Disability and Labour Fource Participaton in Ireland 1995 - 2000

Disability and Labour Force Participation in Ireland 1995-2000 Brenda Gannon* Economic and Social Research Institute and Department of Economics, National University of Ireland, Maynooth Summary This paper aims to analyse the effect of disability on participation in the labour force, using the Irish component of the European Community Household Panel Survey 1995-2000. A range of panel models are considered, but to allow for any unobserved influences or state dependence in labour force participation, our preferred model is a dynamic panel model. We show how the estimates of current disability are changed once we control for the effect of past disability and previous participation. We compare base estimates of disability with those controlling for unobserved heterogeneity and past participation. The results suggest that the base effect of disability is overestimated by between 40-60 per cent for men and by 5-10 per cent for women. Keywords Disability; labour force participation; static and dynamic panel models 6 October 2004 *Correspondence to: Brenda Gannon, Economic and Social Research Institute, 4 Burlington Road, Dublin 4, Ireland. Email: [email protected] WORK IN PROGRESS: This paper is not for publication and should not be quoted without prior permission from the author. 1 I. Introduction People with disabilities face many barriers to full participation in society, not least in the labour market, and the extent and nature of participation in the labour market has a multitude of direct and indirect effects on their living standards and quality of life. In studying the effect of disability on labour force participation, we are faced with a variety of analytical challenges, such as the effect of unobserved characteristics of disabled individuals and the effect of their past participation in the labour market. This paper uses panel data methods to control for these factors and we estimate the impact of disability on participation, controlling for unobserved heterogeneity and past participation. Internationally, the first generation of econometric studies on the effect of disability on labour force participation emerged around the late 1970’s. Bartel and Taubmann [1] estimate an OLS model of weekly hours worked to analyse the effect of health on earnings and labour supply, whereas Chirokos and Nestel [2], estimate a Tobit model relating annual hours worked to health history by looking at the degree of poor, good, improved or deteriorating health over the previous ten years. More recent research emphasises the importance of the way health and limitations are captured, with the type of health status variable used leading to different patterns in terms of labour force participation. Wolfe and Hill [3], for example, measure health status using an index of limitation in daily activities, Madden and Walker [4] measure health in terms of those who report a longstanding illness or disability, while Mete and Schultz [5] also measure health status using a health index. Using Labour Force Survey data, Kidd, Sloane and Ferko [6] analyse the effect of health limitations on the kind of paid work possible in the UK. They confirm the presence of substantial wage and participation rate differences between disabled and non-disabled individuals. The focus of previous policy for disabled people has been on the provision of services, whereas more recently, there is a campaign for civil rights and the provision of legislation for equality and full participation. Employers and policy makers are therefore interested in whether or not disability has an effect on participation. In this paper, we aim to determine whether it is disability that determines the participation decision, or if there may be some other unobservable characteristics involved that distort the disability effects. 2 Previous studies analysing unobserved individual effects in this context emerged in the mid-eighties. Sickles and Taubman [7] were one of the first to use longitudinal data in estimating retirement decisions, and allow for unobserved heterogeneity in the retirement function. Estimating a binary random effects probit model, they allow for unobserved affects by simultaneously estimating the health and retirement equation, allowing the errors to be correlated. They allow for correlation of the unobserved effect with the disability variable, but they do not include the effect of labour market history. Their findings show that moving from poor health to good health decreases the probability of retirement, but they do not show how the health effect changes as a result of allowing for unobserved effects. Bound [8] looks at a retirement equation in the cross sectional context, and shows that if the errors in the health and retirement equation are correlated, then there is an upward bias in the effect of health. He aims to identify the effects of financial incentives on reporting behaviour and retirement decisions, and investigates if objective measures may be used as a proxy for subjective measures of health. The author concludes that the self-reported measure is not reliable in estimating the effect of health