Healthy, Wealthy and Working: Retirement Decisions of Older New Zealanders (WP 10/02)

Abstract#

Print this section

Health status is an important element in the decision to continue working or retire among older workers. Given the demographic projections for the next four decades, there will be increasing opportunities for older workers to remain in the workforce. However, an individual’s decision is likely to be influenced by both their health status and their accumulated wealth. This study analyses the influence of health and wealth on the decision to participate in the labour force amongst older New Zealanders, aged 55 to 70. It is based on the first wave of data collected in a longitudinal survey of Health, Work and Retirement conducted by researchers at Massey University.

The study employs a range of measures of health including the results from the international Short Form (SF36), self-reported health status and the prevalence of chronic illness. Regardless of the measures tested, a significant reduction in labour force participation is associated with poorer health status. It is widely recognised that health status itself may partly be determined by labour market characteristics. Attempts to deal with this statistically were not successful. Perhaps surprisingly, wealth did not appear to be strongly related to the decision to retire. A marked fall in participation is associated with the receipt of New Zealand Superannuation at age 65, arguably masking the effect of privately held forms of retirement wealth. The paper reports associations between health, wealth and retirement which do not necessarily constitute evidence of causality.

Acknowledgements#

Print this section

The authors acknowledge the support and guidance of Fiona Alpass and Jack Noone of the Department of Psychology, Massey University in making available the data from Wave 1 of the Health, Work and Retirement Longitudinal Survey. Throughout the study they willingly answered questions about the data. Dean Hyslop of Victoria University provided guidance on the use of econometric models, and the authors benefited from discussions with and helpful suggestions made by Heather Holt and Katy Henderson of the Treasury. Judith Davey of the Institute of Policy Studies, Des O’Dea from the Wellington School of Medicine, Adam Tipper from the Ministry of Social Development and David Law and Tony Burton of the Treasury made helpful suggestions on a draft. Kathy Anderson of Vanderbilt University assisted with access to earlier papers.

Access to the data used in this study was provided by Massey University. The results in this study and any errors contained therein are those of the authors, not Massey University.

Disclaimer#

Print this section

The views, opinions, findings and conclusions or recommendations expressed in this Working Paper are strictly those of the authors. Theydo not necessarily reflect the views of the New Zealand Treasury. The Treasury takes no responsibility for any errors or omissions in, or for the correctness of, the information contained in these Working Papers. The paper is presented not as policy, but with a view to inform and stimulate wider debate.

 

1  Introduction#

Print this section

New Zealand, in common with many other countries, faces an ageing population. Projections by Statistics New Zealand indicate that by 2051 those over 65 will comprise more than a quarter of the total population. Significant economic changes are expected to follow as a result of these demographic shifts. These in turn will have implications for economic growth, productivity, income, fiscal costs and living standards.

One factor that underlies many of these issues is the extent to which older people will continue in the workforce. What are the key determinants that influence the so-called “retirement decision” of older workers? That decision will reflect an individual's preferences together with such factors as age, health, expected income, family status and savings.

The primary objective of this paper is to assess the importance of a person's health status on their participation in the labour market and on the related decision as to retirement. It is reasonable to expect that health status would in some cases directly affect a person's ability to participate in the labour market. In addition there can be indirect channels, as the health status of a person or their family members may well influence the extent of wealth accumulation, expected expenses in retirement and the need for precautionary saving. These factors in turn will influence decisions about workforce attachment and retirement.

Earlier studies on retirement in New Zealand have not been able to incorporate measures of health status, as there were no unit record data at a national level which incorporated income, wealth, socio-economic information and health measures. Recently this situation has changed and two longitudinal surveys are now available. The first of these is the Survey of Family, Income and Employment (SoFIE) conducted by Statistics New Zealand.[1] A recent study by Holt (2010) analyses the relationship between health and labour market participation using data from SoFIE.

The second of these surveys, and the data on which the present study is based, is a survey on Health, Work and Retirement (HWR) based at Massey University. This rich data source allows us to analyse the relation of overall measures of physical and mental health with income, wealth, living standards and labour force participation. In addition, information on 19 chronic diseases allows their individual effects to be studied. As the sample is for 55 to 70 year-olds, it is possible to obtain population estimates for the health and workforce attachment specifically for this older group.

In Section 2 we present a brief account of the data, followed by an overview of the results from the HWR survey (Section 3). Section 4 discusses methodological issues, and Sections 5 to 10 summarise the results for wealth, income, living standards, health, labour force participation and chronic diseases, respectively. Section 11 contains a summary and conclusions, and highlights some key limitations.

The paper examines associations between health, wealth and retirement; it does not provide evidence that, for example, poor health “causes” retirement. Health and retirement are almost invariably jointly determined and the question of the direction of causality is a challenging statistical issue which we address without resolving.

2  The data#

Print this section

The HWR survey was established by the School of Psychology at Massey University in collaboration with the Health Research Council of New Zealand, the New Zealand Institute for Research on Ageing and the Centre for Māori Health, Research and Development at Massey University, to investigate the factors surrounding work and retirement for those aged 55 to 70.[2] The first wave was run in 2006, with a total 6,662 respondents to a postal questionnaire.[3] The sample was drawn from the Electoral Roll[4], and those identifying as being of Māori descent were deliberately over sampled.

…equal probability sampling procedures were used to select both the general and Māori sub-samples for the HWR study. Both sub-samples were treated independently and random selection was used to select 55 to 70 year-old New Zealanders participants from each population of interest (the general and Māori respectively) to their respective sub-sample. (Towers, 2007)

A summary of the sample size is given in Table 2-1.Table 2-2 summarises the sample by age groups, and shows the number of usable observations. A total of 822 records were dropped as they did not report at least one of age, gender, ethnicity, physical or mental health status. The sample was further pruned of eleven cases by eliminating those whose reported age fell below 50. The sample was reweighted to reflect these adjustments. The weighting was done by ethnicity, as the population totals were provided for those values (Towers 2007). All results in Section 3 onwards are the weighted results unless otherwise indicated. Additionally, the wider age ranges are used, including those under 55 and over 70, unless otherwise indicated.

Note:
¹ Excluded from the study as they were never able to participate (eg they were unable to be contacted, were deceased or had been institutionalised).

Note that not all observations can be used, as they may be dropped from models if they are missing values for variables other than the ones listed above.

Notes:
¹  Includes those observations aged less than 55 in the first two columns.
²  Includes those observations aged more than 70 in the first two columns.
³  Includes 161 observations with no determinable age, sex or ethnicity.

A complete listing of all the variables used in this study together with their definitions, means, medians and inter-quartile ranges, is set out in Appendix B.

3  Overview#

Print this section

This section provides a series of basic tabulations with the aim of providing the reader with an overview of the survey results for key variables relevant to this study.

3.1  Wealth#

Print this section

Respondents were asked to indicate which assets they owned (real estate, farms, businesses, holiday homes, financial investments etc). Where the respondent failed to indicate either yes or no, for this analysis the assumption is made of “no”. The present value of expected income from New Zealand Superannuation (NZS) is a legitimate part of the wealth of an individual and could be counted as part of total wealth; this has not been done in this study.[5] The results are summarised in Figure 1, for both those working and retired.[6]

Figure 1: Asset ownership

 

Source:  Health, Work and Retirement Survey

Home ownership is at high levels for working and retired respondents. The ownership of farms, businesses and rental properties declines after retirement. This result is consistent with the findings of Coile and Milligan (2006) for the USA. They note that there are marked shifts in the composition of asset holdings with ageing. The share of assets held in banks and term deposits rises from 11% of total assets at ages 60 to 64 to 28% at ages 80 to 84. They find that the effect is more pronounced where the person has suffered a health shock. They conclude that the standard risk versus return models of portfolio selection need to be augmented with ageing and health status. Older people and those with physical or mental disabilities tend to hold their assets in a more liquid form to be able to more readily meet unexpected costs. In addition, older people place more emphasis on the ease of portfolio management.

However, Love and Smith (2007) have questioned whether in fact there is a causal relation between health status and portfolio choice, such that those in poorer health tend to chose less risky assets. They argue that both health status and financial decisions are “driven by characteristics such as risk preference and impatience that are unobserved by the researcher” (p2). The consequence of this unobserved heterogeneity, if inadequately accounted for, is to severely bias the estimates of the demand for different classes of assets. Using data from the Health and Retirement Study for the USA, they attempt to correct for this unobserved heterogeneity and as a result find no statistically significant relation between health status and asset choice. Their findings underscore the difficulties in establishing robust relations between health status and other variables of interest.

Respondents in the HWR survey were given an option of providing an approximate value for each of the asset classes they owned. These were summed to give an estimate of total wealth. As not all respondents either indicated which assets were owned or chose not to provide estimates of values, the total number of observations was 3,966. The means, medians and inter-quartile ranges by age groups are summarised in Table 3-1.

Note: IQR is the inter-quartile range defined as the difference between the observations at the 25th and 75th percentile points of the distribution and is one measure of the dispersion of wealth.

In order to increase the sample size available for analysis, a value for total wealth was imputed for those cases where the values were missing. This was done by first estimating a regression of total wealth on an extended series of explanatory variables, and then using the estimated coefficients, values of total wealth were predicted for those individuals with missing values. This results in an expanded set of observations which are summarised in the right hand side of Table 3-1. Owing to not all variables in the imputation model being present for all observations, this increased the sample size to 5,093 out of the 5,339 observations available.

In making imputations there is a risk that the group who did not answer the question are systematically different from those who did. To test for this, a comparison was made of the characteristics of those with actual observations with the group for which total wealth was imputed by this method; for observable characteristics no significant differences were found.[7] The mean wealth level for the original and expanded samples is very similar, although the median is higher once the imputed values are included, indicating the distribution became somewhat more skewed toward higher wealth values.

In addition to the results in Table 3-1, total wealth levels for the working and retired groups were calculated. The average wealth is $676,000 for the working group and $427,000 for the retired group.

Unfortunately, it is not possible to compute net wealth based on the survey data as there is no corresponding estimate of total liabilities. The only information available is whether the respondent had a mortgage or a loan. This applied to 52% of the working group and 15% of the retired group, indicating that many people plan to pay off any outstanding debt by the time they retire.

3.2  Income#

Print this section

Total household income was estimated by summing the before tax income of the respondent and the combined income of all other members of the household over the previous 12 months.[8] The results are summarised by age range in Table 3-2, and by working status in Table 3-3. As the distributions of both income and wealth are typically highly skewed owing to a few high income or wealth individuals, the tables present both means and medians. Again, a similar procedure was used for imputation to expand the usable sample size to 5,054 where variables in the imputation model were present. In this case the means, medians and inter-quartile ranges for the original and expanded samples are strikingly close, in part owing to a smaller number of missing values.

 

 

 

 

The ratio of mean wealth to mean income rises almost 50% from around six times for those working to nine times for those who are retired. This is consistent with the widely recognised phenomenon “of asset rich but cash poor” that characterises many retired individuals. Income drops significantly with retirement, and while assets are run down (the “dissaving” phase of the life cycle), the decline is much less marked. This is accentuated by the fact that the principal residence is a major share of total assets and owner occupied rates remain high for many retirees in New Zealand.

The HWR survey also asked respondents for the sources of income. The categories were wages and salaries paid by an employer; self-employment; income from investments (interest, dividends, rent); regular payments from the Accident Compensation Corporation (ACC); New Zealand Superannuation (NZS) or veterans' and war pensions (VP); other superannuation; unemployment benefits (UB); domestic purposes benefit (DPB); invalids' benefit (IB); student allowance; other benefits and other sources; and an option to record no source of income in the preceding 12 months. Respondents could indicate more than one source so the totals can exceed 100%. The results are depicted in Figure 2 (Panel a).

The majority of the working group received income from wages and salaries, while amongst retirees investment income and NZS were the most common sources. Very few of the retired group reported self-employment income. The average household income for the working group was found to be $114,000, while for the retired group the average income was $49,000. As this latter estimate is more than double the married allowance under NZS, it is apparent that investment income is an important source for many retirees.

Panel b of Figure 2 summarises the assessments of respondents in relation to the adequacy of their incomes. In large measure the results for both the working and retired groups are similar, with around 80% perceiving their current incomes to be adequate. The proportion of retired respondents who felt their income was not enough (about 10%) is comparable to the proportions reported by Fergusson et al (2001: Table 9) who found 12% of singles and 10% of older couples felt their income was not enough.

Figure 2: Income: sources and adequacy

 

Source:  Health, Work and Retirement Survey

3.3  Living standards#

Print this section

Standard of living was rated on two different scales. The first was self-assessment of the level of their material standard of living, while the second assessed the respondent's satisfaction with their living standards. Both measures are depicted in Figure 3. Just over half of both groups consider their standard of living to be at medium level, with more of the working group then the retired group rating their standard as high. Around the same percentage of the working and retired groups consider their living standard to be low. The results for respondent’s satisfaction are very similar for all categories across the retired and working groups. These findings are consistent with those from a national survey in which only 8% of those over 65 years old reported any degree of hardship Jensen et al (2006).

