Discrete time transition rates into retirement for both social security records and survey data are estimated. All relevant information from social security records is used while the resulting models are replicated using survey data. Baseline estimations serve for comparative purposes and it is important to keep in mind that these data sources differ by the type of information that they provide. Thus, the definition of retirement differs across data sources (VSKT: benefit claiming; SOEP: self-reported retirement) and dif- ferent assumptions for the calculation of benefit entitlements are necessary. Moreover,
are provided in table 3.3.19
Table 3.3: Baseline Estimation: Benefit Reductions and Retirement Transitions.
Variable Logit Probit Compl. Log-Log
m.eff. s.e. m.eff. s.e. m.eff. s.e.
Social Security Records (VSKT)
BRR –.068 (.007) –.079 (.007) –.067 (.007) EPDV .000 (.000) .000 (.000) .000 (.000) Male –.008 (.001) –.007 (.001) –.006 (.001) West –.009 (.001) –.009 (.001) –.009 (.001) + Eligibility-Type-Dummies + Year-Dummies + Duration-Dummies
Mean Transit. Rate (%) 3.48 3.50 3.47
Obs.(Person-Month-Obs.) 14660(407663) 14660(407663) 14660(407663)
Survey Data (SOEP)
BRR –.190 (.030) –.198 (.030) –.183 (.030) EPDV .000 (.000) .000 (.000) .000 (.000) Male .008 (.002) .009 (.002) .007 (.002) West –.023 (.003) –.021 (.002) –.023 (.003) + Year-Dummies + Duration-Dummies
Mean Transit. Rate (%) 2.92 2.92 2.92
Obs.(Person-Month-Obs.) 1527(43245) 1527(43245) 1527(43245)
Source: Own calculation based on SUFVSKT2002-SUFVSKT2010 and SOEP (1995-2011). Note:
Reported values are average marginal effects. For factor variables, reported values are the discrete change corresponding to the reference category. Standard errors in parentheses. Mean transition rates
are predicted from respective models and reported in per cent. EPDV is the expected present discounted value and BRR is the benefit reduction rate.
The key regressor BRR reflects individual adjustment rates, which are exogenously de- termined by year and month of birth. This variable varies between 0 for individuals at ages where no adjustment applies up to 0.18 (i.e. 18%) for individuals at ages where the maximum adjustment applies. For social security records, the marginal effect for BRR suggests, that increasing the adjustment rate by one percentage point reduces the probability to retire by 0.07 percentage points on average for a given point in time. While this effect seems small in absolute terms, evaluated at the predicted mean tran- sition rate of 3.47%, this is an average decrease of 2% in the probability to observe a transition into retirement in a given period at risk. For survey data, the marginal effect for “BRR” indicates that increasing the adjustment rate by one percentage point reduces the probability to retire by 0.18 percentage points on average for a given point
19All binary choice models as presented across columns, i.e. logit, probit and complementary log-log, are robust over corresponding distributional assumptions. For this reason, the subsequent discussion concentrates on results as obtained from the discrete time proportional hazard model (complementary log-log).
in time. Evaluated at the predicted mean transition rate of 2.92%, this is an average decrease of 6% in the probability to observe a transition into retirement in a given period. For all estimated models in the baseline scenario, the probability that this result occurs by chance is very small (<0.001). Thus, in terms of conventional error probabilities, the null hypothesis of no impact of benefit reductions is rejected. This is in favour of Hypothesis 1, that benefit reductions lower the attractiveness of early retirement and thus induce postponed retirement. While the results are very similar with respect to their sign, they do differ by magnitude across data sources.
Differences in magnitude of estimated coefficients and average marginal effects are not surprising when taking into account that the underlying data sources differ substantially by quality and quantity of information that they provide.
Omitted variable bias in regressions using the VSKT may attenuate estimates towards zero and explain the smaller marginal effects. Retirement decisions are outcomes from a complex mix of determinants. Aside from financial resources as discussed in section 3.3, the literature is clear about other influential determinants such as marital status and health. Married individuals may condition their retirement entry decision on their spouses retirement behaviour (see e.g. Blau and Riphahn, 1999). Individual health may play an important role in the timing of retirement, which has been subject to many previous studies (see e.g. Berkovec and Stern, 1991; McGarry, 2004).
For survey data (SOEP), measurement error in regressors may attenuate estimated coefficients towards zero. For both the VSKT and the SOEP the estimated response in retirement timing is a lower bound for the true (unknown) response.
3.5.2 The Heterogeneous Response of Manual and Non-Manual Work-
ers
Estimating a richer model with further information on worker heterogeneity shows that the marginal effect in the baseline estimation on survey data seems to be biased towards zero to some extent (table 3.4). Consequently, the marginal effect of benefit reductions as measured by “BRR” is larger in absolute terms (i.e. more negative).
Table 3.4: Benefit Reductions, Retirement Transitions, and Worker Heterogeneity.
Variable Logit Probit Compl. Log-Log
m.eff. s.e. m.eff. s.e. m.eff. s.e.
Survey Data (SOEP)
BRR –.226 (.034) –.234 (.033) –.220 (.034) EPDV .000 (.000) .000 (.000) .000 (.000) Manual .003 (.002) .002 (.002) .002 (.002) BRR X Manual .063 (.026) .068 (.026) .067 (.026) Male .006 (.002) .007 (.002) .005 (.002) West –.022 (.003) –.020 (.003) –.022 (.003) Married –.010 (.003) –.010 (.003) –.009 (.003) Years of Education .000 (.000) .000 (.001) .001 (.000) Moderate Health .001 (.002) .001 (.002) .002 (.002) Good Health .001 (.002) .001 (.002) .001 (.002) + Year-Dummies + Duration-Dummies
Mean Transit. Rate (%) 2.91 2.91 2.92
Obs.(Person-Month-Obs.) 1497(42353) 1497(42353) 1497(42353)
Source: Own calculation based on SOEP (1995-2011). Note: Reported values are average marginal
effects. For factor variables, reported values are the discrete change corresponding to the reference category. Standard errors in parentheses. Mean transition rates are predicted from respective models
and reported in per cent. EPDV is the expected present discounted value and BRR is the benefit reduction rate.
at the predicted mean transition rate of 2.92%, this is an average decrease of 7.5% in the probability to observe a transition into retirement in a given period. However, the central finding in this model is the positive marginal effect for the interaction “BRR * Manual”. This result indicates, that manual workers respond to a much lower degree to benefit reductions compared to non-manual workers. For both, “BRR” and “BRR * Manual”, the probabilities that these results occur by chance are not larger that 0.01 and thus the null hypothesis is rejected supporting Hypothesis 1a. Further results sug- gest that living in West Germany and being married are in favour of retiring at higher ages.
Explicitly assuming and modelling an unobserved heterogeneity distribution following Prentice and Gloeckler (1978); Meyer (1990) yields coefficient estimates that are robust in magnitude and sign to all previous results (see appendix 3.B) suggesting that these findings are robust unlikely to be confounded by unobserved factors.