on retirement. Kreider [9] also analyses work participation with cross section methodology and arrives at the same conclusion. He finds that when the true measure of disability is used, the effect on participation is lower, by 17.2% for men and 24.9% for women. Both Bound [8] and Kreider [9] use cross section data to estimate the effect of the true effect of disability on participation, but identification of their models requires a variable, that affects health but that is not correlated with participation. Kidd, Sloane and Ferko [6] apply a decomposition approach to cross section data and find that approximately 50 per cent of the difference in participation rates is due to unexplained effects. Our data offer the possibility of analysing the relationship between disability and labour force participation over a significant period rather than just at a point in time, and allow us to use panel data techniques in our estimation. Using panel data, we capture the effects of variables that are particular to an individual and are constant over time. We can control for these unobservables by using a panel model and we therefore do not need to include an identifying variable. Labour force participation may also be influenced by past participation, where non-participants in the previous 3 year may be less likely to participate in the current year. Although this may be true for all individuals, it may also be a specific characteristic of disabled people and lead to an incorrect interpretation of the disability effect. It may be that disability reduces the probability of previous participation, and therefore indirectly influences current participation. Using panel data, we can incorporate this state dependence effect and re-estimate the effect of disability on participation. It may also be that individuals report a disability as an ex-post justification for not being in work in the previous year. Again, we would expect the effect of disability to be misinterpreted, and can use the results of the dynamic model to disentangle the unobserved heterogeneity and past participation effects. More recently, Lindeboom and Kerkhofs [10] also include the effect of past labour market outcomes on current health in their retirement model. They find that for elderly people, working in the previous period only slightly decreases the value of health. They estimate a multinomial logit model, to facilitate the three different labour market states compared to working, available to individuals nearing retirement age in the Netherlands. Although they only have two waves of panel data, by using information on previous labour market history, they specify an equation for initial participation and estimate the probability of working initially. This is included into the overall likelihood function from which unobserved effects are integrated out. They find that the effects of health are exaggerated for elderly people in a simple multinomial model, compared to their preferred model. In this paper, we follow a different approach to Lindeboom and Kerkhofs [10] mainly because we use six waves of panel data and can therefore identify the effect of past participation within a less complicated model. The main focus in this paper is to model two labour market outcomes – participation and non-participation – and hence we concentrate on a binary response variable. In contrast to Lindeboom and Kerkhofs [10], we follow an approach by Wooldridge [11] that allows us to avoid specifying a distribution for the initial participation. The likelihood function from our approach is easier to estimate and serves the same purpose in terms of looking at the effect of unobserved heterogeneity. Our findings using Irish data are similar to those of Lindeboom and Kerkhofs [10] among others, in that their reported disability variable over estimates the impact of disability on participation in the Netherlands. In addition, 4 we show exactly how much unobserved heterogeneity contributes to variation in participation and how this changes the effect of disability. Finally, we show the effect of past disability (via it’s effect on previous participation), on current labour force participation. II. Theoretical Framework We model participation firstly as a static process with current participation, and secondly as a dynamic process. In our dynamic model, we account for the fact that the choice between consumption and leisure is considered as a lifetime decision, so we assume that individuals maximise their expected utility over their lifetime. Following Bound et al [12], in general, the participation equation is based on the assumption that individuals maximise a utility function given by: ),,(max jjj tj T tj tZLCUE     [1] where Cj and Lj are consumption and leisure in period j respectively. Zj is a vector of taste shifters and includes disability. The utility function is maximised subject to an intertemporal budget constraint: jjjjjj ArCHWA )1()( 11     [2] where Wj is the wage, Hj is hours of work, Aj represents assets, and rj is the rate of interest. In this paper, our empirical model shows how individuals compare the utility between two states – participation and non-participation. Solving this model provides an expression for optimal leisure as a function of W, H, Aj and Zj. Much of the literature on the effect of health on labour force behaviour has treated health as an exogenous taste shifter. We take this approach, and hence do not specify a health production function. In this context, we obtain estimates from a reduced form modela and concentrate on how the disability effect changes once we allow for unobserved individual effects and state dependence in labour force participation. 