The proportion of both the working and retired groups who regard their living standard as low or expressed dissatisfaction with their standard of living is typically less than 10%. This accords with the results of Fergusson et al (2001: Table 9) who reported that among older New Zealanders, 10% of singles and 5% of couples assessed their living standard as low or fairly low.

Respondents were asked how they expected their living standards might change when they retire. For those actually retired the question related to the perceptions of living standards as they approached retirement; ie the working group was asked: “How do you expect your living standards to change when you retire?” while the retired group was asked: “How did you expect your living standards to change when you retired?” The results are shown in Figure 4 (Panel a). Respondents were also asked how they saw their future retirement years in comparison to their working years, with the results in Figure 4 (Panel b). It is not immediately clear how retired people would view this question.

Figure 3: Standard of living

 

Source:  Health, Work and Retirement Survey

Only a very small proportion of either the working or retired group anticipated a rise in their living standards at retirement. Over 40% of both groups expected their living standards to decline. Yet while they anticipated a decline, the majority of those actually retired expressed satisfaction with their living standards in retirement (Figure 3). This implies that in planning for retirement a large proportion of people expect to take a fall in living standards, but as it was anticipated, they do not subsequently express dissatisfaction when they actually experience a drop in living standards. An alternative explanation could be that the actual fall in living standards was less than anticipated.

The willingness to accept a fall in living standards may influence their decision to leave the workforce or reduce their hours. Those wishing to maintain a standard of living in retirement closer to that enjoyed in their working years could be expected to have higher rates of labour force participation.

However the results in Figure 4 (Panel b) provide an interesting contrast. The vast majority of respondents (around 80%) did not expect to see a fall in their living standards in future retirement years. This suggests that people adjust their expectations of living standards to match the reality of their circumstances.

Figure 4: Expected changes in living standards

 

Source:  Health, Work and Retirement Survey

One way to summarise the overall standard of living is through the use of an Economic Living Standards Index (ELSI). A short form of this index is designed

to measure levels of consumption, social activity, and asset ownership, rather than the economic resources that enable them. The scale assesses restrictions in ownership of assets (8 items), restrictions in social participation (6 items), the extent to which respondents economise (8 items), and a self-rated indicator of standard of living (3 items). The ELSI-SF scores on each of the items were combined to form a continuous variable ranging from 0 to 31 (higher scores reflect higher economic living standards) and as an ordinal variable with 7 levels from severe hardship to very good.[9] (Stephens et al, 2008)

Figure 5 summarises the distribution of the ELSI for the working and retired groups. While the working group has a little higher mean value (23.7) than the retired (22.9) (not shown on the figure), the medians were in fact the same and the overall range is broadly similar. Were it the case that ELSI were a truly objective measure, then this result would provide support for the argument that the actual fall was minimal. In contrast, if ELSI is viewed as a more subjective measurement, then this would tend to lend support to the argument that people adjust their expectations.

Figure 5: Economic Living Standards Index

 

Source:  Health, Work and Retirement Survey

3.4  Health#

Print this section

The HWR survey includes an extensive series of questions related to health. The following extract describes the use of this information:

The measure of health used for the HWR study was the Australian and New Zealand version of the SF-36 which has become one of the most widely used (both in New Zealand and internationally) standard questionnaires for measuring physical and mental health status. The SF-36 includes 36 items measuring physical and mental health in relation to nine health scales: general health (self assessment of health overall), physical functioning, role physical (how much physical health has affected daily activities), mental health, role emotional (how much emotional health has affected daily activities), social functioning (how health has affected social activities), health transition (perceptions of health changes), bodily pain, and vitality. All scores have been weighted so that they may be interpreted in the same direction: higher scores mean better health. The subscales have also been combined to provide two summary scores for physical and mental health respectively. The general health status of this cohort is indicated by scores on the SF-36 which have been transformed using New Zealand population norms and standardised so that the population mean is 50. (Stephens and Noone, 2007: p3)

In the present study we focus on the summary measures for physical and mental health (see Figure 6).

The SF-36 also yields two psychometrically-based physical and mental health summary measures: the Physical Component Summary (PCS) and the Mental Component Summary (MCS). The PCS and MCS are computed following a three-step standardised procedure. First, all eight domains are standardised using a linear z-score transformation. Z-scores are calculated by subtracting the domain means for the general population (by age and sex) from each individual's domain score and dividing by the standard deviation of the NZ population (Ministry of Health, 1999). Second, the z-scores are multiplied by the domain factor score coefficients for PCS and MCS and summed over all eight domains. Finally t-scores are calculated by multiplying the PCS and MCS sums by 10 and adding 50 to the product to yield a mean of 50 and a standard deviation of 10 for the NZ population.[10]

Scott, Sarfati, Tobias and Haslett (2000) analyse the SF-36 health survey for New Zealand, to test whether the two-dimensional structure of physical and mental health is applicable to all major ethnic groups (European, Māori and Pacific). They find that while the structure applied to European and Māori aged less than 45, for Pacific peoples and older Māori the structure did not clearly differentiate between the physical and mental health components. For these groups the two components are not seen as independent. In subsequent analyses based on the HWR survey we have included ethnicity as an explanatory variable in an attempt to correct at least in part for this problem.

Table 3-4 summarises the pattern of the physical and mental health scores for the New Zealand population.

Source: Scott, Sarfati, Tobias and Haslett. (1999).

Figure 6: Structure of the SF-36 scheme for the measurement of health

 

Source:  Ware (2000).

Note: This figure does not include the category of health transitions referred to in the extract from Stephens and Noone (2007) cited above although these are in the summary measures used here.

A summary of the levels and distributions of the physical and mental health scores for both the working and retired groups is depicted in Figure 7. The average levels are similar for both groups although the ranges for the retired group are wider, especially in the case of the physical health measure.

As well as determining an objective measure of health, the survey asked respondents to self-assess their health. Table 3-5 reports this, broken down by sex and work status. The broad pattern of working aligning with better health is continued here.

Figure 7: Distribution of physical and mental health scores

 

Source:  Health, Work and Retirement Survey

Note:
¹ The estimates for those unemployed are based on a small sample of approximately 40 observations for each gender.

3.5  Labour force participation#

Print this section

In the context of the present study, an important aspect of the HWR dataset is the split between those who are still working, and those who are retired. Unfortunately the distinction is not clear cut.[11] In part this reflects the fact that increasingly there is no universal “age of retirement” implying a clean break from the labour market. There are a number of possible criteria for defining “retirement”. Banks and Smith (2006) identify the following:

• complete and permanent withdrawal from any paid employment
• being in receipt of a public or private pension
• a state of mind in which individuals perceive themselves as retired.

This section describes in some detail the classification of the data and the derivation of the working and retired subsets. There are two critical questions in the survey that could be used to define this split. The first is:

Do you consider yourself partially retired, completely retired or not retired? (Q49)

for which respondents were given the choice of:

• Not retired at all
• Partly retired
• Completely retired

At a later stage of the questionnaire respondents were asked:

What is your current work situation? (Q74)

to which they could respond according to the following options:

• Full-time paid employment including self-employment
• Part-time paid employment including self-employment
• Retired, no paid work
• Full-time homemaker
• Full-time student
• Unemployed and seeking work
• Not in the workforce

The results of these questions overlap significantly. Some of those who responded that they are completely retired subsequently recorded that they were in full- or part-time paid work. It is possible that they may have viewed being retired as meaning to have attained the age of eligibility for NZS. Furthermore, some who did not consider themselves retired were nonetheless retired as they undertook no paid work.

Table 3-6 summarises the responses from these two questions, and shows the population estimates in each cell.

Note: The unweighted counts are reported in Appendix Table A.1.

The category of “Other” includes: homemaker, student, unemployed and seeking work, not in the workforce and non-response.

We have used these data to form a split between working and retired, as follows:

Working = Full-time or part-time or unemployed but seeking work. Technically the correct term is “participating”, although the more informal term of “working” will be used.

= 273,612 + 129,617 + (those recorded as not retired (20,546) or partly retired (19,273) from the “Other” category of work status who were unemployed but seeking work = 8,830)

= 412,059

Retired = Retired (no paid work) or those who recorded retired from the “Other” category of work status who were not unemployed but seeking work)

= 135,405 + (22,311 out of 23,291)

= 157,716

This leaves (609,720 – 412,059 – 157,716) = 39,945 people for whom we do not have information, other than they are not in the workforce. These people are dropped from the population considered in this paper, except for where needed to compare to other datasets (Section 9.7). However, the effect of their exclusion is explored in Section 9.2.

Having constructed the two groups, working and retired, it is instructive to compare some of their basic characteristics. Table 3-7 summarises the differences.[12] A higher proportion of retired people considered their health to be poor.

The overall labour force participation rates for males and females in the age range 55 to 70 years appear high, both by historical and international standards.[13] As shown in

Table 3-8, 76% of all males and 66% of all females in this age range are in either full-or part-time work. Remaining attached to the labour market is clearly an important aspect for those over 55 as they age. For males, almost one in four over 65 years old is in full-time work; one in four females over 65 is in either full-or part-time work. Females tend to have higher rates of part-time employment.

All respondents were asked whether or not they planned to continue some form of paid work after retirement. It is assumed that those currently working responded based on their intentions. However it is less obvious how those retired would have responded. On one hand their response could have been based on their intentions prior to retirement. Alternatively, their responses could reflect their current status. With this caveat in mind, the results are depicted in Figure 8.

 

The growing trend toward a transition from full-time work to fully retired is strikingly evident, as over two-thirds of all those in the working group stated that their intention was to continue some paid work after “retirement”. Nearly 35% of those in the retired group stated they intended to work or were working after retirement. This accords with the earlier results presented in Table 3-6 where 35% of those who were partly or totally retired stated they were in some form of paid employment.

Both the working group (WK) and the retired group (RT) have broadly similar attitudes towards retirement. Retirement is largely seen as a time to slow down, start on a new part of life and nearly all of them have some idea of what they want to do. It is not seen as frustrating, nor unwanted, nor as a problem as a replacement for work (see Figure 9).

Figure 8: Intentions with regard to paid work after retirement

 

Source:  Health, Work and Retirement Survey

Figure 9: Attitudes toward retirement

 

Source:  Health, Work and Retirement Survey

4  Modelling approaches#

Print this section

The preceding section presented some preliminary results from the survey. They were intended to capture the broad patterns; for example, the difference in wealth between the working and retired groups. The limitation of this type of analysis is that other factors might lie behind the observed differences and these are not being held constant. In this example it might be that the real reason for the differences lies in the health status of the two groups and is not directly related to their employment status. Regression models, a form of multi-variate analysis, can be used to test the effects of particular variables while holding constant the influence of other variables. In some cases there may be interaction effects between explanatory variables and these can be readily incorporated.

In this study we use two forms of regression: ordinary least squares (OLS) where the dependent variable is a continuous variable (eg, wealth or income) and logistical regressions where the dependent variable of interest is a binary variable (eg, working as distinct from retired; or self-rated good health as distinct from poor health). The next section provides a relatively non-technical discussion aimed at providing the reader with a basic understanding of the models employed.[14]

4.1  OLS regression models#

Print this section

OLS models are of the general form:

 

(1)

where

Y_i= the value of the dependent variable for the i-th respondent

X_ij= the values of the explanatory variables for the i-th respondent

α, β_j= the regression parameters to be estimated

ε_i= a random error term assumed to follow a normal distribution.

4.2  Logistical regression models#

Print this section

The logistical or logit regression is applied when the dependent variable takes only the values of zero or one (eg, we could assign one to those in good health and zero to those in poor health). The OLS model could potentially be applied in this case with the continuous dependent variable on the left hand side merely replaced by a binary variable. Such a linear probability model, while generating the correct sign and significance of the coefficients, is generally not appropriate. There are at least three reasons for this. First, the variance of the dependent variable is not independent of the values of the explanatory variables (the problem of heteroskedasticity). Second, the predicted probabilities could be negative or greater than 1, which is counter-intuitive. Finally these models assume the marginal effects are constant. For these reasons it is preferable to employ a logit model.

Logit models are built on the notion of probability. Suppose the probability denoted by p, of some event is 0.8 (eg, the probability that a respondent is in good health). The probability that they are not in good health is then simply 1–p = 1–0.8 = 0.2. The odds of the event are defined as the ratio of the probability of “success” to the probability of “failure”:

 

(2)

In the above example, this would be 0.8/(1–0.8) = 4. In other words, the odds of a person being in good health (relative to being in poor health) are four to one. Note that it is quite arbitrary which way round p and 1–p are assigned.

