27
additional controls. There is no other statistically significant cause of death before age
65. Nor can we find any single cause that is statistically significant up to age 70.
24
In
order to study heterogeneous treatment effects further, we estimate effects on the
number of days in inpatient care, subdivided into different diagnoses in the next section.
Table 7: The effect (hazard ratio) of being offered early retirement on the conditional
probability of dying, based on a discrete-time Cox regression model, by cause of death
Censoring at 71
Censoring at 66
Acute myocardial infarction
0.5889
0.5767
0.2670*
0.2776*
(0.2189)
(0.2197)
(0.1515)
(0.163)
Ischemic heart diseases and Stroke
0.9183
0.917
0.4994
0.4955
(0.4065)
(0.4225)
(0.2863)
(0.2964)
Neoplasms (tumors)
0.691
0.7057
0.5796
0.586
(0.1675)
(0.1708)
(0.1971)
(0.1985)
Alcohol or narcotics
0.7851
0.8266
0.7446
0.8054
(0.4813)
(0.5377)
(0.5371)
(0.6329)
External causes
1.643
1.3309
2.1868
1.7189
(0.937)
(0.7809)
(1.6744)
(1.3388)
Other
0.6533
0.7402
0.8499
1.0908
(0.2489)
(0.287)
(0.4502)
(0.591)
Controls
No
Yes
No
Yes
Note. Estimation is performed with maximum likelihood. Standard errors are adjusted for 19, 906 clusters
(individuals): † p<.1; * p<.05; ** p<.01. All models include a military dummy and dummy for cohort 1938-1939.
„Controls‟ are income, income interacted with dummy for cohort 1938-1939, and education. The number of
observations is 283,208 and 232,087 for the estimates that are censored at age 71 and at age 66 respectively. The
exact ICD-codes for groups are given i
6.4.1 Effects on number of days in inpatient care due to different diagnoses
provides the results from the estimation of the effect of the offer on the number
of days in inpatient care for six diagnosis groups (mental and behavioral disorders,
injuries and poisoning, acute myocardial infarction, other ischemic heart diseases,
stroke and tumors, and the remaining (other). There is no statistically significant effect
on acute myocardial infarction or other ischemic heart diseases. One potential problem
or concern regarding the interpretation of this result is the above-observed reduced
mortality in these diseases. The implication is that an effect on the number of days in
inpatient care could be biased towards zero (or even be positive). There is, however a
large reduction in the number of days in inpatient care due to external causes (e.g.
injuries and poisoning), by 76-89 percent (see columns (a) and (b)). This indicates that
24
The assumption made in the competing risk model is that, conditional on the observed covariates, the risks between
the 6 causes of death should be independent. This assumption may be too restrictive, as unobserved health may
jointly affect all exits. We tried estimating mixed proportional hazards models (that is, we allowed for unobserved
heterogeneity). However, we did not obtain convergence in these models.
