Il | REGRESSION | | PANEL DATA 17 | moves imarily used eith, d in.the preceding J6 chapters primarily ae Ba i ain CY cic tao ofthese types of data has its unique features =that is, models that In this chapter we discuss. th Pm thet over (Some well-known examples of panel data sets are: 1. "The Panel Study of Income Dynamics (RSID): This is conducted by the “of Social Research at the University of Michigan, Started in 1968, each year the | institute collects data on some 500 families abut various socioeconom: demographic variables, 2_ Survey of Income and Progeam Participation (SIP): This survey is cond by the Bureau of the Census of the US Department of Commerce. Fou 4w\ year respondents are interviewed about their economic conditions. 3 ‘The German Socio-Economic Panel (GESOEP): It studied 1,761 individua every year between 1984 and 2002, It collected information on each in about year of birth, gender, life satisfaction, marital status, labor earni annual hours of work, # ‘There are several such surve 8 that are conducted by ee ‘es in many countries, ) fed by governments and pr of panel data the advantages of pane d lata over puri ai Swe Baliagilitsthe folowing facto °C"! ue dal, firms, states, countries and so on ove! data ty In these units, which may be ofte” techniques can take such heterogen'"Y AeEFexsION models, see Gujarati/Porter, oP “** New York, 1995, pp. 3-6,PANEL DATA REGRESSION MODELS | 327 aenit explicitly into account by allowing for subject-specific variables, ax we shall “ show shortly. We use the term subject generically to include microunits such as > inatividuuals, firms or states, © 2 By Combining time series of Cros-veetional observations, panel data gives » “more informative data, more vartability, less collinearity among variables, more : degrees of freedom and mote effictoney” 3 By studying the repeated cross-sections of observations, panel data are better suited to study the dynamics of change, Spells of unemployment, job turnover, ten duration of unemployment, ind labor mobility are better studied with panel of ye data. sider! ie data can better detect and measure effects that cannot be observed in pure Mew ress > -sectional of time seties data, Thus the effects of minimum wage laws on employment and earnings can be better studied if we follow successive waves of © nereases in federal and/or state minimum wages. heno a such as economies of scale and technological change can be better studi by panel data than by pure cross-sectional or pure time series data.) tombe example: charitable giving -on the companion website) gives data on charitable giving by 47 period 1979-19883 ned as follows: dependents claimed on the tax return d from the 1979-1988 Statistics of Income (SO1) Panel of d panel because the number of time observations (10) is If that were not the case, it would be an example of an ‘here are also called a short panel. In a short panel the individual units N (here 47) is greater than the number I fest panel, on the other hand, T'is greater than N. e 4 model of charity giving in relation to the variables ion. How do we proceed? We have five options: Rese ‘and Panel Data Analysis and Applications in ‘New York, 2004.928 | seuecTED TOPICS IN ECONOMETRICS charity functions: We can estimate by OLs 47 ‘charity functions, one for each individual i the in for 10 years series ty’ ciple we can estimate these functions, we wil have Very fey, ‘Although in princip! do meaningful statistical analysis. This is because we j, degrees SO in all, five for the five explanatory variables ang ri anequene ot Besides, these individual charity functions neglect the informs, sont ec individuals’ charity contributions because they all operate in : the same regulatory environment, eR charity functions: We can estimate by OLS 10 €T0SS-Sectiona) 2, charity functions, one for each year. There will be 47 observations er year tp ut again, we neglect the dynamic aspect of charitabje i inctions. Bi mug Soe ene contributions made by individuals over the years wi © Gepend on factors like income and marital status. ~~ Pooled OLS charity function: We can pool all 470 observations (47 10) and . ‘a “grand” charity function, neglecting the dual nature of time series ang naired od cs data. Not only would we be neglecting this if we were to run 2 Fooled model, but such a pooling assumes that