By working with the logarithm of the odds, we circumvent the problem of the restricted range for the probability. The transformation to logarithmic odds maps the underlying probability whose range is from zero to one, into a variable with range from negative infinity to positive infinity. This is referred to as a logit transformation.[15]

We now sketch the use of this transformation in the estimation of the coefficients associated with the independent or explanatory variables in the underlying model. Let p be the probability that a respondent is a member of the category of the dependent variable as assigned the code of 1. All the remainder who do not belong to this category are assigned zero. Then:

 

(3)

 

(4)

We can now show that Z is the log of the odds. Rearranging equation (3) to solve for Z yields:

 

(5)

Hence Z is the log odds or logit. Table 4-1 sets out the relation between probabilities, odds and log odds.

We can now proceed to estimate equation (4), in this case using Maximum Likelihood Estimation (MLE), an iterative procedure which searches for that set of values for α and the set of β such that the probability of observing the dependent variables in the sample of data is maximised.

4.3  Interpreting the logistical regression#

Print this section

We turn now to the interpretation of the coefficients in the logit equation. The estimated values of each of the coefficients describe the amount by which Z, the log odds, changes in response to a one unit change in the corresponding X_i, where X_i is continuous. In the event that X_j is itself a binary variable describing a particular category (eg, male or female) and taking only the values of zero or one, the value of the estimated coefficient on X_j (denoted

 

) describes the change in the log odds of moving from the category coded 0 to the alternative category coded (1).

 

However, this interpretation is not especially intuitive. A preferable approach is to consider the impact of a unit change in a particular X_j on the odds rather than the log odds. The odds ratio is defined as

 

(6)

ie, it is the ratio of two odds. Consider the following which involves finding the odds ratio for a one unit change in say X₁:

 

(7)

This expression reduces to

 

as all other terms cancel out. This result simply states that the ratio of the odds for a one unit increase in X₁ is given by

 

; ie, a one unit change in X₁ results in a

 

change in the odds ratio. It is constant; specifically it does not depend on the values of the other variables (X_j). Note that while the odds ratio is constant, this does not imply that the odds themselves are constant at various values of the X_j. In fact owing to the multiplicative effect, the actual change in the odds depends on the starting point. An odds ratio of two would increase odds of one to two, and odds of two to four.

 

Table 4-2 shows the direction of change in the odds of an event occurring for a one unit change in a variable, based on the sign of the corresponding coefficient.

It is worth noting that while the distribution of the coefficients is symmetrical about zero, the odds have a lower bound of zero and so are not symmetrical (as shown in Table 4-1).

4.4  An example#

Print this section

Suppose the event of interest is the probability of females working full-time (W). This is coded (1) for those who are in full-time work and 0 if not working full-time (NW). For simplicity we postulate that whether a female is working depends solely on her age (A) and a continuous measure of health status (H), and that these variables are independent. The first step is to fit the logit model. Suppose this resulted in the following equation:

 

(8)

Case A:

Let the means of the independent variables be A = 50 and H = 50 so that evaluating Z at the means yields:

Z = 0.03 – 0.015(50) + 0.006(50 )= -0.42

so that Pr (W) = 1/(1+e^-z) = 1/(1+e^0.42 ) = 0.40

and Pr (NW) = 1 – 0.40 = 0.60 and

 

 

= or the odds of an average 50-year-old female working are 0.66 to 1.

 

Log odds = ln(0.66) = –0.42, where ln is the natural logarithm.

Case B:

Now consider the change when we allow the age to increase by one unit (a year in this case) from 50 to 51 while holding H at its original level of 50; ie, we are interested in estimating the effect of an increase in age on the probability of working holding constant a female's health score.

Z = 0.03 – 0.015(51) + 0.006(50) = -0.435

so that Pr (W) = 1/(1+e^-z ) = 1/(1+e^0.435 ) = 0.39

and Pr (NW) = 1 – 0.39 = 0.61 and

 

 

or the odds of an average 51-year-old female working are 0.65 to 1.

 

So that the odds ratio =

 

 

which we can verify as

 

= e^-0.015 = 0.98.

 

This implies that a one year increase in age decreases the odds of working by a factor of 0.98, or similarly decreases odds by 2%. In Case A, the odds of a 50-year-old female working full-time were 0.66 to 1. For a 51-year-old female the odds are (0.66*0.98) to 1, or 0.65 to 1. Note this is not the same as the probability of a female working full-time which, for a one year increase in age, fell from 0.40 in Case A to 0.39 in Case B. The difference in the latter probabilities is the marginal effect which we discuss in the following section.

Case C:

In some instances we might be interested in a non-marginal change. Consider the case of a 65-year-old female, again isolating the effect of a change in age by holding the health status constant at its initial mean level. In this case the predicted value of Z is given by:

Z = 0.03 – 0.015(65) + 0.006(50) = -0.645

so that Pr (W) = 1/(1+e^-z ) = 1/(1+e^0.645) = 0.34

and Pr (NW) = 1 – 0.34 = 0.66 and

 

 

or the odds of an average 65-year-old female working are 0.52 to 1.

 

Log odds = ln(0.52) = –0.645,

and the odds ratio =

 

 

We can now compute the change in the log odds corresponding to a change in age from 50 to 65:

(Log Odds Case C) – (Log Odds Case A) = -0.645 – (-0.420) = -0.225.

Note that this result is simply 15 times the value of β₁; ie, 15 * -0.015 = -0.225.

4.5  Measuring marginal effects in binary models#

Print this section

The impact of a unit change in one of the independent variables on the probability of an event is typically referred to as the marginal effect. The first step is to compute the value of the probability using equation (3) when all the values of the X_i are set at some predetermined value; ie, the marginal effect can be evaluated at different points depending on the set of X_i values chosen. Generally the mean values for the whole sample are used, but the choice is arbitrary.

The second step involves repeating the calculation of the probability with a different set of X_i values. Typically this will involve leaving all but the variable of interest at their mean values and incrementing the variable of interest by one unit.

In the above example (Case A) the probability of working with both independent variables set at their means:

 

 

 

In the second stage we compute the probability of working when age is increased by one year and health status remains at its mean value:

 

 

 

The marginal effect is then given by

 

 

 

From this we conclude that the marginal effect of increased age is to reduce the probability of working by one percentage point.

To this point we have been using independent variables which are continuous. What happens when there is an independent variable that itself is binary? In the current context this could be a variable such as: does the female have a partner who is working (yes or no), or is the person a migrant (yes or no); or does the person have a tertiary qualification (yes or no). The presence of such variables has two implications. To calculate the marginal effect we would have to compute the value of Z setting the binary variable to zero and then to one. We then simply need to state clearly the base for the calculations. In addition, if there are other binary variables that we wish to hold constant, then there are a number of alternatives. We could (for apparent consistency with continuous variables) use the mean which is nothing more than the raw proportion. Alternatively we could assign the variable either zero or one. There is no “correct” way, and the option chosen will depend on the particular context being analysed. In the empirical work in later sections of this report we use the mean values.

We conclude this section with a summary in Table 4-3 of the different interpretations of a logistical regression coefficient.

Table 4-3 Interpreting logistical regression coefficients

 

Note:
¹ There is a third case for a categorical variable X_j which is not discussed here.

4.6  The logit vs. probit transformation#

Print this section

In formulating the logit model, an implicit assumption has been made about the variance of the error term. To arrive at equation (3) requires the var(ε) = π²/3. An alternative assumption is that var(ε) = 1. In this case the model is referred to as a probit model. In short, while the assumption about the var(ε) is arbitrary, it does not affect the estimated values of the probabilities.[16] While the choice between the probit and logit models is to a large extent arbitrary, the effects in the logit model, unlike the case of the probit, can be interpreted as changes in the odds. Additionally, logit and probit models are generally asymptotically equivalent – so in a study such as this the coefficients should be effectively the same. We have chosen to use logit models in this study.

4.7  Potential problems#

Print this section

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

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

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

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

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

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

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

5  Wealth#

Print this section

In this section we analyse key factors associated with the level of total wealth based on an OLS regression model of the type shown in equation (1). As the survey did not collect data on liabilities, the results relate only to total, as distinct from net wealth.[21] A summary of the significant results is given in Table 5-1 while the complete results of the regression model are set out in Appendix Table C.1.

Of primary interest is the relation between health status and wealth accumulation. Clearly there is a likelihood that those with poorer physical health would tend to have less connection with the labour market. As a result of lower earnings and possibly higher medical expenses, they could be expected to have lower wealth accumulation. An advantage of the regression model used here is that it is possible to control for a wide range of other variables including specifically whether the person was working and whether in receipt of a benefit.

Notes:
¹ Only those explanatory variables that were statistically significant are shown in this table, rounded to the nearest thousand. The full results are in Appendix Table C.1.
² *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level.

The results indicate that a one unit improvement in the physical health score is associated with an additional $6,000 of total wealth. It is considered that a five unit change in the score is clinically significant (Ware, 2000). Extrapolating from the regression results yields an estimate of an additional $28,000 in total wealth associated with a five unit improvement in the physical health score. This represents a 10% increase at the mean level of the health score.

That those with better health have higher accumulated total wealth is not surprising. In the first place, those with poorer health may have had less working years and greater medical costs. Furthermore, poor health may be seen as signalling lower life expectancy, and with that less incentive to accumulate as much retirement wealth.

Males have significantly more wealth than females, and Māori and other ethnic groups have less wealth than NZ Europeans. Individuals who are separated or who never married have less wealth than those who are married with a non-working spouse. Those on a benefit have less wealth, and those with a superannuation scheme have higher wealth. Those planning to stop work totally at the time of retirement have higher wealth. Those who place importance on a range of negative aspects of retirement have significantly less wealth.

In the present study no association was found with mental health. Those with higher mental health scores did not report significantly higher levels of total wealth. It is possible that any relationship that might exist was obscured by the fact that in this study total, rather than net, wealth was measured. High levels of liabilities for a given level of total wealth may be associated with poorer mental health.

Carter et al(2008) examine the relation between mental health and net wealth using data from the SoFIE. They use a measure of psychological distress and explore the extent to which this is associated with the level of net wealth, after holding constant a range of socio-demographic characteristics. They find that the odds of reporting high psychological distress are much greater amongst those in the lowest wealth quintile compared to the highest. They conclude that policies which enhance wealth accumulation may have positive benefits for mental health. However, in drawing this conclusion the authors acknowledge that the potential exists for reverse causation such that those with poor mental health may accumulate less wealth through lower levels of labour force participation, lower earnings or higher levels of expenditure.

Anastasiadis (2010) has analysed the association between health and net wealth based on Waves 1 to 3 of the SoFIE data. He finds that poorer health is associated with lower net wealth after controlling for a range of other factors. Those with greater wealth are less likely to suffer health shocks.

Rich people tend to be healthier and live longer. With greater life expectancy they have more incentive to save for retirement and accumulate greater wealth. The finding in this study that better physical health is associated with higher wealth is consistent with this argument, and with similar findings for the USA by Nardi, French and Jones (2006).

6  Income#

Print this section

In this section, interest focuses on the association of a range of factors with both the level and adequacy of income. Three models were estimated. The first, a regression model, is for the level of reported income. The second, a logistic model, refers solely to those in the working group and relates to whether in their judgement they are satisfied with the level of family income they expect to have in retirement. The third model, a logistic model, relates to both working and retired groups, to whether the respondent felt his or her total income was enough to meet basic everyday needs (including accommodation, food and clothing and other necessities). The results of each of the three models are summarised in Table 6-1, while the complete results are given in Appendix Tables C.2–C.4. The direction and level of significance of the association of each of the variables with income is indicated by plus and minus signs for a positive and negative association respectively. The number of signs indicated the degree of significance: three is associated with a variable whose regression coefficient was significant at the 1% level while two and one are associated with 5% and 10% significance levels respectively.

Notes:
¹ Only those explanatory variables that were statistically significant are shown in this table.
+++ or --- = significant at the 1% level; ++ or -- = significant at the 5% level; and + or - = significant at the 10% level.
² The table of full results is given in Appendix Tables C.2-4.

For those in the working group, their income is higher than among the retired group by some $33,000. The results indicate that after controlling for other factors we fail to reject the hypothesis that neither the physical nor mental health scores are associated with the level of income. While improved health, both physical and mental, is found to be positively related to income, the relationships are not statistically significant. Working respondents with better mental and physical health reported that they are less likely to be satisfied with their expected income in retirement. This could imply they were more concerned to increase their retirement savings to achieve a better standard of living in retirement. However, in contrast, amongst all respondents, there was a strong positive association between their health scores and the ability to meet basic needs with their current income. While higher health scores were not associated directly with higher incomes, better health was associated with a positive view of the ability to meet basic needs.

While there is a highly significant association between better health and the respondent reporting they had enough income to meet basic needs, it is important to assess the absolute magnitude of the effect. In some cases this might be very modest even though it is statistically significant. One way to assess this is by the use of the odds ratio. Based on this measure, a five point increase in the physical and mental health scores would result in an increase in the odds of being able to meet basic needs of 11% and 14% respectively.

In a similar manner, one can compute the impact of being on a benefit. For those on a benefit, the odds of claiming to have enough income to meet basic needs is 65% lower than the odds for those not on a benefit. Likewise, being Māori implies that the odds of claiming to be able to meet basic needs from reported income is 25% less than for NZ Europeans, after controlling for a wide range of other variables.

