. that with fixed effects) fixed effects is consistent but random effects is not. The test is based on the following Wald statistic: W = [ FE - RE] -1[ FE - RE] where Var[ FE - RE] = Var[ FE] - Var[RE] = W is distributed as 2 with (K-1) degrees of freedom where K is the number of parameters in the model. If W is greater than the critical value obtained from the table then we reject the null hypothesis of that both estimators are consistent i.e. of “no correlation between the right hand side variables and the ‘random effects’” in which case the fixed effects model is better. The intuition behind the test is relatively simply if both estimates are consistent then FE - RE should not be too great, i.e. the two should be close together. [ FE - RE] [ FE - RE] would be equivalent to summing the squares of the differences between the two sets of estimators. Hence the greater this is the more unlikely the null hypothesis is to be valid. The insertion of -1 effectively weights these differences in inverse proportion to the variance Var[ FE - RE]. If this is great then the measure tends to downplay the difference between FE and RE. On the other hand if this variance is small than any difference between FE and RE is given substantial weight. In terms of which is more appropriate I tend to favour on intuitive terms the random effects model in most cases. Take the firm example below, most of the difference in for example capital stock is between firms rather than within firms over time. To use fixed effects in this case would be to lose a great deal of power from the capital stock variable. The same is true with distance from London. This worry becomes much less if we change our mode of our analysis, our unit from e.g. the firm to the region the firm is operating within or the industry. But this is then very easy to do within conventional regression analysis by the use of dummy variables. The key question is whether the unobservable heterogeneity is correlated with the right hand side variables, if yes then fixed effects has a case as anything else will induce bias – i.e. some of the impact due to the unobservable heterogeneity will be wrongly ascribed to the firm specific variables such as capital stock. But in solving this problem we lose a lot of explanatory power from the regressions as argued above with capita stock. This raises the question as to why we expect the unobservable characteristics to be linked with the other variables. If e.g. it is due to quality of entrepreneur, does this then have a linkage with capital stock? EXAMPLES Example 1: The dependent variable is [the log of] Gross value added for firms. The data base is on firms operating in the UK. Note: the number of firms in this data set is 49,027. The panel is unbalanced. Wald statistic has 35 degrees of freedom as K, the number of parameters being estimated is, 36. The critical value from the 2 table with 35 degrees of freedom is 57.34. We can see that the Wald value is massively greater than this and hence we would reject the random effects model and accept the fixed effects model. There are R2 figures relating to ‘within-groups’ and ‘between groups’. Basically the within groups R2 is the explanatory power due to the right hand side variables explaining changes in GVA for individual firms, this is relatively low compared to the R2 between groups. This is not surprising as the bulk of the differences in GVA come from differences in capital stock and size of labour force between firms. The output is from a STATA program. Random-effects GLS regression Group variable (i): dlink_ref22 R-sq: within = 0.0254 between = 0.8312 overall = 0.8395 Random effects u_i ~ Gaussian corr(u_i, X) = 0 (assumed) Number of obs = 69349 Number of groups = 49027 Obs per group: min = avg = 1.4 max = 5 Wald chi2(35) Prob > chi2 = 1 = 244114.96 0.0000 -----------------------------------------------------------------------------lgvafc | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------lemp | .5558439 .0040942 135.76 0.000 .5478195 .5638683 Log of Employed Labour force lcap | .4022117 .002785 144.42 0.000 .3967532 .4076702 Log of capital stock lpropnoqua~l | -.0506918 .0092815 -5.46 0.000 -.0688833 -.0325004 Log of proportion with no qualifications lprophighq~l | .0625173 .0086047 7.27 0.000 .0456524 .0793822 Log of proportion with high