Analytical Studies Branch Research Paper Series
Health and Work in the Family: Evidence from Spouses’ Cancer Diagnoses
by Sung-Hee Jeon, Social Analysis and Modelling Division, Statistics Canada
and R. Vincent Pohl, University of Georgia
Skip to text
The authors thank Sara Allin, Sandra Milicic, and conference participants at Canadian Health Economists’ Study Group and Canadian Economics Association for their helpful comments.
Using Canadian administrative data from multiple sources, this study provides the first nationally representative estimates of the effect of individuals’ cancer diagnoses on the employment and earnings of their spouses and on total family income. This effect is theoretically ambiguous, but clear evidence for a reduction in employment and earnings among men and women is found. These results are interpreted as individuals taking time off work in order to care for their sick spouse (caregiver effect). In order to give a causal interpretation of the results, a combination of matching and individual fixed effects are employed.
Key words: spousal cancer diagnosis, spousal employment, spousal earnings
Changes in health status may affect not just the individuals who experience such changes, but also their family members. For example, if the main earner in a family loses his or her ability to generate income due to a health shock, it invariably affects the financial situation of the spouse and other dependents. In addition, spouses and working-age children may themselves increase or reduce their labour supply to make up for the lost income (“added worker effect”) or care for a sick family member (“caregiver effect”). Since consumption smoothing and self-insurance occur at the household level, the financial effects of health for other family members have important policy implications. To shed light on such effects, this study analyzes how one spouse’s cancer diagnosis affects the employment and earnings of the other spouse and (before-tax) total family income using administrative data from Canada.
The data used in this paper come from five administrative sources and are combined in the 1991 Census–Longitudinal Worker File (LWF) linkage data. The sources include the Canadian 1991 Census of Population, the Canadian Cancer Database, the Canadian Mortality Database, the LWF, and the T1 Family File. This study combines matching methods with a generalized difference-in-differences (DID) strategy to control for observed and unobserved heterogeneity. Coarsened Exact Matching is first applied to the data to make the treatment group—individuals whose spouses have been diagnosed with cancer—and the control group—individuals whose spouses have never been diagnosed with cancer—observably similar. To make the DID approach more robust, individual fixed effects are included. The study data and methods allow for interpretation of the estimation results in a causal manner, which is essential for an informed policy discussion.
The study finds that both men and women reduce their employment rates by about 2.4 percentage points in the years following their spouse’s cancer diagnosis. Since women have lower average employment rates, this decrease represents a larger relative decline for them. Furthermore, annual earnings decrease by about $2,000 for men and $1,500 for women. In relative terms, this decline corresponds to 3.4% for men and 5.9% for women. Finally, the results show substantial decreases in family income caused by the decline in the earnings of individuals diagnosed with cancer and a parallel decline in the earnings of their spouses. For men whose wives have been diagnosed with cancer, family income decreases by up to 4.8% while a corresponding reduction for women amounts to 8.5%.
While the labour market effects of spousal health shocks are theoretically ambiguous, the empirical results of this research clearly reject the added-worker hypothesis in favour of the caregiver hypothesis. Individuals whose spouses are diagnosed with cancer (i.e., a sudden and severe negative change in health) experience a decrease in employment and earnings. A cancer diagnosis also often implies that the affected spouse’s life expectancy is suddenly reduced, in which case the other spouse may want to work less temporarily in order to spend time with the sick husband or wife. Therefore, the larger negative employment and earnings effects that were found for both men and women in response to spousal cancer may be due to both caregiving needs and the desire to enjoy more leisure time together after the cancer diagnosis.
Overall, this study’s results provide novel and important evidence on the intra-family labour market effects of one family member’s severe health shock. The magnitude of these effects is substantial, suggesting that a cancer diagnosis has the potential to change labour supply from a level that is optimal when both spouses are healthy to a scenario that can strongly affect the family’s financial well-being—in addition to the ensuing psychological costs of dealing with a health shock.
Changes in health status may affect not just the individuals who experience such changes but also their family members. For example, if the main earner in a family loses the ability to generate income due to a health shock, the financial situation of the spouse and other dependents is invariably affected. In addition, spouses and working-age children may themselves increase their labour supply to make up for the lost income or reduce it to care for a sick family member. Since consumption smoothing and self-insurance occur at the household level, the financial effects of health shocks for other family members have important policy implications. To shed light on such effects, this study analyzes how one spouse’s cancer diagnosis affects the employment and earnings of the other spouse and total family income using administrative data from Canada’s national statistical agency, Statistics Canada.
As in other developed countries, cancer is one of the leading causes of mortality and morbidity in Canada. Almost 200,000 individuals were diagnosed with cancer in 2014. According to 2011 data, it is the leading cause of death, accounting for 30% of all deaths.Note 1 With recent medical advances, however, survival chances following a cancer diagnosis have improved. For instance, the average five-year survival rate for all cancers in Canada increased from 56% in 1993 to 63% in 2007. This shift to longer survival emphasizes the importance of considering the medium- and long-term effects of cancer on survivors’ labour market outcomes and financial well-being. For example, Bradley, Bednarek and Neumark (2002a, 2002b), Bradley et al. (2005, 2007b) and Bradley, Oberts and Schenk (2006) use data from the Health and Retirement Study (HRS) and small samples of breast and prostate cancer survivors to estimate the effects of cancer diagnoses on patients’ labour supply and earnings. For Canada, Jeon (2014) finds a substantial decrease in earnings among cancer survivors.
When an individual experiences a negative health shock, the labour supply of his or her spouse is subjected to two opposing forces. First, the spouse’s labour supply may increase to make up for the sick individual’s lost income. This “added worker effect” has been documented after non-health related job losses (e.g., Stephens 2002). In the United States, access to employer-sponsored health insurance provides an additional incentive for the healthy spouse to continue working or increase labour supply (Bradley et al. 2007a). By using data from Canada, which has a universal health care system, this reason for changing spousal labour supply after a health shock can be disregarded and therefore lower added-worker effects can be expected. The Canadian environment also limits the role of selection into health insurance plans and hence makes it more straightforward to give causal interpretation to this study’s results. Moreover, these findings may shed light on the effect of cancer diagnoses on spousal labour market outcomes in most European countries with universal health care systems.
Second, family members, and spouses in particular, may reduce their labour supply in order to care for their sick spouses diagnosed with cancer. The “caregiver effect” has been generally documented in situations where one family member requires long-term care and another family member acts as an informal caregiver (see Van Houtven, Coe and Skira 2013; Heger 2014; and Skira 2015). Healthy spouses may also reduce their labour supply if both spouses wish to spend more leisure time together after a cancer diagnosis (complementarity of leisure). Which of the two potential effects dominates is theoretically ambiguous. In this paper, this question is answered empirically in the context of spousal cancer diagnoses among Canadians.