5 III. Data The data on disability and labour force participation in Ireland are from the Living in Ireland Survey 1995-2000.b The Living in Ireland Survey is the Irish component of the European Community Household Panel, conducted by the ESRI for Eurostat. We wish to focus on individuals of working age, hence we exclude those aged 65 and over. In the Living in Ireland Survey, detailed information on current labour force status was obtained. For current purposes this allows us to distinguish between those who were at work, or unemployed but seeking work – who we will count as active in the labour force – and all others, whom we will count as inactive. The percentage of those unemployed but seeking work is quite low ranging from 7.5% in 1995 to 2.8% in 2000, giving a panel average of 5.1%. For this reason, we do not include them as a separate category in our dependent variable. A measure of disability can also be constructed from the Living in Ireland survey on the basis of individual responses to the following question: “Do you have any chronic, physical or mental health problem, illness or disability?” It may well be, that not only the presence of such an illness or disability but also the extent to which it limits or restricts a person, is important. To capture this, we use responses to a follow-up question concerning the impact of the disability to distinguish a) those reporting a chronic illness or disability and saying that it limits them severely in their daily activities b) those who report a chronic illness or disability and saying it limits them to some extent, and c) those who report such a condition but say it does not limit them at all in their daily activities. The extent to which respondents say they are limited relates to their daily activities rather than work, but similar measures have been shown to have significant discriminatory power in terms of labour force participation in research elsewhere (e.g. Malo [13]). Furthermore, as Table 1 shows there are different rates of participation for 6 each sub-group, so it is important that we distinguish between the different levels of disability, in our analysis of labour force participation. The effects of disability on labour force participation may differ among individuals, depending on other characteristics, for example age or education. Since disability may be correlated with other variables, we include measures of age, education, region, unearned income, age of youngest child and marital status. These variables are defined in detail in Table 2 and summary statistics are provided in Table 3. The youngest individuals in this sample are aged 16 and the number of observations of males and females are 7,188 and 7,670 respectively. IV. Panel models and Results Static Pooled Probit Model: Using the Living in Ireland Survey 1995-2000, we estimate a range of panel models to capture the effect of disability on participation. We exclude 1994 because the questions regarding health problems and limitations differed from 1995 and subsequent years. Firstly, we estimate a static pooled model, assuming that the errors are independent over time and uncorrelated with the explanatory variables. This model provides us with base estimates, with which we can compare results from models that incorporate unobserved heterogeneity and state dependence. The log likelihood function for the pooled panel data is similar to that of the cross sectional probit: ))(1log()1()(log)(log ' 111 ' 1  it T tit N i T titit N ixFyxFyL    [3] and maximising this across all i with respect to , we obtain the pooled probit estimator. The standard errors are adjusted to account for serial correlation in the errors at the individual level. The main variables of interest are, disability and the associated limitations in daily activities, but we also control for other factors that may 7 be correlated with disability, as mentioned earlier. In addition, it is likely that past disability has a direct effect on current participation, so we include lagged variables for the three types of disability. Pooling all available data for the years 1995 to 2000, and estimating a standard probit model, we obtain estimates from the pooled balanced sample.c We present results from this pooled static model in Table 4, Columns 1 and 4, for men and women respectively. These results are presented as parameter coefficients, but we will later discuss some of the main results in terms of percentage effects. The effects of current disability are quite high for both men and women, reducing the probability of current labour force participation significantly. At a first glance, disability has a greater negative effect on the labour force