qualifications private | .3906842 .0198204 19.71 0.000 .3518369 .4295314 Private sector firm uk | .0878014 .0082682 10.62 0.000 .0715961 .1040068 UK multinational usa | .2033561 .0206409 9.85 0.000 .1629006 .2438116 US multinational llunit | -.027734 .0059808 -4.64 0.000 -.0394562 -.0160119 Log of number of plants mfd1 | -.1846409 .0191904 -9.62 0.000 -.2222535 -.1470284 Multi-region multi-plant firmdummy nw1 | -.0646788 .0270416 -2.39 0.017 -.1176794 -.0116782 REGIONALDUMMIES yorks1 | -.1105119 .0276669 -3.99 0.000 -.164738 -.0562858 ne1 | -.0278356 .0386535 -0.72 0.471 -.103595 .0479238 wmids1 | -.1370831 .0217464 -6.30 0.000 -.1797053 -.0944609 wales1 | -.0999943 .0455664 -2.19 0.028 -.1893029 -.0106858 beds1 | -.1495485 .0214604 -6.97 0.000 -.19161 -.107487 sw1 | -.111522 .0249593 -4.47 0.000 -.1604412 -.0626028 emids1 | -.1523312 .0224785 -6.78 0.000 -.1963883 -.1082742 east1 | -.1660794 .0222715 -7.46 0.000 -.2097307 -.1224282 se1 | -.1366562 .0163141 -8.38 0.000 -.1686313 -.1046812 Time from London timenmfd1 | -.0009853 .0001045 -9.43 0.000 -.0011902 -.0007805 INDUSTRY DUMMIES miningdum | -.3116256 .1813403 -1.72 0.086 -.6670462 .0437949 manufactur~m | -.354197 .1377291 -2.57 0.010 -.6241411 -.0842528 powerdum | .0229343 .1653576 0.14 0.890 -.3011605 .3470292 constructi~m | -.0548109 .138434 -0.40 0.692 -.3261366 .2165148 wholesaler~m | -.1450245 .1376412 -1.05 0.292 -.4147963 .1247473 cateringdum | -1.151165 .1382849 -8.32 0.000 -1.422199 -.8801321 transportdum | -.5101581 .1383287 -3.69 0.000 -.7812774 -.2390388 realestate~m | -.2839032 .1375022 -2.06 0.039 -.5534027 -.0144038 educationdum | -.5616366 .139509 -4.03 0.000 -.8350691 -.288204 socialwork~m | -.6742467 .1385968 -4.86 0.000 -.9458914 -.402602 communitydum | -.7754511 .1382927 -5.61 0.000 -1.0465 -.5044024 year2 | .0236857 .0097134 2.44 0.015 .0046478 .0427236 YEAR DUMMIES year3 | .0075121 .0095382 0.79 0.431 -.0111825 .0262066 year4 | .0635178 .0086933 7.31 0.000 .0464793 .0805563 year5 | .1018157 .0088747 11.47 0.000 .0844216 .1192097 _cons | Included but not reported -------------+---------------------------------------------------------------sigma_u | .73642326 sigma_e | .47167868 rho | .7090994 (fraction of variance due to u_i) ------------------------------------------------------------------------------ Example 2: This is from LIMDEP. We are regressing growth on lagged growth aid per capita lagged one period (APCYL1), positive aid volatility APCYRS1P and negative aid volatility (APCYRS1N) and a time trend. The groups are countries, of which there about 66. We have data over time back to about 1961, but it is an unbalanced panel as we do not have full observations for all countries. We begin with the OLS equation without group dummy variables. There then follows least squares with group variables (Fixed effects). Finally we have the random effects model. Prior to this being printed out we have test results for the Lagrange multiplier test and also the Hausmann test. The test statistics are: | Lagrange Multiplier Test vs. Model (3) = 29.62 | | ( 1 df, prob value = .000000) | | (High values of LM favor FEM/REM over CR model.) | | Fixed vs. Random Effects (Hausman) = 66.54 | | ( 5 df, prob value = .000000) | The value of29.62 suggests that FE/RE are appropriate. The Hausman test suggests that of the two the fixed effects is appropriate. Turning to the fixed effects printout we can see that lagged growth impacts on current growth as does lagged aid. With respect to aid volatility, positive volatility (roughly, unexpected upward shifts in aid) have a positive impact on growth, but negative volatility has no discernable adverse effects.Worryingly there is a negative time trend. +-----------------------------------------------------------------------+ | OLS Without Group Dummy Variables | | Ordinary least squares regression Weighting variable = none | | Dep. var. = GROWTH Mean= 3.676811935 , S.D.= 5.599609225 | | Model size: Observations = 2017, Parameters = 6, Deg.Fr.= 2011 | | Residuals: Sum of squares= 59973.73230 , Std.Dev.= 5.46103 | | Fit: R-squared= .051243, Adjusted R-squared = .04888 | | Model test: F[ 5, 2011] = 21.72, Prob value = .00000 | | Diagnostic: Log-L = -6283.1290, Restricted(b=0) Log-L = -6336.1784 | | LogAmemiyaPrCrt.= 3.398, Akaike Info. Crt.