There are two main challenges to analyzing the effects of changes in one spouse’s health status on the other spouse’s labour supply decisions. The first one is data availability. Analyzing the effects of individuals’ health shocks on their family members requires that families in the data be identified. This is relatively easy in household-level survey data. However, the number of individuals suffering from severe health problems that can change families’ economic well-being is usually small in such surveys. A considerably larger number of individuals with severe health conditions can be observed in administrative data (such as hospital records), but information about family members in such data is usually unavailable. To deal with these problems, this study uses a unique data set linking data from several Canadian administrative sources. In addition to containing a large number of individuals suffering from a health problem (specifically, who were diagnosed with cancer), this study identifies married couples and examines their labour market outcomes.
The second challenge is methodological. In particular, it is related to the interpretation of the estimates. While the effects of one spouse’s ill health on the other spouse’s earnings (or income) can be easily estimated, finding causal links between the former and the latter is usually a much harder task. Family formation is not random—couples are matched based on observable and unobservable characteristics that also affect the health and earnings of the spouses later in life. An obvious example is the correlation in couples’ attitude towards a healthy lifestyle (e.g., spouses are likely to have a similar attitude towards smoking) and their health outcomes. This study controls for time-invariant unobserved heterogeneity by including individual or couple fixed effects in the regression models. In addition, changes in individuals’ health status may not occur independently of their own or their spouses’ labour supply and income even conditional on individual fixed effects (i.e., they may not be strictly exogenous). For example, it is possible that one spouse’s job loss leads to the other spouse’s psychological stress, and subsequently to mental or physical health problems. This study uses cancer diagnoses as an example of a substantial and unanticipated, hence strictly exogenous, health shock. It is unlikely that one spouse’s labour supply or work preferences directly or indirectly affect the other spouse’s likelihood of a cancer diagnosis. Therefore, these findings can provide a causal interpretation.
Given these challenges, only a small number of studies have investigated the effect of changes in one spouse’s health status on the other spouse’s labour supply. Hollenbeak, Short and Moran (2011) use survey data on spouses of cancer survivors in Pennsylvania matched to a control group drawn from the Panel Study of Income Dynamics. They find a decrease in wives’ employment but no effect for husbands. The sample of cancer survivors is not nationally representative and labour supply is self-reported. Nahum (2007) finds evidence for caregiver effects using Swedish administrative data on spouses’ sickness absence. She also finds a more pronounced negative effect among wives. Coile (2004) uses data from the HRS to analyze the effect of health shocks and finds small added-worker effects for husbands but not for wives. Given the decrease in the earnings of the sick spouse, this result implies that the household’s financial situation significantly deteriorates after one spouse’s health shock. In a recent article, García-Gómez et al. (2013) use administrative data from the Netherlands to analyze the effects of individuals’ hospitalizations on their own and their spouses’ labour market outcomes. They find no significant effects for wives and negative effects for husbands of sick individuals.
This study contributes to the literature by providing new evidence on the effect of severe health shocks on spousal employment and earnings and family income. In contrast to most of the existing literature, it combines objective health shock measures from the Canadian Cancer Registry with administrative and nationally representative earnings data from Canadian longitudinal income tax records. In particular, this study’s health shock measure differs from what other studies have used. For example, García-Gómez et al. (2013) and others use acute hospitalizations as a measure for health shocks, but an individual could have experienced declining health prior to being hospitalized. Hence, it is not clear that a hospital admission constitutes an unanticipated shock. In contrast, this study uses an individual’s cancer diagnosis to measure a sudden and unanticipated change in a person’s health status. It is unlikely that an individual and his or her spouse adjust their labour market behaviour because they are aware of the illness before the diagnosis. Besides using this novel source of exogenous variation in individuals’ health, this study combines matching methods with a generalized difference-in-differences (DID) strategy to control for observed and unobserved heterogeneity. Coarsened Exact Matching (CEM) is first applied to the data to make the treatment group (individuals whose spouses have been diagnosed with cancer) and the control group (individuals whose spouses have never been diagnosed with cancer) observably similar. To make the DID approach more robust, individual fixed effects are included. The study data and methods allow the estimation results to be interpreted in a causal manner, which is essential for an informed policy discussion.
To preview this study’s findings, the estimates show that both men and women reduce their employment rates by about 2.4 percentage points in the years following their spouse’s cancer diagnosis. Since women have lower average employment rates, this decrease represents a larger relative decline for them. Furthermore, annual earnings decrease by about $2,000 among men and $1,500 among women. In relative terms, this decline corresponds to 3.4% and 5.9% for men and women, respectively. Finally, substantial decreases in family income—which are due to lower earnings among individuals diagnosed with cancer and an additional decline in earnings among their spouses—are estimated. For men whose wife was diagnosed with cancer, family income decreases by up to 4.8% while the reduction for women amounts to 8.5%.
The rest of this paper is organized as follows: Section 2 describes the multiple data sources and how they are combined; Section 3 discusses the empirical strategy; Section 4 explores the estimation results; and Section 5 concludes.
The data used in this paper come from five administrative sources and are combined in the 1991 Census–LWF linkage data. The sources include the Canadian 1991 Census of Population, the Canadian Cancer Database (CCDB), the Canadian Mortality Database (CMDB), the LWF, and the T1 Family File (T1FF). The LWF and the T1FF are derived from individual tax returns. A brief description of each data source is provided in Appendix 1.
Statistics Canada linked these data sources in multiple steps. Initially, selected personal information from the CMDB and CCDB was linked to individual records of those 25 and over in the 1991 Census file.Note 2 This initial data linkage is called the ‘1991 Canadian Census Cohort: Mortality and Cancer Follow-Up.’ Individuals’ death records up to 2006 and individuals’ cancer records up to 2003 were initially obtained from both the CMDB and CCDB.Note 3 Subsequently, the 1991 Census cohort was linked to the LWF, which is a random 10% sample of Canadian tax return files from 1983 onward, to create a subset of the data containing demographic characteristics, cancer diagnoses and death records, as well as longitudinal individual income profiles. More recently, spousal and total family incomes from the T1FF have also been added to the 1991 Census–LWF data. The current version of the data includes incomes of individuals, incomes of their spouses, and total family income from 1983 to 2010. In addition to death records from the CMDB, income tax files also provide information about individuals’ death years up to 2010.Note 4 The final 1991 Census–LWF linkage data represent approximately 1.4% of the Canadian population aged 25 and over as of 1991.
