بیماری و بازنشستگی در بریتانیا: یک تجزیه و تحلیل مبتنی بر داده های پانل

We examine the effect of ill health on retirement decisions in Britain, using the British Household Panel Survey (1991–1998). As self-reported health status is likely to be endogenous to the retirement decision, we instrument self-reported health by a constructed ‘health stock’ measure using a set of health indicator variables and personal characteristics, as suggested by Bound et al. (Bound, J., Schoenbaum, M., Stinebrickner, T.M., Waidmann, M., 1999. The dynamic effects of health on the labor force transitions of older workers. Labour Economics 6, 179–202). Using a range of econometric techniques, we show that adverse shocks to individual health stocks predict individual retirement behaviour among workers aged from 50 until state pension age. We compare responses of economic activity to constructed health measures with that arising using direct indicators of functional limitations and specific health problems. We also examine the dynamics of health shocks and whether adverse and positive health shocks have symmetric effects on transitions in and out of economic activity.

The number of people on disability benefits more than doubled between the late 1970s and the end of the 1990s in the UK, although thereafter the number stabilised. Throughout this period around half of claimants were aged between 50 and the state pension age. The substantial growth in claimants aged 50 and over after 1980 is shown in Fig. 1. ‘Ill health’ is a major reason for retirement among British men, especially for men without access to an occupational pension (Tanner, 1998, Table 7) – indeed ‘own ill health’ is the most commonly cited reason for retirement among both men and women in the early 1990s (Disney et al., 1997, Table 2.19). Similar trends have been observed in other countries such as The Netherlands and the US (Bound and Burkhauser, 1999).At any point in time there is a strong correlation between observing a person not working and their self-reported overall poor health status, but this may give a misleading impression of the impact of health state on retirement. First, individuals who are inactive often have an incentive, for self-esteem if nothing else, to report worse-than-actual health. Second, differences in reported self-assessed health are large, even for individuals in identical labour market states – individual heterogeneity is important. Third, individuals with permanent and very poor health may never have worked, so they cannot be observed ‘retiring’. Fourth, ill health may impact on other labour market attributes of the worker (for example, the wage they earn – see Meghir and Whitehouse, 1997) which implies that there are both income and substitution effects on labour supply arising from shocks to the worker's health status. Finally, the health stock may be endogenous to the labour market state of the individual ( Kerkhofs et al., 1999). The potential measurement error and endogeneity of self-reported health status has led some economists to reject the use of such general measures in retirement models completely (such as Myers, 1982) even though they have been, and continue to be, commonly used in this field for want of better measures.1 A further problem, in the UK at least, is that for those individuals with no private pension rights, disability benefits are the only ‘route’ into early retirement through the social security programme since the social security pension cannot be received before the state pension age (currently 65 for men and 60 for women). Consequently, there is an inducement for early retirees to utilise the ill health route and their self-assessed health status will correlate with preferences for early retirement (Blundell and Johnson, 1998).2 For the researcher interested in the link between ill health and retirement, one obvious strategy is to substitute more objective measures of ill health (if available) for self-reported health status in the model ‘explaining’ retirement.3 Some studies have argued for the intrinsic superiority of this approach, since it eliminates the errors-in-variables and biases arising from the subjective health measure (many such studies are cited in Quinn et al., 1990). But as Bound (1991) points out, we cannot be sure that such proxies are any better predictors of (in)activity than self-reported health status, as the researcher thereby assumes some link between work status and these other health measures.4 Such a strategy does not eliminate the errors-in-variables problem but replaces it with a similar problem on a proxy variable, and may thereby lose any additional information on the ‘true’ association between health and behaviour that might be intrinsic to the self-reported ‘subjective’ measure.5 Another pertinent suggestion, explored by Anderson et al. (1986) and Bound et al., 1998 and Bound et al., 1999 (hereafter Bound et al.) is that changes in labour market status e.g. ‘retirement’ (whether permanent or temporary) should be associated with ‘shocks’ to the individual's underlying ‘health stock’. Bound et al.'s strategy is to construct a latent health stock or index of health for each individual as a function of personal characteristics and health indicators. This constructed variable is used to instrument self-reported health in a panel data model of economic activity in order to explore the relationship between time variations in health and changes in work status (see also Stern, 1989). 6 Modelling health ‘shocks’, it can be argued, eliminates any person-specific association between characteristics and labour market outcomes (such as fixed preferences for work, or longstanding disability), whilst proxying self-reported health status by time-varying health and personal characteristics should ameliorate any reporting bias in the former. This paper follows the general strategy suggested by these authors. It exploits the panel element of the data set to construct individual ‘health stocks’, and uses time variation in these ‘stocks’ as an explanatory variable in reduced form models of labour market (in)activity amongst a sample of older people in Britain. Two econometric approaches are used. In the first, linear and non-linear fixed effects estimators are used to examine the impact of changes in health and lagged individual health on the economic activity rates of individuals aged between 50 and state pension age. The subjective health question invites the respondent to compare their health to that of people of a similar age, so this approach pins down the association between transitions in and out of paid work and current variations in relative health and allows for respondent heterogeneity. The second approach uses the same data and constructed variables but estimates a hazard function with non-parametric duration dependence. Relative to the fixed effects model, such a specification has a more intuitive interpretation as a retirement model and permits greater flexibility both in examining the dynamic impact of health status on retirement and in examining whether the impact of health on movements in and out of the labour market differs according to current work status. These advantages come at the cost of imposing restrictions on the structure of individual heterogeneity, which could be important in this context (see Section 3.2). Whichever approach is used, we find robust evidence that individual health deteriorations lead to a greater likelihood of transitions into economic inactivity later in the working life. Several questions arise from this finding. First, does this two-step procedure do better than, for example, simply augmenting the reduced form retirement equation with individual indicators of ‘objective’ health limitations and difficulties? We therefore also show that the predictive power of (linear) combinations of objective health limitations in explaining ill health retirement is limited relative to our chosen method of constructing an individual time-varying health stock. Second, both current and lagged changes to the health stock affect current transitions in and out of economic activity. How do we interpret these findings, which are somewhat different from the structure of current and lagged parameter estimates found by Bound et al. (1999)? Third, are relative individual health improvements associated with a greater likelihood of reverse transitions out of economic inactivity? The fixed effect models do not differentiate between types of state transitions, although symmetry of behavioural responses to positive and negative ‘shocks’ to health can be tested. In contrast, in the baseline hazard specifications, we utilise the last reported exit from economic activity (if observed) during the period as the indicator of ‘retirement’. To examine symmetry of activity responses, we then utilise all transitions between economic activity and inactivity, and examine hazards of entry and exit to economic activity separately. We show some evidence of symmetric response: that is, the probabilities of individual transitions from economic activity to inactivity are strongly associated with deteriorations in relative health status and that the reverse transition is observed with relative improvements in health status. We do not, in this paper, explore the ‘feedback’ of labour market activity or inactivity on the evolving health stock. In summary, our range of methods and tests suggests that this method of modelling ill health and economic (in)activity is reasonably robust. We also attempted to test whether the UK's 1995 disability benefit reform, which tightened some eligibility conditions and cut real disability benefits, had any impact on the link between work-related disability and economic inactivity. We find no evidence of any impact in the data. This may of course be the ‘true’ answer,7 but it may simply be that our ‘test’ is inappropriate in part because we only observe receipt and not applications for disability benefits. These results are not discussed further here (see Disney et al., 2003 for further discussion of this issue). The structure of the paper is as follows. Section 2 describes the construction of our health stock variable. Section 3 uses this variable as an instrument for health status in reduced form labour market models. Section 4 describes our interpretation of the coefficients on lagged health, and sensitivity analyses to asymmetriesin the health-economic activity relationship and to alternative health measures. Section 5 concludes.

This paper represents one of the first attempts to examine the impact of ill health on retirement in some detail in the UK, using the British Household Panel Survey. The focus of the paper is on the nature of health measures that are utilised in reduced form retirement models of the type described here. It argues that reporting bias is intrinsic to self-reported measures of general health (especially, in questions that explicitly link health to economic activity status), but that there is a lack of ‘fit’ between objective measures of disability and functional limitations on the one hand and health as it relates to economic activity on the other. We therefore follow the approach of Bound et al. (1999) in constructing an underlying ‘health stock’ of the individual, and in treating temporal variations in this measure as proxying individual-specific ‘health shocks’ that affect retirement behaviour, using a fixed effect specification for the labour market model. We also input our health stock measure into a hazard specification of the retirement model that has a relatively flexible dynamic structure and captures duration dependence non-parametrically. The paper shows that a constructed proxy variable of this type has a powerful explanatory effect for transitions between economic activity and inactivity in a reduced form model that incorporates other time-varying covariates. Moreover, the approach seems superior, in terms of explanatory power, to the application of linear reduced form models incorporating disability index-type health measures. The results show that lagged health as well as current health shocks affect decisions concerning economic activity. An analysis of health dynamics provides some intuition as to why this might be the case. Finally, the paper tested for symmetry in labour market transitions to changes in health. Little evidence of asymmetry was found. Richer models of the link between retirement and ill health are being developed, but we believe that the approach used here is capable of generating important insights into that link.