participation probability of men, compared to women. Although the effect of a severely limiting disability is less for women than men, it is still substantial. In the case of men, even those with no limitations have a slight reduction in the probability of participation. For women, we see that the probability of participation for those with no limitations, is not significantly different from women with no disability. The gap between the effects of severe and some limitations is quite large for men and even more pronounced for women, suggesting that severe disability has a more negative effect on women’s participation. Past disability, in the previous year, also has a substantial effect on current participation, and is not much lower than the effect of current disability. This applies in the case of severe and some limitations, for both men and women. Similar to current disability and severe limitations, we see that individuals who previously had a severely limiting disability have a much lower probability of current participation, compared to those with no previous disability. In terms of the other explanatory variables (see Table A1), we see that labour force participation increases with age up to 34 (compared to those aged 55-64), but the effect falls slightly after the age of 44. Those with secondary or third level education have a greater probability of participating in the labour market. As expected, we see that women with children are less likely to participate, and this effect gets smaller as the youngest child is older. The opposite effect is found for men, where children increase the probability of participation, in particular when the youngest child is either aged less than 4, or in the older age group of 12-18. 8 The results from the static pooled model raise two important questions. The first interesting question is whether or not past disability affects current participation directly, or does it work through a separate channel by negatively affecting past participation? If so, we would expect to see that past participation influences current participation, and the effect of past disability should disappear. This would suggest that past disability still does have an effect on current participation, but does so by a) directly influencing past participation and therefore, b) indirectly affecting current participation. The second question arising from these results is whether or not the control variables appropriately account for any unobserved characteristics of disabled people that also influence their labour force participation decision? Again, if this were not true, we would expect that the actual effect of current disability should be lower. We now explore a dynamic model of participation that incorporates both past participation and unobserved effects. State Dependence and Unobserved Heterogeneity: In order to distinguish between the two effects – unobserved individual effects and past participation - we now include a lagged dependent variable into the model.d In general terms the following likelihood is derived and maximised; iiiiiiiitti T t it iiiiiTiiTiiTiiTi dxfxyfxyyf dxfxxyyfxxyyf   )|(),,|()],,,|([ )|(),,,...|,...,(),,...,|,...,( 001, 1 1000           [4] We must specify ),|( 0iii xyf  - known as the initial conditions problem. Heckman [14] suggests approximating ),|( 0iii xyf  and then specifying )|( ii xf  . Then )|,...,( 0iiTi xyyf is obtained by integrating out the unobserved effect. The main difficulty in this approach is in specifying the distribution of initial participation. We therefore follow an alternative approach suggested by Wooldridge [11] where we consider: iiiiiii T iiiiiTi dxyfxyyyfxyyyf  ),|(),,|,...,(),|,...,( 0010 1   [5] 15 disability and participation, and people with these disabilities may re-join the labour force. The incentive effects of disability benefits may also play a role here and these factors will be investigated in future research. V. Conclusions People with disabilities face many barriers to full participation in the labour market, with serious implications for living standards and quality of life. This paper has analysed the factors associated with participation or non-participation in the labour market, using data on people reporting chronic illness or disability in a large-scale Irish representative survey. The results of the panel analysis presented in this paper, bring out the scale of the impact on labour force participation, of having an illness or disability that limits the individual severely in their daily life. We controlled for state dependence and unobserved heterogeneity by estimating a dynamic model with correlated random effects. The results show that unobserved heterogeneity contributes substantially to the base effect of disability for men, and to some extent for women. In our preferred model, (pooled dynamic) disabled men with a current severe limitation are now only 9 percentage points less likely to participate compared to non-disabled men. However, the effect of past participation is quite high, at 40 percentage points. For women, our preferred model is the dynamic model with correlated random effects. Those with a severely limiting disability have