The 1991 Canadian Census Cohort: Mortality and Cancer Follow-Up can be used to track cancer histories until 2003 of those who were married to individuals enumerated in the 1991 Census–LWF. Their death records are available up to 2006. However, the marital status of individuals in the 1991 Census–LWF data can change over time. To study the impact of their spouse’s cancer diagnosis on these individuals’ labour market outcomes, it first had to be verified that they were still married to the person who was identified as their spouse in the 1991 Census at the time their spouse was diagnosed with cancer. Before the study sample was selected, continuously married couples had to be identified. To that end, the following steps were taken. For 1991, all married individuals aged 59 and under were selected and only individuals never diagnosed with cancer up to the end of 1991 were retained. In this paper, these people will simply be referred to as “individuals” and their spouse will be called “spouse.”Note 5 Individuals were dropped if their spouse was aged 60 or over in 1991 or if their spouse had been previously diagnosed with cancer. Next, individuals’ marriage spells were constructed using family status information from the annual T1FF. In any year, individuals are treated as being continuously married to the same spouse if their marital status has not changed between any two consecutive years from 1991 to that year. If individuals separate, their marriage spell ends. However, in the event the spouse dies, the marriage spell is coded as continued until the individual is remarried, hence widows and widowers are retained in the sample as long as they do not remarry. Once all continuous spells from 1991 onward are identified, changes in individuals’ marital status from 1991 back to 1983 are tracked in order to identify the starting year of the continuous marriage spells. Within the identified continuous marriage spells that span years before and after 1991, individuals are presumed to have been married to the person identified as their spouse in the 1991 Census.Note 6
The marriage spell data contain 107,921 married individuals aged 59 and under in 1991 whose spouse was also aged 59 and under in 1991 and both of whom had no cancer history prior to 1992. The average length of the marriage spells that cover part of or the whole period from 1983 to 2010 is 21.4 years, and 94% of the spells are 10 years or longer. The average age of the individual is 39.7 years in 1991 and the average age of their spouse is 39.8 years. Spouses of 3,665 individuals were diagnosed with cancer for the first time between 1992 and 2003. The age of these individuals at the time of their diagnosis ranges from 28 to 64 years.
In the next step, further restrictions were imposed on the marriage spell data to obtain the treatment and control samples for this study. In each year from 1992 to 2003, individuals who had yet to reach the age of 60 were selected so they would still be of working age. The same age restriction was imposed on the spouse. Individuals who were never diagnosed with cancer up to the end of year and who lived for at least five years following year were kept regardless of the length of their marriage spells.Note 7 The sample is restricted to individuals whose employment status (working or not) can be determined in at least two years prior to year . The individuals are presumed to have worked in each year in which they had non-zero annual earnings.Note 8 The treatment group that satisfies these restrictions consists of 2,636 individuals (1,501 men and 1,135 women) whose spouses were diagnosed with cancer for the first time from 1992 to 2003.Note 9 The most common cancer sites for male spouses are prostate (16.7%) and lung and bronchus (12.7%), and the most common for female spouses are breast (39.2%) and cervix uteri (11.8%). The control sample consists of individuals whose spouse was not diagnosed with cancer at any time from 1992 to 2003. In the control sample, individuals satisfying the above sample restrictions in each year may appear more than once. The total number of observations from 1992 to 2003 in the control sample is 932,970 (450,763 for men and 482,207 for women). This is the pooled number of observations for 100,449 individuals (48,583 men and 51,866 women).
The analysis is conducted separately for men and women because male and female age profiles of labour supply differ. The sample is not restricted to individuals who worked prior to their spouse’s cancer diagnosis to allow for the inclusion of all possible changes in employment due to a spousal health shock.Note 10
Table 1 shows differences in the characteristics of the treatment and control samples for men (Columns 1 and 2) and women (Columns 6 and 7). The same patterns are observed for both men and women. The most notable difference is in the average ages between the treatment and control samples.Note 11 Individuals in the treatment sample are older than those in the control sample. The age differences also seem to be associated with differences in other characteristics. Individuals in the treatment sample are less likely to work but, on average, their annual earnings and total family income are higher than those of their counterparts in the control sample. They have fewer children at home and the youngest child in the treatment sample is generally older than in the control sample. For both men and women, there are also fewer members of visible minorities among the treatment sample than among the control sample. Not surprisingly, individuals’ age is positively correlated with the probability of their spouse’s cancer diagnosis and their own labour supply. However, other differences in the characteristics of the treatment and control samples such as the number of children and family income can be also highly associated with individuals’ labour supply decisions.
In order to balance the covariates shown in Table 1 between treatment and control samples, CEM is first applied to the data before estimating the effect of spousal cancer diagnoses on individuals’ employment and earnings. The next section describes the matching approach.
3. Empirical strategy
To balance treatment and control group covariates, CEM is used—a multidimensional exact matching algorithm applied to cells generated by dividing continuous variables into discrete intervals or by regrouping categorical variables into fewer coarsened categories.Note 12 The CEM algorithm creates a set of strata with the same coarsened values of matching variables; it also restricts the matched data to areas of common empirical support by pruning unmatched observations from both the treated and control samples. For each stratum, the CEM algorithm returns weights that can be used to reweight observations in the matched control sample and balance the empirical distributions of the matching variables between the two samples.Note 13 Later, these matching weights are used in the regression analysis of work status, annual earnings and family income.Note 14
Increasing the number of matching dimensions by adding extra matching variables decreases the probability of finding matches between the treatment and control because the CEM requires exact matching in all coarsened categories of the matching variables. Therefore, it is ideal to have a relatively small set of matching variables sufficient to control for observable differences between the treatment and control samples and, at the same time, small enough to reduce the number of unmatched individuals from the treatment sample. Here the set of matching variables includes individuals’ own and family characteristics, but spouses’ characteristics are not included in the matching variables.Note 15 The personal and family characteristics of individuals chosen as matching variables are likely to be direct determinants of individuals’ labour market outcomes before and after their spouses’ cancer diagnoses. Matching on these variables therefore controls for selection on observables in the outcome variables of interest, which are individuals’ employment, earnings and family income.
Individuals in the treatment and control samples are matched using pooled data from 1992 to 2003 with calendar years used as one of the matching variables. The matching variables also include age (coarsened into 5-year intervals), education, visible minority status (coarsened to three categories), and province of residence. Family characteristics included in the matching variables are the number of children in the family (coarsened to four categories), age of the youngest child (coarsened to three categories), and total family income in the previous year (coarsened to quintiles). The share of the individual’s earnings in the total family income in the previous year (coarsened to two categories) is also included as a matching variable to account for the individual’s earnings contribution to the total family income prior to the spouse’s cancer diagnosis. To account for individuals’ attachment to the labour market prior to their spouse’s cancer diagnosis, the first and second lags of their employment status (i.e., working or not working) are also included as matching variables.
Columns 4, 5, 9, and 10 in Table 1 show the characteristics of the matched samples for men and women in the treatment and control samples, respectively. Not all individuals in the treatment sample could be matched to comparable individuals in the control sample. For 306 men (20.4%) and 211 women (18.6%) in the treatment samples, no comparable matches could be found in the control samples. Most characteristics of unsuccessful matches among men and women in the treatment samples are very similar. Those in the treatment sample who did not work in the two years prior to their spouse’s cancer diagnosis have a smaller chance of being matched to someone in the control sample than those who worked in those two years. Consequently, the matched individuals in the treatment sample have higher average individual earnings and total family income than individuals in the pre-matched treatment sample shown in Columns 1 and 6 of Table 1. Individuals not identified as a visible minority in the treatment samples are more likely to be matched with someone from the control samples than those identified as a visible minority. However, other characteristics such as average age, education, and age of the youngest child are similar for the pre- and post-matched treatment samples. Overall types of spousal cancer are also similar in the pre- and post-matched treatment samples. The types of spouses’ cancer in the matched treatment samples are presented in Table 2 for both male and female spouses.
