Health and the Future Course of Labor Force Participation at Older Ages Michael D. Hurd Susann Rohwedder Introduction For most of the past quarter century, the labor force participation rates of the older population have been increasing. As Figure 1 shows, the proportion of men 65 to 69 years of age has increased from onefourth to one third since 199, while the proportion of women the same age doing so has increased from about one sixth to one fourth. Furthermore, a majority of men near retirement age, i.e., 62 to 64, are now in the labor force, as are nearly half the women of such ages. Such trends carry several important public policy implications. First, as the baby boom generation continues to enter retirement age, increasing labor force participation of persons who otherwise might enter retirement will result in a larger workforce than might otherwise be expected. Second, should workers continue to work past retirement age, they will continue to contribute to Social Security programs and thereby help maintain solvency for such programs. Third, should older workers continue in employment that also provides health insurance, these trends will help reduce pressure on Medicare, which acts as a second payer of costs when an individual has other health insurance. Forecasting labor force participation of older persons therefore can help in shaping public policies in future years. Continued increases in labor force participation of older workers will boost the size of the labor market and the solvency of Social Security and Medicare programs. Understanding the size of these changes is an important concern for policymakers. Current Social Security Administration forecasts envision continued increases in labor force participation at older ages. An important component of these forecasts is past trends. That is, as long as labor force participation continues to increase, Social Security Administration forecasts will generally predict future increases as well. An alternative to these forecasts and one that might better anticipate future changes is available in trends of the subjective probability of individuals for working past age 62 or 65. If there are changes in this subjective probability, i.e., if younger workers or workers approaching retirement age become more or less likely to say they will work past age 62 or age 65, then we may be able to better anticipate changes in labor force population of older workers, and the consequences for the overall labor force and social welfare programs for older individuals. Subjective probability of working past age 62 or 65 is determined by asking individuals at younger ages their chances for working past a target age, e.g., asking workers, and non workers, in their 5s about their chances of working past age 62 or 65. Asking these questions of respondents in a panel survey and later comparing these responses with actual outcomes for these same individuals yields insights into how subjective probabilities can predict future workforce participation. Past research has shown that subjective probability of work at older ages is a good predictor of both individual work past a certain age 1
and overall labor force participation at older ages. Indeed, since the early 199s, subjective probability of working to age 65 generally increased presaging the increase in labor force participation at older ages. In recent years, however, there have been some possible exceptions to these increases in subjective probability of working to age 65. Changing health of older workers may account for some recent changes in these trends. Health is an important correlate of subjective probability of working. The health of the population in their 5s does not appear to have been increasing raising questions about whether the subjective probability of working past future target ages will continue to increase, as well as whether actual labor force participation will increase at older ages. If the health of workers in their 5s is worsening, then we may not expect labor force participation at older ages to continue to increase regardless of what trends in actual labor force participation have indicated to date. This paper seeks to quantify the relationship between health and the subjective probability of working past age 65, and particularly how that relationship has changed over time. In some aspects, this approach can be superior to a study of actual retirement. Changes to health or other determinants of retirement take many years to affect actual trends in retirement, but can be seen more immediately in subjective probability. Put another way, deteriorating health of workers in their 5s today won t have an effect on labor force participation rates of workers in their 6s for at least another ten years. But deteriorating health can have an immediate effect on subjective probability of working past a given age and provide a predictor today of labor force participation rates of older workers in a decade. To assess the relationship between health and the subjective probability of working past age 65, we use panel data from the Health and Retirement Study (HRS). Among questions of the HRS asked of individual respondents over time are several on health in addition to those on subjective probability of retirement at a future age. These data allow us to examine how changes in health conditions over time may affect probability of working at future ages. Data Description of HRS to be added Results The HRS asks respondents, Thinking about work in general and not just your present job, what do you think the chances are that you will be working full time after you reach age 62? It also asks this same question with regard to a target age of age 65. Responses to these questions yield the subjective probability of working past a target age. We call the subjective probability of working past age 65 P65, and the subjective probability of working past age 62 P62. Figure 2 shows P65 increased for both male and female workers 55 to 59 years of age from 1992 through 28. And these trends do appear to presage the increases in actual work after age 65 evident in panel data and as shown in Figure 1. For example, among women 55 to 59 years of age, P65 increased from 21 percent in 1994 to 39 percent in 28. During that same time, the actual labor force 2