7  Living Standards#

Print this section

In this section we summarise the key factors associated with a range of measures designed to capture overall living standards. Table 7-1 reports two measures: the Economic Living Standard Index (ELSI) described in Section 3.3, and a measure of hardship. The latter is defined in terms of the number of measures taken to reduce costs in the previous year and includes elements such as staying in bed to reduce heating costs, postponing visits to a doctor, not buying fresh fruits, etc.

Notes:
¹ Only those explanatory variables that were statistically significant are shown in this table.
+++ or --- = significant at the 1% level; ++ or -- = significant at the 5% level; and + or - = significant at the 10% level.
² The table of full results is given in Appendix Tables C.5-6.

Better physical and mental health are strongly associated with higher living standards as captured by these two measures. Māori have a significantly lower ELSI than NZ European, and take significantly more measures to control household expenses than do NZ Europeans. Recipients of NZS have a significantly lower score for the ELSI, in contrast to those with other forms of superannuation. Higher incomes and wealth predictably are associated with higher living standards and less deprivation.

Table 7-2 summarises the most salient results associated with a person’s self-rated standard of living and the extent to which they are satisfied with their current material standard of living, expanding on the results in Figure 3. Both measures are binary variables formed from categorical responses. Those rating their living standards as high, fairly high or medium were coded as 1, while a 0 was assigned to those responding that they rated their living standards as fairly low or low. Similarly, those who were very satisfied or satisfied with their material standard of living were assigned 1, while those who were neutral, dissatisfied or very dissatisfied were coded as 0.[22]

Notes:
¹ Only those explanatory variables that were statistically significant are shown in this table.
+++ or --- = significant at the 1% level; ++ or -- = significant at the 5% level; and + or - = significant at the 10% level.
² The table of full results is given in Appendix Tables C.7-8.

A person's assessment of their living standards is a subjective judgement and depends on a wide set of personal attributes, many of which are unobservables. Of the variables included in the models represented in Table 7-2, many are found to have no significant association with how an individual views their standard of living. The main exceptions are that higher income and wealth are strongly associated with a higher assessment of living standards. Better physical and mental health are associated with greater satisfaction with an individual’s material standard of living. In short, health and wealth matter. Other observable attributes show no systematic association with subjective assessment of living standards.

The next analysis relates to the impact of retirement. This is based on two questions. First, all respondents were asked whether they expected their living standards in retirement years to change (relative to their current living standards). Those expecting increased or similar living standards were coded 1, while those expecting their living standards to decline were assigned a code of 0. The results are presented separately for those who are working, and for respondents who were retired at the time of the survey. In the case of working respondents, the question relates to their expectations about their future retirement. In the case of those who are retired, the responses refer to the expectations they held before retirement. Highlights of the results for both groups are summarised in Table 7-3.

Notes:
¹ Only those explanatory variables that were statistically significant are shown in this table.
+++ or --- = significant at the 1% level; ++ or -- = significant at the 5% level; and + or - = significant at the 10% level.
² The table of full results is given in Appendix Tables C.9-10.

 

For both the working and retired groups the expectations of an increase in living standards was positively associated with better physical health. In the case of those working at the time of the survey, those with better mental health expected an increase in living standards. Working Māori respondents expected an increase in living standards, compared to NZ Europeans. Those placing weight on negative aspects of retirement expected a decline in living standards compared to those who did not.

Among the working respondents there was no significant effect of income and only a marginally significant effect of wealth on their expected living standards. This finding is consistent with the implications of a life cycle model of consumption smoothing in which individuals are assumed to adjust their spending and saving during their working years so as to avoid a major drop in living standards following retirement.

The second question asked whether retirement years would be better, the same (coded 1) or not as good (coded 0). The results of this model for the working group are given in the last column of Table 7-3. Given that this question is closely allied to the previous one relating to changes in living standards, it is not surprising that the results are very similar. Physical and mental health matters – better health is associated with a higher expectation that retirement years will be better. Māori expected a decline in retirement years, while those currently on a benefit or receiving NZS expected an improvement in retirement years. Positive and negative feelings towards retirement correlate with positive and negative impacts on expectations. Respondents reporting higher levels of total current wealth expect retirement years to be better.

8  Health#

Print this section

In the first part of this section we present summary measures of health status (Section 8.1). The second part analyses the factors associated with health status.[23]

8.1  Measures of health status#

Print this section

As described in Section 3.4, the HWR has extensive coverage of health. From these questions a physical and mental score can be computed as well as a combined, or overall, score. In addition, respondents were asked to rate their health as excellent, very good, good, fair or poor.

We first consider the component scores for mental and physical health. Figure 10 summarises the distribution of these health status scores for males and females. Overall there is little difference between the sexes. Considering both genders together, and taking the mean as 50, 45% of all respondents are below the mean physical score, 37% are below the mean mental score and 35% below the mean overall, combined score.

Figure 10: The distribution of health status scores: by sex

 

Source:  Health, Work and Retirement Survey

Second, we examine the extent to which there is a systematic pattern of these component scores in relation to the self-reported health categories. Table 8-1 summarises the mean scores for each self-reported health category from the HWR survey respondents.

The mean scores for both mental and physical health decline systematically with poorer self-reported health status. For those whose self-reported health is given as poor, the component scores for mental and physical health fall to 80% or below of the overall mean values.

Source: Health, Work and Retirement Survey

Clearly these various measures are related and to the extent they were highly correlated their use as separate independent measures would be compromised. However, as indicated in Table 8-2, the cross-correlations with the self-rated measure are modest, as is the correlation between the mental and physical component scores. Logically, the component scores are more highly correlated with the overall score.

8.2  Factors associated with health status#

Print this section

In this section we analyse factors associated with health status. In each case (physical, mental and overall) the continuous score (H^p, H^m and H^o) was regressed on a common set of explanatory variables (Z_i), with a view to identifying the factors that are most strongly associated with health status, either positively or negatively.

H^p = β^mH^m + Σβ_i^pZ_i + ε^p (9)

H^m = β^pH^p + Σβ_i^mZ_i + ε^m (10)

H^o = Σβ_i^oZ_i + ε^o (11)

Figure 11: Factors associated with health status

 

Source:  Health, Work and Retirement Survey

Notes:
¹ The numerical scores refer to positive and negative effects at the 1% (=3), 5% (=2) and 10% (=1) levels of statistical significance.
² A summary of the regression results is given in Appendix Tables C. 11-13. Complete results are available on request.

Summary results from these three regression models for the different health scores, for the working group[24] and broken down by male and female are in Appendix Tables C.11–13.

The overall results are summarised in Figure 11 where bars to the right are factors that are associated with better health and those to the left with poorer health. The length of the bars is proportional to their statistical significance, so that a bar of three units in length is associated with a variable whose regression coefficient was significant at the 1% level. Bars of lengths one and two refer in turn to the 10% and 5% significance levels.

More education and labour force participation are associated with a higher probability an individual will have better physical and mental health. Finding such associations does not carry implications of causality. Working is significantly associated with better health, but, as we will show in Section 9, those with better mental and physical health are more likely to be working. In short, poorer health is associated with a lowering of the likelihood of working, but working is associated with better health. Those individuals with higher income and wealth are more likely to be healthy. Again, however, their ability to accumulate wealth may well be associated with their health status (recall Table 5-1). In each of the three models reported in Figure 11, being Māori relative to NZ European increases the likelihood of poorer health.

Stephens et al (2008) report similar findings based on the HWR data. In particular they find wealth is positively associated with physical but not mental health. They also report that retirement is associated with poorer scores for both mental and physical health.

Benzeval and Judge (2001) use longitudinal data from the British Household Panel Survey to study the relation between income and health. They distinguish between permanent and transitory income as well as addressing the possibility of reverse causation. They find that long-term income is more important for health than current income; that the level of income is more important than changes in income; that decreases in income are more important than increases; and persistent poverty is more harmful to health than occasional episodes.

The question of causation between health and income has been the subject of extensive debate. A clearer understanding of the relationship is important for public policy. Would policies that improved the incomes of a target group be expected to lead to improved health status? The possibility always exists that the measured effect of income on health might be owing to reverse causality; ie, those with better health will have higher labour force participation and command higher wages, resulting in greater income. The standard approach to dealing with this endogeneity has been to use a two stage instrumental variables model. It is expected that this would purge income of the effect of health status. The challenge has always been to find suitable instruments; ie, variables associated with income but not health. Ettner (1996) uses the state unemployment rate from cross-sectional data for the USA, together with work experience, parental education and spouse characteristics. Her results support the hypothesis that the direction of causality is from income to health. However, such models remain susceptible to the choice of instruments, and as Currie and Madrian (1999).note, the evidence remains mixed

Dave, Rashad and Spasojevic (2006) use longitudinal panel data from the Health and Retirement Study in the USA to examine the effects of retirement on physical and mental health. Partly as a result of changes in lifestyle after retirement, including less physical activity and fewer social interactions, they report robust evidence that both physical and mental health deteriorates. Complete retirement is associated with 5% to 16% increase in difficulties of mobility and daily activity, a 5% to 6% increase in illnesses and a 6% to 9% decline in mental health.

9  Labour Force Participation#

Print this section

This section analyses the labour force participation patterns of older people. What are the factors associated with continuing in the workforce on a full- or part-time basis as distinct from retirement? In particular we are interested in the extent to which a respondent's health status influences the decision to remain in the workforce.

9.1  Overview of the probabilities of being in the labour force#

Print this section

The initial analysis presents the results of a simple model relating the probability of working to age. Formally the logit model is of the form:

Pr(W) = α + β(age) + ε (12)

Figure 12 provides a snapshot of the probabilities that an individual of a given age will be in full- or part-time work. Recall from Section 3.5 that those not in the workforce are classified as retired (with the exception of a small residual for whom the data was incomplete). The figure plots the predicted probabilities derived by regressing full- or part-time participation solely on age and an indicator as to being age 65 or over (no other variables were involved), with all points lying on the continuous line joining them. The actual probabilities for each age group are shown and imply an over- or underestimate of the probability of working, based on age alone.

Figure 12: Unconditional probabilities of working full- or part-time: by age and sex

 

Source:  Health, Work and Retirement Survey

For males, the probability of participation in full-time work falls with increasing age. Despite the predictable drop at age 65 (given eligibility for NZS), almost 20% of males remain in full-time employment at the age of 70. In contrast, part-time work by males rises with increasing age, with only a slight fall at age 65. Even so, by age 66, more than 40% of all males remain in full-time or part-time work. Clearly, while eligibility for NZS has a notable impact, a significant share still participates in the labour force after age 65.

The underlying pattern for females is broadly similar. The principal difference is in the absolute levels. The probability of being in part-time work is higher than for males, up to age 65. The corollary is that full-time rates for females are some 20 percentage points lower up to age 65, but they show the same downward trend as was the case for males.

Smeaton and McKay (2003) report a similar deterrent effect on labour force participation of the British state pension. They find a drop in participation of some 10 percentage points once an individual reaches the age of eligibility. This contrasts with a drop in the New Zealand participation rates at age 65 of more than twice that amount. This reflects the more generous level and universal eligibility for NZS compared to the British state pension.

Many factors other than age contribute to the decision to remain in the workforce. Critical amongst these is health. Table 9-1 summarises the population estimates of labour force status by self-reported health status.[25] There is a very marked decline in labour force participation for both males and females in the poorer health categories. While 82% of all males aged 55 to 70 reporting excellent health are working (either in full-time, part-time or seeking employment), only 53% in fair health and 28% in poor health remain working. A striking finding is that while only 28% of males who state they are in poor health are working, 45% of females in poor health are reported as working. This could arise as a result of different expectations about health and working being applied by males and females, or by males having a greater preponderance of conditions that result in both a self-rating of poor health and preclude labour force participation.

9.2  Factors influencing the labour force: retirement choice#

Print this section

It is well recognised that the decision to leave the workforce and retire is influenced by a wide range of personal and family circumstances. In this study we are particularly concerned to identify the extent to which an individual's health status influences that decision.

Figure 13 illustrates schematically the process of modelling the effect of health on the decision to work. The principal avenue explored in this study is the contemporaneous effect of health status on working (as indicated by the bold arrow). The models endeavour to isolate this effect while holding constant a range of other “confounding” variables indicated around the periphery. The dashed arrows indicate potentially important linkages from both previous health states and earlier labour market experience. However, as the data used here are cross-sectional, inclusion of these lagged effects will need to await data from future waves of the HWR survey so that longitudinal observations can be incorporated.

Figure 13: Modelling the effect of health on labour force participation

 

A key message from this schematic view is that health status may influence the decision to work in a number of ways, as argued by Dwyer and Mitchell (1998). Poorer health may result in lower earnings owing either to reduced productivity or fewer hours worked, or some combination of the two. Poorer health may alter the preferences for leisure versus consumption, or make work more demanding. All these factors would tend to raise the utility of leisure and lead to earlier retirement.