Finally, comparing Columns 4 and 5 for men and 9 and 10 for women, respectively, in Table 1 shows how similar the characteristics of the matched treatment and the matched CEM-weighted control samples are. There are virtually no differences in characteristics between the two matched samples with matching weights.
As the final step, a regression sample for the matched individuals in the treatment and control samples is constructed. The year of the spouse’s first cancer diagnosis is year and t is the number of years elapsed from the year of the diagnosis ( in year ). In the matched control sample, can be equal to 0 in any year from 1992 to 2003 depending on the in the matched treatment sample, so that is the same year in both samples. Individuals’ longitudinal profiles are constructed from to as long as these time periods fall within individuals’ continuous marriage spells.Note 16
3.2 Generalized difference-in-differences regressions with individual fixed effects
To control for time-invariant unobservable individual characteristics potentially correlated with individuals’ labour market outcomes and their spouses’ health behaviour, a DID model with individual fixed effect model is applied.Note 17 The effects of one spouse’s cancer diagnosis on the other spouse’s labour market outcomes are allowed to vary over time (generalized DID). The results from these regressions can be interpreted as causal effects when combined with strictly exogenous health shocks, specifically cancer diagnoses.
Matching and DID are combined by estimating a fixed effect model with interactions between Treatment group ( ) and time ( ) dummies and applying the CEM matching weights in the estimation (Jacobson, LaLonde and Sullivan 1993; Hijzen, Upward and Wright 2010; Boden and Galizzi 2003) as follows:
where is the labour market outcome variable (work status, annual earnings or family income) for individual in time period . The time-invariant individual fixed effect is . Vector consists of individuals’ time-varying characteristics. Each is a dummy variable equal to 1 if and 0 otherwise. is a dummy equal to 1 if the individual’s spouse was diagnosed with cancer and 0 otherwise (treatment dummy). The reference period is . Hence, is an estimate of the difference in between treatment and control groups in different time periods relative to the difference in between the two groups at (which is a year prior to the year of the spouse’s cancer diagnosis). In other words, is the generalized DID effect of spousal cancer on individuals’ labour market outcomes for time period after the cancer diagnosis.
In order for the DID parameters to have a causal interpretation, the pre-trends of the outcome variables have to be similar between treatment and control groups. That is to say the have to be close to 0 and not significant for . Since data on individuals’ labour market outcomes before the (placebo) cancer diagnosis are available, this common trends assumption can be easily tested. See Subsection 4.1 for graphical evidence for this assumption.
Individual panels are unbalanced as the start and end of the continuous marriage spell can differ for different individuals. However, all marriage spells are continuous. The length of individual panels varies from 4 to 11 consecutive time periods. The sample restrictions described in Section 2 imply that the minimum number of time periods in individual panels is 4 since each panel includes at least the time periods . One male individual in the treatment group and five people (one male and four females) in the control groups are present in the sample for only 4 time periods. Overall, 83% of individuals in the treatment groups (82% males, 86% females) and 92% of the control groups (92% males, 93% females) have 11 full time periods of data.
In addition to the generalized DID regressions (1), basic DID regressions that restrict the effect of spousal cancer diagnoses to be constant over time are also run:
where is the coefficient of interest. Regressions (2) are also used to estimate the effects of specific cancer diagnoses (e.g., lung, colon, breast, and prostate cancer) on spousal employment and earnings. Because of smaller sample sizes, there is not enough statistical power to estimate time-specific effects . As in Equation (1), the CEM weights are also applied to the regressions in Equation (2). In combination with individual fixed effects, the coefficient can therefore be interpreted as a causal parameter.
4.1 Graphical evidence
Graphs depicting the outcomes of interest are presented before the discussion on the regression results. Each of the following graphs shows annual averages for the treatment group and the control group. These averages are plotted over time relative to the year when the (placebo) cancer diagnosis occurred ( ) and separately for men and women.Note 18 In this section, the common trends assumption that is necessary for the difference-in-differences (DID) estimation below is assessed.
The following four outcomes are considered: employment (defined by non-zero earnings in a given year), annual earnings, annual earnings conditional on employment, and total family income.Note 19 Chart 1 shows the graphical results for men using the CEM weights. Chart 1-(a) (“Employment”) shows clear evidence for a decrease in men’s employment after their wives are diagnosed with cancer. One year after the diagnosis, average employment is about 2 to 3 percentage points lower among the treatment group than the control group. This difference remains mostly stable during the five-year follow-up period. To assess the common trends assumption, employment rates across treatment and control groups before the cancer diagnosis are inspected. Weighted employment rates in the two years before the diagnosis are exactly equal because these two variables enter the CEM weights. Going back in time up to 10 years prior to the diagnosis also shows closely aligned trends. In particular, there is no evidence for a dip in employment among the treated, which may have indicated potential endogeneity of the wife’s cancer diagnosis with respect to the husband’s employment status in the pre-diagnosis period.
Chart 1-(b) (“Annual earnings”) shows unconditional annual earnings (i.e., they include men who were not employed in a given year). While earnings prior to the cancer diagnosis are higher among men whose wife is diagnosed with cancer, the trends are roughly similar, and slightly increasing, for both groups. After the cancer diagnosis, earnings of the treated group decline relative to control-group earnings. Hence, the associated DID estimate is negative. With average annual earnings of around $59,000 in the matched treatment group before the cancer diagnosis, this decline amounts to about 3% of annual earnings. Comparing this finding to Chart 1-(c) (“Annual earnings conditional on employment”) shows that a large part of the decline in earnings is due to a decrease in labour supply at the extensive margin. When only men who work are included, the DID estimate for annual earnings is about $1,000. Again, men in the treatment group have higher conditional earnings before the diagnosis, but the trend is similar to control-group earnings. Hence, the conclusion is that the caregiver effect dominates the added-worker effect among men whose wife was diagnosed with cancer.
Finally, Chart 1-(d) (“Total family income”) (i.e., the joint income of men and their wives diagnosed with cancer, and possibly other household members) is considered. The parallel trends assumption in this case is also verified and a substantial drop in total income among the treatment group following the wife’s cancer diagnosis is found. Two years after the diagnosis, this decline reverses a bit, but even five years after the cancer diagnosis, family income is about $4,000 lower among the treated group relative to the pre-diagnosis income difference.
Data table for Chart 1
|Years relative to cancer diagnosis (year 0)||Employment||Annual earnings||Annual earnings conditional on employment||Total family income|
|No spousal cancer||Spousal cancer||No spousal cancer||Spousal cancer||No spousal cancer||Spousal cancer||No spousal cancer||Spousal cancer|
Note: Authors' calculations.
