Cognitive functioning and labour force participation ... - CiteSeerX

Cognitive functioning and labour force participation among older men and women in England David Haardt∗ 15 April 2007

DRAFT: PLEASE DO NOT QUOTE OR CIRCULATE Abstract In this paper, I analyse the relationship between cognitive functioning (CF) and employment among older men and women using data from the English Longitudinal Study of Ageing (ELSA), 2002 and 2004. Regression analysis for individuals aged 50–70 shows that the change in CF over time does not have any statistically significant effects on the probability to exit or enter employment, or on working hours. These results are not sensitive to the definition of ‘work’. My findings differ from earlier research on younger age groups in the US and Germany where modest effects of CF on labour market outcomes were found.

1

Introduction

Increasing the labour force participation of older men and women may be an important channel through which to finance the costs of an ageing population. As figure 1 shows, the proportion of men and women working in the UK falls dramatically over age, with ∗

Institute for Social and Economic Research (ISER), University of Essex, Wivenhoe Park, Colchester, Essex CO4 3SQ, United Kingdom. E-mail: [email protected]. This will be the third paper of my PhD thesis. It would not have been possible without the outstanding supervision which I have been receiving from Stephen P. Jenkins. Thanks are also due to my second supervisor John F. Ermisch, to Martin Browning, Philip J. Cozzolino, Brenda McWilliams, Vegard Skirbekk, and seminar participants at the Vienna Institute of Demography, at ISER, University of Essex, and at the Irish Centre for Social Gerontology, National University of Ireland, Galway. I am grateful to the data depositors of ELSA (National Centre for Social Research) and to the UK Data Archive, University of Essex, for providing access to the data. I would also like to thank Ben Jann for providing the Stata command estout and for drawing my attention to the dcolumn package for LATEX. Finally, I would like to thank the Austrian Academy of Sciences, the Economic & Social Research Council, the University of Essex, and the Provincial Government of Upper Austria for funding. I alone am responsible for errors and opinions.

the decline starting to set in at a rather young age when compared to other countries such as the USA. Therefore, it is important to understand the factors influencing older people’s work decision.

Figure 1: Proportion of men and women working in the UK by age (source: own analysis using BHPS data, cf. Haardt, 2006: 5). Previous research has studied the effects of many different factors on participation, including health and financial incentives. In a recent paper, Haardt (2006) showed that potential income out of work and health status are the two most important determinants of older men and women’s labour market transitions in the UK, with effects that are larger than found in previous studies for British and US men. However, these need not be the only important factors in predicting whether older people work or not. Cognitive functioning (henceforth CF) may be another important factor—even low-skilled manual jobs have certain requirements on memory and other aspects of cognition. Workers not being able to meet these requirements may find it difficult to retain their job, or to find one. It is therefore important to analyse the effects of cognition on labour force participation. Since cognitive performance crucially depends on early-life circumstances, policy may be able to affect CF among future cohorts and, hence, their propensity to work in older age. Understanding how cognition affects older people’s work decision will also be a first starting point in assessing the economic value of initiatives such as ‘skills for life’ among 2

older people. This initiative is aimed at improving adult literacy and numeracy skills and has been launched by the Department for Education and Skills (DfES). In the past, good data on cognitive functioning have rarely been available in data sets which also offer comprehensive information on socio-demographic characteristics and, in particular, employment. However, this has recently changed by the introduction of the English Longitudinal Study of Ageing (ELSA). Most of the literature on the impact of CF on labour market outcomes focuses on the impact of schooling achievement on labour market entry; there is virtually nothing on older people. However, there are also a few papers (and at least one book) on the general working-age population. Pryor and Schaffer (2000), in a book which received a lot of attention in the academic community and beyond, use the Current Population Survey and the National Adult Literacy Survey to explore recent transformations of the US labour market. Among other things, they stress the role of cognitive skills, going beyond formal educational attainment, and their effects on labour market participation, occupational mobility, and wages. They argue that while educational credentials have become more and more widespread, functional literacy (numeracy, literacy, and writing skills in a work context) has not grown equally fast, causing wage rates for high-CF jobs to increase further. Cawley, Heckman, and Vytlacil (2001) as well as Heckman and Vytlacil (2001) study US men and women aged 15–37 and the impact of CF (as measured by the Armed Services Vocational Aptitude Battery or ASVAB, measuring work-related and general cognitive skills), on (log) wages. They encounter identification problems caused by the large correlation between education and ability (as measured by the ASVAB results) and find only modest effects of CF on wages.1 Anger and Heineck (2006) study German men and women aged 20–60 and the effects of CF (measured by speed of cognition and animal names) on the unemployment probability and (log) earnings, finding that CF affects the unemployment probability but has only weak effects on (log) earnings. Even though these studies tend to focus on wages and earnings as labour market outcomes, I will focus on work, for two reasons. First, the drop in participation as people become older is more pronounced than the decline in average earnings in the late life-cycle. Second, income inequality increases sharply over the life-cycle, mostly due to individual effects, requiring long panels to estimate reliable wage or earnings regressions for older people.2 I will, however, also report the effects of CF on earnings and wages in Appendix C for the sake of comparison with other studies. 1

Another focus of their paper is how to separate age and time effects. Cf. Deaton and Paxson (1994) for theoretical arguments as well as empirical evidence for Great Britain, the USA, and Taiwan. 2

3

I would like to answer the following three research questions in this paper: 1. Is there a relationship between CF and employment among older men and women in England? 2. If so, which measures of CF show particularly strong links? 3. Can we say anything about causality?

2

How does cognitive functioning affect work?

In order to address these research questions, it is essential to decide on how to model the relationship between cognitive functioning and the work decision. Even when deciding to use a binary definition of ‘work’, the question remains whether to model (1) the effects of the level of CF on the probability to work or (2) the effects of changes in CF on the transition probabilities out of and back to work. I will model the effects of changes in CF on exit and entry probabilities and on the change in working hours, arguing that it is advantageous to employ the latter approach since it is a very strong assumption to say that people with lower CF are in general less likely to work. There are jobs with different requirements of CF, and somebody who loses a high-CF job due to a decline in CF may well still have a higher level of CF than somebody who had a low-CF job for their entire working life. Since it is difficult to measure the ‘cognitive requirements’ of a job, attempting to explain the employment probability will suffer from much more noise than explaining the exit from or the entry into employment. Therefore, exit and entry regressions can be expected to be superior to static employment regressions. Nevertheless, it will be reassuring to see that my key results with respect to CF hold regardless of whether I use this specification or the other. In principle, it is of course also possible to extend the exit and entry approach to a more general ‘hours change’ approach. However, I observe in the data that changes in employment patterns with age are mostly due to changes on the extensive margin (work), not due to changes on the intensive margin (working hours). Working hours remain highly stable among men aged 50–64 and women aged 50–59, with only a slight decline in median working hours among men who report non-zero hours, and only a slight inverse U-shape among women who report non-zero hours.3 Moreover, measurement error and missing values are much more a problem for exact working hours than for whether somebody worked or not. Therefore, exit and entry regressions can be expected to yield more clear-cut results than hours change regressions. 3

This empirical observation is true both cross-sectionally using ELSA as well as longitudinally using the BHPS.

4

My choice of exit and entry regressions over other alternatives is also supported by a recent paper by Disney et al. (2006). Using the BHPS, they estimate economic activity equations and hazard models.4 They show that using a fixed-effects logit model for the economic activity equations (and hazard models) is superior to other approaches since it links changes in health to changes in the employment probability rather than levels to levels, thereby eliminating person-specific effects. This conclusion can be seen as related to earlier findings by Bound et al. (1999) who also find that changes in health rather than its level appear to drive labour market behaviour. Modelling the exit from and the entry into employment will also help to alleviate a number of other problems. These will be discussed later on in this paper. First, however, I will present the data and sample selection criteria used with particular emphasis on cognitive functioning and work.

3

Data and sample selection

3.1

The English Longitudinal Study of Ageing

ELSA is a survey of people aged 50 or above living in households in England, and their partners regardless of their age. The first ELSA wave included 12,100 respondents, thereof approx. 11,500 men and women aged 50 plus and approx. 600 younger partners. The ELSA sample is based on respondents from the cross-sectional Health Survey for England (HSE) 1998, 1999, and 2001. This has the advantage of providing baseline data (such as on health, for instance) before people enter the ELSA observation window.5 Currently, data are available from the first two ELSA waves (2002 and 2004).

3.2

Sample selection

ELSA does not have an upper age limit; in this paper, I focus on the age group 50–70. This is due to the fact that it may be useful to look at a few more years beyond state pension age which is 65 for men and 60 for women. Employment rates are very low beyond the age of 70 which is why I drop such observations. Another issue is whether to exclude the self-employed or not. It could be argued that the self-employment decision is governed by a different process than the employment decision; in particular, somebody who is self-employed may continue to work even though his or her level of CF is lower than necessary to make a living in that job. On the other 4

Their main concern is that health stock has much strong predictive power as an explanatory variable than self-reported health status. However, this is not crucial for the point which I want to make here. 5 Taylor et al. (2004) provide a comprehensive overview of the study’s sample and methodology.

5

hand, excluding the self-employed implies that it becomes difficult to define the at-risk groups for exit from employment or entry into employment. Furthermore, almost one fifth of people working in my subsample are self-employed, implying that excluding the self-employed reduces sample size noticably. This is particularly worrying when modelling entry since this event is quite rare among this age group. Therefore, I decided to include the self-employed in my exit, entry, and hours change regressions. I do however exclude the self-employed from my wage, earnings, and wage change regressions in Appendix C since there is the problem of negative self-employment income and, hence, negative ‘wage rates’ among the self-employment. In line with these considerations, I observe that these regressions have markedly higher explanatory power when excluding the self-employed. However, the key results with respect to CF hold true also when including them.

3.3

Cognitive functioning in ELSA

The key advantage of ELSA is that it collects a huge amount of information specifically related to ageing, including highly detailed information on work, health, and pension saving. Among the information on cognition are (all collected in both waves, unless otherwise mentioned): • self-assessed memory performance • self-assessed change in memory performance compared to two years ago • date score (knowing the date and the day of the week) • immediate recall (respondents are asked to repeat a four-item word list) • delayed recall (respondents are asked to repeat the same word list a bit later during the interview) • verbal fluency (mentioning as many animal names as possible within one minute) • prospective memory score (respondents are asked to carry out a certain task at a specified later point during the interview) • numeracy (wave 1 only) • literacy (wave 2 only) Since ELSA provides longitudinal data, it will become increasingly easy to model change over time as more waves become available. Currently, there is no measure of the 6

change in numeracy or literacy; this will require waves 3 and 4, respectively. I therefore do not use numeracy or literacy in my exit, entry, or hours change regressions. One important advantage of the CF data in ELSA is that there are subjective as well as more objective measures of CF. This enables a comparison between self-assessed and tested memory. Furthermore, it reduces measurement error and avoids problems are different perceptions of the question across respondents (some respondents may for instance answer the question about self-assessed memory performance compared to the population as a whole, others compared to people of similar age). I will not use selfassessed CF in my analysis; however, it is important to keep in mind that dven objective CF may of course be the result of past employment, an issue which I will address in Section 3 of this paper. Along similar lines, it is also beneficial that ELSA does not only collect information on self-assessed health, but also a comprehensive range of items on more objective measures of physical health. There is, for instance, plenty of information on functional limitations. In addition to that, a nurse visit was carried out in wave 2 where a blood sample as well as several basic measurements (such as waist circumference or gait speed) were taken. This implies that it is easier to address the potential endogeneity of physical health as addressed by Bound (1991) compared to many other data sets. Since there are many CF variables in ELSA, it may be useful to create an index of CF to provide a convenient summary measure. Analysis of Cronbach’s alpha shows that only immediate and delayed recall can be meaningfully combined, with an alpha of 0.79, whereas the alphas for other combinations of CF variables are much lower. This suggests that there are several underlying constructs of cognitive performance. Even under these circumstances, it may still be useful to create an index of CF to reduce the dimensionality. Steel et al. (2004) and Huppert et al. (2006) use CF indices which are ad hoc combinations of the underlying variables, rescaled in a rather arbitrary way. My approach is slightly less arbitrary since I standardise the CF variables to have a mean of zero and a standard deviation of one and then add them up. I create two CF indices, the regular index which is based on the date score, immediate recall, delayed recall, animal named, and the prospective memory score, and a second index which excludes the date score (since this variable is most likely to be influenced by whether somebody works or not). Figure 2 shows the value of the CF index by whether a person is working or not, where ‘working’ means working 16 or more hours a week. (I will soon explain why I chose a threshold of 16 hours per week. It is important to emphasise that the choice of the threshold does not affect my point here.) One can see only a very slight difference between the two panels of the figure – the mean is slightly higher for those who work and 7

Figure 2: CF index by whether respondent is working full-time or not (source: own analysis using ELSA data, pooled waves 1 and 2). the variance slightly lower. Of course, this figure does not take any control variables into account. Schaie and Strother (1968) showed that notions of ‘cognitive decline’ over age may well be due to cohort effects rather than due to cognitive decline for a given person. Even though my two observations are only two years apart, it may still be interesting to look at changes in the CF variables and how mean and median changes vary by age group. Table 1 shows that there is indeed less decline than one may have thought. For prospective memory and delayed recall, even in the oldest age group (which is not included in my regression analysis) neither the mean nor the median is negative. The only CF variable for which both mean and median are negative in the oldest age group is the number of animal names mentioned within one minute. For the remaining two CF variables, date score and immediate recall, the mean is negative in the oldest age group but the median still equals zero. This confirms the early results of Schaie and Strother (1968) for the US and is an interesting finding in its own right.

