A principle in economic modelling is that an individual’s well-being increases with consumption of goods and leisure and decreases with hours spent at work. In health economics the workhorse model by Grossman (1972) on health demand starts by assuming a trade-off between investing time in one’s health and allocating it elsewhere. These two modelling frameworks suggest that work hours and ill-health are positively associated. However, the empir-ical applications on this relationship are scarce. The fourth essay in this thesis evaluates this relationship in the context of the elderly workforce.
Analysing the health or well-being effects of work are difficult because of the asymmetry of information. Individuals have heterogeneous preferences towards work and choose occupation and work hours accordingly. Hours spent at work affect health differently in different occupations. Also there is reverse causality: health affects the work hours decision. To estimate the causal relationship, either an experiment or a quasi-experiment is required. In the next subsections I will discuss the literature on the theoretical modelling of health and estimating the causal effects in this context.
1.3.1 Health stock model
The human capital model assumes that the investments in knowledge raise productivity and this leads to increases in the monetary returns in the labour markets. For this reason, an individual has an incentive to allocate time and resources into education, either in the form of formal schooling or in
on-the-job training. If the expected returns are higher than the direct costs and the opportunity cost of time spent studying, the individual invests in acquiring education. The optimal quantity of investment varies in different phases of life and between individuals of the same age but with varying abilities. A similar kind of modelling can be extended to health.
We can think of "good health" as a commodity which is demanded. Each individual has an initial stock of health when they are born. The health stock produces healthy time which increases well-being. However, this health stock is depreciating with age but an individual can make investments which in-crease the stock. Investments are not costless and their (shadow) price is as-sumed to be increasing with age. This leads to different amounts of health (and medical care thereof) demanded at different points of the life-cycle. The shadow price for health is affected by many factors, like the price of medical care and the cost of exercise. (Grossman, 1972).
In this modelling framework initiated by Grossman (1972) health is de-manded because it is on one hand a direct source of utility but also an in-vestment commodity which makes more healthy time available for market and non-market activities. Unlike in the human capital investments, the in-creases in the stock of health does not only affect the productivity and the wage rates but health also affects the time constraint one can spend in market or household production. This is an important difference between the health stock model and the human capital formation. As an example, an individual working as an economist is usually considered a high-productive worker who has invested a lot in the human capital formation. However, the initial health stock and the depreciation rate of this stock affect how well those returns can be collected in the form of wages. The economist’s number of projects con-ducted are dependent on this health constraint.
The work itself can also affect the stock of health and the depreciation rate.
While this latter point is more discussed for example in the field of sociol-ogy, economists have just recently started to explore the effects of work on health (see chapter 5 and references therein). In the Grossman model the to-tal amount of time in any period is divided between time at work (market production), time at leisure, time producing health and time lost from mar-ket and non-marmar-ket activities due to sickness. If the time producing health
increases, for example because work hours are reduced, keeping other inputs like income fixed, the model predicts improving health. However, one can also argue that reduced working hours can increase leisure activities which deteriorate health. To my knowledge, this type of mechanism has not been incorporated in the Grossman model. Finally, this is an empirical question.
1.3.2 Causal inference between work hours and health
It is particularly challenging to identify causal effects between work charac-teristics and health. Work characcharac-teristics are endogenous to the underlying health stock. In the example of work hours’ effect on health, there is reverse causality when health affects the decision of working hours. Experiments or quasi-experiments are required in order to reveal the causal effects between work and health.
The methods for policy evaluation are reviewed in Abadie and Cattaneo (2018). In the ex-post programme evaluation the goal is to find exogenous variation. If a randomized experiment is run, exogenous variation comes di-rectly from the setting and treatment effect comes from the comparison of the treatment and control groups. Unfortunately, running randomized ex-periments is seldom available and sources of exogenous variation need to be looked at from the social environment.
One popular method in programme evaluation is using a difference-in-differences approach. To use this method we need to find a treatment group, whose work hours changed exogenously, and a control group, who do not experience any shock in their working hours. For example such a setting took place in France in the 1990s when statutory weekly work hours were reduced by employer characteristics (Bietenbeck and Berniell, 2017).
An important question in the difference-in-differences setting is to what extent the control group is credible17. Naturally, we want to compare in-dividuals who are as similar as possible to the treatment group. Otherwise we do not have a counterfactual with which to compare the outcomes of the treatment group. Beside difference-in-differences approach, policy evaluation
17A necessary condition for difference-in-differences method to work is the parallel trends assumption before the treatment. However, this is not a sufficient condition.
have used for example regression discontinuity design and instrumental vari-ables successfully.
Instrumental variables (IV) have been widely used both in randomized and quasi-experiment evaluation when there is imperfect compliance. The idea behind instrumental variables is to use the variation in the explanatory variable that is caused by the instrument. That is, in the first-stage the ex-planatory variable is regressed against the instrument. If the instrument is associated with the explanatory variable and does not affect the outcome vari-able of interest directly (exclusion restriction assumption), in the second stage one utilizes the first-stage’s predicted values of the explanatory variable with the controls to get a causal estimate for the explanatory variable.
The commonly used instruments in studies exploring the effect of retire-ment or reduced working hours include the statutory work-week regulation (Cygan-Rehm and Wunder, 2018; Ahn, 2016) or statutory retirement ages (Kantarci, 2016). Recently, in the health economics literature, the instrumen-tal variables have been accompanied with fixed effects if panel data is avail-able (Cygan-Rehm and Wunder, 2018; Ahn, 2016). Fixed-effects instrumen-tal variable estimation has the advantage that it removes the time-invariant unobserved variables while taking into consideration the endogeneity of the explanatory variable.
The aim of the programme evaluation is to give some recommendations to the policy makers. The object of interest in experiments is often the aver-age treatment effect (ATE). ATE represents how the averaver-age outcomes differ in the whole population if we could move everybody from the inactive to active treatment. A some-what more limited effect is the average treatment effect for the treated (ATET). (Abadie and Cattaneo, 2018). In the difference-in-differences setting the effects are average treatment effects for the treated if all study units take-up the treatment. This means that with the example of work hours reduction policies, all individuals need to actually reduce hours worked. However, it is seldom that the effects can be interpreted this way.
Many treatments often have only partial compliance to the assignment and so the effects are intention-to-treat (ITT) effects. For some policy evaluation this is enough because the policy maker is interested in knowing whether provid-ing an access to a certain policy is beneficial.
Especially with IV estimates, the effects often lack the external validity and are instead interpreted as local average treatment effects (LATE) (Angrist et al., 1996). LATEs mean the effect on the sub-population of compliers who change their behaviour because of the instrument or the average effect for those who would always be in compliance regardless of the treatment. So these effects cannot be extrapolated to other parts of the population. The re-searchers should carefully consider what sort of effects they are observing and what kind of policy advice they can draw from those effects.