Source: Statistics Canada, 1991 Census–Longitudinal Worker File.
Chart 2 contains the same set of results for women. Chart 2-(a) indicates that women reduce their employment by about 2 to 3 percentage points after their husband is diagnosed with cancer. Hence, the raw effects are similar to those for men. Annual earnings among women in the treatment group decrease by about $2,000 compared with the control group. Given the lower earnings levels among women, this drop constitutes a larger relative earnings change for women than men. A smaller decrease is found for earnings conditional on employment than for unconditional earnings, implying that women reduce their labour supply mostly at the extensive margin. Overall, the graphical results show that the caregiver effect dominates among women too.
Finally, family income drops substantially when husbands’ cancer diagnoses are considered (Chart 2-[d]). This effect consists of a drop in earnings among the affected husbands, with an added decrease in earnings by women who reduce their labour supply in order to potentially act as caregivers. The negative effect on family income increases over time and reaches about $15,000 (15%) of average annual family income three years after the husband’s cancer diagnosis. Hence, a husband’s cancer diagnosis has important implications for the financial situation of affected families.
Using the graphs in Chart 2, the common trends assumption for women’s labour market behaviour can also be assessed. While the pre-diagnosis employment trend is noisier for the female treatment group than the control group, the assumption overall holds in this case. The two earnings measures also have parallel trends in the treatment and control groups before the (placebo) cancer diagnosis. For family income, annual averages are similar between treatment and control groups before . This result is remarkable since only family income in the year prior to the cancer diagnosis, along with number of children and the age of the youngest child, enters the CEM weights, but the pre-trends are very close for at least seven years before the diagnosis. For this study sample, the variables that enter the CEM algorithm are sufficient to control for observable differences between the treatment and control groups for the extended pre-diagnosis period.
Data table for Chart 2
|Years relative to cancer diagnosis (year 0)||Employment||Annual earnings||Annual earnings conditional on employment||Total family income|
|No spousal cancer||Spousal cancer||No spousal cancer||Spousal cancer||No spousal cancer||Spousal cancer||No spousal cancer||Spousal cancer|
Note: Authors' calculations.
Source: Statistics Canada, 1991 Census–Longitudinal Worker File.
4.2 Regression results
In this section, the following sets of results are presented for men’s and women’s employment (using a dummy variable that equals 1 if an individual has non-zero annual earnings during a given year), annual earnings, and family income: DID regressions with time-invariant effects (Tables 3 and 4); generalized DID regressions with time-varying effects (Tables 5 to 10 show the main results); the same set of results restricting the sample to individuals whose spouse survived at least five years after their cancer diagnosis (Table 11); and DID regressions for different types of cancer diagnoses (Table 12).
4.2.1 Time-invariant effects
First, regression results from estimating Equation (2) are briefly presented using a standard DID framework with time-invariant effects of spousal cancer diagnoses on labour market outcomes. Tables 3 and 4 contain the results for the following outcomes: employment, annual earnings, and family income for men and women. Three separate regressions—which differ by the definition of the post-diagnosis period—are provided for each outcome. In particular, how time period (year of the cancer diagnosis) is treated varies as follows: included in the pre-period, included in the post-period, and excluded from the estimation sample. The estimates show that men and women reduce their employment by about 2 percentage points after the cancer diagnosis of their spouse. The decrease is slightly larger among women in both absolute and relative terms. Annual earnings decline by $1,600 to $2,100. Women’s earnings decrease less in absolute terms than men’s, but more relative to average pre-diagnosis earnings (about 3.4% for men and 5.2% for women). Finally, family income decreases substantially. This reduction is largest among women, where total income declines by about $7,000 (6.5%) relative to pre-treatment levels. Part of this reduction is due to women’s decline in earnings, but a larger contribution comes from their husband, whose earnings decline after the cancer diagnosis (see Jeon 2014).
4.2.2 Time-variant effects
The main results are presented below (i.e., the results from estimating the generalized DID regressions in Equation (1) separately for men and women). As described in Subsection 3.2, these regressions are estimated using the CEM weights and include individual fixed effects. Each regression is estimated first without additional controls and then again with different sets of controls. These controls include indicators for whether the sick spouse was diagnosed with an additional cancer during the five years after the initial diagnosis, whether the individual became a widow (widower), and whether the individual received Canada Pension Plan disability benefits or disability tax credits in a given year. They also include individuals’ non-labour income, the number of children, and self-employment status.Note 20 Standard errors in all regressions are clustered on the individual level, which is equivalent to clustering on the family level because one observation is included for each couple and year.
Table 5 contains the results for men’s employment.Note 21 Column 1, which shows the results without controls, confirms the graphical results in Chart 1. Men whose wife was diagnosed with cancer reduce their employment in the subsequent years. In particular, they are 2.2 to 2.4 percentage points less likely to work in the first three years after the diagnosis compared with men whose wife has never been diagnosed with cancer. In the fourth year, this negative effect increases to 3 percentage points, but no statistically significant effect is present in the fifth year. Overall, these results suggest that men significantly adjust their labour supply at the extensive margin for about the first four years after their wife’s cancer diagnosis. After four years, cancer patients have likely recovered or may have passed away, so the need for caregiving is reduced and these men return to work.
Columns 2 to 6 in Table 5 contain results for men’s employment with added control variables. Overall, the estimates are stable, but less precise. Column 2 adds an indicator for an additional cancer diagnosis during the five-year follow-up period. While an additional diagnosis has a large negative effect on husbands’ employment, this effect is not statistically significant (not shown). However, the DID estimates for the initial diagnosis are unchanged. Column 3 controls for lagged widowhood (i.e., an indicator that equals 1 if the individual’s spouse passed away one year before or earlier is included).Note 22 Becoming a widower has a large negative but statistically insignificant effect on employment (not shown). The effects of the cancer diagnosis on employment become smaller in absolute value, but the overall pattern is similar to that estimated in Column 1. In Columns 4 and 5 of Table 5, controls are in place for non-labour income, number of children, and individuals receiving disability benefits, but no significantly different estimation results are found. Finally, when all of the above variables in Column 6 are controlled for, smaller decreases in employment after the cancer diagnosis are found, which are also estimated less precisely. The overall pattern is preserved, however.
All regressions in Table 5 contain interactions between the treatment variable (the individual’s wife was diagnosed with cancer in ) and time periods before the diagnosis. The effects of these pre-treatment interactions on the outcome variable allow placebo tests to be conducted (i.e., the common trends assumptions can be formally assessed). None of these interactions had a significant effect, therefore it can be concluded that wives’ cancer diagnoses do not affect husbands’ employment before they occur. This finding confirms the study’s assumption that an initial cancer diagnosis changes a family’s information set and spousal employment does not change in anticipation of such a health shock.