8

change in date score, mean change in date score, median change in immediate recall, mean change in immediate recall, median change in animal names, mean change in animal names, median change in prospective memory, mean change in prospective memory, median change in delayed recall, mean change in delayed recall, median change in CF index, mean change in CF index, median change in CF index w/o date score, mean change in CF index w/o date score, median

50–54 0.0090 0 0.1419 0 0.0423 0 0.2897 0 0.2097 0 0.3512 0.3199

55–59 0.0238 0 0.1389 0 0.3544 0 0.2660 0 0.1941 0 0.4030 0.4191

60–64 65–69 70 plus 0.0604 0.0445 –0.0871 0 0 0 0.0424 0.0569 –0.0277 0 0 0 0.1859 –0.0044 –0.4391 0 0 –1 0.0797 0.2293 0.0117 0 0 0 0.1402 0.1421 0.0318 0 0 0 0.2135 0.2797 –0.1816 0.1920 0.3378 –0.0637

Table 1: Means and medians of the change in the CF variables by age group.

3.4

Potential endogeneity of cognitive functioning

A potential problem concerning the use of CF as an explanatory variable is that it may be endogenous. Banks (2006: 299) emphasises that not only the impact of CF on employment is important, but also the reverse impact of employment on CF. Implicitly, my specification is based on the former direction rather than on the latter. If CF is endogenous, i.e., if being in work helps to build and maintain cognitive performance, an endogeneity bias may occur. This means that my measures of CF are positively correlated with the error term, leading to an upward bias of the coefficients of the CF variables, i.e., an overstatement of the estimated effects of CF on work (cf. Wooldridge 2002: 62). However, it is very difficult to think of potential instruments which affect cognitive functioning but do not have direct effects on the work decision. Given these considerations, it is interesting to compare whether, when using data from the second wave only, the results are systematically different when using lagged CF variables compared to when using contemporaneous CF variables. If there are such systematic differences, it is interesting to check whether they correspond to our expectations in the presence of an endogeneity bias as just explained. The results of this comparison can be seen in table 2.6 It can be seen that there are important differences depending on whether or not lagged CF variables are used. In particular, the absolute size of the coefficients of immediate recall and delayed recall decreases when using lagged CF variables. As mentioned previ6

This table uses a threshold of 16 hours per week as in my main analysis. However, the key findings are the same when using different thresholds, such as 10 hours per week or 22 weeks per annum.

9

workt=2 = f(CFt=2 ) workt=2 = f(CFt=1 ) Men Date score Immediate recall Animal names Prospective memory Delayed recall Women Date score Immediate recall Animal names Prospective memory Delayed recall

–0.0266 –0.0105 0.0015 0.0050 0.0129

0.0826** 0.0054 0.0003 0.0251** 0.0056

0.0722** 0.0173* –0.0009 0.0131 –0.0177*

0.0445 0.0012 –0.0018 0.0147* –0.0006

Table 2: Marginal effects from the probit model, comparing contemporaneous versus lagged CF variables. Includes full controls. Full results in table 14, Appendix A. ously, this behaviour is consistent with what would be expected if these variables were endogenous. More generally, women’s coefficients follow more closely the endogeneity hypothesis than men’s. However, and this is independent of sex, the same two CF variables are always statistically significant when using lagged CF variables: the date score and the prospective memory score. Moreover, the differences as a whole do not seem to follow any particular systematic pattern. This can be interpreted as another hint that linking changes in CF to exits and entries may be more fruitful than linking the level of CF to the probability of work. Furthermore, I also argue that CF is the result of an accumulation of cognitive capital over the lifecourse. There is now some evidence that the key effects are during childhood, and that the stock of CF as well as its development during adulthood crucially depend on child development (cf. for instance Heckman 2000). Similarly, Simpson (1980: 306), in a survey article, stresses that empirical results on whether work has short-run effects on cognition are rather mixed. Here, the conceptual framework of a model of changes proves again useful. If the observations of Heckman and Simpson are true, CF will not instantaneously react to changes in work status, working hours, or occupation. 7 However, there may well be instantaneously effects of changes in CF on employment outcomes, particularly when considering large sudden declines. (These declines may have physiological or other causes, but this goes beyond the scope of this paper.) 7

The only CF variable for which such an effect is plausible is the date score; I will therefore run all regressions twice, once with and once without this variable.

10

3.5

Defining ‘work’

Traditionally, most research in labour economics has analysed economic activity (also referred to as labour force participation) rather than work. This is mainly due to the fact that economic activity is a decision of the economic agent alone, whereas whether somebody will actually work or not is also subject to factors beyond his or her control. McFadden (1981) showed how to derive a number of (nowadays) common econometric models of binary choice from utility maximisation theory. I argue that in my context it is better to model work than economic activity, for at least three reasons: first, if we are interested in CF as a potential factor through which to finance population ageing, work rather than economic activity matters. Second, work can be objectively measured whereas economic activity in the labour market cannot (because of the problems in defining those looking for a job). Third, the unemployment rate in this age group is very low which means that differences between the two approaches can be expected to be minor. In principle, it may be worthwhile to model both processes. However, this would probably require a larger sample, in particular more people who are economically active yet do not work, i.e., more unemployed. Having decided to use work rather than economic activity, it is also important to think about what ‘work’ exactly means. Much of the literature also regards those who only work one or two hours per week as working (cf., e.g., Harkness 1993). However, I think that imposing a certain minimum amount of hours will be useful in my context. Looking at the distribution of weekly working hours in ELSA (pooled data from both waves) shows that 15 hours per week is the first major peak in the data. However, using 16 or more hours per week to define work may be more appropriate as this is the threshold commonly used to distinguish between part-time and full-time work for the purpose of benefit income in the UK.8 Using 16 hours per week as a minimum classifies only 11.15% of those who call themselves working as not working (i.e., 11.15% of those who call themselves working work 15 or less hours per week). I therefore use 16 hours per week as my minimum threshold. Somebody is, therefore, defined as working in my analysis if he or she works at least 16 hours per week, and as not working if he or she works less than 16 hours per week, including zero hours due to unemployment, self-reported retirement, or any other reason. Table 3 shows the work transition matrix when adopting the ‘16 hours’ definition. It can be seen that the risk of panel attrition is markedly higher for those who did not work in wave 1 than for those who did. Also, the exit probability is five times as large as the 8

More specifically, this threshold is used for the following benefits: jobseeker’s allowance, incomebased jobseeker’s allowance, job grant, income support, and working tax credit. Cf. Phillips and Sibieta (2006) for further details.

11

t+1 0 1 Missing Total 0 63.28 3.26 33.46 100.00 1 17.27 63.60 19.13 100.00 Total 43.55 29.13 27.32 100.00

t

Table 3: Work transition matrix, row percentages. entry probability. I also carried out tests to see whether the results are robust to the choice of the cut-off point and to the choice of weekly working hours as such. I carried out these tests with three different types of variable groups: • Self-reported work status • Certain minimum hours per week • Certain minimum weeks per annum Fortunately, the key results of my regressions remain qualitatively unchanged regardless of the cut-off point or variable group chosen. The coefficients of the CF variables remain virtually unchanged when doing so.

4 4.1

Empirical implementation Econometric models

I now present my three econometric models: the exit model, the entry model, and the hours change model. Let us start with the exit and entry models which are essentially identical. with which I aim to address the five issues raised. Suppose that the latent propensity of individual i to exit from (enter) the labour market is given by p∗i = α + β∆ci + γhi + δxi + ²,

(1)

where ∆c represents the change in cognitive functioning, h physical health, x is a vector of other (exogenous) variables, and ² an error term. We do not observe the latent propensity p∗i , only the outcome. In the case of the exit event, a 1 means that somebody had 16 or more hours per week in t = 1 and less than 16 hours (including zero hours) in t = 2, and a 0 means that somebody had 16 or more hours per week in both t − 1 and t.

12

In the case of the entry event, a 1 means that somebody had less than 16 hours (including zero hours) per week in t = 1 and 16 or more hours per week less than 16 hours in t = 2, and a 0 means that somebody had less than 16 hours (including zero hours) per week in both t = 1 and t = 2. As mentioned before, c may include all measures of cognitive functioning as available in ELSA, and h may include ‘objective’ information about functional limitations or medical conditions. Finally, x may include ‘classic’ variables such as age, education, marital status, housing tenure, ethnicity, occupation, or information on other people living in the household. Note that even though ci appears in changes, hi and xi are still included in levels. For most variables, there are arguments for both approaches and it is not immediately clear which one is to be preferred. In the survival analysis literature it is common to explain conditional event probabilities by variables in levels whereas in the dynamic panel literature it is common to explain changes in the dependent variable by changes in the explanatory variables. I will adopt the former approach. These regressions are estimated using probit models which are run separately by sex to allow for different parameters. I will report marginal effects. Let us now turn to modelling the change in working hours in a standard OLS framework: ∆Li = α + β∆ci + γhi + δxi + ²,

(2)

where L represents labour supply measured in weekly hours. The right-hand side of equation 2 is equivalent to that of equation ?? before. Among these models, the exit and entry regressions are my preferred specifications from a substantive and a methodological point of view, as detailed in Section 2 of this paper.

4.2

Controlling for sample selection

Controlling for sample selection may be necessary in this context. I argue that there are a few different forms of sample selection which should be considered in this study. First, there is the problem of panel attrition. It is interesting to observe that CF scores in wave 1 are systematically lower for those who do not participate in wave 2 than for those who do. This is true even when limiting the sample to, say, the age group 50–54, implying that this phenomenon is unlikely to be linked to death or severe health problems. If the process governing panel attrition and the process governing the work decision 13

are linked, i.e., if ρ, the correlation between the error terms, is different from zero, this may bias my estimates of the effects of CF on work. I therefore also carry out Heckman probit models of the exit from and the entry into work and Heckman selection models of the change in working hours. I use four additional regressors in the selection model which I exclude from the main equation: whether respondents agreed to provide an additional contact address through which to get in touch with them, whether respondents agreed to record linkage for economic and health data, whether respondents consulted their documents during the ELSA section on income and assets, and the interviewer’s assessment of the reliability of the respondent’s answers in this section.9 It can be argued that these variables have an impact on panel attrition but not on labour force participation. Second, there is the common problem of selection into work. We observe labour income, and thus wages, only for those who work. This may again cause a selection bias. In most studies on the general working-age population, the number of dependent children is used as the instrument for the selection into work. However, this is not advisable in my context. The children of my sample age group will in most cases already be too old to need high levels of supervision. However, it may well be that respondents in this age group are providing care for their spouse or for their parents. I therefore use information on the hours spent giving care to others during the last week as the instrument for selection into work. This variable includes giving care to one’s children, spouse, or parents. Third, the two selection problems may of course appear jointly. In regressions explaining the change in the wage rate, respondents are only included in the regressions sample if they worked and survived to the second wave of ELSA. However, one may argue that the work selection bias is limited to the wage level rather than to its change. I therefore do not consider such joint selection which would only complicate the analysis further.