a lower probability of participation by 26 percentage points, compared to women with no disability. The effects of some and no limitations are less substantial. The effect of past participation is lower in the model for women, reducing current participation by 13 percentage points. The interaction of disability, education and participation of women, should be explored further. In this paper, we aimed to provide more accurate estimates of the effect of disability on participation. However, we acknowledge some limitations. In particular, if the reporting of disability in the survey is prone to measurement error, we cannot estimate the true effect of disability on participation. This may help to explain the substantial contribution of unobserved individual effects, but without extending the model to allow for measurement error in reporting behaviour, our results on the effect of disability on participation are not conclusive. Again, this will form part of future 16 research where we will model labour force participation and disability, while controlling for reporting behaviour. 17 Footnotes: a. Our specification includes a measure of unearned income but does not include a control for wages. Correctly accounting for the relationship between disability and wages is a topic for future research. b. Another data source is a special module on disability included with the Quarterly National Household Survey (QNHS) in the second quarter of 2002, which focused on the extent and nature of restriction of activities for people with disabilities and their labour force status. Similar analyses of disability labour force participation in a cross sectional context, were carried out using QNHS data and we arrive at similar conclusions obtained from the Living in Ireland 2000 data. c. In this paper, we are assuming that although there is attrition in the sample between 1995 and 2000, it does not bias the results of the effect of disability on participation. This is especially evident in the pooled model, where we follow Wooldridge [11] and test for the effect of attrition using inverse probability weights on the pooled model for the unbalanced sample. The probability of being in each wave is not influenced by disability, for both men and women. In the participation model for men there was no change in the overall co-efficients. For women, we find that there is a slight overestimation of the effect of severe disability, changing the co-efficient in the unbalanced sample from –0.3678 in the original pooled model to –0.4203 in the weighted pooled model. However, in this paper we assume overall that attrition is not a problem in biasing estimates of disability, and focus on the balanced sample throughout. d. We could introduce a lag of two years for participation, and then include initial participation and previous participation as the two initial values. However, this increases data requirements and without a larger T, we cannot afford to be so flexible in the dynamics of the model. Furthermore, the transition matrix probabilities of participation in each year show that the rate of change from participation to non-participation or vice versa is the same for each pair of years. The correlation between participation and previous 18 participation is 0.79, likewise the correlation between previous participation and lagged previous participation is 0.79. 19 References 1. Bartel, A. and Taubman, P. Health and Labour Market Success: The Role of Various Diseases. The Review of Economics and Statistics 1979; LXI (1): 1-8. 2. Chirokos, T. and Nestel, G. Further Evidence on the Economic Effects of Poor Health. The Review of Economics and Statistics, 1985; LXVII(1): 61-69. 3. Wolfe, B. and Hill, S. The Effect of Health on the Work Effort of Single Mothers. The Journal of Human Resources 1995; 30[1]: 42-62. 4. Madden, D. and Walker, I. Labour Supply, Health and Caring: Evidence from the UK. UCD Working Paper 1999; 99/28. 5. Mete, C. and Schultz, T. Health and Labour Force Participation of the Elderly in Taiwan. Economic Growth Center Yale University Discussion Paper 2002: 846. 6. Kidd, M, Sloane,P. and I. Ferko. Disability and the labour market: an analysis of British males. Journal of Health Economics 2000, 19:961-981. 7. Sickles, R. and P. Taubman. An Analysis of the Health and Retirement Status of the Elderly. Econometrica 1986; 54[6]: 1339-1356. 8. Bound, J. Self-Reported versus Objective Measures of Health in Retirement Models. The Journal of Human Resources 1990; XXVI[1]:106-138. 9. Kreider, B. Latent Work Disability and Reporting Bias. The Journal of Human Resources 1999; XXXIV[4]:734-769. 10. Lindeboom, M. and M. Kerkhofs. Subjective Health Measures, Reporting Errors and the Endogenous Relationship between Health and Work. Working Paper 2002. 20 11. Wooldridge, J. (2002), Econometric Analysis of Cross Section and Panel Data. 12. Bound, J, Schoenbaum, M., Stinebrickner, T.R. and T. Waidmann. The Dynamic Effects of Health on the Labor Force Transitions of Older Workers. Labour Economics, 1999, 6(2): 179-202. 13. Malo, M. and Garcia-Serrano, C. An Analysis of the Employment Status of the Disabled Persons Using the ECHP Data, Second Draft; 2003. 