4.7  Potential problems

In addition to the usual problems that arise from missing observations, measurement error[17] and misspecification of the model, the models applied in this study are subject to the question of endogeneity. This arises when the assumption that the right-hand side variables are exogenous is violated. Exogenous variables are independent of the values of the dependent variable. For example, if health status is a variable to be used in a model to explain labour force participation, it is assumed that the values adopted by the health status variable are independent of labour force participation. The problem of endogeneity can arise for a number of reasons:[18]

1. While health status may in fact influence a person’s decision to participate in the labour force, it may well be that their participation affects their health status; in short, health status and labour force participation are simultaneously determined and health status is no longer a truly exogenous variable.[19]
2. However comprehensive a set of survey data, there will be many variables that simply remain unobserved. Examples could include a person’s attitude toward risk, or their preference for current over future consumption. To the extent that any one of a host of unobservable characteristics influences both the measured health status and the labour force participation, the problem of endogeneity is again present.
3. It is possible that some people not in the labour force might be inclined to report that their health status is poor as a way of rationalising their lack of participation both to themselves and to the interviewer. Once again, health status as observed is being influenced by the labour force participation and cannot be regarded as exogenous.

The presence of endogeneity may lead to the estimates of the coefficients in equation (4) to be biased from their true values, although typically the direction of the bias is uncertain. In addition the estimates may not be consistent; ie, their values may not necessarily converge to the unknown population value as the sample expands.

While the existence of these problems and their implications is well known, it is much less clear as to the magnitude of the distortions, and even less certain about whether there are approaches to successfully mitigate the effects. Ideally one would address the problem of simultaneity by purging the health variable of the effect of the influence of current labour force participation. This leads to the simultaneous estimation of equations for labour force participation and for health status. What is required, however, is a variable that influences the health status but not the labour force participation decision. Such instruments are not easy to identify. For example, drinking or smoking might be argued to affect health but not labour force participation. As a consequence they would be included in a supplementary equation for health status but excluded from the labour force participation equation, the principal equation of interest.[20] In a comprehensive survey of studies of health and labour force participation, Currie and Madrian (1999) conclude:

…estimates of the relationship between health and labour force outcomes vary widely and are sensitive to the identification assumptions employed. Many of the studies discussed above either ignore endogeneity issues altogether or rely on exclusion restrictions that are not easy to justify. (p. 3352)

In this study we tested a range of potential instrumental variables (including ethnicity and smoking) but concluded that none of the results obtained were satisfactory. We have therefore relied on a single equation approach. In the following sections we analyse the factors associated with key variables of interest with particular emphasis on the role of health. In each case we fit a regression model: an OLS in the case of continuous variables and a logit model where the dependent variable is binary.

Finally, a note of caution: whenever statistically significant associations are identified, these should not necessarily be taken to imply causation.