4.3

Issues concerning the explanatory variables

There are also three issues concerning the explanatory variables. 4.3.1

Cognitive functioning and education

The first issue is the potentially high correlation between CF and educational attainment. Cawley, Heckman, and Vytlacil (2001) as well as Heckman and Vytlacil (2001) show that these two variables are highly correlated in the US. They do for instance not have any college graduates from the lowest CF quartile in their data. More generally, many cells in 9

The first two of these four variables are not part of the public ELSA release.

14

their education-CF matrix are empty. Therefore, they argue that the effects of education and CF cannot be separated without strong parametric assumptions. To investigate whether the same problem exists in my UK data, I created Table 4, resembling their tables. Looking at the upper left cell, one can see that there are not few university graduates in the lowest CF quartile. There is certainly a positive correlation between education and CF also in my UK data, but it is much weaker than in the US data of Heckman and Vytlacil. The correlation is slightly higher for men than for women. A potential explanation why their problem is not an issue in my data is that my measure of CF consists of measures of memory and verbal fluency whereas the CF measure of Heckman and Vytlacil also includes factors which are more likely to be closely related to education, including numeracy and literacy (not included in the CF index used to compute in Table 4 since I do not use these variables in my regression analysis). However, even though the correlation of numeracy and literacy with education is indeed somewhat higher than that of other CF variables with education, I never observe correlations as high as Heckman and Vytlacil do. I am therefore confident to be able to separate the effects of education and CF. Degree Other HE A levels O levels NVQ1/CSE Other None

1st 8.68 16.06 13.69 15.85 36.94 25.45 41.45

2nd 16.99 22.83 23.30 24.31 30.93 28.78 28.31

3rd 29.38 27.11 27.57 30.22 19.59 26.26 19.78

4th 44.96 34.00 35.44 29.63 12.54 19.51 10.47

Total 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Table 4: Cross tab of educ. qual. and CF index quartile, row percentages (both sexes, n=13,740). It is also interesting to cross-tabulate education and the quartile of the change in the CF index. Table 5 shows that there is virtually no correlation between these two variables. In other words, education is correlated with the level of CF, but not with changes in CF. 4.3.2

Cognitive functioning and age

The second issue is rather similar to the first: should age be included in the analysis? CF and age, like CF and education, may be highly correlated and, moreover, may indeed measure similar things. Just computing simple correlations of all the CF variables with age shows that the correlation is most negative for immediate recall and delayed recall (around –0.4 in both 15

Degree Other HE A levels O levels NVQ1/CSE Other None

1st 24.55 23.75 22.10 23.87 30.60 22.27 27.09

2nd 26.96 26.56 27.01 26.22 23.28 25.17 22.32

3rd 23.46 26.82 23.66 26.31 24.57 25.61 24.34

4th 25.03 22.86 27.23 23.61 21.55 26.95 26.25

Total 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Table 5: Cross tab of educ. qual. and change in CF index quartile, row percentages (both sexes, n=5,666). cases). Interestingly, it is much lower for self-assessed memory performance (–0.05) than for any objective CF measure. 50–54 55–59 60–64 65–70

1st 16.42 19.65 26.94 36.48

2nd 22.02 24.35 26.72 26.74

3rd 27.58 25.90 25.88 21.81

4th 33.99 30.10 20.46 14.97

Total 100.00 100.00 100.00 100.00

Table 6: Cross tab of age group and CF quartile, row percentages (both sexes, n=13,798). Table 6, another ‘Heckman-Vytlacil table’ as I call it, shows that even though CF and age are correlated, there is certainly no danger of empty cells. The correlations among the CF variables themselves are in a similar range in absolute terms as the correlations of the CF variables with age, except for the correlation between immediate and delayed recall which is markedly higher (about 0.70). This lines up well with the earlier result on Cronbach’s alpha. Age alone explains up to 18% of the variation in the CF variables (as measured by the adjusted R squared in an OLS regression). This explanatory power is strongest for the immediate recall variable. 50–54 55–59 60–64 65–70

1st 25.88 22.68 26.22 26.10

2nd 23.99 25.70 25.31 24.43

3rd 22.99 25.81 25.10 24.98

4th 27.14 25.81 23.37 24.49

Total 100.00 100.00 100.00 100.00

Table 7: Cross tab of age group and change in CF quartile, row percentages (both sexes, n=5,668). However, it is important to note that these results all consider age cross-sectionally. As shown earlier on in table 1, the correlation of longitudinal decline with age is much lower 16

than cross-sectional data may suggest. The same is shown by table 7 which cross-tabulates age group and the quartile of the change in CF. It can be seen that the correlation is only very slight. Furthermore, I argue that it makes sense to include age in addition to CF since age may also measure many other things, such as changing preferences for or attiudes towards work. After experimenting with different specifications of age and with excluding age, I therefore settled on using an age spline in all my regression analysis. When excluding age, the results with respect to the CF variables are qualitatively similar (and similar with respect to statistical significance), but the coefficients are larger. 4.3.3

Endogeneity of physical health

Third, there is the well-known problem of potential endogeneity, or more generally mismeasurement, of physical health. Lambrinos (1981), using the US 1972 Social Security Survey Survey of Disabled and Nondisabled Adults (SDNA), examines the effects of different specifications of health in labour supply models, arguing that specification matters a lot, and that using more ‘objective’ health measures, such as on functional limitations, is a simple way of reducing endogeneity bias. In a similar vein, Stern (1989), using a simultaneous equations probit model (modelling the participation decision and the decision whether to report a disability) to estimate the effects of disability on labour force participation, finds that the health limitations variable is particularly powerful in explaining the ‘true’ effect of disability. The main implication which I draw from these papers is that using information on functional limitations is a simple but effective measure to prevent endogeneity bias with respect to physical health. Such variables are readily available in ELSA, and using them helps to avoid making the modelling structure even more complicated or adding even more groups of explanatory variables. However, it is important to keep in mind the wellknown study by Bound (1991) in which he shows that using objective health variables may in fact underestimate the impact of health on retirement behaviour.

4.4

Explanatory variables

I use all objective CF variables that are available in both waves, i.e., the date score, immediate recall, delayed recall, verbal fluency (animal names), and the prospective memory score. Since the data score may itself be affected by whether somebody works, I run all regressions twice, one with and once without the date score. I also create two indices of 17

CF, one which includes the date score and one which does not. To check whether only very low or very high CF has an impact on labour market outcomes, I also run all regressions using deciles of the two CF indices. Of course I also use a number of controls, the first of which being information on wealth. These data come from the financial derived variables which are part of the ELSA release. Since the main home may have a different effect than other wealth, I include net housing wealth (main home) and net non-housing wealth (anything else) separately. Since these variables can take negative as well as positive values, I use the inverse hyperbolic sine of them. This ensures that the weight attached to extremely high or extremely low wealth is reduced and provides a better fit for my data than linear or other non-linear specifications.10 The interpretation of the inverse hyperbolic sine is very similar to that of the logarithm as long as values are not too close to zero. After having experimented with different specifications of age, I settled for an age spline with five-year intervals. This ensures reasonable flexibility whilst making sure that there are not too many parameters which have to be estimated. In terms of measures of physical health, I use a dummy for health limitations. As already mentioned, Lambrinos (1981) argues that information on functional limitations is much less prone to endogeneities and mismeasurement than other measures of health or disability. It would be possible to include separate variables for different functional limitations, but since this paper focuses on the role of CF, I use a parsimonious specification with just one dummy variable. As for education, I use the seven categories used in the Heckman-Vytlacil tables before: degree, other higher education, A levels, O levels, NVQ1/CSE, other qualification, and no qualification (base category). I also include a dummy variable on marital status which equals 1 for married individuals (which may or may not be their first marriage) and 0 for everybody else (single, divorced, or widowed). Since housing tenure is an important measure of not only wealth but also geographical mobility and status, I include two dummy variables for this domain: one for outright owners and one for owners with a current loan or mortgage. The base category includes everybody else, i.e., mainly renters (council housing, housing societies, or private). As far as information on occupation are concerned, I merge the occupational groups professional and managerial/technical into a single category. I do the same for the groups partly skilled and unskilled. This is due to the fact that there are only few observations in the groups professional and partly skilled, causing collinearity problems in the regression 10

Burbidge et al. (1988) discuss the advantages of the inverse hyperbolic sine compared to an extended Box-Cox transformation.

18

analysis. This leaves me with four occupational groups: • professional, managerial, and technical • skilled non-manual • skilled manual • partly skilled and unskilled (includes those who never worked) The last of these four is the base category. Finally, I also use household size, i.e., the total number of people living in the household. There are a number of variables which I could not use for various reasons. I had to exclude information on whether the respondent was born in the UK or elsewhere. This is due to the fact that only less than 2% of the sample are born outside of the UK. I cannot use information on the region of residence since, unfortunately, this is not included in the ELSA release for reasons of confidentiality. For the same reason, and due to the fact that I effectively use only a single transition, I cannot use the regional or national unemployment rate. I do not include the number of children, which is often used in related analyses for the general working-age population since having to provide childcare is much less likely to be an issue for my age group.

5 5.1

Discussion of the results Exit from work

The regression results from the exit probit regressions are shown in tables 8 for men and 9 for women. As all the following tables, these tables have four columns: the first column reports the results when using all five individual CF variables, the second column the results when excluding the date score which is very likely to be endogenous, the third column the results when using the full CF index, and the fourth column the results when using a CF index which has been computed without the information from the date score. A quick look at table 8 reassures us that the results are highly robust, regardless of which CF variables I use. Essentially, men’s exit from the labour market is determined by wealth, age, health, and housing tenure. First, it can be seen that there are strong non-linear age effects. Those men who are still in work at ages 55–59 have lower exit rates than comparable men aged 50–54. This can most likely be attributed to selection on unobservables—those with high ability 19

remain in work during their early 50s whereas those with low ability are more likely to lose their job at this age. Another interesting selection effect can be observed for men beyond state pension age whose exit rates are markedly lower than for those aged 60–64, and actually declining up to age 70. These are men who are highly attached to the labour market and therefore remain in work until a high age. The effects of non-housing wealth are much smaller than those of housing wealth. The latter must of course be interpreted jointly with the coefficients of housing tenure. For outright owners with high housing wealth, the wealth effect is larger than the tenure effect, implying that these men are more likely to exit the labour market than other men. However, for outright owners with low or average housing wealth, the opposite is true. For owners with a current loan or mortgage, housing wealth has to be very high for the wealth effect to exceed the tenure effect; in other words, such owners whose housing wealth is high but not very high, the overall effect on exit is still negative which is consistent with previous theoretical and empirical evidence. Finally, there is of course the effect of physical health; men with a health limitation are almost ten percentage points more likely to exit between waves 1 and 2. This implies that, on average, their exit probability is more than 50% higher than that of comparable men without such a health limitation. There are no statistically significant effects of education, marital status, occupation, or household size (at the 10% level or lower). However, most coefficients are of the expected sign. This leads to the key variables of interest; those on CF. It can be seen that no CF variable has a statistically significant effect on the predicted probability of working men to exit the labour market as defined earlier on. Table 9 shows a slightly different picture for women. There are no effects of wealth and housing tenure; these appear to operate via male labour supply only. An age effect is only visible for the five years prior to women’s state pension age. The impact of a health limitation is smaller than for men, both in terms of percentage points as well as when related to the observed exit probability which is higher for women than for men. There are, however, statistically significant education effects for women: women with a degree or with ‘other’ qualifications are much less likely to exit than comparable women with different or no qualifications. In terms of the CF variables, the ‘effect’ of the date score is larger than for men (pointing towards the importance of work experience for women) whereas the other coefficients tend to be smaller than for men. As for men, no CF coefficient is statistically significant at the 10% level.

