5. Incentive effects in the pension system of Hungary
5.2. Differences between pensioners and non-pensioners
The decision to retire depends on the relative weights of the advantages and disadvantages of pensioner and non-pensioner status. It is assumed to be a genuine decision, although – some may rightly object – there may be no al-
58 With the exception of food expenditure, which only appears in monthly data.
59 This working definition of start of retirement is clearly not perfect and, although it is suitable for our purposes, it cannot be used to enumerate people entering retirement in the given period. Since in theory there is no seasonal effect in retirement, people may enter the pension programme in any month of the year. If, however, people are only observed in one month of the year, the odds of someone retiring without the change being observed is exactly 0.5 (the month of retirement is smaller than the month of the observation). The fact that half of the retiring events are left unobserved may be a significant loss, but unfortunately, we have no other option. The main rea- son for this is that a method of almost complete capture would require the use of yearly income data, which would not allow us to determine pensioner and non-pensioner incomes over the entire period with reasonable accuracy.
incentive effects ternative to retirement under the given circumstances due to various pres-
sures. Strictly speaking, it is the different forces of the various constraints which are investigated in this case. Financial incentives are important, but of course not the only factors that contribute to the level of welfare that may be attained in either of the two situations. Other considerations include first of all the temporal distance from retirement age, as well as leisure activities and family commitments.
If we are to estimate the probability of pensioner status (i.e., the effects of individual factors on the probability of an arbitrarily selected person being retired), we can use cross-section data from a given period. The probability of pensioner status (the state, being part of the stock), however, may deviate substantially at a given point of time from the probability of entering retire- ment (the change, being part of the flow). Since, however, in most cases only cross-section data are available, for reasons of comparability, we shall first use a simple model with cross-section data to investigate the effects of individual factors on someone being retired or not. The set of independent variables used in the equations includes the usual human capital type variables and also oth- ers which may have an effect on labour market success due to different condi- tions on regional labour markets.
Table 5.1: Estimate of the probability of pensioner status among people aged 25–64 based on cross-section data (average marginal effects from probit regression)
Explanatory variable (1) (2) (3) (4)
Activity rate of region of residence (%) –1.043*** (0.17) –0.846*** (0.17) –0.342** (0.15) Activity rate of region of residence squared 0.761*** (0.16) 0.612*** (0.16) 0.238* (0.14)
Education: vocational training 0.00216 (0.0058) 0.0283*** (0.0042)
Education: upper secondary –0.0333*** (0.0059) 0.0237*** (0.0047)
Education: higher education –0.0844*** (0.0057) 0.00152 (0.0054)
Female –0.0512*** (0.0025) –0.0390*** (0.0021)
Partner is retired 0.0893*** (0.0032) 0.0691*** (0.0027)
Working –0.311*** (0.0034)
N 113,348 112,854 112,854 112,854
R2 0.461 0.472 0.496 0.645
Control variables not shown + control vari- ables of distance from retirement age
+ control vari- ables of distance from retirement age, region, set- tlement type and year
+ control variables of distance from retire- ment age, region, set- tlement type and year
+ control variables of distance from retire- ment age, region, settlement type and year
Robust standard errors corrected for arbitrary heteroscedasticity and the recurrence of observation units (clus- tering) are given in brackets.
*** p<0.01, ** p < 0.05, * p < 0.1.
The estimates over the observations of every person between the ages of 25 and 64 in the sample are shown in Table 5.1 using four different specifications.
The figures in the table indicate the magnitude of change in the probability of an arbitrarily selected person being retired due to a change in a given vari- able, independently of the effects of every other observed trait. Column (2), for instance, shows that given a group of people of the same age, the same ed- ucational attainment and living in regions with similar unemployment rates, the probability of being retired is 8 per cent lower for a female member of the group than it is for a male member. Individual effects appear independently of each other. Thus, the effect of temporal distance from retirement age, for instance, is not mixed with the effects of educational differences in the com- position of various age cohorts.
It is clear that the distance from retirement age is perhaps the most impor- tant indicator that distinguishes pensioners from non-pensioners. The expo- nentially rising curve alone (shown in Figure 5.1) accounts for 46 per cent of variability in pensioner status. To achieve the greatest possible flexibility, this curve was estimated with a non-parametric method (using a different indicator variable for each distance in number of years). As it is highly space- consuming to display the results in a table, this information is not shown in the data columns of Table 5.1, although it was included in each estimation.
The outcome of this estimation alone can be seen in Figure 5.1. The variable has such a strong explanatory power that the model’s success in predicting retirement only improves by one percentage point if we add the variables of local activity rate, region, settlement type and year of observation (see Col- umn [2] in the table). Including the variables of education, sex and partner’s activity in the model has a more significant but still moderate effect (Column [3]). The results indicate that relatively highly educated people and women are less likely to be pensioners, similarly to those who live in a region with better employment prospects or whose partners are not pensioners. The observed effects of these factors correspond to our expectations to some extent, since employment prospects are related to training, and partners or spouses often synchronize their retirement. It may come as a surprise, however, that wom- en are less likely to be pensioners than men. The explanation is that although the statutory retirement age is lower for women (which should increase the odds), there is a lower incidence of disability pension among women compared to men (which decreases the odds) and there is a large number of untrained women, the effects of which are captured by the inclusion of the variable of education. With these effects controlled for, the probability of pensioner sta- tus is reduced. A major change can be observed with the inclusion of indi- viduals’ employment status (Column [4]): working people are less likely to be pensioners compared to non-working people. Now we are in a much bet- ter position than before to explain pensioner status; it is not clear, however, what was first: the pension or the lack of employment.
incentive effects Figure 5.1: Probabilities of pensioner status and retirement as a function
of the number of years left until retirement age (differences are relative to retirement age as reference value)
Source: Author’s computations based on cross-section and panel data from the HCSO Household Budget Survey.