Next, the estimation results for men’s earnings (reported in Table 6) are explored. The six regressions contain the same sets of control variables as the employment regressions described above. Starting in the year following the wife’s cancer diagnosis, husbands earn about $2,000 less per year. With mean pre-diagnosis annual earnings of $59,000 in the matched treatment group, this corresponds to a 3.4% reduction in earnings. This negative effect remains stable for the following three years and disappears in the fifth year after the cancer diagnosis. Hence, the pattern is the same as for the employment effects. Given the graphical results described above, these results are not surprising. Most of the decline in labour supply occurs at the extensive margin, so employment and annual earnings exhibit similar patterns. The regressions results including control variables, which are reported in Columns 2 to 6 in Table 4, confirm the basic result on men’s earnings. While the estimates become less precise as controls are added, the point estimates show a decrease in earnings of about $2,000 per year across specifications. The placebo tests also show that earnings do not change significantly in anticipation of a cancer diagnosis.
The third outcome considered is total family income before taxes. Table 7 contains the results for men. In this case, the changes in the outcome variable due to a spousal cancer diagnosis potentially operate through two channels. First, the sick wife may reduce her employment or hours worked and therefore have lower earnings. Second, the husband may reduce his labour supply to act as a caregiver, which contributes to an overall decrease in family income. The results in Table 7 are consistent with the previous results in that they show the largest effect for the year immediately following the spousal cancer diagnosis. Depending on the specification, family income declined by about $4,000 to $5,000 (or 3.8% to 4.8%) for men whose wife was diagnosed with cancer.Note 23 This decline is highly statistically significant. In subsequent years, the reduction in family income becomes smaller and is not statistically significant. Hence, after a wife’s cancer diagnosis, the family is able to re-adjust total income in a relatively short time.
Next, regression results for women are reported. The treatment group consists of women whose husband was diagnosed with cancer between 1992 and 2003, and the control group contains women whose husband was never diagnosed with cancer. Table 8 displays the estimation results for women’s employment using the same specifications in Columns 1 to 6 as for men. Overall, women reduced their employment by about 2.5 percentage points during the five years after their husband was diagnosed with cancer. In contrast to the results for men, women do not increase their employment rates in the fifth year after the diagnosis. They either care for their husband longer or do not return to the workforce for other reasons.Note 24 However, these effects are estimated less precisely than for men and are only statistically significant at the 5% level in the first year and at the 10% level in the second and third years. The point estimates do not change substantially when the control variables are included. As for men, the cancer diagnosis has no effect on women’s employment in the pre-treatment periods; hence the common trends assumption is satisfied.
Table 9 contains results for women’s annual earnings. The reduction in earnings amounts to $800 to $1,700 in the five years after the husband’s cancer diagnosis in the baseline regression (Column 1). The earnings loss is highest in the first two years and then becomes less significant in both economic and statistical terms. Given average annual earnings of about $29,000, the initial earnings loss is substantial, however, at about 5.9%. As with the previous results, the effects of a cancer diagnosis become smaller in absolute value when control variables are added. The overall pattern persists, however, with negative point estimates in the first three years for all specifications. Lastly, it is possible to verify the common trends assumption for this set of results.
Finally, regression results for family income for women whose spouse was diagnosed with cancer are reported in Table 10. A husband’s cancer diagnosis has a direct and an indirect effect on family income. The estimates in Table 10 show that the initial drop in family income is substantial, with $5,000 to $6,500 depending on the specification. In contrast to the results for men in Table 7, family income further drops in subsequent years. The largest reductions amount to $8,000 to $9,000 three years after the husband’s cancer diagnosis. These large effects are due both to a reduction in employment and earnings of the husband who was diagnosed with cancer and the wife who works less in response. These results show the large negative effects of husbands’ cancer diagnoses on the entire family’s economic situation.
Overall, this study’s regression results strongly suggest that men (women) reduce their employment and experience earnings losses after their wife (husband) is diagnosed with cancer. Hence, the caregiver effect dominates the added-worker effect in the context of cancer diagnoses in Canada. The reduction in employment rates is similar across genders at about 2.4 percentage points, although men have higher average pre-treatment employment rates than women. In addition, men lose about 3.4% of their annual earnings while women lose about 5.9%. These findings suggest that women reduce their employment and earnings more than men in response to their spouse’s cancer diagnosis.
4.2.3 Robustness and heterogeneity
Table 11 contains regression results for the same outcomes as Tables 5 to 10, but restrict the sample to individuals whose spouse survives for at least five years after his or her cancer diagnosis; hence, they exclude all individuals in the treatment and control groups who become widowed during the sample period.Note 25 Excluding window(er)s from the sample allows this study to consider effects of spousal cancer that are not confounded by the spouse’s death (but the surviving spouses also have less severe cancer diagnoses). Comparing these results to the main results discussed above, the decrease in women’s employment and earnings is found to be less pronounced in the no-widow sample. Also, the decline in the women’s family income in the no-widow sample is smaller than in Table 10, and it is not statistically significant. The women’s sample contains proportionally more spouses diagnosed with cancer types in the low-survival category than the men’s sample, therefore excluding widows from the women’s sample results in a smaller negative effect of cancer on all outcomes. In other words, it appears that women whose spouse is diagnosed with cancer reduce work and earnings mostly in cases where the diagnosis is particularly severe and their husband does not survive for at least five years after being diagnosed with cancer. In these cases, women’s family income loss is likely to be persistent and substantial.
The results so far do not distinguish between different cancer sites. It is likely, however, that spouses react differently according to how severe a cancer diagnosis is. Four specific types of cancer that are the most common in each severity level are considered: lung cancer (low survival probability), colon cancer (medium survival probability), and breast and prostate cancer (high survival probability).Note 26
Panel A of Table 12 contains the results for lung cancer. A comparison of the DID parameters with those in Tables 3 and 4 show that the decline in men’s employment in response to their wives’ lung cancer diagnosis is larger than the overall average response to all cancer types (9.6 percentage points instead of 1.6 percentage points). Women’s employment is not significantly affected by their husbands’ lung cancer diagnosis. Qualitatively similar results hold for annual earnings of men and women. For family income, a large and significantly negative effect is found only for women. Families in which the husband is diagnosed with lung cancer lose about $11,000 per year of before-tax family income.
In response to a spouse’s colon cancer diagnosis, there are overall negative effects on employment, earnings and family income, but most of these effects are not statistically significant due to the smaller sample size (Panel B of Table 12). Finally, negative point estimates are also found for the effect of spouses’ breast and prostate cancer diagnoses on individuals’ labour market outcomes in Panels C.1 and C.2 of Table 12. Some of these estimates are not significant or are only marginally significant, but they provide suggestive evidence for the negative impact of these more common but less severe cancer diagnoses on spousal labour market outcomes. Taken together, the point estimates in Table 12, in particular for men, show a clear positive relationship between the severity of the cancer diagnosis or survival probabilities of wives and their husband’s decrease in employment and earnings, but not for family income.