20

5.2

Entry into work

Tables 10 and 11 show the results of the entry regressions. The key problem with these regressions is of course the low number of events as can be seen from the low observed average entry probability. However, some interesting results can be seen. For men, in table 10, only age, physical health, and occupation have statistically significant effects on the entry into work. The age effect is again non-linear; however, from age 60 onwards there is a near-linear decline in the effect of age on the entry probability. Men with a health limitation and/or with a high-grade occupation have very low entry rates. The latter result may be caused by their exit having been a result of choice rather than redundancy. No other explanatory variables have statistically significant effects on men’s entry probability. Interestingly, many CF variables, including the CF indices, have consistently negative effect on the entry probability (though not statistically significant). Again, this could in principle be the result of voluntary exit among high-CF men and involuntary exit among low-CF men even though this is not substantiated by my exit regressions. For women, shown in table 11, entry is even less common. Only age and physical health have statistically significant effects. The effect of physical health is negligible, but is should be kept in mind that this may be a lower bound to the real effect as shown by Bound (1991).

5.3

Change in working hours

Finally, tables 12 and 13 then report the results of the hours change regressions. For reasons discussed earlier on, these regressions are not my preferred specifications, and indeed it can be seen that the (adjusted) R-squared is extremely low. Hardly any of the variance in the change in working hours can be explained by my explanatory variables. The effect sizes in terms of the predicted hours change are mostly small, and may be driven by a few outliers. However, my key result that CF does not appear to have any statistically significant effects on older men and women’s labour force participation in England holds.

5.4

Some remarks on panel attrition

I also controlled for panel attrition using Heckman probit models (exit and entry) or standard Heckman selection models (change in working hours). For my preferred specifications, i.e., for the exit and entry regressions, ρ is never statistically significant, implying that the normal probit models above are to be preferred. For the hours change regressions, 21

ρ is statistically significant, large, and negative, implying that those whose unobservables make them more likely to be retained in the panel have a lower hours change than predicted by observables. Full results of all these selection models can be found in Appendix B.

6

Further tests

Even though there is substantial variation in most of the CF variables, it could still be that measurement error is driving the results. However, they also prevail when using deciles of the CF index, either with or without the date score, on the right-hand side. In other words, not even having very low or very high CF appears to affect work. One interesting question is whether the relationship between CF and work changed over time. This is of course only possible for the levels-levels model with contemporaneous CF on the right-hand side since there are only two ELSA waves available as of yet. When interacting CF with wave, I find that the impact of CF on participation did not change between 2002 and 2004. Heckman and Vytlacil also argue that the interaction of education and CF is more important than the separate effects. This is likely to be related to the fact that CF and education are highly correlated in their US data, which is not true in my UK data. When including such an interaction term in my models, it is hardly ever statistically significant at a reasonable margin. Also, the effect of CF does not change significantly when estimating the models separately by education (or occupation, or education-occupation) group. This is of course only possible for the biggest groups because of sample size limitations. Finally, it is also important to note the difficulties of interpreting the coefficients of interaction terms in binary regression models (cf. Ai and Norton 2003), making it more appealing to refrain from using such interaction terms too often. One potential criticism against this paper is that I analysed mainly the effects of memory (and verbal fluency) on work outcomes whereas earlier studies on the general working-age population used CF variables more similar to intelligence tests. In particular, numeracy and literacy may have strong effects on labour market outcomes. I excluded these variables since they are only available in one single wave. However, it is of course possible to analyse the effects of wave 1 numeracy on the probability to work in wave 1 or wave 2 or on the exit and entry probabilities, and the effects of wave 2 literacy on the probability to work in wave 2. The corresponding regressions do not reveal any statistically significant impact of literacy or numeracy on labour force participation, corroborating my key finding that CF does not appear to have any strong effects on labour market outcomes in the UK. 22

7

Summary and conclusions

There are a number of conclusions which can be drawn from this research, going back to the three research questions stated at the beginning. Summing my results up, there are virtually no effects of the change in CF on the exit from work, the entry into work, or the change in working hours, and only minor effects of the level of CF on the probability to work. However, these effects are restricted to the date score and the prospective memory score. The direction of causality is particularly disputed for the date score. Coming back to the three research questions stated at the beginning of this paper it can be said that there appears to be only a rather weak link between CF and labour market outcomes. This link is present only for the date score and the prospective memory score, and it disappears in models of changes, implying that the link is likely to be caused by a joint process rather than by CF affecting work. It is important to point out that the size of the link is small, perhaps surprisingly small compared to my expectations before embarking on this research. It is also important to emphasise that my results are not driven by my focus on measures of memory and verbal fluency. When including the lagged numeracy score in employment, entry, wage, or earnings regressions (i.e., in all regressions in which I used lagged CF rather than the change in CF), there are no statistically significant effects either. As soon as the third ELSA wave comes out, it will be possible to run regressions including the change in the numeracy score between 2002 and 2006. There are a number of reasons why my results may be observed even in the case where there is an impact of CF on employment. For instance, it may be that the employment change takes place later, implying that a model including lagged explanatory variables would be necessary. This would of course require more ELSA waves. It could also be that ‘normal’ CF decline is accepted by employers in an implicit contract framework. However, since I do not even observe effects for the lowest deciles of CF change, this alone is not likely to be the case. As Banks (2006) writes, not much research has been done on cognition and work among elderly people; this paper is one of the first. Avenues for future research include finding out more about occupation, both about the impact of CF on occupational choice as well as about the CF requirements of different jobs, and, of course, improving the dynamics of the model and the treatment of the apparent endogeneity of CF as more and more ELSA waves will become available.

23

−1

sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Number of obs Log likelihood Observed P

(1) 0.0201* 0.0039** 0.0296** 0.0125* 0.0521*** −0.0598*** 0.0969*** 0.0053 0.0045 −0.0007 −0.0008 −0.0030 0.0140 −0.0046 −0.0311 −0.0116 −0.0082 0.0091 −0.0146 −0.2138** −0.3036** −0.0114 0.0120 −0.0190 −0.0135

1459 −590.1124 0.1851

(2) 0.0201* 0.0039** 0.0296** 0.0126* 0.0521*** −0.0601*** 0.0967*** 0.0045 −0.0006 −0.0008 −0.0030 0.0139 −0.0046 −0.0314 −0.0119 −0.0089 0.0093 −0.0145 −0.2133** −0.3030** −0.0116 0.0114 −0.0190 −0.0133

1459 −590.171 0.1851

(3) 0.0202* 0.0039** 0.0298** 0.0126* 0.0522*** −0.0600*** 0.0961***

(4) 0.0202* 0.0039** 0.0298** 0.0126* 0.0522*** −0.0602*** 0.0962***

0.0117 −0.0043 −0.0313 −0.0109 −0.0060 0.0088 −0.0157 −0.2143** −0.3041** −0.0107 0.0119 −0.0192 −0.0125 0.0001

0.0117 −0.0043 −0.0311 −0.0108 −0.0061 0.0091 −0.0156 −0.2145** −0.3044** −0.0107 0.0120 −0.0191 −0.0125

1459 −590.6019 0.1851

Table 8: Marginal effects from the exit probit model for men.

24

−0.0004 1459 −590.5965 0.1851

−1


(1) −0.0034 0.0032 0.0047 0.0661*** 0.0145 −0.0279 0.0750** −0.0282 −0.0001 −0.0008 0.0003 0.0006 −0.0963** 0.0038 0.0174 −0.0160 0.0157 −0.0881** 0.0398 0.0952 0.0436 −0.0429 −0.0517 0.0294 −0.0191

1216 −563.9808 0.2401

(2) −0.0035 0.0032 0.0046 0.0659*** 0.0140 −0.0275 0.0738** −0.0003 −0.0009 0.0004 0.0007 −0.0948** 0.0069 0.0183 −0.0156 0.0213 −0.0863** 0.0408 0.0977 0.0455 −0.0432 −0.0517 0.0289 −0.0195

1216 −564.5521 0.2401

(3) −0.0033 0.0032 0.0046 0.0657*** 0.0141 −0.0275 0.0744**

(4) −0.0034 0.0032 0.0045 0.0659*** 0.0140 −0.0274 0.0746**

−0.0948** 0.0074 0.0193 −0.0157 0.0197 −0.0849** 0.0406 0.0965 0.0439 −0.0432 −0.0514 0.0303 −0.0193 −0.0023

−0.0943** 0.0078 0.0190 −0.0155 0.0212 −0.0852** 0.0410 0.0966 0.0441 −0.0430 −0.0515 0.0293 −0.0194

1216 −564.5434 0.2401

Table 9: Marginal effects from the exit probit model for women.

25

−0.0006 1216 −564.645 0.2401

−1

sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Number of obs Log likelihood Observed P

(1) 0.0025 0.0004 0.0023 −0.0063* −0.0037 −0.0146** −0.0337*** 0.0069 −0.0012 −0.0009 −0.0004 −0.0026 0.0095 −0.0052 0.0047 −0.0115 −0.0034 −0.0144 −0.0023 0.0277 −0.0243* −0.0029 −0.0014 0.0071

1047 −181.2473 0.0544

(2) 0.0025 0.0003 0.0024 −0.0063* −0.0038 −0.0150** −0.0342*** −0.0012 −0.0008 −0.0002 −0.0027 0.0091 −0.0052 0.0042 −0.0124 −0.0051 −0.0144 −0.0021 0.0278 −0.0234* −0.0011 −0.0004 0.0069

1047 −181.7797 0.0544

(3) 0.0026 0.0003 0.0028 −0.0064* −0.0041 −0.0149** −0.0351***

(4) 0.0025 0.0003 0.0028 −0.0065* −0.0039 −0.0150** −0.0344***

0.0088 −0.0053 0.0035 −0.0128 −0.0070 −0.0138 −0.0022 0.0272 −0.0240* −0.0015 −0.0002 0.0069 −0.0015

0.0093 −0.0048 0.0044 −0.0119 −0.0056 −0.0137 −0.0006 0.0294 −0.0239* −0.0017 −0.0006 0.0069

1047 −182.952 0.0544

Table 10: Marginal effects from the entry probit model for men.

26

−0.0028 1047 −182.2222 0.0544

−1


(1) 0.0023 0.0001 0.0037 −0.0112*** −0.0024 −0.0023 −0.0178** −0.0062 0.0005 −0.0009 0.0020 −0.0003 0.0153 0.0176 −0.0002 0.0084 0.0061 0.0041 −0.0085 −0.0229 −0.0155 −0.0098 0.0152 0.0294 0.0008

1814 −282.7835 0.0430

(2) 0.0023 0.0001 0.0036 −0.0110*** −0.0026 −0.0022 −0.0177** 0.0004 −0.0009 0.0020 −0.0003 0.0148 0.0178 −0.0000 0.0086 0.0054 0.0044 −0.0085 −0.0231 −0.0153 −0.0099 0.0150 0.0293 0.0006

1814 −283.2108 0.0430

(3) 0.0023 0.0001 0.0036 −0.0113*** −0.0024 −0.0025 −0.0174**

(4) 0.0023 0.0001 0.0035 −0.0112*** −0.0025 −0.0025 −0.0173**

0.0153 0.0168 −0.0015 0.0077 0.0063 0.0044 −0.0072 −0.0233 −0.0162 −0.0095 0.0163 0.0293 0.0003 −0.0005

0.0153 0.0167 −0.0012 0.0079 0.0065 0.0046 −0.0071 −0.0236 −0.0163 −0.0097 0.0162 0.0294 0.0002

1814 −284.9902 0.0430

Table 11: Marginal effects from the entry probit model for women.

27

0.0000 1814 −285.05 0.0430

−1

(1) −0.2034 −0.1412*** −0.4244 −0.3129 −0.2094 1.8188*** 1.0698 0.2798 −0.1780 −0.0818 0.0654 0.0383 1.5908 1.2549 0.6789 0.9493 0.1881 0.5181 −1.3226 2.8369 2.2884 −1.4555 −0.7356 −1.3062 0.1763

sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Constant 20.6894 Number of obs 2554 Adjusted R-squared 0.0141

(2) −0.2040 −0.1411*** −0.4288 −0.3065 −0.2123 1.8116*** 1.0543 −0.1767 −0.0804 0.0693 0.0413 1.5889 1.2541 0.6530 0.9268 0.1520 0.5187 −1.3145 2.8485 2.2958 −1.4607 −0.7329 −1.3031 0.1744

20.9249 2554 0.0144

(3) −0.2005 −0.1415*** −0.4374 −0.3094 −0.2161 1.8326*** 1.0153

(4) −0.2014 −0.1413*** −0.4346 −0.3161 −0.2122 1.8285*** 1.0258

1.5441 1.2430 0.5486 0.8594 0.1241 0.4136 −1.3240 2.7878 2.2472 −1.4546 −0.7209 −1.3092 0.1701 −0.0720

1.5380 1.2355 0.5748 0.8800 0.1395 0.4302 −1.3202 2.8149 2.2695 −1.4419 −0.7153 −1.3057 0.1693

21.4871 2554 0.0142

−0.1229 21.3321 2554 0.0144

Table 12: OLS coefficients from the hours change model for men.