the coefficients of the charity function remain constant across time and cross-section. The pooled OLS esti- ‘mation is also known as the constant coefficient model, for we are assuming that coefficients across time and cross-section remain the same. 4 Fixed effects least-squares dummy variable (LSDV) model: As in Option 3, we ‘pool all 470 observations, but allow each individual to have his or her individual Jintercept dummy. A variant of this is the within estimator, which we will explain 1. Individual time series of time ls are random drawings from a much larger population of individuals ‘of fact, the SOI Panel is a subset of the IRS Individual Tax Model ons 3, 4 and 5 sequentially. ; regression of charity function we aS He 7.) 47; t=1,2,...,10 1) cripts on the med that the term. It Notice that we have put two subs: Hf ‘Section unit, and ¢, the time. It is assw stochastic, are uncorrelated with the error ¢ Satisfies the usual classical assumptions- eit as thor ition, which we will not discuss here bec" periods. For a brief discussion of this ™*PANEL DATA REGRESSION MODELS | 329 es A priori, we would expect age, income, price, and marital status to have a posi- tive impact on charitable giving and the number of dependents to have a negative ___ impact, The reason the price variable, as defined, is included in the model is that it represents the opportunity cost of giving charitable contributions ~ the higher the marginal tax, the lower the opportunity cost. < ‘Using Eviews 6, we obtained the results of Table 17.2, Assuming that pooling of the data is valid (a big assumption), the results show that Age, Income, and Price have significant positive impact on charitable donation, and MS has negative but ally insignificant effect on charitable contributions, Surprisingly, DEPS has a and significant impact on charitable giving. The low Durbin—Watson in the tance is probably mote an indication of specification error than spatial or ~4.674219| _1.298134| _.-3.600722 1547275 | 0.216955 |... 7.131788 1.035779] 0.128944 | 8.032766 0.0000 0.483092 | 0.207703 || 2.325875 0.0205 0.175368 | 0.042642] 4.112556 0.0000 0.008036 | 0.184849] 0.043476 0.9653 Mean dependent var 6.577150 S.D. dependent var 1.313659 Akaike info criterion 3.152681 Schwarz criterion 3.205695 Durbin-Watson stat 0.701077 Prob(F-statistic) (0.000000 may exist among the 47 individuals, Perhaps the ‘is subsumed in the composite error term, up. AS & error term is correlated with some of the regressors is indeed the case, the estimated coefficients in Table the first-order serial correlation on the330 | SELECTED TOPICS IN ECONOMETRICS 17.4. The fixed effects least squares dummy variable (LSDV) model account the heterogeneity that may exist among 47 ing; ike into , Sharer abasic al to have his or her own intercept, a8 in the following viduals is to allow each individu‘ ‘equation: Cy = By + By Age + By Income, + By Pricey, + BsDEPSip + BoMSie + tit $51,2,..)47; £212.10 Notice that we have added the subscript i to the intercept to indicate that the inter. cept of the 47 individuals may be different. The difference may be due special features of each individual, such as education or religion. Equation (17.2) is known as the fixed effects regression model (FEM). The term “fixed effects” is due to the fact that each taxpayer's intercept, although different from the intercepts of the other taxpayers, does not vary over time, that is, it is time-in- variant. If we were to write the intercept as B, ,, the intercept of each taxpayer would pomaetaa note that in Eq. (17.2) we assume that the slope coefficients are _ But how do we make Eq, (17.2) operational? This can be done easily by introducing differential intercept dummies, which we first discussed in Chapter 3 on dummy i ally, we modify Eq. (17.1) as follows: By + B,Dy + B3Dy, +...