5. Discussion and conclusion
In this paper, unique and nationally representative administrative data are employed to estimate the effect of one spouse’s cancer diagnosis on the other spouse’s subsequent labour market outcomes. The results show that individuals reduce their employment, and their earnings decline, due to their spouse’s health shock. In addition, the overall effect of one spouse’s cancer diagnosis on his or her family’s financial situation can be quantified by looking at the changes in before-tax family income. This study finds negative effects that are both statistically and economically significant. For example, both men and women whose spouses are diagnosed with cancer reduce their employment by about 2.4 percentage points, and a husband’s cancer diagnosis leads to a total family income loss of up to 8.5% annually. The average annual income loss of $2,700 for men and $6,900 for women is substantial when compared to other types of costs incurred after a cancer diagnosis. For example, average out-of-pocket costs associated with cancer treatment amount to about $2,900 per year in Ontario (Longo et al. 2006), and the average cost to the health care system is about $26,000 per cancer diagnosis in the year following such a diagnosis (de Oliveira et al. 2013).Note 27
The labour market effects of spousal health shocks are theoretically ambiguous, but this study’s empirical results clearly reject the added-worker hypothesis in favour of the caregiver hypothesis. Individuals whose spouses are diagnosed with cancer (i.e., a sudden and severe negative change in health) decrease their employment and experience earnings losses. This result is mostly in line with the existing literature. For example, Hollenbeak, Short and Moran (2011) find negative employment effects for wives of cancer survivors, but not for husbands. On the other hand, Coile (2004) finds small positive labour supply effects for men, but not for women. In contrast to this study, the above-mentioned studies use data from the United States, where access to employer-sponsored health insurance may lead to stronger added-worker effects. For a country with a universal health insurance system similar to the Canadian one—the Netherlands—García-Gómez et al. (2013) find negative effects from spousal hospitalizations on employment and earnings for men. Their estimated effects are smaller than those in this study, which may be explained by the fact that they consider all types of hospitalizations whereas this paper considers only more severe health shocks associated with cancer diagnoses.
In contrast to other existing studies, this analysis finds a smaller difference in labour market effects between male and female spouses. Specifically, only slightly larger (in relative terms) negative effects for women’s employment and earnings are found relative to those for men. However, Coile (2004) finds that wives’ employment increases less than husbands’ employment. Hollenbeak, Short and Moran (2011) find that wives’ employment decreases more than husbands’ employment, while García-Gómez et al.’s (2013) results show the opposite. Several reasons can explain these discrepancies. First, men and women in the United States, the Netherlands and Canada may have different labour supply profiles. For example, if men’s employment rates are higher than women’s, men may reduce their labour supply in response to their wives’ health shocks while women are more likely to replace their husbands’ lost income. García-Gómez et al. (2013) observe this pattern in the Netherlands, while this study finds slightly stronger negative effects for women although men’s employment rates are also higher in Canada.
Second, different caregiving options may play a role in employment effects. More easily available or more affordable institutionalized care can replace informal caregiving by spouses, which would lead to smaller or non-existing caregiver effects. Differences between the Canadian and Dutch health care systems may explain why negative employment effects are found for women, in contrast to García-Gómez et al. (2013).
Finally, differences in the types of health shocks considered in these studies can also explain different findings. More particularly, individuals may reduce their labour supply to care for a disabled spouse after his or her hospitalization, but a cancer diagnosis often implies that the affected spouse’s life expectancy is suddenly reduced. In this case, the other spouse may want to work less temporarily in order to spend time with the sick husband or wife.Note 28 Therefore, the larger negative employment and earnings effects that are found for both husbands and wives in response to spousal cancer may be due to both caregiving needs and the desire to enjoy leisure time together after a cancer diagnosis. Moreover, the combination of caregiving needs and leisure complementarities also explains why men and women reduce employment and earnings to a similar extent—the latter are likely independent of the gender of the diagnosed spouse.
Overall, this study’s results provide novel and important evidence on the intra-family labour market effects of one family member’s severe health shock. The magnitudes of these effects are substantial, suggesting that a cancer diagnosis has the potential to change labour supply from a level that is optimal when both spouses are healthy and can strongly affect a family’s financial well-being—in addition to the psychological costs of dealing with such a health shock.
Appendix 1: Data description
The 1991 Census–Longitudinal Worker File (LWF) is a unique dataset that combines data from five sources: Canada’s 1991 Census of Population, the Canadian Mortality Database (CMDB), the Canadian Cancer Database (CCDB), the LWF and the T1 Family File (T1FF).
The CMDB contains individual death records from 1950 onward. Provincial and territorial Vital Statistics offices provide these records annually to Statistics Canada for national-level analysis.
The CCDB is a databank combining two cancer-related data sources: the Canadian Cancer Registry (CCR) and the National Cancer Incidence Reporting System (NCIRS). The former is a person-oriented tumor database that includes clinical and demographic information about Canadian residents with cancer since 1992 (Statistics Canada 2008). The latter is a historical tumor-oriented database containing cancer cases diagnosed as far back as 1969 (Carpenter et al. 2008). Individual cancer records from the CCR are used in the analysis; historical information from the NCIRS is used to verify that individuals in the CCR had no prior cancer history.
The LWF represents 10% of the random sample of Canadians who either filed a personal income tax form (Form T1 General, Income Tax and Benefits Return) or received a statement of remuneration (Form T4, Statement of Remuneration Paid (Slip)) from their employer in each year from 1983 onward. Once individuals are selected into the LWF, they are followed regardless of their employment status for as long as they file a tax return (Form T1 General) or their income is reported to the Canada Revenue Agency by their employers. The current version of the LWF contains information on wages, salaries and net self-employment income as well as firm-level information.Note 29 Wages and salaries are obtained from T4s issued by employers. Net self-employment income and basic personal information (marital status, province of residence, etc.) are obtained from the personal income tax files (T1).Note 30
The T1FF is a family tax file that is built annually based on the information included in the personal income tax files (T1) and supplementary files such as the Canada Child Tax Benefit. A tax unit in Canada is an individual. Using a combination of information available in the T1 along with family benefit information, Statistics Canada constructs the T1FF on an annual basis. The taxfiler’s spouse is primarily identified based on the spouse’s social insurance number (SIN) in the T1, while children are identified based on their parents’ tax return and child benefits program files. Individuals can be followed over time using their SIN, and their family income can be constructed in each year using their family identifier from the T1FF.
Statistics Canada’s Health Analysis Division initially linked selected personal information from CMDB and CCDB to the individual records of individuals 25 and over in the 1991 Census file. This initial data linkage is called ‘1991 Canadian Census Cohort: Mortality and Cancer Follow-Up.’ Individuals’ death records up to 2006 and individuals’ cancer records up to 2003 were obtained from both the CMDB and the CCDB. Subsequently, the LWF records were linked to the 1991 Canadian Census Cohort to provide the crucial income component. The T1FF was added later to provide the spousal and total family income components.