In contrast, the needs of a person with poorer health might be such that they opt to continue working at least part-time to maintain consumption, thereby postponing retirement. The onset of a decline in health may lead an individual to alter the perception of their life expectancy. This may lead to an earlier retirement than would otherwise have occurred had their health not deteriorated. However, for some, continuing to work may provide a psychological boost which itself could mitigate, at least in part, the effects of poorer health. [26] In short, the effect of health on the decision to remain in the labour force is theoretically ambiguous. We are left to appeal to the evidence, recognising that the limitations of both the data and the models may not necessarily result in an unequivocal outcome.

There has been a long-run trend throughout the 20th century toward earlier retirement. Clearly, many factors shape the final decision. What effect does better health have? Does it mean that life expectancy is extended and hence the retirement date is shifted out so that working life simply remains proportional to life expectancy? Bloom, Canning and Moore (2004) extend the standard life cycle model to show that improved health hasbeen associated with a less than proportional increase in working life; ie, more time is spent in retirement as life expectancy rises. On one hand, better health means that the disutility of work is less, so retirement age could be expected to rise. But these authors note that a longer working life means a greater effect of compound interest earnings on savings, creating a wealth effect leading to more leisure (earlier retirement) and higher consumption (implying lower savings).

In a UK study, adverse health shocks are found to be an important predictor of individual retirement behaviour (Disney, Emmerson and Wakefield, 2003). This study relied on self-rated health status, but attempted to deal with the problem of endogeneity by using a constructed measure of a health stock Similar findings are reported by Banks and Tetlow (2008) who demonstrate that it is the onset of a major health condition rather than a chronic illness that increases the odds a person will leave full-time work.

Clearly, many factors shape the final decision to continue participating in the labour force. The final decision is an amalgam of health status (including that of family members), wealth levels, interest rates, the disutility of work, the extent of publicly provided pensions, health services and long-term care, and the strength of the bequest motive.[27] For this reason, it is important to use a multi-variate framework, if there is to be any chance of isolating the effect of a particular variable on the decision to retire.

We now turn to the application of logit models whose aim is to isolate the effect of health on the decision to remain in the labour force, while holding constant the influence of a wide range of other factors that may potentially influence the decision taken by an individual. These effects were estimated by fitting a logit model of the form:

Pr(W) = W(H^p,H^m,Z) + ε (13)

where the probability of working is postulated to depend on the physical and mental health component scores (H^p, H^mrespectively) and a vector (Z) of control variables (including age, gender, marital status, region, migrant status, education level, benefit status, income, wealth, etc) plus a random error term (ε).

A summary of the main significant factors associated with the decision to work (defined as either full-time, part-time or seeking work) for males and females is given in Table 9-2. Estimates of the marginal effects are given in Section 9.3. Both physical and mental health are critical factors influencing the decisions by males whether or not to work. However, only their physical health status appears to influence the decisions of females.

While Māori men and women were less likely to be in the workforce, the differences were not significant. Having a tertiary education significantly raised the probability that both males and females would be in the labour force.

Relative to being married with a non-working spouse, virtually all other marital categories are associated with a higher chance of being in work. For both men and women, being separated or widowed significantly raised the probability of working, as did having a working spouse.

The receipt of a benefit or NZS lowers significantly the chance of being in the workforce for both males and females. While the level of total wealth typically has no effect, males are more likely to be in work (full- or part-time or seeking) when the income of other household members is higher. For males this is also true if they are in receipt of other sources of retirement income. Only half as many women as men have other superannuation schemes, and the effect for females, while still negative, is very much smaller and not significant.

It is notable that the effect of wealth is not significant. One might have expected that those with a taste for work are likely to have accumulated a greater stock of wealth. If this were the case, we would expect to find a positive relation. While it is true that the effect is typically found to be positive in this study, the coefficients for wealth are not statistically significant. There are at least three issues, however, that may have mitigated against finding a significant role for wealth. Ideally, we require a measure of the net stock of wealth but as the HWR survey did not collect estimates of liabilities we are restricted to using gross wealth. Second, the response rate on estimates of the value of major assets was relatively low, as respondents were invited to provide this information only if they so wished. Third, the NZS system provides a lifetime defined benefit on a universal basis, free of income or asset testing. As shown by Scobie et al, (2005) the implied stock of wealth associated with NZS forms a major share of the total retirement wealth of many New Zealanders. The incentive to accumulate other forms of wealth is reduced, relative to that which would prevail in the absence of such a state pension scheme. As a consequence it is possible that some of what is actually a wealth effect is being picked up by the highly significant effect of receiving NZS. We found some tentative evidence to support this explanation. As shown in Table 9-2, both males and females in receipt of NZS tend to have a lower probability of working, after controlling for their level of wealth and an extensive set of other variables.

A notable finding is that health status, as measured by the mental component score, has no significant effect on the labour force participation decisions of women. This stands in marked contrast to the results for males, whose decision to work is strongly related to both their physical and mental health scores.

Notes:
¹ Only those explanatory variables that were statistically significant are shown in this table.
² +++ or --- = significant at the 1% level; ++ or -- = significant at the 5% level; and + or - = significant at the 10% level.
³ The table of full results is given in Appendix Table C.14.

 

To explore further this apparent anomaly the labour force equation for women was re-run including only physical and mental health scores. Again, mental health was not significant. We conclude that this is a robust finding; however, the reasons for it remain as a matter for further research.

In Section 3.5, a group of people amounting to some 40,000 out of 609,000 were dropped from the analysis as there was insufficient information about their status. A comparison of their characteristics with those who identified as retired, revealed that they tended to have a higher proportion of females, be less healthy, have lower median income and almost no wealth. When the logistic model for working was re-run including this group the results remained broadly as shown in Table 9-2. The only notable changes were that for males on a benefit, there was a significant reduction in the probability of working, and for females, wealth became a significant factor positively associated with working.

The question as to when to retire (ie, withdraw from paid employment) is a critical decision that older workers must make. One of the shortcomings of many studies of retirement choices is that they potentially suffer from a bias when subjective measures of health are used. This arises when the respondent provides a more pessimistic assessment of their health status in part to justify their decision to retire.

As people age, the probability that they will choose to retire increases. Clearly it is expected that the decision to retire will, at any given age, be influenced by the person's physical health status. To explore this, Figure 14shows the declining pattern of labour force participation for three points taken from the range of physical component scores.[28] These correspond to the 25th percentile, the median and the 75th percentile. Those at the 25th percentile of the physical health score have, unsurprisingly, the lowest rates of labour force participation. At age 55, males at the 25th percentile of the physical health score are 3 percentage points less likely to be working than those at the 75th percentile. However, by age 70, the effect of health status is considerably more pronounced and the inter-quartile range is 12 percentage points. For females, a similar pattern holds; however, the decline with age owing to poorer health is less marked. At age 55 the difference is 2 percentage points rising to almost 8 percentage points at age 70.

Figure 14: Probability of working at three quartile levels of physical health: by sex

 

Source:  Health, Work and Retirement Survey

Note: P25, P50 and P75 refer to the 25th percentile, the median and the 75th percentile of physical health respectively.

There are two steps that can be taken to overcome this problem. The first is to restrict the analysis to those who are currently in the labour force, and focus on their expectations about their retirement age. The second is to use more objective measures of health status. The HWR survey allows us to incorporate both these steps. A logit regression of the following form was fitted:

Pr(Working at 62 or 65) = α + β_pH^p + β_mH^m + β_zZ + ε (14)

where the probability of expecting to be still working full-time at age 62 (or 65) is expressed as a function of health measures (H^p and H^m) and a set of conditioning variables (Z).

Table 9-3 summarises an analysis based on the expectations of respondents about future retirement plans. Males are significantly more likely to expect to remain in full-time employment through age 65 than females. Again, marital status is an important factor affecting the decision to continue in full-time work. Attitudes to retirement clearly matter; those with positive attitudes being less likely to plan on being in full-time work at 62 or 65.

It is to be expected that a worker's health will influence the expected retirement age. While better health as measured by the physical and mental health scores does increase the likelihood that a person will remain in employment at 62 or 65, the effect was not statistically significant. In contrast, while the health scores do not appear to directly affect retirement expectations, respondents considered that their own health and that of their family members did have a bearing on their retirement plans.

Notes:
¹ Only those explanatory variables that were statistically significant are shown in this table.
² *** = significant at the 1% level; *** = significant at the 5% level; * = significant at the 10% level.
³The complete results are given in Appendix Table C.19.

McGarry (2002) reports similar findings for the USA based on an analysis of data from the Health and Retirement Study (HRS).[29] She finds that health is a more critical factor in retirement expectations than income and wealth, a result paralleled in the present study. A UK study reports that the main reason both men and women give for early retirement is their health or that of others (Banks and Smith 2006).

Cobb-Clark and Stillman (2006) use the first three waves of the Household Income and Labour Dynamics in Australia (HILDA) survey, covering the years 2001 to 2003, to examine the retirement plans of middle-aged men (aged 45 to 55) and women (aged 45 to 50) in their prime working ages. It is of note that these results for Australia, like the present study, found no effect of health status. Because the authors were able to use Waves 1 and 3 of HILDA they could further explore the effect of a change in health status; here, a deterioration in women's health between the two waves did result in a significant reduction in the age of expected retirement. Poorer health is likely to lead to reduced life expectancy and deteriorating health tends to be associated with reduced consumption expenditures (Olsho, 2005). As a result, individuals are likely to leave the workforce earlier with lower accumulated retirement savings.

The HWR survey asked those respondents whether, when they retire, they wish to stop any paid work or continue with some paid work. It is of interest to ask whether the characteristics of those who plan to continue differ from those of the respondents who plan to stop entirely. A summary of some elected characteristics is given in Table 9-4. Tertiary-educated males are more likely to continue in paid work, as are respondents who were separated or widowed. Those whose self-assessed health was poor are more likely to stop entirely, although the objective physical and mental health scores are virtually the same for the two groups.

9.3  Measuring the marginal effects on labour force participation#

Print this section

Table 9-2 provided a ready overview of those factors significantly related to the decision to work. However, these results, taken alone, do not indicate the magnitude of the effects. It is possible that a variable has a significant effect, yet compared to another, its absolute effect on the probability of working might be quite small. To assess the magnitudes, we calculate the marginal effects.[30]

The probability of remaining in the workforce (either full- or part-time or actively seeking work) as distinct from being retired (with no paid work) is calculated by setting all variables except the one of interest to their mean values in the case of continuous variables, and to zero in the case of binary variables. The mean values were chosen to correspond to the group of interest (eg, mean values for male and female splits are different).

The calculation is then repeated with a change made in the variable of interest. In the case of a continuous variable such as age, the second calculation is made with age increased by one year while all other variables are held at their means. In the case of a binary variable such as whether or not the respondent was separated, the variable is set to one. Exceptions to this apply in the case of the physical and mental health scores (both continuous) where the results are reported for a five unit change deemed to be clinically significant. In addition, for migrants the number of years in New Zealand is increased by five rather than one, simply to scale up what would otherwise be a very small, but significant, effect.

Notes:
¹ Only variables whose coefficients were statistically significant are listed in the table.
² The complete results are given in Appendix Table C.14(a).

Notes:
¹ Only variables whose coefficients were statistically significant are listed in the table.
²The complete results are given in Appendix Table C.14(b).

The results for males and females are shown in Table 9-5 and Table 9-6 respectively, and shown graphically in Figure 15. Overall, the pattern of results is broadly similar for males and females. The largest absolute effects on increasing the probability of working stem from a respondent’s marital status and whether the spouse is in the workforce. For example, the probability that males who are not separated or divorced are in the workforce is 76%. For those who are separated or divorced the probability of working rises to 91%, so the marginal effect is +14 percentage points. The corresponding figures for females are 50% rising to 92% for those separated or divorced, a marginal effect of +42 percentage points.

Figure 15: A summary of the factors that change the probability that a person will be retired: by sex

 

Source:  Health, Work and Retirement Survey

Note: The numerical scores refer to positive and negative effects at the 1% (=3), 5% (=2) and 10% (=1) levels of statistical significance.

In contrast a clinically significant unit improvement in physical health only raises the probability by +2 percentage points for both males and females. However, it is not surprising that improvements in health above the mean level have a modest effect on participation, which is already reasonably high. Also, given that there is some correlation between mental and physical health scores, it is reasonable to expect that the combined effect would be greater than changes in either alone.

To more fully explore the effect on participation rates we estimated the impact of simultaneous changes in physical and mental health scores above and below their respective means. The results are given in Table 9-7and Figure 16. An increase of five units in both scores now leads to a 3 percentage point increase in participation rates for males and females. In contrast, consider a 20 point decrease in the health scores. This corresponds approximately to a change in self-reported health from excellent to poor. At this level, male participation rates fall by 26 percentage points, from 90% to 64% (assuming all other variables are set at their mean values). The corresponding fall for females is from 81% to 68%.