= 6.236 | | Panel Data Analysis of GROWTH [ONE way] | | Unconditional ANOVA (No regressors) | | Source Variation Deg. Free. Mean Square | | Between 4350.44 62. 70.1683 | | Residual 58862.5 1954. 30.1241 | | Total 63212.9 2016. 31.3556 | +-----------------------------------------------------------------------+ +---------+--------------+----------------+--------+---------+----------+ |Variable | Coefficient | Standard Error |b/St.Er.|P[|Z|>z] | Mean of X| +---------+--------------+----------------+--------+---------+----------+ GROWTHL1 .1471741628 .21153013E-01 6.958 .0000 3.6459885 APCYL1 -.1928354626E-02 .18453840E-01 -.104 .9168 7.8342771 APCYRS1P .1781995479 .52169562E-01 3.416 .0006 1.1362761 APCYRS1N .3243939601E-01 .57797666E-01 .561 .5746 -1.1312713 TREND -.6280410686E-01 .11491455E-01 -5.465 .0000 24.482895 Constant 4.527164152 .32255659 14.035 .0000 (Note: E+nn or E-nn means multiply by 10 to + or -nn power.) +-----------------------------------------------------------------------+ | Least Squares with Group Dummy Variables | | Ordinary least squares regression Weighting variable = none | | Dep. var. = GROWTH Mean= 3.676811935 , S.D.= 5.599609225 | | Model size: Observations = 2017, Parameters = 68, Deg.Fr.= 1949 | | Residuals: Sum of squares= 55490.25413 , Std.Dev.= 5.33584 | | Fit: R-squared= .122169, Adjusted R-squared = .09199 | | Model test: F[ 67, 1949] = 4.05, Prob value = .00000 | | Diagnostic: Log-L = -6204.7693, Restricted(b=0) Log-L = -6336.1784 | | LogAmemiyaPrCrt.= 3.382, Akaike Info. Crt.= 6.220 | | Estd. Autocorrelation of e(i,t) .026811 | +-----------------------------------------------------------------------+ +---------+--------------+----------------+--------+---------+----------+ |Variable | Coefficient | Standard Error |b/St.Er.|P[|Z|>z] | Mean of X| +---------+--------------+----------------+--------+---------+----------+ GROWTHL1 .8840072682E-01 .21389867E-01 4.133 .0000 3.6459885 APCYL1 .7405384230E-01 .27813228E-01 2.663 .0078 7.8342771 APCYRS1P .1319179295 .57626292E-01 2.289 .0221 1.1362761 APCYRS1N -.1050606242 .63889238E-01 -1.644 .1001 -1.1312713 TREND -.1006714311 .12309034E-01 -8.179 .0000 24.482895 (Note: E+nn or E-nn means multiply by 10 to + or -nn power.) +------------------------------------------------------------------------+ | Test Statistics for the Classical Model | | | | Model Log-Likelihood Sum of Squares R-squared | | (1) Constant term only -6336.17838 .6321293692D+05 .0000000 | | (2) Group effects only -6264.26752 .5886250009D+05 .0688219 | | (3) X - variables only -6283.12895 .5997373230D+05 .0512427 | | (4) X and group effects -6204.76924 .5549025413D+05 .1221693 | | | | Hypothesis Tests | | Likelihood Ratio Test F Tests | | Chi-squared d.f. Prob. F num. denom. Prob value | | (2) vs (1) 143.822 62 .00000 2.329 62 1954 .00000 | | (3) vs (1) 106.099 5 .00000 21.723 5 2011 .00000 | | (4) vs (1) 262.818 67 .00000 4.048 67 1949 .00000 | | (4) vs (2) 118.997 5 .00000 23.689 5 1949 .00000 | | (4) vs (3) 156.719 62 .00000 2.540 62 1949 .00000 | +------------------------------------------------------------------------+ +--------------------------------------------------+ | Random Effects Model: v(i,t) = e(i,t) + u(i) | | Estimates: Var[e] = .284711D+02 | | Var[u] = .135170D+01 | | Corr[v(i,t),v(i,s)] = .045324 | | Lagrange Multiplier Test vs. Model (3) = 29.62 | | ( 1 df, prob value = .000000) | | (High values of LM favor FEM/REM over CR model.) | | Fixed vs. Random Effects (Hausman) = 66.54 | | ( 5 df, prob value = .000000) | | (High (low) values of H favor FEM (REM).) | | Reestimated using GLS coefficients: | | Estimates: Var[e] = .285879D+02 | | Var[u] = .212761D+01 | | Sum of Squares .602189D+05 | +--------------------------------------------------+ +---------+--------------+----------------+--------+---------+----------+ |Variable | Coefficient | Standard Error |b/St.Er.|P[|Z|>z] | Mean of X| +---------+--------------+----------------+--------+---------+----------+ GROWTHL1 .1135835519 .21085842E-01 5.387 .0000 3.6459885 APCYL1 .1991765749E-01 .21725640E-01 .917 .3593 7.8342771 APCYRS1P .1733946894 .53760692E-01 3.225 .0013 1.1362761 APCYRS1N -.2436091344E-01 .59905717E-01 -.407 .6843 -1.1312713 TREND -.7883073300E-01 .11664972E-01 -6.758 .0000 24.482895 Constant 4.895058779 .36527924 13.401 .0000 (Note: E+nn or E-nn means multiply by 10 to + or -nn power.)