The 1991 Census–LWF data sample contains 263,674 individual records corresponding to about 1.4% of the Canadian population aged 25 and over in 1991. Approximately 58.8% of the 1991 Census–LWF cohort was observed in all 28 years of the LWF (from 1983 to 2010). Individuals were present in the sample for an average of 24.8 years. Tax filing rates were slightly lower in the 1980s compared with more recent decades (from 1990 to 2010), and 66.9% of the 263,638 individuals were observed in all 21 years, for an average of 18.5 years.
Appendix 2: Inverse propensity score weighting
As an alternative to Coarsened Exact Matching weighting, regression results are also provided where the data are weighted using estimated inverse propensity score weights before the effect of spousal cancer on individuals’ labour market outcomes is estimated. Propensity scores are obtained by estimating a probit regression of treatment status (the individual’s spouse was diagnosed with cancer) using the following independent variables: individual’s and spouse’s age (both in five-year bins), individual’s and spouse’s education categories, a visible minority indicator, number of children, age of the youngest child, employment indicators for the individual and the spouse for five years prior to the diagnosis, individual’s and spouse’s earning quintiles for five years prior to the diagnosis, individual’s non-earned income quintiles and family income quintiles for five years prior to the diagnosis, and year and province dummies. To assign a placebo-diagnosis year to individuals in the control group, a year between 1992 and 2003 is randomly drawn for each control observation. After estimating probits of treatment status separately for men and women, inverse propensity score weights are obtained as , where is an indicator for treatment status (spousal cancer diagnosis) and is the predicted treatment probability for individual based on the probit regression described above.
Boden, L.I., and M. Galizzi. 2003. “Income losses of women and men injured at work.” Journal of Human Resources 38 (3): 722–757.
Bradley, C.J., D. Neumark, H.L. Bednarek, and M. Schenk. 2005. “Short-term effects of breast cancer on labor market attachment: Results from a longitudinal study.” Journal of Health Economics 24 (1): 137–160.
Bradley, C.J., D. Neumark, Z. Luo, and H.L. Bednarek. H.L. 2007a. “Employment-contingent health insurance, illness, and labor supply of women: Evidence from married women with breast cancer.” Health Economics 16 (7): 719–737.
Bradley, C.J., D. Neumark, Z. Luo, and M. Schenk. 2007b. “Employment and cancer: Findings from a longitudinal study of breast and prostate cancer survivors.” Cancer Investigation 25 (1): 47–54.
Bradley, C.J., H.L. Bednarek, and D. Neumark. 2002a. “Breast cancer survival, work, and earnings.” Journal of Health Economics 21: 757–779.
Bradley, C.J., H.L. Bednarek, and D. Neumark. 2002b. “Breast cancer and women’s labor supply.” Health Services Research 37 (5): 1309–1328.
Bradley, C.J., K. Oberst, and M. Schenk. 2006. “Absenteeism from work: The experience of employed breast and prostate cancer patients in the months following diagnosis.” Psycho-Oncology 15 (8): 739–747.
Carpenter, M., M.E. Fair, C. Poliquin, and P. Lalonde. 2008. Canadian Cancer Data Base, 1969 to 1991 History and Development. Occupational and Environmental Health Research Section, Report no. 16. Health Statistics Division. Ottawa: Statistics Canada.
Coile, C.C. 2004. Health Shocks and Couples’ Labor Supply Decisions. NBER Working Paper 10810.
de Oliveira, C., K.E. Bremner, R. Pataky, N. Gunraj, K. Chan, S. Peacock, and M.D. Krahn. 2013. “Understanding the costs of cancer care before and after diagnosis for the 21 most common cancers in Ontario: A population-based descriptive study.” CMAJ Open 1 (1): E1–E8.
García-Gómez, P., H. van Kippersluis, O. O’Donnell, and E. van Doorslaer. 2013. “Long term and spillover effects of health shocks on employment and income.” Journal of Human Resources 48 (4): 873–909.
Heger, D. 2014. Work and Well-being of Informal Caregivers in Europe. Ruhr Economic Papers 512.
Hijzen, A., R. Upward, and P.W. Wright. 2010. “The income losses of displaced workers.” Journal of Human Resources 45 (1): 243–269.
Ho, D., K. Imai, G. King, and E. Stuart. 2007. “Matching as nonparametric preprocessing for reducing model dependence in parametric causal inference.” Political Analysis 15 (3): 199–236.
Hollenbeak, C.S., P.F. Short, and J. Moran. 2011. “The implications of cancer survivorship for spousal employment.” Journal of Cancer Survivorship 5 (3): 226–234.
Iacus, S.M., G. King, and G. Porro. 2011. “Multivariate matching methods that are monotonic imbalance bounding.” Journal of the American Statistical Association 106 (493): 345–361.
Iacus, S.M., G. King, and G. Porro. 2012. “Causal inference without balance checking: Coarsened Exact Matching.” Political Analysis 20 (1): 1–24.
Imbens, G.W., and J.M. Wooldridge. 2009. “Recent developments in the Econometrics of Program Evaluation.” Journal of Economic Literature 47 (1): 5–86.
Jacobson, L.S., R.J. LaLonde, and D.G. Sullivan. 1993. “Earnings losses of displaced workers.” American Economic Review 83 (4): 685–709.
Jeon, S.-H. 2014. The Effects of Cancer on Employment and Earnings of Cancer Survivors. Analytical Studies Branch Research Paper Series, no. 362. Statistics Canada Catalogue no. 11F0019M. Ottawa: Statistics Canada.
Longo, C.J., M. Fitch, R.B. Deber, and A.P. Williams. 2006. “Financial and family burden associated with cancer treatment in Ontario, Canada.” Supportive Care in Cancer 14 (11): 1077–1085.
Michaud, P.-C., and F. Vermeulen. 2011. “A collective labor supply model with complementarities in leisure: Identification and estimation by means of panel data.” Labour Economics 18 (2): 159–167.
Nahum, R.-A., 2007. Labour Supply Response to Spousal Sickness Absence. Institute for Futures Studies. Working Paper 2007:2.
Skira, M.M. 2015. “Dynamic wage and employment effects of elder parent care.” International Economic Review 56 (1): 63–93.
Statistics Canada. 2008. Canadian Cancer Registry System Guide, 2007 Edition. Canadian Cancer Registry Redesign Team. Statistics Canada Catalogue no. 82-225-X. Ottawa: Statistics Canada.
Stephens, M. 2002. “Worker displacement and the added worker effect.” Journal of Labor Economics 20 (3): 504–537.
Van Houtven, C.H., N.B. Coe, and M.M. Skira. 2013. “The effect of informal care on work and wages.” Journal of Health Economics 32 (1): 240–252.
Wilkins, R., M. Tjepkema, C. Mustard, R. Choinière. 2008. “The Canadian census mortality follow-up study, 1991 through 2001.” Health Reports 19 (3): 25–43.
- Date modified: