Unemployment and psychosocial adjustment in young adults: causation or selection?

Abstract

This study investigates the relationships between unemployment following school leaving and psychosocial adjustment problems (mental health, substance use, crime, suicidal behaviours and teenage pregnancy) in a birth cohort of over 1000 New Zealand born young people. The data were gathered during the course of the Christchurch Health and Development Study (CHDS). The CHDS is a longitudinal study of a birth cohort of 1265 children born in the Christchurch (NZ) urban region who have been studied from birth to age 21. Data were gathered by personal interview on: (a) exposure to unemployment and (b) personal adjustment over the period from age 16 to age 21. Measures of personal adjustment included mental health (depression, anxiety), substance use, crime, suicidal behaviours and (for females) teenage pregnancy. Data were analysed using a fixed effects regression model that took into account both observed and non-observed sources of confounding and the possibility of reverse causal associations between personal adjustment and unemployment. Before adjustment for confounding and reverse causality there were significant (p<0.001) associations between exposure to unemployment and measures of mental health, substance use, crime, suicidal behaviours and teenage pregnancy. Adjustment for confounding factors and reverse causality reduced these associations quite substantially and after control for sources of confounding a number of associations became non-significant. Nonetheless, after such control, exposure to unemployment remained significantly (p<0.05) associated with suicidal ideation, substance abuse and criminal behaviours. It is concluded that, in part, the associations between unemployment and personal adjustment are spurious and reflect the presence of confounding factors that are related to both unemployment and adjustment. Nonetheless, the findings suggest that exposure to unemployment may be associated with increased risks of suicidal thoughts, crime and substance use.

Introduction

In recent years there has been increasing research into, and concern about, a range of adolescent mental health and adjustment difficulties that include mental health problems; alcohol and illicit drug abuse; crime and suicidal behaviours. Concern about rising rates of these disorders in many developed societies has led to growing interest and speculation about the aetiological factors that may encourage the development of these conditions (see for example Farrington, 1995; Loeber, 1990; Rutter, 1995; Rutter & Smith, 1995).

One factor to which considerable attention has been paid is youth unemployment. Interest in this area can be traced back to research conducted in the 1930s during the Great Depression. As early as 1938, Eisenberg and Lazarsfeld (1938) were able to review the results of over 100 studies examining links between unemployment and personal adjustment. Their conclusion from this review of evidence was that “unemployment tends to make people more emotionally unstable than they were previous to unemployment” (p. 359).

Since this time there has been continued interest in and debate about this issue. Attention has increased in the last two decades as processes of economic restructuring that have led to rising youth unemployment rates in many societies. As a result of these concerns, a growing number of contemporary studies have reconsidered the extent to which exposure to unemployment increases rates of psychosocial disorders in young people. Such studies have focused on mental health problems (Fergusson, Horwood, & Lynskey, 1997a; Hammer, 1993; Kessler, Turner, & House, 1989; Peck & Plant, 1986; Tiggemann & Winefield, 1984; Winefield, Winefield, Tiggemann, & Goldney, 1991), crime (Farrington, Gallagher, Morley, St. Ledger, & West, 1986; Fergusson, Lynskey, & Horwood, 1997b; Fleisher, 1963; Freeman, 1983), substance use and abuse (Hammer, 1992; Janlert & Hammarstrom, 1992), suicidal behaviours (Iversen, Andersen, Andersen, Christoffersen, & Keiding (1996), Beautrais, Joyce, & Mulder (1998); Iversen, Andersen, Andersen, Christoffersen, & Keiding, 1987; Jones, Forster, & Hassanyeh, 1991; Morrell, Taylor, Quine, & Kerr, 1993; Platt & Kreitman, 1984) and other similar adverse outcomes. In general, these studies have reported associations between exposure to unemployment and psychosocial adjustment problems, suggesting a possible causal link between unemployment and personal adjustment in young people.