14. Heckman, J. J. The Incidental Parameters Problem and the Problem of Initial Conditions in Estimating a Discrete Time-Discrete Data Stochastic Process, in Structural Analysis of Discrete Data with Econometrics Applications, Manski, C.F. and D. McFadden. Cambridge, MA:MIT Press, 179-195. Formatted 21 Table 1 Labour Force Status by level of restriction for those with Chronic Illness or Disability, age 15-64, Living in Ireland Survey 1995-2000 Severe limitation Some limitation No limitation No chronic illness or disability Men Participation 34.92 58.02 81.45 91.59 Non-participation 65.08 41.98 18.55 8.41 N 189 655 318 6026 Women Participation 13.82 31.82 44.65 55.15 Non-participation 86.18 68.18 55.35 44.85 N 123 707 318 6522 22 Table 2 Variable definitions for Dependent and Independent Variables Variable Definition LFP =1 if participating in the labour market, =0 otherwise Disabled with severe limitation =1 if disabled and severely limited in daily activities, =0 otherwise Disabled with some limitation =1 if disabled and limited to some extent in daily activities, =0 otherwise Disabled with no limitation =1 if disabled and not limited in daily activities, =0 otherwise (Base category=No disability) Age 15-24 =1 if aged 15-24 years, =0 otherwise Age 25-34 =1 if aged 25-34 years, =0 otherwise Age 35-44 =1 if aged 35-44 years, =0 otherwise Age 45-54 =1 if aged 45-54 years, =0 otherwise (Base category=aged 55-64 years) BMW =1 if living in Border, Midlands, West region, =0 otherwise (Base category=Rest of Country) Secondary Education =1 if highest level of education completed is secondary, =0 otherwise Third Level Education =1 if highest level of education completed is third level, =0 otherwise (Base category=No qualifications or highest level of education completed is primary) Married =1 if married or living with a partner, =0 otherwise Age Youngest Child<4 =1 if age of youngest child is less than 4, =0 otherwise Age Youngest Child>=4 and <12 =1 if age of youngest child is greater than or equal to 4 and less than 12, =0 otherwise Age Youngest Child>=12and <18 =1 if age of youngest child is greater than or equal to 12 and less than 18, =0 otherwise (Base category=No children) Unearned Income =Net Household Income – Net Individual Disposable Income (Net Individual Disposable Income includes net incomes from work, social welfare payments and child benefit. Net Household Income aggregates individual data to household level) Note: The regional classifications are based on the NUTS (Nomenclature of Territorial Units) classification used by Eurostat. 23 Table 3 Summary Statistics for all Variables Variable Percentage of Sample in each Category Men Women LFP 86.6 51.9 Disabled with severe limitation 2.6 1.6 Disabled with some limitation 9.1 9.2 Disabled with no limitation 4.4 4.1 No Disability 83.8 85.0 Age 15-24 12.3 10.1 Age 25-34 16.4 17.2 Age 35-44 26.2 27.1 Age 45-54 24.4 25.6 Age 55-64 20.7 20.0 BMW 24.7 21.9 Secondary Education 51.8 59.0 Third Level Education 16.7 13.3 No education or primary only 31.4 27.6 Married 68.7 73.3 Age Youngest Child<4 12.5 13.3 Age Youngest Child>=4 and <12 21.3 24.5 Age Youngest Child>=12and <18 15.2 17.7 Unearned Income 228.64 (240.13) 389.5 (307.7) N 7188 7670 Note: For unearned income we present the mean and standard deviation (in parentheses) 24 Table 4 Panel Model Results Men (co-efficients) Women (co-efficients) Pooled Static Random effects dynamic (re-scaled) Pooled Dynamic Pooled Random effects dynamic (re-scaled) Pooled Dynamic Lag LFP 0.7511** (0.1194) 1.687** (0.0918) 0.7494** (0.0835) 1.7974** (0.0623) Disabled with severe limitation -1.2368** (0.1314) -0.6639** (0.2653) -0.5653** (0.2218) -0.9173** (0.1736) -0.8256** (0.2827) -1.1359** (0.2393) Disabled with some limitation -0.7886** (0.0814) -0.5159** (0.1594) -0.4757** (0.1285) -0.3296** (0.0755) -0.3137** (0.1283) -0.4210** (0.1106) Disabled with no limitation -0.2066** (0.1042) -0.3464** (0.2161) -0.3397** (0.1380) -0.0175 (0.0928) -0.1811** (0.1497) -0.2732** (0.1326) Lagged Disability Disabled with severe limitation -1.0555** (0.1275) -0.2534 (0.2593) -0.0765 (0.2465) -0.6203** (0.1626) -0.1470 (0.2863) 0.0102 (0.2643) Disabled with some limitation -0.5802** (0.0783) 0.0259 (0.1592) 0.1796 (0.1302) -0.2742** (0.0714) -0.0056 (0.1303) 0.0514 (0.1177) Disabled with no limitation -0.0925 (0.1175) 0.0887 (0.2254) 0.1298 (0.1461) -0.0290 (0.0962) -0.0495 (0.1566) -0.0464 (0.1363) Initial condition LFP in 1995 1.2059** (0.2096) 0.6399** (0.0944) 0.8984** (0.1353) 0.6315** (0.0626) Random effect (time averages) Disabled with severe limitation -0.8815** (0.5948) -0.9013** (0.4588) -0.3077 (0.7211) -0.2653 (0.5607) Disabled with some limitation -0.7265** (0.3237) -0.7146** (0.2371) -0.1387 (0.2744) -0.1209 (0.2041) Disabled with no limitation 0.3616 (0.5068) 0.2146 (0.3297) 0.4464* (0.3844) 0.5171* (0.3087) Constant 0.4642** (0.1332) -0.8210** (0.2167) -1.0449** (0.1332) -0.5446** (0.1074) -0.1118** (0.1595) -1.5214** (0.0945) N 5930 5930 5930 6330 6330 6330 Pseudo R2 0.2772 0.5371 0.1700 0.5303 Rho 0.4684** 0.3984** Note: 10.0,*05.0**  pp . Note: 10.0,*05.0**  pp (Significance in random effects models are based on t-stats on base co-efficients).