participation of women age 65 to 69 years of age increased from 16 percent to 26 percent. These numbers are not directly comparable because of the differences in age bands, but it is still striking the subjective probability of working past age 65 and the actual probability of doing so both increased at least 1 percentage points. The increase in actual labor force participation, however, is slightly puzzling in light of apparently declining trends in self rated health. The HRS asked respondents to rate their health on a five point scale: excellent, very good, good, fair, poor. Figure 3 shows that self rated health has declined over time for workers 55 to 59 years of age. In particular, the overall proportion of individuals in this age group rating their health as fair or poor increased from 23 percent in waves 2 to 4 to 28 percent in waves 1 to 12. There is a substantial difference between respondents 55 to 59 years of age who are working and those who are not working. In particular, in waves 2 to 4, only 13 percent of those 55 to 59 years of age who were working were in fair or poor health, although this proportion, too, increased in later waves. Of course, there are many other determinants of labor force participation, such as wage rate, wealth, life expectancy, and pension, that are correlated with health. Still, in a fixed effects estimation holding these other determinants constant (not shown), we found P65 reacts to changes in self rated health: a decline in P65 accompanies a decline in self rated health in panel data. For additional insight into the slowdown in labor force participation growth among older workers we also looked at annual changes in labor force participation among workers 65 to 69 years of age. Figure 4 shows first differences: beginning in the early 199s labor force participation for both men and women increased.6 percent to.8 percent each year until 28. These increases largely stopped in 28, at least initially likely triggered by the Great Recession. That same year, P65 also peaked. Yet Figure 4 indicates the minimal growth in labor force participation among older workers after 28, which might have been considered statistical noise, now seems more like a break in the trend and not just a hangover of a few years from the recession. There has been no resumption in the previous trends since then and for men possible even a decline. The coincidence of the Great Recession and the cessation of growth in labor force participation is a reminder that there are many other influences on labor force participation at older age besides health. Nonetheless, health is an important determinant, and we would like to understand how labor force participation increased so strongly in the 199s to the mid 2s even as health worsened. Possibly the relationship between labor force participation and self reported health may have changed, and further the relationship between self reported health and true health may have changed, masking the relationship between an individual s actual health and labor force participation at a given age. To better understand how expectations of future work affect actual labor force participation, we use data on the cohort of individuals 55 to 59 years of age. We focus on workers of this age because data on their subjective probability of working at age 65 is available in all waves, unlike that for persons age 51 to 54, for whom such data is only available on new recruitment in waves 1, 4, 7, and 1. One drawback to our analysis is the P65 question was only asked of workers prior to 26, while we want to consider the entire population over time. Our solution is to use the question asked since 1994 of all respondents about the chances that you will be working for pay at some time in the future, which we call PWork. 3
We will predict P65 prior to 26 among non workers based PWork. Specifically we regress P65 on PWork for non workers from 26 to 214 and then use the regression estimates to predict P65 for nonworkers prior to 26. Table 1 shows the results of this regression analysis. The question on probability of future work, or PWork, is a highly significant predictor of P65 for nonworkers in waves 8 to 12 of the HRS. In particular, as PWork increases from. to 1., P65 increases by.45. This coefficient is nearly the same for males and females. Conditional on PWork neither education nor age is a significant predictor of P65. We use this relationship to impute P65 for nonworkers in HRS waves 2 to 7. Figure 6 shows the actual P65 value for workers and the predicted P65 for nonworkers, both male and female, in waves 2 to 7 for respondents 51 to 59 years of age. Among workers, P65 is flat or increasing with age (particularly among females) for workers of older ages, while decreasing among nonworkers. By age 59, actual P65 for workers in earlier waves is 26 percent, while that for nonworkers is about 6 percent. There is a similar difference among males. Under stationarity, these trends are expected because with each year of age there are fewer future years of being at risk of a transition out of the labor force among workers and a transition into the labor force among nonworkers. Given the variation in P65 over cohorts of 55 to 59 year old workers and nonworkers, we grouped waves to reduce noise in the data. Figure 7 shows trends by groups of waves for 55 to 59 year old respondents in P65 by employment status of respondent. For workers, non workers, and the total population, P65 has increased over time. Yet looking at individual waves more closely, as in Figure 8, shows trends in P65 for individual waves flattening after 28. If P65 accurately forecasts trends in labor force participation six to ten years into the future, then the large increases we have seen in labor force participation at age 65 will soon come to a halt. One possible explanation for the leveling of P65 and actual labor force participation trends is that there has been a widening of health differences between workers and non workers, and that widening is only imperfectly captured by self reported health. For example, the actual health of those self reporting fair health may have worsened over time. If this were to occur, then the gradient of P65 by self rated health adjusted for age would decrease over time. To test this explanation, we regressed P65 for respondents 55 to 59 years of age on self reported health by sex in four groups of waves: 2 4, 5 7, 8 9, and 1 12. We controlled for individual years of age and year of HRS survey. The reference group in these regression analyses is those self reporting very good health. Figure 9 shows the results of these regressions, with each bar showing deviation from the very good category. Not surprisingly, in any given year, P65 is lower for persons in poorer self reported health categories than in better ones. Yet these gradients have indeed become worse with time. For example, among males, those in poor health in wave group 1 (1994 1998) had P65 that was 19 percentage points less than that for those in very good health. By wave group 4 (21 214), this difference had increased to 32 percentage points. Similarly, for females, those with poor health in wave group 1 had a P65 that was 15 percentage points lower than that for those with very good health; by wave group 4, this 4