+ BigDag; + Bay Age, *BsgIncomey; + Big Price: + B3yDEPS, + Bs)MSy + uy, j= 1 for individual 2, 0 otherwise; Ds, o& (17.2) 73) for individual 3, 0 otherwise; and nt to note that we have used only 46 dummies to represent 47 indi- id the variable trap (perfect collinearity). In this case the 46 ma pene itercept dummy coefficients — that is, they cept coefficient of the individual that is assigned a from the benchmark category. We are treating the first or reference category, although any individual can be notiee about the re of the individual d results in Table 17.3 is that the table does not ‘estimat liferential intercept coefficients, although theY h dias the model. However, the differential intercePt ; See Exercise 17.1), Secondly, if you compare the the FEM results, you will see substantial difet in the values of the Coefficients, but also in theit ion the coefficient of DEPS is not only positive But is also highly significant. The MS coeticien tis not statistically significant. Why shou! Pooled OLS estimates. If you examin® ‘will find that several of them #"°| PANEL DATA REGRESSION MODELS | OLS charity regression with individual dummy coefficies ki aie 10 sasha indinie: 47 2.089970 8 0.102249 0.208039 0.491490 0.6233 0.838810 0.111267 7538725 0.0000 0.366080 0.124294, 2.945265 0.0034 .086352| 0.053483 | 1.614569 0.1072, 0.199833 0.263890 0.757257 04493 | fixed (dummy variables) 0.763177 Mean dependent var 6.577150 0.734282 S.D.dependentvar ‘1.313659 0.677163)... Akaike info criterion 2.162215 i 191.6735 Schwarz criterion 2.621666 »)7456.1204» Hannan—Quinn crite, 2.342975 2641239 Durbin-Watson stat 1.234015 uly significant (see Exercise 17.1), suggesting that the pooled estimates geneity among the 47 charitable donors, le a test to find out if the fixed effects model is better than the OLS ven in Table 17.2, Since the pooled model neglects the heterogeneity explicitly taken into account in the fixed effects model, the pooled ed version of the fixed effects model, Therefore, we can use the We discussed in Chapter 7, which is: (17.4) are unrestricted and restricted coefficients of determination, » meters omitted from the restricted model (46 here), » is the in the sample, and k is the number of parameters estimated ion (here a total of 52), ‘the restricted and unrestricted R2 Tables 17,2 and 17.3, respectively. numbers from Tables 17.2 and 17.3, we obtain the following = 20.672 418 dfin the denominator, this F is highly significant, is model is superior to the pooled regression model. features of the fixed effects model are worth not- h as a one-way fixed effects model, for we have cross-sections (the 47 individuals), but not mmies to represent 10 years (again to avoid 331 aON the dummy variable trap) along with the 46 cross-section dummies. In that case model that emerges is called a two-way fixed effects model. Of course, if we add these time dummies, in all we have to estimate 46 cross... tion dummies, nine time dummies, the common intercept and five slope coefficien,, of the five regressors: in all, a total of 61 coefficients. Although we have 470 obseryq. tions, we will lose 61 degrees of freedom. ‘We have assumed that the slope coefficients of the charity function remain the same, But it is quite possible that these slope coefficients may be different for al) 47 individuals, To allow for this possibility, we can introduce differential slope coeff. cients, multiplying the five slope coefficients by 46 differential intercept dummies, which will consume another 230 degrees of freedom. Nothing prevents us from interacting the 10 time dummies with the five explanatory variables, which wil ‘consume another 50 degrees of freedom. Ultimately, we will be left with very fey, degrees of freedom to do meaningful statistical analysis. 332 | SELECTED TOPICS IN ECONOMETRICS z aa 17.5 _ Limitations of the fixed effects LDV model Although easy to implement, the LSDV model has the following limitations: = 1 Every additional dummy variable will cost an additional degree of freedom. j ‘Therefore, if the sample is not very large, introducing too many dummies will __ eave few observations to do meaningful statistical analysis. _ 2 Too many additive and multiplicative dummies may lead to the possibility of tmulticollinearity, which make precise estimation of one or more parameters in estimates with desirable statistical properties, we need to pay careful tention to the error term 1. The statistical results presented in Tables 172 ‘17.3 are based on the assumption that the error term follows the classical mptions, namely uj, ~ N(0,0). Since the index i refers to cross-sectional subject, we can assume that there is no autocorrelation ove! ti? assume autocorrelation of the AR(1) type. time, we can allow the error term of individual #1 to be some ith the error term for say, individual #2, or we can