Despite this evidence, the issue of whether exposure to unemployment is a causal factor in the development of psychosocial disorder remains contentious. The major threat to the validity of causal inferences linking unemployment and personal adjustment comes from the possibility of selection effects (Hammarstrom, 1994; Morrell, Taylor, & Kerr, 1998; Schaufeli, 1997) that arise because the selective processes by which young people become unemployed may also be related to (or involved in) the processes by which young people develop psychosocial disorders. Selection effects may operate in at least two ways to produce a misleading impression that unemployment causally influences psychosocial adjustment.

1. Confounding: First, it may be argued that the associations between unemployment and personal adjustment are partially or wholly spurious and arise because young people who are prone to adjustment problems prior to work force entry show greater rates of unemployment and are also more prone to develop psychosocial disorders. This explanation implies that the association between unemployment and psychosocial adjustment arises because both outcomes are related by virtue of common or correlated risk factors and causal processes. There is considerable indirect prior evidence to support this hypothesis; extensive literature exists suggesting overlap between the risk factors and life processes that are associated with youth unemployment and the risk factors and life processes associated with the development of crime, mental health problems, substance use and suicidal behaviours (Caspi, Wright, Moffitt, & Silva, 1998, Farrington et al., 1990; Fleming & Offord, 1990; Hawkins, Catalano, & Miller, 1992; Loeber, 1990; Winefield, Tiggemann, Winefield, & Goldney, 1993).

2. Reverse causality: A further explanation of links between unemployment and personal adjustment in young people is that these associations may reflect a reverse causal association in which the development of adjustment problems in young people increases their exposure to unemployment rather than unemployment leading to the onset of disorder (Hammarstrom, 1994; Morrell et al., 1998; Schaufeli, 1997).

A growing number of studies have examined causal links between unemployment and personal adjustment by adjusting the associations using regression models in which the associations between unemployment and rates of disorders are corrected for observed confounding factors that are present prior to the individual's exposure to unemployment (Banks & Jackson, 1982; Iversen, Andersen, Andersen, Christoffersen, & Keiding (1996), Beautrais, Joyce, & Mulder (1998); Farrington et al., 1986; Fergusson, Horwood, & Lynskey (1997a), Fergusson, Lynskey, & Horwood (1997b); Graetz, 1993; Kessler et al., 1989; Peck & Plant, 1986). The adjusted estimates from these studies are then treated as giving, at least preliminary estimates of the possible causal contributions of unemployment to the personal adjustment of young people. All of these studies share the common limitation that associations between unemployment and psychosocial outcomes have been adjusted for known and observed covariate factors. However, it may be suggested that failure to control unobserved sources of confounding could lead to a mis-estimation of the contributions of unemployment to risks of psychosocial disorders in young people. In this paper, we develop an approach to addressing this issue through the use of longitudinal data and fixed effects regression models. The statistical background to the proposed analysis is developed below.

An approach to the control of observed and non-observed confounding factors that is being used increasingly in social research designs that have repeated measures or longitudinal data is the fixed effects regression model (see for example Duncan, Yeung, Brooks-Gunn, & Smith, 1998; England, Farkas, Kilbourne, & Dou, 1988; Hersch & Stratton, 1997). The basic logic of this approach is described below.

Consider a longitudinal study in which a sample of N subjects has been observed on T occasions on measures of some outcome variable Y (e.g. crime) and an exposure variable X (e.g. unemployment). It is assumed that Y and X are related by the model

$$\text{Y}{}_{\mathrm{it}}\text{=B}{}_0\text{+B}{}_1\text{X}{}_{\mathrm{it}}\text{+u}{}_{\mathrm{i}}\text{+e}{}_{\mathrm{it}}\text{,}$$

where Y_it, X_it are the outcome and exposure measures for the ith subject observed at time t, u_i denotes non-observed sources of fixed variation that influence the outcome of the ith subject and e_it is a random error term. The fixed term u_i summarises all non-observed fixed sources of variation that influence the score Y_it. These sources of variation will include all non-observed fixed confounding factors that influence Y_it. Thence, to estimate B₁ in a way that takes into account these sources of confounding, it is necessary to estimate B₁ in a way that takes into account any covariation between u_i and X_it. It can be shown (Greene, 1993; Hsiao, 1986) that an estimate of B₁ corrected for the covariation between X_it and u_i can be obtained by fitting the model