Participation rates drop off at an increasing rate as health scores deviate further below their mean values. However, the decline is much more marked for males. For example, if the mean health scores fall from 10 to 20 units below their means, female participation rates decline only by 7 percentage points. In contrast, male participation for the same decline in health scores falls by 16 percentage points. So while male participation rates are higher than those for females, they decline more rapidly as health scores fall further below their means (Figure 16).

Figure 16: Health scores and participation rates

 

9.4  The choice between full- and part-time work#

Print this section

The decision to work rather than retire can be broken down into two stages. The first is whether the person chooses to be in the labour force at all; the second, having decided that they will be in the labour force, the subsequent decision is that of working full- or part-time.

Given that a person is employed, what factors are associated with their choice to work either full- or part-time? To analyse this, a logit regression model was estimated with a binary dependent variable set at 1 = full-time and 0 = part-time, and restricting the sample to those who were employed (as distinct from seeking employment). In broad measure the results (Table 9-8) were similar to those reported in Table 9-5 and Table 9-6. The notable exception was the health measures. The probability that a person in the workforce would chose full-time employment was not significantly related to either the physical or mental health scores. While physical health status has a significant effect on whether to work or retire, the evidence is that, given a person is employed, their choice about full- or part-time work is not a function of their health status.

Notes:
¹ ns = not significant at the 10% level; ++ = significant at the 5% level; +++ = significant at the 1% level.
² Complete results are given in Appendix Table C.18.

Both males and females have a lower probability of working full-time as they age, receive a benefit or have income from superannuation. In contrast they are more likely to be in full-time employment if they are widowed or have dependants. For men, having a working spouse increases the probability of working full-time (compared to having a non-working spouse), and having a tertiary education or higher income from other household members decreases their probability. For females, they are more likely to be in full-time employment if they are Māori, are separated or have a superannuation scheme.

The phasing of retirement has been documented in a survey conducted by the Equal Employment Opportunities Trust of workers in New Zealand aged between 45 and 64 (2006). Almost 50% of those surveyed chose doing part-time work as their ideal transition to retirement. Finances and health were given as the two most prevalent factors affecting the decision when to retire.

 

9.5  Using a self-reported health measure#

Print this section

To this point we have analysed the impact of health status on labour force participation using the SF-36 measures of physical and mental health. An alternative approach is to use the self-reported health measure based on the response to the question:

In general would you say your health is: Excellent; Very Good; Good; Fair; or Poor?

Previous researchers have undertaken studies aimed at assessing the relative merits of so-called “objective” measures and self-rated measures. Bound (1989) identifies four reasons to be suspicious of the survey responses to a question soliciting the indivdual’s own rating of their health status. First, as this is a subjective judgement, there is no reason to suppose that each person would use the same subjective scale, thus raising the possibility that the results are not strictly comparable across respondents. Second, the responses may be influenced by the very labour market outcomes that are to be explained. Third, a person may use limitations owing to health as a reason for their not participating (the so-called justification bias). Finally, some may have a financial incentive to qualify for a disability allowance as a means of taking early retirement.[31]

A number of studies have explored the justification bias. Anderson and Burkhauser (1985) use early mortality data derived from longitudinal panel data, and find this more objective measure had a smaller effect on labour supply than did self-rated health status.

One approach has been to use objective measures as an instrumental variable, hoping thereby to overcome the problem of endogeneity; eg, Stern (1989). However Bound (1989) reports that while this has appeal, it leads to understating the effect of economic variables on retirement. He concludes that neither of the measures of health status will generate reliable estimates of the effect of health on labour force attachment of older males. However, as each tends to produce estimates biased in opposite directions, results from utilising both measures may well bound the actual effect.

Arguably, the risk of justification bias is lower with the SF-36 measures, given they are based on an extensive set of questions. In contrast, the single question soliciting self-reported health status may be subject to a much greater risk of bias by those who feel they need to justify their withdrawal from the labour force on health grounds. In other words the apparent health status based on a self-reported measure may not necessarily be the true underlying health status of the respondent.

Notwithstanding, it is of interest to test whether labour force participation is associated with the self-reported health measure. The results are summarised in Table 9-9. The estimated coefficients show that there is a significant association between health status and the probability of working. This reinforces the results obtained using the SF-36 measures. In both cases, health status has a significant association with the decision to retire or remain in the workforce.

For each health status, bar “excellent”, there is an odds ratio based on the estimated coefficients from the logit regressions. These ratios give the amount by which the odds of working must be discounted relative to the odds of working when self-reported health is excellent. For example, the odds of a female in fair health being in the workforce are 0.36 times those of a female in excellent health. Both males and females reporting either fair or poor health have significantly lower odds of working than those with excellent health. In the case of males, those reporting good health the odds are only half those in excellent health. Recall that these estimates are based on holding all other factors constant (eg age, income, education, marital status, ethnicity, etc).

Notes:
¹ ns = not significant at the 10% level; ** = significant at the 5% level; *** = significant at the 1% level.
² The odds ratios are found as

 

where

 

is the estimate of the k-th coefficient from the logit regression.
³Complete results are given in Appendix Table C 15.

 

We now supplement our earlier analysis of the marginal effects on participation (Section 9.3) by incorporating self-reported health measures. Recall that while the marginal effects of a five unit change in the component scores for physical and mental health were typically significant, their magnitude was apparently small (of the order of 1 to 2 percentage points).

As a first step we repeat the estimation of equation (13) replacing the health scores with the self-reported measures (denoted SR):

Pr(W) = W(SR,Z) (15)

while maintaining the same set of control variables (Z). As the self-reported measures are a set of categorical variables, it is necessary to delete one category (in this case Excellent) such that the estimates for the remaining categories are measured with respect to the deleted category. The results are summarised in Table 9-10, showing that males in poor health have a participation rate 11 percentage points lower than those in excellent health with a corresponding 12 points for females.

We now estimate the marginal effects on labour force participation by utilising the changes in health scores associated with the self-reported categories, and applying these changes to the logit model estimated on the basis of health component scores (ie, equation (13)). For example, for males, the average physical health component score for those reporting excellent health is 57 (see the first cell of Table 8-1), while those reporting very good health have a mean score of 53.

Note:
** = significant at the 5% level; *** = significant at the 1% level.

Based on this four point fall in the physical component score, participation would fall by an estimated 2 percentage points (Table 9-11). Note that this result closely mirrors that found for the five point change in the physical score described in Section 9.5. In the right-hand panel of Table 9-11, the marginal effects on participation are shown for the combined effect of both physical and mental scores. This is done as when respondents provide an estimate of their self-reported health status they presumably give an assessment derived from their perception of both their mental and physical health.

9.6  Using estimated wage rates#

Print this section

It is critical that the effect of the economic variables be captured accurately in order that the underlying relation between health and participation in the labour force is correctly revealed. If the model is misspecified, some of the effect owing to economic forces might be incorrectly attributed to health status. Up to this point the analysis has used either respondent income or the income of other family members.

Anderson and Burkhauser (1984) demonstrate that different measures of health status can alter the interpretation of economic variables. In this case they were concerned with the effect of wage rates. Two questions arise: Are wage rates an alternative to the income measures we have used to this point? Is the effect of wage rates modified by using different health measures?

To incorporate wage rates in our analysis we undertook the following steps. First we estimated an average hourly rate (w) based on the respondent's reporting of hours worked (h) and their total pre-tax income (Y).

 

(16)

We then fitted a wage equation of the form

w = α + Σβ_kZ_k + ε (17)

to all observations that reported non-zero hours, and from this used the predicted values (^ŵ) as the reservation wage for those not in the workforce. The vector Z of explanatory variables included age, education, marital status, etc. The final step was to use the log wage as an independent variable in the univariate logistic model for labour force participation.

On theoretical grounds alone, it is not certain whether a higher wage would increase or decrease the probability of working as distinct from retirement. There is both an income and a substitution effect. A higher wage raises the opportunity cost of leisure and would therefore tend to diminish the demand for leisure (ie, increase the propensity for labour market work). However, a higher wage also implies a higher income (for any given hours worked) with a greater opportunity to accumulate wealth for retirement. Whether the phenomenon of the “backward bending supply of labour” prevails is then an empirical question.

Notes:
¹ ns = not significant at the 10% level; ** = significant at the 5% level; *** = significant at the 1% level.
² Complete results are given in Appendix Tables (C.14-17)

The results are summarised in Table 9-12. Four separate logit regressions for labour force participation were run for both males and females. The first two, reported in the upper half of the table, used the estimated wage rate as the economic variable. It is intended to capture the effect of the wage on the decision to retire or not. In addition, separate runs were made using the SF-36 health measures and the self-reported health status. The latter results are given relative to excellent.

In the lower half of the table the runs of the logit model are repeated but this time using the income of other family members as a proxy for the effect of economic incentives on the decision to retire. There are a number of important conclusions to be drawn from this analysis. First, in all cases health has a significant influence on the decision to remain in the workforce or retire. Second, when the health measures from the SF-36 are used, both physical and mental health scores are significant for males; higher scores (meaning better health) are associated with a higher probability of remaining in the workforce. However for females, there is no significant effect on labour force participation from mental health scores. Third, based on the self-reported health status, both males and females reporting good (when the wage rate model is used for females), fair or poor health are significantly less likely to be in the workforce than those reporting excellent health.

Fourth, the wage effects are similar regardless of which measure of health status is used. For both males and females, higher wage rates lower the probability that the person would remain in the workforce. Finally, the overall conclusion that health has a significant influence on the decision to remain working holds for males and females regardless of which measure of health or which economic variable is used. It appears to be a robust finding.

9.7  Comparisons with results from SoFIE#

Print this section

In this section we present a brief comparison of the factors that affect the probability of working drawn from the present study based on the HWR survey and a similar study using data from the Survey of Family Income and Employment (SoFIE).[32] Data in SoFIE cover a wide range of ages so that the first step was to select only those respondents aged 55 to 70 years, thereby matching the sample in the HWR survey. Similarly, using SoFIE it is not easy to distinguish between “not in the labour force” and “retired”. As a result, the HWR population was widened to include the others in the survey, who were all placed in the “not in labour force” group with the retired group. This involved those identifying themselves as students, homemakers and others. A logit regression for the probability of working was estimated for this group using a set of explanatory variables common to both surveys. This inevitably meant that some of the explanatory variables used earlier in this study could not be included, as there was no comparable measure in the SoFIE data set.

In broad measure the results, summarised graphically in Figure 17, are comparable. Marital status (having a working spouse and being single), better health (physical and mental) and being educated to the tertiary level significantly increase the probability that a person will remain in the workforce. Being on a benefit reduces the probability in both surveys.

Figure 17: Factors influencing the probability of working: a comparison of HWR (Massey) with SoFIE

 

Source:  Health, Work and Retirement Survey and Survey of Family, Income and Employment

Notes:
¹ The numerical scores refer to positive and negative effects at the 1% (=3), 5% (=2) and 10% (=1) levels of statistical significance.
² The complete results are given Appendix Tables C.21 and C.22

10  Chronic diseases#

Print this section

As well as the measures for physical and mental health, the HWR survey records whether a respondent has ever been told by a health professional that have any one of 19 separate so-called “chronic” conditions. The purpose of this section is to explore the extent to which the presence of any one or more of the 19 chronic diseases are associated with a lower probability of an individual being in the labour force. A secondary objective is to assess the overall impact on the labour supply of older workers from particular chronic diseases. This will depend not only on the lower probability that an individual sufferer will be in the workforce but, in addition, the prevalence of the condition in the population.

10.1  Extent of chronic disease#

Print this section

A summary of the chronic diseases, their prevalence rates amongst this sub-population of 55 to 70 year-olds and the distribution of the self-rated health status for each illness is given in Table 10-2. Of those reporting no disease, nearly three-quarters of respondents rate their health as excellent or very good. However, in the presence of any one or more of the 19 conditions, this proportion drops sharply to under 40%. Stroke together with liver and kidney disorders are the conditions that lead to the lowest health rankings based on those reporting fair and poor health. In addition, 16% of respondents suffering from chronic liver conditions rate their health as poor, the highest proportion of any condition.

Although correlation is not considered, Appendix Table C.23 has the correlation matrix for the chronic diseases. The largest is 30% between asthma and respiratory conditions, the second largest is 21% between high blood pressure and heart conditions and the third largest is 18% between high blood pressure and diabetes.

Heavy drinking and smoking are associated with a much lower proportion of respondents reporting excellent or very good health, compared with those who have no chronic disease.

The effect of a chronic disease on the distribution of self-rated health status is only one measure of its importance. Prevalence as well as severity matters. So while stroke and kidney and liver conditions are severe in their impact on health status, their prevalence is low relative to skin cancer, heart conditions, asthma, high blood pressure, arthritis and hearing loss. Figure 18 summarises data on prevalence by ethnicity. It is notable in all but four of the 19 conditions that Māori respondents report higher rates of prevalence. The two largest differences are for blood pressure and diabetes.

