13
4
Empirical Modeling and Results
Interest is in analyzing effects on prevalence of employment, sickness, and DB 15 months
after the experiment was conducted. An individual can be part time on sickness and
disability benefits and part time working/unemployed. As a consequence we do not
consider the three states as mutually exclusive. Unfortunately, there is no spell
information on employment. The implication of this is that if an individual is not observed
as being unemployed, sick, or disabled we do not know if the individual is working a
certain month. Potentially the individual has left the labor force instead. A reasonable
assumption, however, is that the vast majority of those in this “residual state” among
those employed at the beginning of the sickness absent spell return to work if not
unemployed or on sickness and DB. As we have information on yearly income from work
this maintained assumption, of resumption to employed if no longer being unemployed or
on sickness or disability benefits, is possible to investigate.
Effects from the treatment 15 months after the experiment was conducted are
provided in section 4.1. In section 4.2 we study the dynamics by presenting the effect for
each of the 15 months. We furthermore investigate the assumption of resumption to
employment for employed individuals in the residual state. Finally, section 4.3 provides a
sensitivity analysis.
4.1
Results
Since individuals naturally recover from an illness there is a natural outflow from the
sickness absence for both treated and controls. The weakly hazard rate in the control
group (not reported here) is quite constant for the first 60 days: around 4 to 4.5 percent.
This means that approximately 1,700 individuals in the control group have left sickness
absence before having a chance of receiving Sassam. One consequence of this natural
outflow is that only 45 percent and 31 percent in the treatment and control groups,
respectively, received Sassam. The corresponding shares in the AM-experiment were 30
and 26 percent (the share of treated for both groups each month in the 15 months follow-
up period is provided in Appendix B).
This natural dropout from sickness absence makes it possible to estimate the effect
of Sassam and AM by making use of a WALD or a two-stage least squares estimator
(2SLS). In the situation of heterogeneous treatment effects that depend on the length of
the sickness absence spell, a consequence of the design is that the 2SLS estimator does
not estimate the average treatment effect from Sassam or AM for the population of
sickness benefits recipients. Instead the estimator provides an estimate of the local
average treatment effect (LATE) (Imbens and Angrist, 1994) for the population who
