Fixed-Effects Results Empirical Results

are small. Therefore, it appears that life satisfaction remained fairly constant in the West over the 15-year period.

B. Fixed-Effects Results

Table 3 provides the causal estimates from the Ordered Logit model with fixed effects for East and West German, males and females, respectively. The Table also shows the values of our test-statistic for the appropriateness of the fixed-effects model relative to the random-effects specification. For men, the number of restricted parameters is 17, for which the 1 percent critical value of the chi-squared distribution is 33.4. For men in West Germany, this means the null of no difference between the random- effects and fixed-effects model is clearly rejected. For men in East Germany, the null is not rejected, though for men in East and West combined it is rejected. For females, the number of restricted parameters is 19 and the appropriate 1 percent chi-square critical value is 36.2, from which we can see that the null hypothesis is rejected for females in West Germany at the 5 percent level and for East German females at the 1 percent level. As a test of total changes with the fixed effects, it holds that the sum of the test statistics is under the null chi-square distributed with 72 degrees of freedom. The sum of the test-statistics is 152.8 and the 1 percent critical value of the chi-square distribution with 72 degrees of freedom is 114.8. Bearing in mind that the test statis- tic was biased towards accepting the null, our specification tests hence clearly point to the presence of fixed effects. Unfortunately, the fixed-effect model does not provide estimates of the probabili- ties of having a particular level of life satisfaction, thus it has no Marginal Effects ME proper. By approximation, however, an increase of 1 in a variable with coefficient The Journal of Human Resources 664 Figure 3 Year-Effects for West German Males Frijters, Haisken-DeNew, and Shields 665 Table 3 The Determinants of Life Satisfaction for East and West Germans: Ordered Logit Models with Fixed Effects East West Male Female Male Female Covariates β t -stat β t -stat β t -stat β t -stat Age squared100 0.080 1.74 0.170 5.03 0.043 3.98 0.035 3.54 Married 0.152 1.01 0.025 0.15 0.237 6.03 0.199 5.32 Separated in last year −0.341 −1.65 −0.593 −3.14 −0.520 −5.97 −0.396 −4.55 Divorced in last year 0.560 1.34 −0.581 −1.89 −0.191 −1.31 0.031 0.24 Spouse died in last year −1.838 −3.57 −1.073 −3.02 −1.130 −4.04 −1.413 −7.65 Death of other family −0.050 −0.18 −0.073 −0.25 −0.217 −1.78 −0.147 −1.14 member in year Disabled −0.093 −0.63 −0.058 −0.38 −0.160 −3.48 −0.207 −3.92 Ln1+number of days in −0.073 −2.00 −0.088 −2.67 −0.103 −7.99 −0.058 −4.99 hospital in last year Number of children 0.236 3.70 0.105 1.63 −0.035 −2.41 −0.075 −5.13 Had a baby in last year 0.124 0.76 0.225 1.19 0.160 3.30 0.372 6.67 Invalid in household −0.165 −0.74 −0.346 −1.69 −0.412 −5.89 −0.442 −6.37 Employed full-time only 0.838 9.60 0.876 10.41 0.876 19.58 0.497 10.26 for females Employed part-time — — 0.711 6.55 — — 0.448 8.68 Maternity leave — — 0.551 3.15 — — 0.607 8.57 Nonparticipant 0.584 4.50 0.317 2.70 0.773 14.70 0.504 10.18 Fired in last year −0.037 −0.37 −0.077 −0.72 −0.180 −2.71 −0.200 −2.44 Log household income 0.430 4.62 0.586 6.75 0.161 6.94 0.151 7.51 post tax Moved within East West −0.022 −0.19 0.101 0.88 −0.054 −0.91 −0.114 −2.12 Germany in last year Moved to West East 0.744 2.90 0.538 2.07 −0.494 −1.97 0.114 0.49 Germany after reunification Live on the border of East −0.047 0.18 0.330 1.47 −0.093 −1.01 0.106 1.04 and West Germany Year controls Base = YES YES YES YES YES YES YES YES 1991, 1985 Mean Log likelihood −2.308 −2.330 −4.195 −4.286 2L βt ML FE -max αt {2L α β t t RE } 22.18 40.13 56.38 34.06 αt 1.39 1.52 1.30 1.29 µt 0.533 0.561 — — Sample individuals 1,777 1,834 6,075 5,976 Notes: Due to the different normalisations of the random and fixed-effects models the fixed estimates esti- mate should be multiplied by αt to allow direct comparison with the random-effects estimates. — means not included in the model. β has an effect of µtβ on expected life satisfaction. The coefficients from the fixed- effect model can furthermore be compared with the coefficients from the random- effects model multiplied by the estimate of α. Although our statistical test favours the fixed-effects approach, an important result is that many of the key relationships found by the random-effects models also hold qualitatively for the fixed-effects models. In particular, we find that the effect of a recent separation, death of spouse or some other family member still have a large neg- ative impact on life satisfaction. The death of spouse in the previous year is the life event with the largest quantitative effect combined with unemployment of all the variables included in the models. The effects are particularly large for East German men ME = −0.290 and women ME = −0.163. Similarly, being disabled, spending time in hospital, having an invalid in the household, and getting fired are all estimated to reduce life satisfaction. However, a number of these effects are no longer statisti- cally significant. The findings with respect to the detrimental effect of unemployment and the positive effect of employment on life satisfaction remain statistically robust for both East and West Germans, which suggests that it is not the most unobservably unhappy who were observed in unemployment. This points to the involuntary nature of unemployment in both locations Gerlach and Stephan 1996; Winkelmann and Winkelmann 1998. In addition, the positive effect of children found for East Germans, and the negative effect of children found for West Germans, which we find with the random-effects approach, remain statistically strong once we control for fixed effects. This is also the case for the positive effect of having a baby in the last year for West Germans. Moving from the East to the West of Germany increases expected satisfaction of East German males by about µt0.744 ≈ 0.40 and for females by about 0.26. Again, this implies that it is not the unobserved happy who moved from East to West. There is a genuine satisfaction gain from living in West Germany, which is independent of the possible associated changes in income and other variables included here. We also find that the detrimental effect of moving from the West to the East following reuni- fication still holds for males, but the effect is no longer present for females. An important difference between the models occurs with respect to the impact of income: the effect of household income which is now identified from income changes on life satisfaction in East Germany is far greater in the fixed-effect case, than it has in the random-effect case, and it is much higher in the East than in the West. An increase of 1 in lnincome increases expected satisfaction of East German males by about µt0.430 ≈ 0.23 and ≈ 0.33 for females. This large effect concurs much more with the economists’ intuition that money must surely matter a lot, even though many other studies find only small effects see, for example, Clark and Oswald 1994. The coefficient in comparison with the coefficients of other variables is amongst the highest found in this literature, even with the same data see Ferrer and Frijters 2003. A potential reason is that our data is very “clean”: measuring incomes by surveys is notoriously difficult because respondents’ underreport transient ele- ments of income, such as bonuses, side-benefits, holiday payments, etc. However, the GSOEP contains information on more than 50 sources of income, detailed at the monthly level. See Burkhauser et al. 1999 for more information on this. Finally, it is important to note that what is captured by the year dummies here is quite different from the random-effects case, because year effects in the fixed-effects The Journal of Human Resources 666 models include age and years in panel effects as well as unobservable variables, mak- ing them incomparable. For East Germans, who are the main focus of this paper, we further investigate the importance of general changes captured by our year and age variables in the post-reunification period in the following decomposition analysis using the estimates from the fixed-effects models highlighted in Section IV.

C. Decomposition Results