$$\text{(Y}{}_{\mathrm{it}}\text{−y}{}_{\mathrm{i}}\text{)=B}{}_1\text{(X}{}_{\mathrm{it}}\text{−x}{}_{\mathrm{i}}\text{)+(e}{}_{\mathrm{it}}\text{−e}{}_{\mathrm{i}}\text{),}$$

where y_i is the mean score of the ith subject on variable Y observed over the T times, x_i is the corresponding mean score on variable X and e_i is the mean of the error terms e_it over the T times. By expressing the model in this form it becomes possible to estimate B₁ in a way that takes account of all fixed sources of confounding (see Greene (1993) and Hsiao(1986) for a more detailed description of the fixed effects model).

Although the fixed effects model above has been widely used in econometric applications with normally distributed data, frequently the data gathered in epidemiological studies do not have the distributional properties suitable for this approach. For example, in many applications, outcome variables may be dichotomous or highly skewed count data. In such circumstances there are generalisations of the fixed effects model for both dichotomous and count data (Hamerle & Ronning, 1995).

It is important to recognise that although the fixed effects model takes into account non-observed sources of fixed confounding, it fails to take into account confounding factors that may vary with time. For this reason, it is often useful to extend the fixed effects model to include observed time-dependent covariates.

Although the fixed effects regression models described above provide a powerful technique for controlling both observed and non-observed fixed sources of confounding, this approach does not address the issue of reverse causality in which the onset of disorder increases risks of unemployment. One means of addressing the issue of reverse causality is to extend the fixed effects models developed above to include lagged time dynamic variables representing the individual's adjustment prior to the onset of unemployment. For example, in the case where the outcome is dichotomous the model is

$$\text{Logit}\text{Pr(}\text{Y}{}_{\mathrm{it}}\text{=1)=α}{}_{\mathrm{i}}\text{+B}{}_1\text{X}{}_{\mathrm{it}}\text{+B}{}_2\text{Z}{}_{\mathrm{it-1}}\text{,}$$

where Z_it−1 is a measure of the individual's adjustment observed in the interval prior to time t and to the assessment of the unemployment measure X_it and the outcome variable Y_it and α_i is a term reflecting fixed sources of non-observed variation influencing the response of the ith subject. In effect this model corrects the estimate of B₁ for: (a) all observed or non-observed fixed covariates that may confound the association between X_it and Y_it and (b) the individual's level of adjustment prior to the observation of X_it and Y_it. It does not, however, take into account the possibility of a simultaneous (reciprocal) relation between Y_it and X_it owing to the fact that it is not technically possible to develop models of simultaneous influence for the logistic regression model (Fienberg, 1980). For this reason the control of reverse causality offered by this approach is less comprehensive than the control of fixed confounding factors.

In the present study we use data gathered over the course of a 21-year longitudinal study of a birth cohort of 1265 New Zealand born children to study the links between exposure to unemployment and the development of a series of psychosocial disorders including: crime; substance abuse; juvenile offending and suicidal behaviours. This study is distinguished by the collection of longitudinal data on both exposure to unemployment and data on a range of psychosocial disorders over the period from age 16 (the minimum school leaving age in New Zealand) to age 21. The analyses address a series of questions for each of these outcomes:

1. Association: To what extent is variation in exposure to unemployment associated with increased risks of the outcome?

2. Control of confounding: Is exposure to unemployment associated with the outcome when due allowance is made for fixed confounding factors? Here, we will contrast two approaches to adjusting confounding. The first approach is via the fixed effects logistic regression model described above. The second is through a more conventional regression approach in which we adjust the association between unemployment and psychosocial outcomes for covariate factors.

3. Reverse causality: Does the association between unemployment and psychosocial adjustment persist after statistical control for: (a) fixed effects and (b) the subject's adjustment prior to the assessment of unemployment and the outcome?

4. Population heterogeneity: Do the associations between unemployment and rates of psychosocial outcomes vary with population features including gender, race and family socio-economic status?