assume correlation.” né associate with LSDV can be alleviated if we consiet below. jon models eT d errors forthe panel data regress unrelated regression (SURE) m0 ‘efficient method of estimating see! ‘American Statistical Association, yu 57, 8-68PANEL DATA REGRESSION MODELS | 333 the fixed effect within group (WG) estimator ‘since the LSDV model may involve estimating several coefficients, one way to elim- the fixed effect in By; in Eq. (17.2) is to express both the regressand and the ssors in this equation as deviations from their respective (group) mean values - and run the regression on the mean-corrected variables. To see what this does, start h Eq, (17.2): Gy =By + BrAgen + Balncomey +B, Pricey +B; Depsy +BgMSy + ty g this equation on both sides and dividing by T (= 10) we obtain: 1s 1 io sg Ba Cie. = jp ear, + ByAge;, + Balncomey, + By Pricey (17.5) +BsDepsi + BG MS, +u;)) the parameters do not change over time, this reduces to lei Bi, +B, Age; +Bs Income: +B Price: 4B Deps; +B, Si +7, TE) ids) 3 ere a bar over a variable represents its aver "ill have 47 averaged values of each v: od of 10 years, i tracting (17.6) from (17.2), we obtain: = By (Age: —Age;)+ Bs (Income;, —Incomet) +B (Price;, —Pricei) + Bs (Deps;, —Deps,) a77) FBG (MS i, ~MSi) + (uj. —;,) the fixed or individual effect intercept term By, drops out. 1 can see from Eq. (17.7), we are essentially running mean-corrected regres- B mean-corrected regressors, Since the average value of the mean-corrected 1&8 18 zero, there is no intercept term in Eq. (17.7). OLS estimators obtained from Eq. (17.7) are known as within group (WG) a they use the (time) variation within each cross-sectional unit. 'to the pooled estimators given in Table 17.2, the WG estimators provide estimators of the slope coefficients, although they are not efficient Ge Variances), ingly, the estimators obtained from the LSDV method and the within d are identical, because mathematically the two models are identical, nin Table 17.4 (the results are obtained from Stata 10). economical than the LSDV model, one drawback of the WG esti- g the fixed, or individual, effects (ie. By), italso removes the it regressors that may be present in the model. For example, in on of wages on work experience, age, gender, education, race, ffect of gender and race will be wiped out in the mean-corrected age value over 10 years. For our exam- ariable, each average value taken over 2 i i jation in the ‘We express variables as deviations from their mean values, the variat ‘much smaller than the variation in the original values of the variables. of the disturbance term u may be relatively large, thus leading to higher coefficients.Be ecserorercsnecononer ‘Table 17.4 Within 1350 up estimators of the charity function, aged 11022493 197458 ined 8388101 11056075 7.94 prid 3660802 1179726, 3.10 2| depd 0863524] __.0507623 =1.70 0.090 | isd 1998327 250468 0.80 0.425 cons 3.15¢-09 0296465 0.00 1 00] Note: The standard errors shown in this table are slightly different from those shown in Table 17.3. Note also the value of the constant term is practically zero, as it should be. values of the regressors, for gender and race will not vary for an individual over time. ‘So we will not be able to assess the impact such time-invariant variables on wag Before moving on, we present the robust standard errors of FEM (Table 17. using White's procedure, which we have discussed in earlier chapters. ~~ If you compare these results with those given in Table 17.3 you will find that the standard errors were substantially underestimated in Table 17.3.9 included: 47 el (balanced) observations: 470 ~2.089970 1.710019 =1.222191 nce G2 RSS/(NT— 2) has tobe in-computing group averases- 0.102249 0.113897 0.897738 0.838810| _0.145653| 5.758977 0366080 | 0.146602 2497102 | _~0.086352 0.069186 1.248111 0.199833 0.712740 0.280373 A Mesh dependent var 6577150 :D.dependent yar 1.313659 Akaike: 2.162215 b é 2.621666 Dusk 1.234015, 0.000000 adjusted Howe et| ! | PANEL DATA REGRESSION MODELS | 335 .random effects model (REM) or error components the fixed effects model it is assumed that the individual specific coefficient B,, is for each subject, that is, it is time-invariant, In the random effects model it is that By; is a random variable with a mean value of B, (no / subscript here) intercept of any cross-section