28

−1

(1) 0.2075** −0.0062 0.1125 −1.0308*** 0.8872*** 0.1776 1.2069*** 0.2900 −0.0157 −0.0104 0.0248 −0.0158 0.0944 −0.0673 −1.3845* −0.2014 0.2396 0.8205 0.2752 −3.2071*** −4.4190*** −1.0386* −0.2861 −1.7984*** 0.3590

sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Constant −6.5073 Number of obs 3030 Adjusted R-squared 0.0303

(2) 0.2077** −0.0064 0.1153 −1.0338*** 0.8916*** 0.1752 1.2057*** −0.0153 −0.0095 0.0238 −0.0160 0.1008 −0.0754 −1.3806* −0.1937 0.2198 0.8174 0.2660 −3.1997*** −4.4128*** −1.0420* −0.2861 −1.7977*** 0.3615

−6.6507 3030 0.0304

(3) 0.2075** −0.0062 0.1125 −1.0335*** 0.8943*** 0.1719 1.2205***

(4) 0.2075** −0.0061 0.1144 −1.0356*** 0.8947*** 0.1723 1.2159***

0.0965 −0.0852 −1.3872* −0.1983 0.2283 0.8203 0.2738 −3.1922*** −4.4059*** −1.0350* −0.2822 −1.8001*** 0.3605 0.0074

0.0964 −0.0820 −1.3839* −0.2012 0.2220 0.8205 0.2713 −3.1947*** −4.4127*** −1.0367* −0.2816 −1.7964*** 0.3619

−6.5264 3030 0.0313

−0.0154 −6.6089 3030 0.0313

Table 13: OLS coefficients from the hours change model for women.

29

Appendix A: Effects of the CF level on work −1

sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation l(Date score) l(Immediate recall) l(Animal names) l(Prospective memory) l(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size Number of obs Log likelihood Observed P

(Men) (Women) 0.0018 0.0039 −0.0030 −0.0018 0.0447 0.0072 −0.0421*** −0.0473*** −0.0626*** −0.0644*** −0.0841*** −0.0270*** −0.3800*** −0.2016*** 0.0382 0.0459** −0.0000 0.0100 0.0013 −0.0014 0.0061 0.0108* −0.0013 −0.0061 −0.0419 0.1015** −0.0333 0.0261 0.0126 0.0374 −0.0189 0.0264 −0.0197 0.0321 −0.0739 0.0922*** 0.0781** −0.0738*** 0.0469 −0.0247 0.2519*** 0.1194** −0.0025 0.1273*** −0.0319 0.1242*** 0.0550 0.0960*** 0.0173 −0.0047 2586 3054 −1177.507 −1359.118 0.4849 0.3294

Table 14: Marginal effects from the levels probit model (dependent variable: whether individual works at least 16 hours or not). In table 14, I report the results of the levels equation which links the level of CF to the probability to work. Of course, the CF coefficients there are the same as those in the last column of table 2. UPDATE! For men, no CF coefficient is statistically significant, whereas for women, I find statistically significant coefficients for the date score and for the prospective memory score. However, because of the endogeneity of the date score and because of the fact that the prospective memory score is statistically significant only at the 10% level, we have to be careful not to attach to much importance to these results. Furthermore, it is also important to note that, even though there are two statistically 30

significant coefficients in the regression for women, the effect size is quite small considering that the date score only ranges from 0 to 4 and the prospective memory score from 0 to 5. Compared to most of the control variables, the impact of CF is small.

31

Appendix B: Controlling for panel attrition As mentioned earlier on, I found strong evidence for increased occurrence of panel attrition among those with low CF scores. If the process of attrition is linked to the process of the exit from work, or the process of changing working hours, this may lead to a bias in my coefficients. Therefore, I also carry out Heckman probit models for the exit from and the entry into work and standard Heckman selection models for the change in working hours. The selection equation is specified in a rather parsimonious way and includes only age, health, CF in wave 1, and the four additional variables about providing a contact address, agreeing to record linkage, document consulting, and interviewer’s response reliability assessment. I argue that these variables indicate how seriously the respondent took the survey, and therefore how likely it is that he or she will continue to participate in wave 2. I also argue that these variables are not linked to the labour market outcome under consideration. Tables 15 and 16 report the results of the Heckman exit probit models for men and women. The top panel displays the exit equation, the bottom panel the selection equation (1 if present in wave 2, 0 if not). First, it is important to note that ρ is never statistically significant, implying that the two equations are independent. This is also shown by the R test of independent equations which is usually not statistically significant. Second, we can see that none of the CF variables is statistically significant in the exit equations, implying that the key result is the same as when not controlling for attrition. In other words, I do find that attrition is explained by the level of CF, but I do not find effects of the change in CF on the probability to exit work, nor do I find a link between attrition and exit. Tables 17 and 18 show the results of the Heckman selection model for the change in working hours for men and women. Again, there is no statistically significant effect of CF on the change in working hours. However, ρ is statistically significant in women’s regressions when using the individual CF variables. It has a negative value, implying that individuals with unobserved characteristics which make them more likely to increase their working hours are also individuals with unobserved characteristics which make them less likely to participate in the second wave (and vice versa). Even though this is an interesting result in its own right it does not bear any importance for my research aims in this study.

32

Exit from work equation sinh−1 (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Sample retention equation Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation Date score Immediate recall Animal names Prospective memory Delayed recall Resp consulted docs Response reliability Contact address given Record linkage agreed CF index CF index w/o date Number of obs Log likelihood ρ

(1)

(2)

(3)

(4)

0.0124* 0.0026** 0.0181** 0.0050 0.0277*** −0.0652*** 0.0270 0.0041 0.0015 −0.0010 −0.0008 −0.0016 0.0146 0.0000 −0.0164 −0.0030 −0.0127 0.0086 −0.0075 −0.1458* −0.1750** 0.0015 0.0159 −0.0039 −0.0097

0.0129* 0.0027** 0.0185** 0.0053 0.0289*** −0.0655*** 0.0294

0.0135* 0.0028** 0.0189** 0.0057 0.0299*** −0.0659*** 0.0321

0.0139* 0.0028** 0.0192** 0.0057 0.0306*** −0.0661*** 0.0336

0.0132 −0.0002 −0.0176 −0.0026 −0.0100 0.0103 −0.0096 −0.1559* −0.1864** 0.0015 0.0172 −0.0053 −0.0096 −0.0010

0.0133 −0.0008 −0.0179 −0.0029 −0.0104 0.0103 −0.0096 −0.1598* −0.1909** 0.0016 0.0176 −0.0050 −0.0098

0.0016 −0.0010 −0.0008 −0.0015 0.0148 −0.0006 −0.0173 −0.0035 −0.0136 0.0091 −0.0079 −0.1507* −0.1807** 0.0013 0.0158 −0.0043 −0.0097

−0.0015 −0.0141 −0.0128 −0.0174 −0.1776*** −0.1867*** 0.0149 0.0231** 0.0093*** 0.0071 −0.0057 0.0451** −0.0376* 0.0633*** 0.0751**

2336 −1667.584 0.7449

−0.0140 −0.0131 −0.0173 −0.1777*** −0.1869*** 0.0232** 0.0093*** 0.0072 −0.0052 0.0455** −0.0372* 0.0634*** 0.0761**

2336 −1667.9 0.7030

−0.0153 −0.0123 −0.0190* −0.1749*** −0.1861***

−0.0149 −0.0127 −0.0187* −0.1751*** −0.1850***

0.0471** −0.0368* 0.0635*** 0.0720** 0.0211***

0.0471** −0.0368* 0.0634*** 0.0728**

2336 −1674.153 0.6608

0.0235*** 2336 −1673.766 0.6370

Table 15: Marginal effects from the Heckman exit probit model for men.

33

Exit from work equation sinh−1 (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Sample retention equation Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation Date score Immediate recall Animal names Prospective memory Delayed recall Resp consulted docs Response reliability Contact address given Record linkage agreed CF index CF index w/o date Number of obs Log likelihood ρ

(1)

(2)

(3)

(4)

−0.0053 0.0030 0.0025 0.0540*** 0.0055 −0.0584** 0.0437 −0.0277 −0.0035 −0.0009 −0.0009 0.0014 −0.0728* 0.0163 0.0140 −0.0013 0.0186 −0.0734** 0.0342 0.1112* 0.0658 −0.0341 −0.0434 0.0159 −0.0165

−0.0053 0.0029 0.0024 0.0524*** 0.0047 −0.0585** 0.0401

−0.0055 0.0031 0.0024 0.0563*** 0.0059 −0.0560** 0.0478

−0.0053 0.0030 0.0023 0.0530*** 0.0049 −0.0580** 0.0417

−0.0741* 0.0204 0.0162 −0.0008 0.0240 −0.0730** 0.0365 0.1166* 0.0682 −0.0366 −0.0458 0.0181 −0.0172 −0.0040

−0.0690* 0.0205 0.0158 0.0005 0.0251 −0.0692** 0.0349 0.1110* 0.0653 −0.0335 −0.0427 0.0162 −0.0165

−0.0037 −0.0010 −0.0009 0.0014 −0.0693* 0.0190 0.0147 −0.0005 0.0237 −0.0702** 0.0343 0.1107* 0.0659 −0.0333 −0.0422 0.0152 −0.0167

−0.0025 0.0077 −0.0214** −0.0314** −0.1603*** −0.1587*** 0.0444 0.0277*** 0.0048** 0.0215*** 0.0140 0.0462** −0.0139 0.0509** 0.1380***

2158 −1587.646 0.4179

0.0073 −0.0218** −0.0316** −0.1602*** −0.1598*** 0.0279*** 0.0049** 0.0225*** 0.0146* 0.0470** −0.0137 0.0502* 0.1375***

2158 −1589.132 0.4424

0.0076 −0.0211** −0.0312** −0.1608*** −0.1579***

0.0070 −0.0215** −0.0315** −0.1606*** −0.1591***

0.0463** −0.0144 0.0518** 0.1377*** 0.0340***

0.0474** −0.0141 0.0505** 0.1375***

2158 −1588.678 0.3710

0.0360*** 2158 −1589.507 0.4307

Table 16: Marginal effects from the Heckman exit probit model for women.

34

d(hours) equation −1 sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Constant Sample retention equation Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation Date score Immediate recall Animal names Prospective memory Delayed recall Resp consulted docs Response reliability Contact address given Record linkage agreed CF index CF index w/o date Constant Number of obs Log likelihood ρ

(1)

(2)

(3)

(4)

−0.1618 −0.0915* −0.4592 0.0190 −0.8851*** 4.6504*** 2.4056*** 0.2591 0.0602 −0.0307 0.0767 −0.0698 1.0576 0.9220 0.0402 0.0437 0.7614 −0.3252 −1.7281** 2.7414 2.7698 −1.6672* −0.5280 −0.9732 0.4083

−0.1640 −0.0902* −0.4643 0.0232 −0.8920*** 4.6499*** 2.3826***

−0.1626 −0.0876* −0.4589 0.0155 −0.8924*** 4.6728*** 2.3816***

−0.1651 −0.0893* −0.4642 0.0182 −0.8874*** 4.6567*** 2.3682***

1.0353 0.9352 −0.0518 −0.0236 0.7783 −0.4817 −1.6404** 2.7249 2.7148 −1.6022* −0.4209 −0.9269 0.4202 0.0583

1.0035 0.9070 −0.0546 −0.0122 0.7636 −0.4910 −1.6695** 2.7827 2.7976 −1.6715* −0.5098 −0.9835 0.4232

25.3604

25.6445

25.2371

0.0200 25.5909

0.0003 −0.0087 0.0867*** −0.4071*** −0.1807*** −0.0405 0.0426** 0.0145*** 0.0061 0.0089 0.0782** −0.0033 0.0638 0.1712***

0.0004 −0.0080 0.0863*** −0.4064*** −0.1790***

−0.0025 −0.0067 0.0818*** −0.4032*** −0.1810***

−0.0022 −0.0073 0.0838*** −0.4041*** −0.1777***

0.0807*** −0.0117 0.0692* 0.1724*** 0.0371***

0.0789*** −0.0100 0.0670 0.1688***

2.8394*** 3366 −11512.61 −0.8566

2.8391*** 3366 −11513.48 −0.8557

0.0551 −0.0292 0.0818 −0.0635 1.0213 0.8969 −0.0185 −0.0127 0.7174 −0.3502 −1.7000** 2.7548 2.7596 −1.6597* −0.4911 −0.9574 0.4061

0.0418** 0.0143*** 0.0058 0.0082 0.0783** −0.0051 0.0651 0.1724***

2.8396*** 3366 −11521.57 −0.8535

0.0452*** 2.8395*** 3366 −11518.21 −0.8545

Table 17: Heckman selection model coefficients from the hours change model for men.

35

d(hours) equation −1 sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Constant Sample retention equation Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation Date score Immediate recall Animal names Prospective memory Delayed recall Resp consulted docs Response reliability Contact address given Record linkage agreed CF index CF index w/o date Constant Number of obs Log likelihood ρ

(1)

(2)

(3)

(4)

0.2328*** 0.0061 0.0934 −1.0105*** 0.3754** 2.2207*** 1.4468*** 0.3402 0.0784 0.0085 0.0943 −0.0068 −0.5675 −0.2789 −1.6055* −0.6087 0.2438 0.6902 0.0716 −3.6638*** −4.4886*** −0.9952 −0.5205 −1.2929* 0.2668

0.2340*** 0.0060 0.0950 −1.0123*** 0.3790** 2.2199*** 1.4416***

0.2289*** 0.0068 0.0941 −1.0131*** 0.3831** 2.2150*** 1.4727***

0.2310*** 0.0071 0.0949 −1.0153*** 0.3843** 2.2121*** 1.4664***

0.0796 0.0099 0.0952 −0.0052 −0.5554 −0.2855 −1.5971* −0.5995 0.2172 0.6861 0.0588 −3.6725*** −4.4958*** −1.0014 −0.5141 −1.2969* 0.2675

−0.5043 −0.2503 −1.5682* −0.5944 0.2321 0.6957 0.0736 −3.5888*** −4.4407*** −0.9731 −0.5173 −1.2754* 0.2860 0.1143

−0.5134 −0.2619 −1.5781* −0.6104 0.2112 0.6876 0.0620 −3.6192*** −4.4729*** −0.9829 −0.5031 −1.2811* 0.2823

−2.0468

−2.1202

−2.1871

0.1019 −2.1930

0.0573** −0.0159 0.0967*** −0.4117*** −0.1051** 0.0378 0.0264 0.0154*** 0.0189 0.0170 0.0554* 0.0506 0.1254*** 0.1008*

0.0571** −0.0162 0.0965*** −0.4116*** −0.1063**

0.0571** −0.0165 0.0965*** −0.4117*** −0.1049**

0.0564** −0.0166 0.0963*** −0.4111*** −0.1062**

0.0556* 0.0526 0.1241*** 0.1020* 0.0470***

0.0571* 0.0511 0.1231*** 0.0996*

−3.2233** 3824 −12463.73 −0.7900

0.0268 0.0155*** 0.0196 0.0177 0.0560* 0.0501 0.1250*** 0.0996*

−3.0772** 3824 −12464.25 −0.7898

−2.4553* 3824 −12467.06 −0.7901

0.0516*** −2.4059* 3824 −12466.77 −0.7889

Table 18: Heckman selection model coefficients from the hours change model for women.

36

Appendix C: Wage and earnings regressions In line with Anger and Heineck (2006), Cawley, Heckman, and Vytlacil (2001), Heckman and Vytlacil (2001), and Pryor and Schaffer (2000), I also use (log) wages and (log) earnings as dependent variables. I use monthly RPI data from National Statistics to convert all amounts into January 2005 pounds.11 Since I also include the self-employed in my analysis, earnings and thus the wage rate may be negative. This implies that the logarithm is not defined for these observations. Therefore, I use the inverse hyperbolic sine instead of the logarithm. The results are not affected by whether negative values are excluded or not. The data on wages and earnings come from the derived financial variable files of ELSA and contain imputed data, implying that item non-response is not a problem. However, there is of course the usual problem of selection into work which is why it may be useful to estimate Heckman selection models rather than standard OLS regressions. As expected, the explanatory power of these models is much lower than that of the participation models, and than that of income regressions for the general working-age population. Tables 19 and 20 show the results of OLS regressions explaining the absolute change in the wage rate for men and women. Again, I do not find any statistically significant effects of CF. The wage change for this age group seems to be driven just by education. Also for the wage models it may be important to control for attrition. Tables 25 and 26 show the results of Heckman selection models explaining the absolute wage change. I find that for men, the difference in delayed recall has an impact on the wage change. In other words, every decline in the delayed recall of the word list is associated with fall of the hourly wage rate by 50 pence. Consider two otherwise identical men, both of whom scored 4 out of 4 on the delayed recall in wave 1. In wave 2, person A is again able to remember all four words whereas person B only remembered two words. Therefore, person B is predicted to experience a wage change that is one pound lower than that of person A. For women, I do not find such an effect. Moreover, for both sexes, ρ is not statistically significant, implying that the wage change model may be estimated separately from the attrition model.

11

The time series used is called CHAW. This is the full RPI including all components such as housing and mortgage repayments.

37

−1

(1) 0.1745 0.0405 0.2112 −0.1394 0.5664 −1.3977 −0.7302 0.3705 −0.4448 0.0289 0.0091 0.4465 3.2470** −0.6758 0.3271 0.1832 0.5923 0.0366 1.0946 −0.8572 −2.6289 0.9422 0.1901 −0.1219 −0.4348

sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Constant −11.1880 Number of obs 1281 Adjusted R-squared 0.0016

(2) 0.1739 0.0422 0.2012 −0.1293 0.5612 −1.4132 −0.7535 −0.4447 0.0302 0.0112 0.4544 3.2582** −0.6536 0.3232 0.1891 0.5865 0.0691 1.1222 −0.8272 −2.6045 0.8950 0.1254 −0.1442 −0.4333

−10.6989 1281 0.0022

(3) 0.1627 0.0361 0.2061 −0.1313 0.5769 −1.4242 −0.8183

(4) 0.1620 0.0369 0.2024 −0.1287 0.5743 −1.4313 −0.8207

3.3569** −0.5877 0.2659 0.0531 0.4706 0.0363 1.1822 −0.6767 −2.4455 0.9183 0.0560 −0.1350 −0.4587 0.0836

3.3522** −0.5902 0.2619 0.0578 0.4588 0.0620 1.1969 −0.6538 −2.4246 0.8999 0.0225 −0.1446 −0.4590

−10.8891 1281 0.0018

0.0566 −10.7009 1281 0.0017

Table 19: OLS coefficients from the absolute wage change model for men.

38

−1

(1) −0.2669 −0.0210 0.1639 0.4266 0.2483 −0.4953 1.1781 −0.8067 −0.2491 −0.0254 −0.0917 0.2763 2.5453 1.6695 1.2151 1.7419 1.5470 0.0508 −0.0528 3.6119 4.1330 −0.2301 −1.4888 −1.7590 −0.5066

sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Constant −8.7263 Number of obs 1215 Adjusted R-squared −0.0037

(2) −0.2755 −0.0225 0.1533 0.4250 0.2485 −0.5034 1.1275 −0.2539 −0.0286 −0.0904 0.2817 2.5844 1.7304 1.2102 1.7278 1.6696 0.0624 −0.0278 3.7423 4.2510* −0.2340 −1.4543 −1.7307 −0.5133

−8.2070 1215 −0.0034

(3) −0.2666 −0.0201 0.1506 0.4329 0.2334 −0.4661 1.1387

(4) −0.2689 −0.0209 0.1441 0.4382 0.2314 −0.4654 1.1319

2.6521 1.8010 1.2149 1.7689 1.7573 0.1956 −0.0581 3.6337 4.1217 −0.2240 −1.4600 −1.7110 −0.4712 −0.0938

2.6540 1.8150 1.2000 1.7752 1.7966 0.1943 −0.0463 3.6605 4.1486 −0.2090 −1.4470 −1.7066 −0.4736

−8.1831 1215 −0.0020

−0.0501 −7.8741 1215 −0.0021

Table 20: OLS coefficients from the absolute wage change model for women.

39

−1

(1) 0.0161* 0.0014 −0.0021 −0.0243 0.0276 −0.0679** −0.1870** 0.0717 0.0277 −0.0053 0.0290* 0.0023 0.3789*** 0.0659 0.2130** 0.0688 0.1619 0.0219 0.1885** −0.0768 −0.1356 0.2437** 0.1803 0.0335 0.0219

sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation l(Date score) l(Immediate recall) l(Animal names) l(Prospective memory) l(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size l(CF index) l(CF index w/o date) Constant 1.6682 Number of obs 1357 Adjusted R-squared 0.0665

(2) 0.0161* 0.0012 −0.0016 −0.0242 0.0274 −0.0681** −0.1889** 0.0276 −0.0051 0.0286* 0.0044 0.3843*** 0.0711 0.2189** 0.0759 0.1649 0.0245 0.1853** −0.0795 −0.1360 0.2497** 0.1883 0.0329 0.0209

1.9049 1357 0.0662

(3) 0.0165* 0.0010 −0.0022 −0.0230 0.0289 −0.0701** −0.1835**

(4) 0.0165* 0.0009 −0.0012 −0.0232 0.0287 −0.0704** −0.1849**

0.3721*** 0.0648 0.2149** 0.0699 0.1474 0.0283 0.1836** −0.0823 −0.1415 0.2378** 0.1803 0.0293 0.0195 0.0189*

0.3784*** 0.0696 0.2206** 0.0751 0.1496 0.0296 0.1821** −0.0831 −0.1414 0.2426** 0.1855 0.0297 0.0190

2.1263 1357 0.0660

0.0176 2.0717 1357 0.0653

Table 21: OLS coefficients from the sinh−1 (wage) model for men.

40

−1

(1) −0.0048 −0.0005 −0.0209 0.0065 0.0050 −0.0446 0.0120 0.0898 0.0202 −0.0021 −0.0007 −0.0002 0.4372*** 0.2779*** 0.1060 0.0089 0.0421 0.1447 −0.0336 0.1655 0.1356 0.2641*** 0.0471 −0.0626 0.0005

sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation l(Date score) l(Immediate recall) l(Animal names) l(Prospective memory) l(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size l(CF index) l(CF index w/o date) Constant 2.7437 Number of obs 1299 Adjusted R-squared 0.0588

(2) −0.0052 −0.0004 −0.0208 0.0059 0.0049 −0.0450 0.0094 0.0204 −0.0019 0.0003 0.0009 0.4435*** 0.2848*** 0.1085 0.0096 0.0484 0.1468 −0.0350 0.1699 0.1436 0.2652*** 0.0530 −0.0617 0.0006

3.0673 1299 0.0585

(3) −0.0059 −0.0006 −0.0216 0.0063 0.0053 −0.0435 0.0124

(4) −0.0060 −0.0005 −0.0220 0.0061 0.0051 −0.0439 0.0122

0.4341*** 0.2769*** 0.1012 0.0053 0.0454 0.1456 −0.0325 0.1798 0.1541 0.2619*** 0.0503 −0.0645 −0.0015 0.0098

0.4408*** 0.2827*** 0.1062 0.0092 0.0477 0.1474 −0.0330 0.1822 0.1568 0.2652*** 0.0542 −0.0626 −0.0012

3.2020 1299 0.0603

0.0072 3.2196 1299 0.0600

Table 22: OLS coefficients from the sinh−1 (wage) model for women.