Figure 19 summaries data on prevalence by work status. The retired group reports a greater prevalence of all diseases relative to the working group (with the exception of hepatitis, where the sample size is very small). Again, arthritis, high blood pressure and heart conditions are the most prevalent conditions and the ones where the difference between the retired and working groups is most marked.

Figure 18: Disease prevalence by ethnicity

 

Source:  Health, Work and Retirement Survey

Figure 19: Disease prevalence by work status

 

Source:  Health, Work and Retirement Survey

Of those reporting no chronic conditions, 80% are working. However, in the presence of any one or more of the diseases listed (Table 10-3) the proportion working falls to 65%. The three most significant conditions seemingly associated with reduced labour force participation are liver conditions, arthritis and heart trouble. However, a truer measure of the importance of diseases on labour force participation requires that the fall in participation be weighted by the prevalence; for example, skin cancer is (0.67-0.80)*10%= -13. When this is done, arthritis, high blood pressure and heart trouble are the three largest prevalence-weighted conditions reducing aggregate labour force participation amongst those aged 55 to 70. Note, however, the estimates of prevalence made by this method have not controlled for the influence of other factors. We return to the question of prevalence in Section 10.3.

To this point the results have been based on simple bivariate comparisons. However, it is important to hold other factors constant and this was done in a logit model for the presence of any disease. Formally we fitted the following:

Pr(respondent has any chronic disease) = α + β^oH^o + β^a (age) + Σβ_kZ_k + ε (18)

where H^o is the overall health score, and Z_k a series of control variables.

The results are presented in Figure 20. As overall health status improves the likelihood of reporting the presence of any chronic condition falls markedly. However, at any given overall health score, the likelihood increases with age.

 

Figure 20: Likelihood of a chronic condition by age and overall health status

 

Source:  Health, Work and Retirement Survey

10.2  Influence of chronic diseases on participation#

Print this section

We turn now to the impact that the presence of chronic diseases have on the likelihood that an individual is in the labour force. One way to assess this is to calculate the odds of working. These findings are summarised in Figure 21. In each case the bar represents the odds ratio, where if the confidence interval includes one, then the odds ratio is not significant. This is interpreted as the factor by which the odds of working are changed owing to the presence of the particular illness. For example, the odds that a woman with high blood pressure would be working are only 70% of the odds of a woman with identical characterisitics who is not afflicted with high blood pressure. In this context this means that there is no specific allowance for the presence of multiple illnesses or the possible impact of interaction effects between various illnesses in the one individual.

For males the presence of cancer (other than skin), epilepsy, heart problems and epilepsy all lower the odds that the individual would be working relative to the odds that an individual with identical characteristics who did not suffer from the chronic illness in question would be working. Perhaps perversely, asthmatics are more likely to be working relative to those who do not have asthma. For females the conditions that lower the odds of working are diabetes, high blood pressure and liver conditions. Those with skin conditions are more likely to be working.

Figure 21: Odds ratios for working and retired: by chronic illness and sex

 

Source:  Health, Work and Retirement Survey

Notes:
¹ Bars coloured red are based on statiistically significant coefficients in the underlying logistic regressions. For each illness, the upper and lower 90% confidence intervals are plotted.
² The graph is stopped after 1.5, although the value and confidence intervals may have larger values.
³ The data for the odds ratios are given in Appendix Table C. 20.

In a USA study using males aged 51 to 61, Dwyer and Mitchell (1998) find that severe physical limitations, stroke and heart problems reduce the expected age of retirement by one to two years. Paralleling the present study, a wide range of other chronic conditions had no significant effect.

Some caution is needed in interpreting these results. In addition to the fact that for both males and females the confidence intervals are in many cases relatively wide, there can be questions about the accuracy of the self-reported incidence. Baker, Stabile and Deri (2004) examine Canadian data in an attempt to assess the accuracy of self-reported disease incidence. They match the self-reported answers to data for the corresponding individual held by the Ontario Health Insurance Plan. They take the medical records in the insurance plan as their reference point (the “truth”) and quantify the incidence of false positives and false negatives in the self-reports for all the major disease categories. False positives are cases where the individual reports the presence of an illness but there is no corresponding entry in the insurance records. In contrast, false negatives are those cases in which the presence of the disease is recorded in the insurance files but not in the self-reported response. Their findings are somewhat disturbing – for many diseases the error rate in the self-reports was up to 50%. They demonstrate that this can lead to significant biases when using these measures as explanatory variables in regression models.

We conclude this section by analysing the marginal impact of particular chronic conditions on the probability of being in the workforce. The results are summarised in Table 10-4. Note that only those variables whose underlying coefficients were significant are included. These results demonstrate the absolute magnitudes of effect on labour force participation of up to 19 chronic conditions. For males, epilepsy, cancer (other than skin) and ulcers are three conditions that have the greatest impact in reducing the probability of working. Table 10-4 also includes two other conditions for males which were almost significant; namely liver conditions and high blood pressure. For females, liver problems, diabetes and high blood pressure are the chronic conditions that had a significant depressing effect on the probability of working.

Notes:
¹ Only variables whose coefficients were statistically significant are listed in the table.
² The full results are in Appendix Tables C.20(a) and (b).

In interpreting the findings in Table 10-4 it must be emphasised that the marginal effects on labour force participation of a particular disease are calculated holding all other variables at their mean values. This allows for the effect of multiple disease occurrence at the population level. In other words the marginal effect of epilepsy in males is to reduce the probability of working by 23 percentage points, assuming that the occurrence of other conditions is at their average levels. An alternative approach is to measure the marginal effect of a given disease by assuming no other diseases are present. This method was tested, and the differences were found to be minor, typically of the order of 1 to 2 percentage points.

Finally in this section we report the multiple occurrence of disease for those diseases that had a significant depressing effect on labour force participation. The results are summarised in Table 10-6 and Table 10-5 for females and males respectively. The bold numbers on the diagonal are the percentages of those reporting that they have or have had a given disease but report no other disease. The remaining numbers in the columns are the percentages of those with a given disease who also have another condition. For example, of all the females who report having a liver condition, 10% also have diabetes and 45% also have high blood pressure. The tables do not show those with more than two conditions.

Note:
The bold numbers on the diagonal refer to those who report solely having the given condition.

 

 

Note:
The bold numbers on the diagonal refer to those who report solely having the given condition.

10.3  Assessing the impact on the total labour supply of older workers#

Print this section

While it is important to identify which diseases have significant marginal effects, a more complete picture of the effect of the disease on the overall level of the labour force requires that we allow for the prevalence of the disease. A disease could have a very debilitating effect leading to a massive drop in labour force participation amongst the afflicted, but the number of sufferers could be very small.

For each of the significant diseases identified in Table 10-4, the marginal effects were weighted by the estimated number of sufferers. This latter number was obtained by using the observed prevalence in the HWR survey multiplied by the marginal percentage. The results are summarised in Figure 22; for example, for males, cancer (other than skin cancer) and heart problems reduce the labour force population by 1,600 each. Note that this does not take any possible correlations into account, as some people report multiple conditions.

Figure 22: Impact of chronic illnesses on labour force participation

 

11  Conclusions and discussion#

Print this section

This study has drawn on the first wave of the Health, Work and Retirement survey (HWR). The comprehensive nature of the data collected in this survey has allowed an extensive analysis of the factors that influence labour supply decisions and by implication retirement decisions of older New Zealanders. The survey has national coverage of those aged 55 to 70 and has heavy over-sampling of Māori.

The primary objective of the study is to assess the effect of health and wealth on the retirement decisions of older workers. In addition, the rich data set allowed for a range of other questions to be explored.

Analysis of living standards was undertaken using a range of measures. Higher values of the Economic Living Standards Index (ELSI) were associated with higher income, wealthier, better educated people with better health. The results confirm that those working had a lower living standard than those retired (Ministry of Social Development, 2007). Those who were Māori, those working and those on NZ Superannuation (NZS) had lower scores on this measure. Likewise, Māori, those working, on a benefit or NZS and in poorer health were forced to reduce costs on essential items more frequently. Among both working and retired, better health measures are associated with greater satisfaction with current material living standards. Similarly, better health is associated with the expectation of higher living standards in retirement. Furthermore, relative to working Europeans, working Māori expect to have higher living standards in retirement. This reflects the fact that moving from a low wage to NZS for many in the lower income brackets constitutes a rise in real income.

The study makes extensive use of logit regression models to analyse the factors associated with whether or not a respondent is in the labour force. The objective is to isolate the effects of health and wealth while controlling for a wide range of other influences. The overall pattern of results is broadly similar for males and females. In all the estimated models, health status is significantly associated with the decision to work. This result holds regardless of which measure of health was used. In contrast, wealth was not identified as having a significant effect, although this may reflect the limitations of the data more than the true underlying effect of net wealth.

A 10% decline in health below the mean score is associated with a fall in labour force participation of 3 to 4 percentage points. A decline of this magnitude is clinically significant. At 20% below the mean score, participation falls by 10 points for males and 6 for females. A 40% fall in the health scores would correspond approximately to a self-reported assessment of poor health. At this level, male participation falls 26 percentage points and females by 13 percentage points. The drop in participation is more than proportional for males, but less for females; in other words, while male participation rates are higher, they decline more rapidly as health deteriorates.

In addition to the effect of health, substantial absolute effects on the probability of working stem from a respondent's marital status. Being divorced, separated or widowed, or having a spouse working all increase the probability that a person remains in the workforce.

The probability that a person in the workforce would chose full-time over part-time employment was not significantly related to either the physical or mental health scores. While physical health status has a significant effect on whether to join the workforce, the evidence is that, given a person is employed, their choice about full- or part-time work is not a function of their health status.

Both males and females have a lower probability of working full-time as they age, receive a benefit or have income from superannuation. In contrast they are more likely to be in full-time employment if they are widowed or have dependants.

There is a marked reduction in labour force participation when respondents receive NZS, typically at age 65. The results suggest that there is a significant “deterrence effect” on labour force participation of NZS, once the effect of a wide range of other influences has been controlled for.

A core model on labour force participation was estimated using data from the HWR survey as well as similar data from SoFIE. In broad measure the results are consistent; for males, both surveys confirm that poorer physical and mental health reduces the probability of labour force participation. Notably, mental health conditions do not appear to influence the labour force participation decisions of females.

There is wide debate about the appropriate measure of health status. In large part this study has used the physical and mental scores from the international standard SF-36 survey. In addition, however, self-reported health status was tested as an alternative. It is generally argued that this measure may suffer from a reporting bias as those not in the workforce may justify their decision by reporting a health status worse than their actual condition. While acknowledging this drawback, the results indicated that for both males and females, those reporting lower standards of health were less likely to remain in the workforce. The odds of a person working if they report fair or poor health status are very much lower than those reporting excellent health. This finding was repeated using two different measures of the key economic variable: the respondent's wage rate and the income of other family members.

Another approach to measure health status is to ascertain if the respondent had ever been diagnosed with a particular chronic condition. The HWR survey identified 19 such illnesses. This measure is unlikely to be biased by the so-called “justification” effect. For males the probability of being retired is much greater where they report cancer (other than skin), epilepsy, blood pressure, heart conditions, ulcers or liver conditions. For females the critical chronic conditions influencing labour force participation are diabetes, high blood pressure and liver conditions.

The impact on the probability of an individual working and the impact on the labour supply of older workers are related but distinct questions. The latter requires information about the prevalence of each chronic illness. From the perspective of health policy, it is important to recognise the extent of the illness in the population as well as its effect on labour force participation. A given illness might be highly debilitating but only affect a very small proportion of the population. When the marginal effects of chronic illnesses were weighted by their prevalence in the survey sample, it was found that high blood pressure was the single most significant condition for males and females. For males, cancer and heart problems were important while, for females, diabetes was the next most important condition reducing the overall labour supply of older females. A limitation of the chronic disease measure is that the person responding positively to having once been diagnosed, may or may not be currently inflicted. No information on the date of diagnosis or whether the condition persisted was available.

Underlying the debate about appropriate measures of health status is the fact that what is really required is a measure of an individual's capacity for work. Their physical and mental health status and the presence of chronic illness are all attempts to provide a proxy for the unobservable capacity to work.

A further way to measure the effect of health is to ask whether a respondent expects to be in full-time employment once they reach a certain age. Two ages were specified: 62 and 65. Key factors that were associated with a significantly greater probability of expecting to be in the labour force at these ages were: being male; separated or widowed, Māori and the health of family members.

Few studies of retirement decisions capture the effects of the macroeconomic and policy environment. For example, how does the expected rate of inflation or the parameters of a public pension scheme influence individual decisions? French (2005) finds that the tax structure of the Social Security system in the USA has a greater effect on explaining the age of retirement than do the level of the pension or the health status of the individual. The experience in New Zealand of raising the age of eligibility from 60 to 65 and the resulting increase in the labour force participation rates of older workers is a clear reminder that the retirement decision is strongly influenced by social policies (Hurnard, 2005).

In all studies of the effect of health on retirement there is a question of causality; specifically is it possible that workforce status influences health. If so health status cannot be treated as a truly independent explanatory variable. Undoubtedly there is some reverse influence; the challenge is whether or not it can be corrected for by appropriate statistical methods. As in many other studies, attempts were made to find suitable instrumental variables that might determine health status but not influence the labour supply decision. These attempts proved unsuccessful. However, unless an unequivocally robust instrumental variable can be identified, it is unclear that the statistical properties of the resulting estimates are necessarily superior to those from a single equation.

This study has been based on the first wave of a longitudinal panel study. As a consequence it is a cross-sectional analysis. This has at least two consequences. First, it is possible that health status in earlier periods may be associated either directly or indirectly with current labour force status. In absence of data over time it has not been possible to allow for this effect. A better understanding of the relationship requires the use of longitudinal panel data.

Clearly there is an association between contemporaneous health status and labour force participation. The interaction between health and retirement is potentially a complex and dynamic process. The work of Bound, Schoenbaum, Stinebrickner and Waidman.(1998) shows that it is the decline in health status that has an equally important effect. The response to a decline could reflect the nature and rapidity of the decline, the expected persistence of a lower health status and the individual’s preference for consumption over leisure time together with their family and financial position.

Second, the results reported here for individuals of different ages assume no cohort effects; ie, a 60-year-old today is assumed to behave in 10 years time as a 70-year-old observed today. Soldo, Mitchell, Tfaily and McCabe (2006) find that, using USA data, there are significant differences in the health status prior to retirement of different cohorts. Hyslop and Dixon (2008) use the Linked Employee-Employer Dataset (LEED) to analyse the employment activity of older New Zealanders born in 1937, 1938, 1939 and 1940. Wage employment rates at age 63 rose consistently across the four one-year birth cohorts, suggesting that even within this short span there may well be significant cohort effects. Fortunately, as the present study is focused on a relatively narrow age range, this problem is minimised. Again, future waves of the HWR survey will allow the use of longitudinal panel data which largely overcomes this limitation.

A further strength of longitudinal data is its value in reducing or eliminating the effect of unobserved individual heterogeneity. In any cross-sectional survey, there are inevitably many personal characteristics of individuals which, while important in shaping their decisions, are simply not observed. By using longitudinal panel data one compares the same individuals through time, largely removing the effect of the unobservable characteristics.

A potentially important influence on the health status of an individual as a child is the socio-economic status (education, income and occupation) of their parents. A second related question is the extent to which childhood health status influences the subsequent education and labour market outcomes of adults. Currie (2009) finds strong evidence of both these links, “suggesting that health could play a role in the intergenerational transmission of economic status.” Clearly only with extensive longitudinal data sets is it possible to address these questions.

References#

Print this section

Anastasiadis, Simon (2010) “ Health and Wealth”. Wellington, The Treasury, Working Paper No. 10/05.

Anderson, Kathryn H. and Richard V. Burkhauser (1984) "The importance of the measure of health in empirical estimates of the labour supply of older men." Economic Letters 16: 375-380.

Anderson, Kathryn H. and Richard V. Burkhauser (1985) "The retirement-health nexus: A new measure of an old puzzle." Journal of Human Resources 20: 315-330.

Baker, Michael, Mark Stabile and Catherine Deri (2004) "What do self-reported, objective, measures of health measure?" Journal of Human Resources 34(4): 1067-1093.

Banks, James and Sarah Smith (2006) "Retirement in the UK." Oxford Review of Economic Policy 22(1): 40-56.

Banks, James and Gemma Tetlow (2008) "Extending working lives." in James Banks, Elizabeth Breeze, Carli Lessof and James Nazroo eds Living in the 21st century: Older people in England (London: Institute for Fiscal Studies): 19-56.

Benzeval, Michaela and Ken Judge (2001) "Income and health: The time dimension." Social Science & Medicine 52: 1371-1390.

Black, Carol (2008) "Working for a healthier tomorrow: Review of the health of Britain's working age population." London, Presented to the Secretary of State for Health and the Secretary of State for Work and Pensions. http://www.workingforhealth.gov.uk/Carol-Blacks-Review/Default.aspx>

Bloom, David E., David Canning and Michael Moore (2004) "The effect of improvements in health and longevity on optimal retirement and saving." Cambridge, MA, National Bureau of Economic Research, Working Paper No. 10919. http://www.nber.org/papers/w10919>

Bound, John (1989) "Self-reported vs. Objective measures of health in retirement models." Cambridge, MA, National Bureau of Economic Research, Working Paper No. 2997.

Bound, John, Michael Schoenbaum, Todd R. Stinebrickner and Timothy Waidman (1998) "The dynamic effects of health on the labour force transitions of older workers." Cambridge, MA, National Bureau of Economic Research, Working Paper No. 6777.

Browning, Martin, Anne Moller Dano and Eskil Heinesen (2006) "Job displacement and stress-related health outcomes." Health Economics 15: 1061-1075.

Cai, Lixin (2007) "The relationship between health and labour force participation: Evidence from a panel data simultaneous equation model." Melbourne, Melbourne Institute Working Paper No. 01/2007.

Cai, Lixin and Guyonne Kalb (2004) "Health status and labour force participation: Evidence from the HILDA data." Melbourne, Melbourne Institute Working Paper No. 04/2004.

Cai, Lixin and Guyonne Kalb (2006) "Health status and labour force participation: Evidence from Australia." Health Economics 15(3): 241-261.

Carter, Kristie N, Tony Blakely, Sunny Collings, Fiona Imlach Gunasekara and Ken Richardson (2008) "What is the association between wealth and mental health?" Wellington, Otago University.

Cobb-Clark, Deborah A. and Steven Stillman (2006) "The retirement expectations of middle-aged individuals." Bonn, Institute for the Study of Labour, IZA Discussion Paper No. 2449.

Coile, Courtney and Kevin Milligan (2006) "How household portfolios evolve after retirement: The effect of aging and health shocks." Cambridge, MA, National Bureau of Economic Research.

Currie, Janet (2009) "Healthy, wealthy, and wise: Socioeconomic status, poor health in childhood, and human capital development. Journal of Economic Literature 47(1): 87-122.

Currie, Janet and Brigette C. Madrian (1999) "Health, health insurance and the labor market." in O. Ashenfelter and D. Card eds Handbook of labor economics (New York and Oxford: Elsevier Science): 3309-3416.

Dave, Dhaval, Inas Rashad and Jasmina Spasojevic (2006) "The effects of retirement on physical and mental health outcomes." Cambridge, MA, National Bureau of Economic Research, Working Paper No. 12123.

Davey, Judith (2008) "What influences retirement decisions?" Social Policy Journal of New Zealand 33: 110-125.

Disney, Richard, Carl Emmerson and Matthew Wakefield (2003) "Ill health and retirement in Britain: A panel data analysis." London, Institute for Fiscal Studies, Working Paper WP03/02.

Dwyer, Debra Sabatini and Olivia S. Mitchell (1998) "Health problems as determinants of retirement: Are self-rated measures endogenous?" Cambridge, MA, National Bureau of Economic Research, Working Paper No. 6503.

Equal Employment Opportunities Trust (2006) "Work and age report: 2006." Wellington, Equal Employment Opportunities Trust.

Ettner, Susan L. (1996) "New evidence on the relationship between income and health." Journal of Health EconomicsI5: 67-85.

Fergusson, David, Bev Hong, John Horwood, John Jensen and Peter Travers (2001) "Living standards of older New Zealanders." Wellington, Ministry of Social Policy.

French, Eric (2005) "The effects of health, wealth, and wages on labour supply and retirement behaviour." Review of Economic Studies 72(2): 395-427.

Haveman, R. (1994) "Market work, wages and men's health." Journal of Health Economics 13(2): 163-182.

Holt, Heather (2010) "Health on labour force participation.".Wellington, The Treasury, Working Paper No. 10/03.

Hurnard, Roger (2005) "The effect of New Zealand superannuation eligibility age on the labour force participation of older people." Wellington, The Treasury, Working Paper No.05/09. http://www.treasury.govt.nz/publications/research-policy/wp/2005/05-09>

Hyslop, Dean and Sylvia Dixon (2008) "The employment and retirement transitions of New Zealanders aged in their 60s: Evidence from Leed." Wellington, Statistics New Zealand. http://www.stats.govt.nz/leed/reports/the-employment-and-retirement-transitions-of-new-zealanders-aged-in-their-60s.htm>

Idler, E. L. and Y. Benyamini (1997) "Self related health and mortality: A review of twenty-seven community studies." Journal of Health and Social Behaviour 38: 21-37.

Jensen, J., Spittal, M., & Krishnan, V. (2005). ELSI Short Form: User Manual for a Direct Measure of Living Standards (Wellington: Ministry of Social Development).

Jensen, John, Vasantha Krishnan, Rob Hodgson, Sathi G Sathiyandra and Robert Templeton (2006) "Living standards of older New Zealanders 2004: An overview." Wellington, Centre for Social Research and Evaluation, Ministry of Social Development. http://www.dsw.govt.nz/publications/docs/livingstandardssummary.pdf>

Laplagne, P., M. Glover and A. Shomos (2007) "Effects of health and education on labour force participation." Melbourne, Productivity Commission, Staff Working Paper. http://www.pc.gov.au/research/staffworkingpaper/healthandeducation>

Long, J. Scott and Jeremy Freese (2006) Regression models for categorical dependent variables using Stata. (College Station, TX: StataCorp LP).

Love, David A. and Paul A. Smith (2007) "Does health affect portfolio choice?" Washington DC, Federal Reserve Board, Finance and Economics Discussion Series: 2007-45 http://www.federalreserve.gov/pubs/feds/2007/200745/>

McGarry, Kathleen (2002) "Health and retirement: Do changes in health affect retirement expectations?" Cambridge, MA, National Bureau of Economic Research, Working Paper No. 9317. http://www.nber.org/papers/w9317>

Ministry of Social Development (2007) "Positive ageing indicators." Wellington, Ministry of Social Development.

Nardi, Mariacristina De, Eric French and John Bailey Jones (2006) "Differential mortality, uncertain medical expenses, and the saving of elderly singles." Chicago, Federal Reserve Bank of Chicago, WP 2005-13.

Olsho, Lauren (2005) "Spend it while you can still enjoy it: Health, longevity, aging and consumption." Madison WI, University of Wisconsin- Madison, Unpublished draft.

Scobie, Grant M., John K. Gibson and Trinh Le (2005) Household wealth in New Zealand. (Wellington: Institute of Policy Studies, Victoria University of Wellington).

Scott, Kate M., Diana Sarfati, Martin I. Tobias and Stephen J. Haslett (2000) "A challenge to the cross-cultural validity of the SF-36 health survey: Factor structure in Māori, Pacific and New Zealand European ethnic groups." Social Science and Medicine 51: 1655-1664.

Scott, Kate M., Martin I. Tobias, Diana Sarfati and Stephen J. Haslett (1999) "SF-36 health survey reliability, validity and norms for New Zealand." Australian and New Zealand Journal of Public Health 23(4): 401-406.

Smeaton, Deborah and Stephen McKay (2003) "Working after state pension age: Quantitative analysis." London, Department of Work and Pensions, Research Report No. 182. http://www.dwp.gov.uk/asd/asd5/report_abstracts/rr_abstracts/rra_182.asp>

Soldo, Beth J., Olivia S. Mitchell, Rania Tfaily and John F. McCabe (2006) "Cross-cohort differences in health on the verge of retirement." Cambridge, MA, National Bureau of Economic Research, Working Paper No. 12762.

Stephens, Christine, Fiona Alpass, Andy Towers, Jack Noone, Jamas Enright, Brendan Stevenson and Eljon Fitzgerald (2008) "Socioeconomic inequalities and health in early old age in New Zealand." Palmerston North, Massey University.

Stephens, Christine and Jack Noone (2007) "Health, work and retirement survey: Summary report for the 2006 data wave: Health." Palmerston North, Massey University. http://hwr.massey.ac.nz/surveys.htm>

Stern, Steven (1989) "Measuring the effect of disability on labor force participation." Journal of Human Resources 24(3): 361-395.

Towers, Andy (2007) "Health, work and retirement survey: Summary report for the 2006 data wave: Methodology." Palmerston North, Massey University. http://hwr.massey.ac.nz/surveys.htm>

Ware, John E. (2000) "SF-36 health survey update." Spine 25(24): 3130-3139.

Wooldridge, Jeffrey M. (2006) Introductory econometrics: A modern approach. (Mason, OH: Thomson South-Western).

Appendix A: Work and retirement status: sample numbers#

Print this section

Appendix B: Variable dictionary#

Print this section

Appendix C: Regression results#

Print this section

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

 

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

 

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

 

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

 

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

 

 

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

 

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

 

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

 

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: This is the same model as used for Appendix Table C.14, with these variables added.

Note: This is the same model as used for Appendix Table C.14, with these variables added.

 

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.

Note: *** = significant at the 1% level; ** = significant at the 5% level; * = significant at the 10% level; ns= not significant.