unit is expressed as: By =B, +8 (17.8) 2g; is a random error term with mean 0 and variance 0%, In terms of our illustrative example, this means that the 47 individuals included in are a drawing from a much larger universe of such individuals and that ‘havea common mean value for the intercept (= B,). Differences in the individ- es of the intercept for each individual donor to charity are reflected in the erm ¢;. Therefore, we can write the charity function (17.1) as: Ce =B, +B, Age, +Bglncome, +ByPricey + Bs Depsy, (17.9) +BgMSip + Wit ie (17.10) e error term wz has two components: e» which is the cross-section or error component and 1», which is the combined time series and error component.!° you can see why the REM model is also called an error components model the composite error term consists of two (or more) error components.!! ‘usual assumptions of ECM are that (0.02) (17.11) Elejui,)=0; Ele;e;)=0 G+ /) Elite) = Bly hj) =Eluyuy,) = 0 + it #9) individual error components are not correlated with each other and are elated across both cross-section and time series units. It is also critical te that wy is not correlated with any of the explanatory variables included in Since ¢; is a part of wy, it is possible that the latter is correlated with one or: "Mote regressors. If that turns out to be the case, the REM will xesut in inconsistent nation of the regression coefficients, The Hausman test, which will be ex will show in a given application if w, is correlated with the regressors ~ REM is the appropriate model. result of the assumptions in Eq, (17.11), it follows that (17.12) Native) = 03 + of ~ (17.13) ay Sometimes called the idiosyncratic term becuse i varies ove cross-section (Je ind Introduce time dummies, there will be time-specific error components (see Exercise 17336 aes TORS IN ECONOMETRICS 178 ob Pq. (17.1) and Bq. (17.9), in wh Re Lifferenee between Eq. (1 vic Now fey = Osh the ebservations and run the pooled regression, ag f are either no subjec we can simply pool all th is is so because in this situation there r eae a for by the explanatory pastas ' ‘ Although Eq. (17.18) shows that the composite error term ~ homoscedast He, itean be shown that wy and wg (ts) are correlated ~ that is, the error terms of y given cross-sectional unit at two different times are correlated. The correlation coetticjen Detween the two can be shown as: of >. Prcortvanal er Tay tes (17.14) ‘Doo points about this correlation should be noted. First, for any cross-sections unit p remains the same no matter how far apart the two time periods are; and s¢. ‘onaily, p remains the same for all cross-sectional tunits. ; IEwe do not take into account p, the OLS estimators of random effects model are inefficient, Therefore we will have to use the method of generalized least squares {GLS) to obtain efficient estimates. Software packages like Stata can compute robust or panel-corrected standard errors, Before we present the REM results for the charity example, it may be pointed ou ‘that in contrast to the fixed effects model (dummy variable, within or first-ditfere version), in REM we can include time-invariant variables, such as gender, geo: location or religion. They do not get washed out as in the FEM model. Returning to our illustrative example, we obtain the REM of Table 17.6. As in the FEM, the estimated coefficients have the expected signs, although D and MS are individually statistically insignificant. From the effects specit box, we see that o3 =(0,9309)? = 08665 and 8% =(0.6771)? = 0.4584, Then tro (17.14), we obtain p = 0.4585/1.3251 = 0.3460, which gives the extent of cor ofa cross-sectional unit at two different time periods, and this correlation Same across all cross-sectional units. This p value is the same as shown Fixed effects model vs. tandom effects model estimators given in Table 17.3 and the random ¢ You will see substantial differ the cross-section specif Fegressors. If it assumed that &; and the eee appropriate, but if they ‘we also have to which is the . error component and the regressors are tincorrelated, REM m2) are correlated, FEM may be ? — estimate fewer parameters, So hear gree erate: In the form appropriate model So how do we decide ina given situation EM do not differ substantially. His test statistic LA : * usual, if the computed chi-square value exceeds the critictl nd the level of significance, we conclude that REMPANEL DATA REGRESSION MODELS | 337 fable 17.6 Random effects model of the charity function with white standard ‘Dependent Variable: CHARITY f panel EGLS (Cross-section random effects) included: 10 “Cross-sections included: 47 otal panel (balanced) observations: 470 _ | Swamyand Arora estimator of component variances period standard errors & covariance (d.f. corrected) 2.370567, i 1.709817 0.277063 | 0.127176 | _2.178577 0.0299 0.852996 | _0.126574| 6.739099 0,0000 0.370199 | 0.140054] 2.643253 0.0085 0.036254 0.064181 | -0.564874 0.5724 0.199669] 0.472666] 0.422432 0.6729 0.930938, 0.677163 0.132701 Meandependent var 1.474396 0.123355 S.D. dependent var 0.731733 0.685116 Sum squared resid 217.7944 4.19881 Durbin—Watsonstat 1.094039 0.000000 0.136789 Meandependent var 6.577150 698.6427 Durbin—Watson stat 0.341055 a propriate because the random error terms e; are probably correlated with one -M is preferred to REM. Hausman test are given in Table 17.7. The for the p value of the estimated chi-square ‘able compares the fixed effects and random, ¢ last probability column of the table shows, | the differences in the Age and DEPS coefficients are statistically highly significant. lipetterHtatiairion test exaraines (Dy ~ bre)? ~ that ls, the squared difference regression coefficients estimated from REM and FEM. the REM model does not seem appropriate in the present example, we \evert to the FEM model. Another alternative is to continue with REM but use 4 (IV) for the individual effect that may be correlated with model, But the use of instrumental variables with panel data. we will not pursue it in this book, although we will dis-_ detail in Chapter 19. However, it may be noted that -regressors. In this case, FE! “our example, the results of the nan test strongly rejects the REM, very low, The last part of this coefficient of each variable. As th [V method in some338. ‘SELECTED TOPICS IN ECONOMETRICS Table 17.7 Results of the Hausman test. Correlated Random Effects ~ Hausman Test Equation: Untitled ‘Test cross-section random effects AGE 0.102249 0.277063 0.003539 INCOME 0.838810 0.852996 0.000830 PRICE 0.366080, 0.370199 0.000087, | DEPS 0.086352 0.036254 0.000487 0.0232) ‘MS 0.199833 0.199669 0.016167 0.9990 | Hausman—Taylor estimator and the Arellano—Bond estimator use the instrumen. tal variables to estimate REM models. For a somewhat accessible discussion of these estimators, see the references.!2 Some guidelines about REM and FEM 7 Here are some general guidelines about which of the two models may be suitable in practical applications:!3 1 IfT (the number of time observations) is large and N (the number of cross-sec- tion units) is small, there is likely to be little difference in eters estimated by FEM and REM. The choice then d convenience, which may favor FEM. 2 Ina short panel (N large and T'small) the estimates obtained from the two mod els can differ substantially. Remember that in REM By = B, + &, where g; is the cross-sectional random component, whereas in FEM By; is treated as fixed. In the latter case statistical inference is conditional on the observed crose sect! units in the sample, This is valid if we Strongly believe that the cross-sectional nits in the sample are not random drawings from a latger population. [0 case, FEM is appropriate, If that is not the case, then REM is appropriate because in that case statistical inference is unconditional, 3 IENis large and T is small, and if the i REM a assumption: erlying REM hold, REM estimators are more efficient than FEM, fons underlying REM h the values of the param lepends on computational FEM controls for all time-invariant variebles 12 See Koop, , Introduction to Eeonomety discussion, see Cameron, A.C. and ‘Trivedi, University Press, Cambridge, 2005, pp, 765. 13 See Judge, G. G,, Hill, R.C., Griffiths, W. and Practice of Econometrics, 2nd edn, Wiley, vd Kay Chichester, UK, 2008, p. 267-8, Foran advanee! Microeconomics: Methods and Applications, Canb™' ¥ Lutkepoh, Hand Lee, 1: Cx dutroduction to the York, 1985, pp. 489-91, NewPANEL DATA REGRESSION MODELS | 339 whereas REM can estimate only those time-inv: Brotiiced in the model, ariant variables that are explicitly 79 Properties of various estimators!4 In this chapter we have discussed several methods of estimating (li oir s ig (linear) panel regres- = sion models, such as pooled estimators, fixed effects estimators (both LSDV and bahiestoae vi ipa beens random effects. What are their statistical properties? concentrate on the consistency property, since panel data usually invol large number of observations. eee Pooled estimators: If the slope coefficients are constant across subjects, and if the error term in Eq. (17.1) is uncorrelated with the regtessors, pooled estimators are consistent. However, it is very likely that the error terms are correlated over time for a given subject. Therefore, we must use panel-corrected standard errors for hypothesis testing. Otherwise, the routinely computed standard errors may be wd It may added that if the fixed effects model is appropriate, but we use the pooled model, the estimated coefficients will be inconsistent, as we saw in our charity ___. Fixed effects estimators: Even if the underlying model is pooled or random-ef- fects, the fixed effects estimators are always consistent. __. Random effects estimators: ‘The random effects model is consistent even if the true model is pooled. But if the true model is fixed effects, the random effects esti- data regressions: some concluding comments ed at the outset, the topic of panel data modeling is vast and complex. We scratched the surface. Among the topics that we have not discussed at ing may be mentioned. ‘data models inwhich thelagged value(s) of the regressand appears y variable. equations involving panel data, dent variables and panel data. +1 data (on unit roots, see Chapter 13). se topics can be found inthe references cite inthis chaptes, 0 consult them to learn more about this topic. These refer- ‘empirical studies in various areas of business and economics regression models. The beginner is well advised to read ron and ‘Trivedi, op cit, Chapter 21,‘340 | SELECTED TOPICS IN ECONOMETRICS some of these applications to get feel for how researchers have actually implemen, ‘ed such models.!5 17.11. Summary and conclusions (Panel data regression models are based on panel data, which are observations on the same cross-sectional, ot individual, units over several time periods. Panel data have several advantages over purely cross-sectional or purely tire series data, These include: (a) increase in the sample size, (b) study of dynamic changes in cross-sectional units over time, and (c) study of more complicated behay. ioral models, including study of time-invariant variable However, panel models pose several estimation and inference problems, such as heteroscedasticity, autocorrelation, and cross-correlation in cross-sectional units at ‘the same point in time. ‘The two' prominently used methods to deal with one or more of these problems are the fixed effects model (FEM) and the random effects model (REM), also known as the error components model (ECM) In FEM, the intercept in the regression model is allowed to differ among individ- uals to reflect the unique feature of individual units. This is done by using dummy variables, provided we take care of the dummy variable trap. ‘The FEM using dum ‘my variables is known as the least squares dummy variable model (LSDV). FEM is wre in eae erento intercept may be correlated more regressors. ivantage of th i f degrees of freedom when N (the number a ee ee faeneine to LaDy, mn tet es within-group (WG) estimator. Here we sub- “ae ee lore) PF te ee SERRE DE, and ASSEESEOT, from their individual aa acme Fee pe ee variables. Although it is eco- le cient pests oes =e nia variables wipe out .; : s e model. — on sp FEM pe In Ee we assume that the intercept value of an See Rr aS much larger population with a constant ei een Sots a eo. a deviation from the constant ee, BEM Seavert i than FEM in terms of the number of param tds okie Serie pene where the (random) intercept of Pe aterien te with the regressors. Another advantage of eee chicicoian a Fegressors. This is not possible in FEM Eivied can be keeles ‘tae x subject-specific intercept. 1 specific problems with panel data, etween FEM and ECM. : sien model need to be kept in mind. The ™°* een over pico whereby for one reason or apothes Mae ime so that in the subsequent surveys ath remain in the panel. Also, over time suble* ng to answer some questions.) ash Ons, se Allon, PD, Fed Efects Regression Mota)” Cary; North Carolina, 2005.