There are two reasons why a cross-section analysis is not suitable for assessing the incentive effects of the pension programme. First, it shows the cumulative result of factors motivating retirements over a relatively long period of time (which is also influenced by deaths) and therefore it cannot provide a reliable estimate of the incentive effects of the current pension system. Secondly, it does not allow us to measure the effects of considerations arising before the decision to retire is made (which is our main concern). To measure that, panel data are needed and flows rather than status need to be observed.
The results of a simple estimation concerning entry into retirement are shown in Table 5.2. In the simplest model, only the factor of temporal dis- tance from retirement age is included as a variable predicting the probability of entering retirement. Similarly to the cross-section estimation, the effects of this factor are only shown in Figure 5.1, together with the raw cross-section probability of retirement. Although the explanatory power of the model us- ing panel data is smaller (accounting for around 6 per cent of variability), the curve displays an exponential shape similarly to the raw cross-section data and corresponding to our expectations based on the standard duration model.60 A notable difference between the two curves is that while the probability of pensioner status based on stocks peaks at retirement age, the raw probability of entering retirement based on flows peaks three years before, at exactly the point when early retirement becomes available without any penalties. The de- terminants of entering retirement are added to the model gradually.
–0.8 –0.6 –0.4 –0.2 0.0
Raw probability of transition to pensioner status Raw probability of pensioner status
Probability of pensioner status purged of the effect of individual characteristics
0 –10
–20 –30
0.2
–40 10
60 Note that we do not expect equivalence: the two curves could only be equivalent un- der highly exceptional circum- stances.
Table 5.2: Estimate of the probability of entering retirement among people aged 25–64 years based on panel data (average marginal effects from probit regression)
Explanatory variable
25–64 year old 40–64
year old
(1) (2) (3) (4) (5) (6) (6)
Activity rate –0.0242 –0.0193 –0.0238 –0.0439* –0.0418* –0.0669
(0.015) (0.015) (0.015) (0.023) (0.022) (0.041)
Education: vocational training –0.00191 –0.00415 –0.00232 0.0335* 0.0484
(0.0035) (0.0036) (0.0079) (0.019) (0.033)
Education: upper secondary –0.00740** –0.00911*** –0.00409 0.0611 0.0733
(0.0035) (0.0035) (0.0078) (0.045) (0.065)
Education: higher education –0.00941** –0.0121*** –0.00459 0.169 0.133
(0.0037) (0.0034) (0.0082) (0.14) (0.16)
Female –0.00732*** –0.00804*** –0.0230*** 0.0915 0.0619
(0.0017) (0.0017) (0.0028) (0.10) (0.12)
Partner is retired 0.00703*** 0.0215*** 0.0296*** 0.0240*** 0.0311***
(0.0020) (0.0022) (0.0034) (0.0032) (0.0054)
Partner is retiring 0.142*** 0.168*** 0.140*** 0.121***
(0.010) (0.015) (0.014) (0.018)
On sick leave in base period 0.0377*** 0.0313*** 0.0289*** 0.0426***
(0.0040) (0.0039) (0.0037) (0.0065)
Net income (log) –0.159*** –0.157*** –0.275***
(0.049) (0.045) (0.094)
Net income (log) squared 0.00652*** 0.00684*** 0.0119***
(0.0023) (0.0021) (0.0044)
Work experience (years, potential) 0.00726** –0.00528
(0.0034) (0.0071)
Work experience squared 0.0000 0.000194***
(0.000019) (0.000059)
Net income (log) × female –0.00535 –0.00482
(0.0049) (0.0091)
Number of years before retirement age –0.000795 –0.0132**
(0.0032) (0.0056)
Number of years before retirement age squared 0.000151*** 0.000652***
(0.000020) (0.000074)
Over retirement age –0.0126 –0.0360***
(0.012) (0.013)
Working 0.0184***
(0.0025)
N 45,385 45,385 45,385 45,385 21,264 20,834 10,298
R2 0.064 0.072 0.076 0.14 0.23 0.223 0.195
Each equation includes the control variables of year, region and settlement type, which are not shown here. With the exception of the last two, each estimation includes a row with the indicator variable measuring the difference between age and retirement age, i.e., people’s temporal “distance” from the effective retirement age. The final equation also includes year indicators which interact with income.
Robust standard errors corrected for arbitrary heteroscedasticity and the recurrence of observation units (cluster- ing) are given in brackets.
*** p < 0.01; ** p < 0.05; * p < 0.1.
incentive effects The explanatory value of the model is substantially improved when individu-
al traits are added which capture temporal changes and motivations predict- ing the decision – this is shown in Columns (2)–(4) in the table. Factors in this group include sick leave status in the first period of observation, which approximates health status, the partner entering retirement and the net in- come received over the first period. Since labour income will also be taken into account in later analyses, the models shown in Columns (5)–(6) use the sub-sample of people in employment. Disregarding the fact that the effect of education disappears, the results are qualitatively similar to our previous findings. The effect of education is replaced by a highly significant effect of income. The last column shows a re-estimation of model (6) where the find- ing that the incidence of entering retirement displays a steep increase from the age of 40 is taken into account and the sample is reduced to people in the age cohort of 40 to 64 years. The results indicate that the effects of most of the variables are highly similar to those observed before. One exception is work experience, which displays a linear effect for this age group.
5.3. A more complete model estimation of entering retirement