41

−1

(1) 0.0123 −0.0038 0.0524 −0.0677* 0.0125 −0.1482** −0.4148*** 0.1701 0.0078 −0.0118 0.0380 0.0059 0.3054* 0.0105 0.2371 0.0928 0.2753 0.1165 0.3843*** −0.0595 −0.1310 0.2784 0.2670 0.0644 0.0080 0.0260***

sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation l(Date score) l(Immediate recall) l(Animal names) l(Prospective memory) l(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size Weekly working hours l(CF index) l(CF index w/o date) Constant 1.0093 Number of obs 1357 Adjusted R-squared 0.0795

(2) 0.0124 −0.0044 0.0534 −0.0674* 0.0120 −0.1489** −0.4195*** 0.0075 −0.0112 0.0371 0.0110 0.3179* 0.0229 0.2510 0.1094 0.2824 0.1226 0.3768*** −0.0662 −0.1324 0.2929* 0.2857 0.0634 0.0057 0.0259***

1.5771 1357 0.0785

(3) 0.0133 −0.0046 0.0490 −0.0645* 0.0140 −0.1523** −0.4096***

(4) 0.0132 −0.0047 0.0522 −0.0656* 0.0132 −0.1537** −0.4128***

0.2902* 0.0069 0.2417 0.1034 0.2622 0.1324 0.3719*** −0.0706 −0.1409 0.2701 0.2727 0.0534 0.0032 0.0260*** 0.0143

0.3070* 0.0186 0.2558 0.1130 0.2656 0.1334 0.3711*** −0.0709 −0.1405 0.2815 0.2822 0.0570 0.0026 0.0258***

1.8206 1357 0.0787

0.0063 1.6535 1357 0.0784

Table 23: OLS coefficients from the sinh−1 (earnings) model for men.

42

−1

(1) −0.0119 −0.0083 −0.0618 −0.0097 −0.0418 −0.0676 −0.0785 0.0999 0.0008 −0.0050 0.0122 0.0092 0.5186*** 0.2352 0.1052 0.0465 0.0986 0.2514* −0.1215 0.2881 0.3137 0.4181*** 0.1808 −0.0453 −0.0072 0.0325***

sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation l(Date score) l(Immediate recall) l(Animal names) l(Prospective memory) l(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size Weekly working hours l(CF index) l(CF index w/o date) Constant 7.1313 Number of obs 1299 Adjusted R-squared 0.1580

(2) −0.0123 −0.0083 −0.0617 −0.0104 −0.0419 −0.0681 −0.0813 0.0011 −0.0048 0.0133 0.0104 0.5256*** 0.2429 0.1080 0.0472 0.1056 0.2537* −0.1229 0.2931 0.3225 0.4191*** 0.1873 −0.0445 −0.0072 0.0325***

7.4914 1299 0.1582

(3) −0.0131 −0.0084 −0.0606 −0.0098 −0.0404 −0.0673 −0.0846

(4) −0.0132 −0.0084 −0.0611 −0.0099 −0.0407 −0.0677 −0.0846

0.5012*** 0.2190 0.0924 0.0399 0.0990 0.2470 −0.1226 0.2974 0.3305 0.4147*** 0.1838 −0.0519 −0.0093 0.0325*** 0.0074

0.5088*** 0.2254 0.0984 0.0446 0.1012 0.2490 −0.1232 0.3000 0.3331 0.4188*** 0.1880 −0.0495 −0.0090 0.0325***

7.4535 1299 0.1598

0.0040 7.4756 1299 0.1597

Table 24: OLS coefficients from the sinh−1 (earnings) model for women.

43

d(wage) equation (abs) sinh−1 (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Constant Sample retention equation Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation Date score Immediate recall Animal names Prospective memory Delayed recall Resp consulted docs Response reliability Contact address given Record linkage agreed CF index CF index w/o date Constant Number of obs Log likelihood ρ

(1)

(2)

(3)

(4)

−0.0025 0.0175 −0.0556 −0.4936 −0.5296 −6.4113*** −6.4631*** −0.0575 0.3511 −0.0665 −0.0868 −0.0667 0.7264 0.8345 0.9898 0.8039 1.0336 0.6552 0.6168 −0.0452 0.1003 −0.1467 −0.5719 −0.1932 −0.3038

−0.0015 0.0165 −0.0537 −0.4947 −0.5266 −6.4105*** −6.4573***

−0.0058 0.0070 −0.0804 −0.4752 −0.5371 −6.3992*** −6.4233***

−0.0058 0.0112 −0.0775 −0.4793 −0.5364 −6.3909*** −6.4165***

0.3501 −0.0671 −0.0905 −0.0693 0.7449 0.8370 1.0051 0.8106 1.0287 0.6553 0.5962 −0.0510 0.0922 −0.1420 −0.5552 −0.1900 −0.3047

0.5897 0.7746 0.9602 0.8514 1.0913 0.6265 0.4244 0.0169 0.1958 0.0138 −0.3438 −0.1300 −0.2831 −0.0486

0.5838 0.7807 0.9832 0.8157 1.1472 0.5917 0.4865 0.0016 0.1732 0.0486 −0.3184 −0.0842 −0.2743

−0.2049

−0.2926

−0.0316 −0.0140 −0.0574** −0.3765*** −0.4534*** −0.0089 −0.0018 0.0108*** 0.0222* 0.0115 0.1574*** −0.0199 −0.0208 0.0789*

−0.0318 −0.0139 −0.0576** −0.3764*** −0.4533***

2.9372*** 2165 −5902.26 0.9791

2.9371*** 2165 −5902.303 0.9791

−0.0018 0.0107*** 0.0221* 0.0111 0.1580*** −0.0200 −0.0209 0.0792*

1.1504

−0.0157 0.8789

−0.0314 −0.0130 −0.0589** −0.3746*** −0.4481***

−0.0302 −0.0132 −0.0584** −0.3750*** −0.4471***

0.1576*** −0.0191 −0.0224 0.0819* 0.0259***

0.1559*** −0.0196 −0.0218 0.0779*

2.9362*** 2165 −5906.323 0.9777

0.0295*** 2.9358*** 2165 −5905.344 0.9778

Table 25: Heckman selection model coefficients from the absolute wage change model for men. 44

d(wage) equation (abs) sinh−1 (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation d(Date score) d(Immediate recall) d(Animal names) d(Prospective memory) d(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size d(CF index) d(CF index w/o date) Constant Sample retention equation Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation Date score Immediate recall Animal names Prospective memory Delayed recall Resp consulted docs Response reliability Contact address given Record linkage agreed CF index CF index w/o date Constant Number of obs Log likelihood ρ

(1)

(2)

(3)

(4)

−0.3358* −0.0206 0.1808 0.3747 0.2317 −0.8467 0.9456 −0.8322 −0.2696 −0.0310 −0.1109 0.2389 2.9611* 2.0207 1.4341 1.9783 1.8105 0.1185 −0.0714 4.4113 4.9380* −0.3503 −1.4943 −1.7097 −0.4573

−0.3456* −0.0224 0.1706 0.3698 0.2316 −0.8687 0.8817

−0.3378* −0.0199 0.1737 0.3794 0.2217 −0.8036 0.9088

−0.3395* −0.0209 0.1671 0.3786 0.2183 −0.8339 0.8756

3.0788* 2.1687 1.4345 2.0139 2.0000 0.2666 −0.0835 4.4482 4.9484* −0.3617 −1.4737 −1.6620 −0.4230 −0.1353

3.0882* 2.1911 1.4265 2.0222 2.0503 0.2673 −0.0747 4.4717 4.9682* −0.3398 −1.4554 −1.6519 −0.4249

−0.2755 −0.0345 −0.1109 0.2448 3.0047* 2.0869 1.4379 1.9609 1.9405 0.1284 −0.0547 4.5470 5.0611* −0.3567 −1.4574 −1.6802 −0.4631

−9.7049

−9.9478

−0.0965 −9.6781

0.0430 −0.0958*** −0.0419 −0.3699*** −0.3857*** 0.0969 0.0613** 0.0141** 0.0497*** 0.0336 0.0871* −0.0213 0.1094* 0.4184***

0.0421 −0.0967*** −0.0421 −0.3696*** −0.3899***

0.0428 −0.0951*** −0.0419 −0.3709*** −0.3836***

0.0416 −0.0963*** −0.0420 −0.3702*** −0.3892***

0.0874* −0.0216 0.1103* 0.4195*** 0.0818***

0.0906* −0.0205 0.1076 0.4152***

2.7255*** 2157 −5923.852 0.0729

2.7259*** 2157 −5924.897 0.0757

−10.2099

0.0618** 0.0144** 0.0518*** 0.0355 0.0896* −0.0208 0.1080* 0.4157***

2.7261*** 2157 −5925.039 0.0695

0.0874*** 2.7264*** 2157 −5925.655 0.0760

Table 26: Heckman selection model coefficients from the absolute wage change model for women. 45

sinh−1 (Wage) −1 sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation l(Date score) l(Immediate recall) l(Animal names) l(Prospective memory) l(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size l(CF index) l(CF index w/o date) Constant select Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation Date score Immediate recall Animal names Prospective memory Delayed recall l(Care hours) CF index CF index w/o date Constant Number of obs Log likelihood ρ

(1)

(2)

(3)

(4)

0.0176** 0.0023 −0.0224 0.0023 0.1048*** 0.0357 0.2470* 0.0644 0.0269 −0.0072 0.0266 0.0018 0.3750*** 0.0358 0.2010* 0.0725 0.1351 0.0254 0.1781** −0.0930 −0.1474 0.2552*** 0.2080* 0.0462 0.0250

0.0178** 0.0022 −0.0233 0.0044 0.1090*** 0.0422 0.2731**

0.0180** 0.0018 −0.0223 0.0030 0.1035*** 0.0301 0.2368

0.0181** 0.0019 −0.0224 0.0044 0.1069*** 0.0353 0.2585*

0.3718*** 0.0386 0.2055** 0.0765 0.1234 0.0312 0.1737** −0.1013 −0.1569 0.2499** 0.2068* 0.0413 0.0220 0.0152

0.3809*** 0.0441 0.2112** 0.0823 0.1237 0.0350 0.1697** −0.1025 −0.1578 0.2560*** 0.2133* 0.0431 0.0221

2.9472 0.0678 −0.0832*** −0.1735*** −0.1812*** −1.0129*** 0.0480 0.0086 0.0096** −0.0068 0.0129 −0.0023***

0.0858* 2586 −3130.347 −0.6540

0.0272 −0.0071 0.0260 0.0031 0.3839*** 0.0419 0.2069** 0.0805 0.1361 0.0318 0.1714** −0.0964 −0.1491 0.2632*** 0.2180* 0.0480 0.0248

3.2340 0.0637 −0.0807*** −0.1727*** −0.1814*** −1.0073*** 0.0094 0.0099** −0.0053 0.0132 −0.0022***

0.0975** 2586 −3131.278 −0.6827

3.3049

0.0134 3.3116

0.0654 −0.0836*** −0.1750*** −0.1803*** −1.0160***

0.0619 −0.0816*** −0.1743*** −0.1806*** −1.0118***

−0.0023*** 0.0227**

−0.0023***

0.0811 2586 −3134.371 −0.6356

0.0238** 0.0905* 2586 −3135.044 −0.6602

Table 27: Heckman selection model coefficients from the sinh−1 (wage) model for men.

46

sinh−1 (Wage) −1 sinh (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation l(Date score) l(Immediate recall) l(Animal names) l(Prospective memory) l(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size l(CF index) l(CF index w/o date) Constant select Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation Date score Immediate recall Animal names Prospective memory Delayed recall l(Care hours) CF index CF index w/o date Constant Number of obs Log likelihood ρ

(1)

(2)

(3)

(4)

−0.0045 −0.0008 −0.0253 0.0079 −0.0045 −0.0417 −0.0207 0.1078 0.0193 −0.0020 −0.0011 0.0013 0.4327*** 0.2920*** 0.1169 0.0178 0.0518 0.1564* −0.0318 0.1641 0.1345 0.2587*** 0.0374 −0.0662 −0.0025

−0.0050 −0.0008 −0.0247 0.0063 −0.0061 −0.0434 −0.0285

−0.0056 −0.0009 −0.0250 0.0063 −0.0064 −0.0416 −0.0276

−0.0057 −0.0008 −0.0255 0.0061 −0.0066 −0.0421 −0.0278

0.4298*** 0.2920*** 0.1111 0.0131 0.0557 0.1572* −0.0306 0.1781 0.1535 0.2563*** 0.0415 −0.0684 −0.0048 0.0116

0.4375*** 0.2987*** 0.1168 0.0175 0.0584 0.1593* −0.0312 0.1809 0.1566 0.2601*** 0.0459 −0.0662 −0.0045

2.9052

0.0197 −0.0017 0.0003 0.0026 0.4397*** 0.2998*** 0.1192 0.0180 0.0590 0.1584* −0.0330 0.1689 0.1436 0.2601*** 0.0446 −0.0653 −0.0027

3.2636

0.0819 −0.1303*** −0.2020*** −0.0756*** −0.7112*** 0.2895*** 0.0413** −0.0005 0.0541*** −0.0103 −0.0031***

0.0828 −0.1309*** −0.2025*** −0.0743*** −0.7135***

−0.0895*** 3084 −3277.208 0.0235

−0.0884*** 3084 −3286.181 0.0368

0.0432** 0.0005 0.0552*** −0.0047 −0.0031***

3.3783

0.0083 3.4011

0.0840 −0.1304*** −0.1987*** −0.0746*** −0.6987***

0.0844 −0.1309*** −0.1998*** −0.0748*** −0.7032***

−0.0031*** 0.0419***

−0.0031***

−0.0882*** 3084 −3286.549 0.0415

0.0369*** −0.0880*** 3084 −3290.12 0.0415

Table 28: Heckman selection model coefficients from the sinh−1 (wage) model for women.

47

ihsearn sinh−1 (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation l(Date score) l(Immediate recall) l(Animal names) l(Prospective memory) l(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size Weekly working hours l(CF index) l(CF index w/o date) Constant select Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation Date score Immediate recall Animal names Prospective memory Delayed recall l(Care hours) CF index CF index w/o date Constant Number of obs Log likelihood ρ

(1)

(2)

(3)

(4)

0.0288*** 0.0064 −0.0182 0.0156 0.2498*** 0.2499*** 1.0833*** −0.0841 −0.0312 −0.0151** 0.0249 0.0300 0.3856*** 0.0027 0.1596 0.1830* 0.1380 0.2045 0.1022 −0.2818* −0.2189 0.3813*** 0.1834 0.1613 0.0521 0.0150***

0.0285*** 0.0070 −0.0183 0.0158 0.2496*** 0.2475*** 1.0820***

0.0290*** 0.0062 −0.0258 0.0183 0.2521*** 0.2448*** 1.1023***

0.0286*** 0.0064 −0.0264 0.0188 0.2523*** 0.2428*** 1.1001***

0.3675*** −0.0297 0.1361 0.1780 0.1166 0.1927 0.0978 −0.2588* −0.2261 0.3675*** 0.1977 0.1396 0.0451 0.0160*** −0.0055

0.3677*** −0.0180 0.1350 0.1810* 0.1329 0.2113 0.0962 −0.2545 −0.2243 0.3579*** 0.1886 0.1328 0.0469 0.0159***

6.5773

−0.0330 −0.0150** 0.0238 0.0286 0.3760*** −0.0009 0.1450 0.1696 0.1328 0.2000 0.1100 −0.2769* −0.2180 0.3698*** 0.1711 0.1512 0.0553 0.0150***

6.2796

0.0822 −0.0380* −0.1287*** −0.1571*** −0.7279*** 0.0110 0.0130 0.0101*** −0.0011 0.0029 −0.0008*

0.0806 −0.0375* −0.1288*** −0.1576*** −0.7285***

0.7867*** 2586 −3693.238 −0.9849

0.7851*** 2586 −3693.889 −0.9842

0.0128 0.0100*** −0.0005 0.0033 −0.0008*

6.3635

−0.0055 6.3978

0.0701 −0.0362* −0.1290*** −0.1578*** −0.7272***

0.0671 −0.0345* −0.1287*** −0.1583*** −0.7254***

−0.0009* 0.0180***

−0.0008*

0.7852*** 2586 −3700.537 −0.9826

0.0209*** 0.7850*** 2586 −3700.145 −0.9827

Table 29: Heckman selection model coefficients from the sinh−1 (earnings) model for men.

48

ihsearn sinh−1 (Housing wealth) sinh−1 (Other wealth) Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation l(Date score) l(Immediate recall) l(Animal names) l(Prospective memory) l(Delayed recall) Education = degree Education = other HE Education = A levels Education = O levels Education = NVQ1/CSE Education = other Married Outright owners Owners with mg/loan Occ = prof/man/tech Occ = non-man skilled Occ = manual skilled Household size Weekly working hours l(CF index) l(CF index w/o date) Constant select Age spline (50–54) Age spline (55–59) Age spline (60–64) Age spline (65–70) Health limitation Date score Immediate recall Animal names Prospective memory Delayed recall l(Care hours) CF index CF index w/o date Constant Number of obs Log likelihood ρ

(1)

(2)

(3)

(4)

−0.0116 −0.0085 −0.0652 −0.0070 −0.0413 −0.0628 −0.0828 0.1129 −0.0001 −0.0047 0.0115 0.0095 0.5133*** 0.2445 0.1101 0.0514 0.1065 0.2582* −0.1213 0.2960 0.3208 0.4140*** 0.1763 −0.0466 −0.0091 0.0328***

−0.0121 −0.0084 −0.0634 −0.0103 −0.0466 −0.0665 −0.1019

−0.0128 −0.0085 −0.0625 −0.0093 −0.0441 −0.0648 −0.1020

−0.0129 −0.0084 −0.0629 −0.0097 −0.0449 −0.0656 −0.1033

0.4961*** 0.2292 0.0966 0.0438 0.1065 0.2535* −0.1222 0.3050 0.3375 0.4107*** 0.1806 −0.0534 −0.0115 0.0329*** 0.0083

0.5047*** 0.2363 0.1032 0.0490 0.1089 0.2558* −0.1228 0.3079 0.3404 0.4153*** 0.1853 −0.0508 −0.0111 0.0329***

7.2526

0.0004 −0.0044 0.0131 0.0111 0.5208*** 0.2529 0.1125 0.0516 0.1131 0.2602* −0.1224 0.3014 0.3304 0.4155*** 0.1844 −0.0459 −0.0094 0.0329***

7.5580

0.0818 −0.1302*** −0.2020*** −0.0756*** −0.7110*** 0.2902*** 0.0418** −0.0005 0.0538*** −0.0107 −0.0031***

0.0828 −0.1309*** −0.2025*** −0.0743*** −0.7134***

0.3879*** 3084 −3894.581 −0.0162

0.3882*** 3084 −3902.966 0.0104

0.0435** 0.0005 0.0551*** −0.0049 −0.0031***

7.5364

0.0045 7.5530

0.0838 −0.1304*** −0.1987*** −0.0746*** −0.6986***

0.0843 −0.1309*** −0.1998*** −0.0748*** −0.7030***

−0.0031*** 0.0420***

−0.0031***

0.3884*** 3084 −3903.434 0.0049

0.0369*** 0.3885*** 3084 −3906.786 0.0071

Table 30: Heckman selection model coefficients from the sinh−1 (earnings) model for women.

49

References [1] Ai, C. and Norton, E. C. (2003). ‘Interaction terms in logit and probit models’, Economics Letters, vol. 80(1), pp. 123–129. [2] Anger, S. and Heineck, G. (2006). ‘Cognitive Abilities and Labour Market Outcomes: First Evidence for Germany’, DIW Berlin Discussion Paper 655, Berlin: German Institute for Economic Research. [3] Banks, J. (2006). ‘Economic Capabilities, Choices and Outcomes at Older Ages’, Fiscal Studies, vol. 27(3), pp. 281-311. [4] Bound, J. (1991). ‘Self-Reported Versus Objective Measures of Health in Retirement Models’, Journal of Human Resources, vol. 26(1), pp. 106-138. [5] Bound, J., Schoenbaum, M., Stinebrickner, T. R., and Waidmann, T. (1999). ‘The Dynamic Effects of Health on the Labor Force Transitions of Older Workers’, Labour Economics, vol. 6(2), pp. 179-202. [6] Burbidge, J. B., Magee, L., and Robb, A. L. (1988). ‘Alternative Transformations to Handle Extreme Values of the Dependent Variable’, Journal of the American Statistical Association, vol. 83(401), pp. 123–127. [7] Cawley, J., Heckman, J., and Vytlacil, E. (2001). ‘Three observations on wages and measured cognitive ability’, Labour Economics, vol. 8(4), pp. 419–442. [8] DaVanzo, J., De Tray, D. N., and Greenberg, D. H. (1976). ‘The Sensitivity of Male Labor Supply Estimates to Choice of Assumptions’, Review of Economics and Statistics, vol. 58(3), pp. 313-325. [9] Deaton, A. and Paxson, C. (1994). ‘Intertemporal Choice and Inequality’, Journal of Political Economy, vol. 102(3), pp. 437–467. [10] Disney, R., Emmerson, C., and Wakefield, M. (2006). ‘Ill health and retirement in Britain: A panel data-based analysis’, Journal of Health Economics, vol. 25(4), pp. 621-649. [11] Haardt, D. (2006). ‘Transitions out of and back to employment among older men and women in the UK’, ISER Working Paper 2006-20, Colchester: University of Essex. [12] Heckman, J. J. (2000). ‘Policies to foster human capital’, Research in Economics, vol. 54(1), pp. 3–56.

50

[13] Heckman, J. and Vytlacil, E. (2001). ‘Identifying The Role Of Cognitive Ability In Explaining The Level Of And Change In The Return To Schooling’, The Review of Economics and Statistics, vol. 83(1), pp. 1–12. [14] Huppert, F. A., Gardener, E., and McWilliams, B. (2006). ‘Cognitive Function’, in: Banks, J. et al. (eds), Retirement, health and relationships of the older population in England: The 2004 English Longitudinal Study of Ageing (Wave 2), London: Institute for Fiscal Studies, pp. 217–242. [15] Lambrinos, J. (1981). ‘Health: A Source of Bias in Labor Supply Models’, Review of Economics and Statistics, vol. 63(2), pp. 206-212. [16] McFadden, D. (1981). ‘Econometric models of probabilistic choice’, in: Manski, C. F. and McFadden, D. (eds.), Structural analysis of discrete data with econometric applications, Cambridge, Mass.: MIT Press. [17] Phillips, D. and Sibieta, L. (2006). ‘A Survey of the UK Benefit System’, Briefing Note No. 13, London: Institute for Fiscal Studies. Internet: http://www.ifs.org.uk/bns/bn13.pdf [18] Pryor, F. L. and Schaffer, D. L. (2000). Who’s Not Working and Why: Employment, Cognitive Skills, Wages, and the Changing U.S. Labor Market, Cambridge, England: Cambridge University Press. [19] Schaie, K. W. and Strother, C. R. (1968). ‘A Cross-Sequential Study Of Age Changes In Cognitive Behavior’, Psychological Bulletin, vol. 70(6), pp. 671–680. [20] Simpson, M. (1980). ‘The Sociology of Cognitive Development’, Annual Review of Sociology, vol. 6, pp. 287–313. [21] Steel, N., Huppert, F. A., McWilliams, B., and Melzer, D. (2004). ‘Physical and cognitive function’, in: Marmot, M. et al. (eds), Health, wealth and lifestyles of the older population in England: The 2002 English Longitudinal Study of Ageing, London: Institute for Fiscal Studies, pp. 249–300. [22] Stern, S. (1989). ‘Measuring the Effect of Disability on Labor Force Participation’, Journal of Human Resources, vol. 24(3), pp. 361-395. [23] Taylor, R., Conway, L., Calderwood, L., and Lessof, C. (2004). ‘Methodology’, in: Marmot, M. et al. (eds), Health, wealth and lifestyles of the older population in England: The 2002 English Longitudinal Study of Ageing, London: Institute for Fiscal Studies, pp. 357–374. 51

[24] Wooldridge, J. M. (2002). Econometric Analysis of Cross Section and Panel Data, Cambridge, Mass.: MIT Press.

52