difference was 25 percentage points. In other words, while the overall level of P65 increased over most years from 1994 to 214, the gradient for persons with lower self reported health decreased, aggravating the trend in lower levels of self reported health. To quantify the change in the gradient, we estimated the regression of P65 on self reported health, and self reported health interacted with wave number and wave number squared (linear and quadratic in time). Our results are shown in Table 2. Table 2 provides two specifications. The first interacts selfreported health with wave number. The second interacts self reported health with both the wave number and the wave number squared to be able to fit the flattening out of the trajectory of P65 as in Figure 8. The table presents results for males and females separately in both specifications. In the first specification for males, we see a negative monotonic relationship between P65 and selfreported health. The difference between males in very good health and those in poor health is about 15 points. That is, the probability of working at age 65 is.15 lower for those in poor health than it is for those in very good health, on an average value of about.28. The differences increase over time. This is evident in the interaction term for the coefficient of those in poor health with wave number. The value of this interaction term, 1.56, indicates that, in each subsequent wave, the P65 value for those in poor health decreased by 1.56 points relative to those in very good health. As a result, by wave 12, the difference in P65 between those in poor health and those in very good health was 3 percentage points, or.3 in probability. The first specification for females yields similar though less dramatic results. The difference between females in very good health and those in poor health is about 11 points. The interaction term between wave and those in poor health is 1.22. Figure 1 summarizes the results from our second specification evaluated at respondents age 57. It shows that, over time, P65 increased substantially for males 57 years of age in excellent, very good, or good health, but less so for those in fair and poor health. Similar patterns are evident for females: P65 increased more for those in excellent, very good, or good health than for those in worse health. As a result, the gradient in anticipated work as a function of health widened over time. Because working longer can have important benefits on economic preparation for retirement, these health differences in future work imply widening economic differences in old age as well. Altogether, reporting fair or poor health predicts a greater reduction in the reported likelihood of working past age 65 in later waves of HRS. But what is the mechanism for these changes? Worsening health may not affect all parts of the population equally. Possibly population worsening health did not affect all parts of population equally. To explore this further, we examined more objective indicators of health for respondents 55 to 59 years of age. Figure 11 shows these results. The first panel shows the prevalence of diabetes among those 55 to 59 years of age, both workers and non workers, as a function of self reported health in the four wave groupings. It shows, for example, that the frequency of diabetes among those reporting good health increased from 11.7 percent in the first three waves to 23.9 percent in the last three waves. There were similarly large 5
increases among those in fair health, from 22.3 percent to 33.5 percent, and among those in poor health, from 32.7 percent to 43.9 percent. Increases among those in excellent or very good health, however, were only 4 to 5 percent, and the overall proportion among these groups with diabetes remains less than 1 percent. An implication of the sharp rise in diabetes especially in the good, fair and poor groups is that, over time, there has been increasing overlap in diabetes prevalence and groups of respondents as defined by self reported health. For example, the prevalence of diabetes among those reporting fair health in 21 214 was greater than that among those reporting poor health in 1992 1996. If diabetes were the only determinant of true health status, then the meaning of fair and poor health would have changed over time: some categorized as fair in 21 214 would have been categorized as poor in 1992 1996. Similar patterns are evident in the other health indicators variables shown in Figure 11. For example, the proportion of those in fair health with a body mass index of at least 35 increased from 13 percent to 29 percent. Limitations in activities of daily living (ADL) and instrumental activities of daily living (IADL) show similar patterns. Cognition score is positively correlated with health, but within self reported health categories it changes over time. Limitations in mobility, large muscles, and motor activity similarly increased over time. We conclude that the meaning of self reported health has changed over time, particularly for those in the lower health categories. HRS respondents reporting fair health in recent waves have more objective health problems than those reporting fair health in earlier waves. This could explain why those in worse self reported health categories have over time reported lower probability of working past age 65. Another reason for changes in P65 may be the longer work life needed to finance a longer retirement. We can assess this through HRS questions about subjective survival to age 75. HRS asks respondents to estimate, what are the chances you will live to be age 75. We refer to this question as subjective probability of living to age 75, or P75. Responses to this question could be an additional marker of respondent health. Figure 12 presents the subjective probability of survival to age 75 for male, female, and all respondents 55 to 59 years of age from wave 2 in 1994 to wave 12 in 214. Note females have expressed a higher subjective probability of survival. Subjective survival reached its maximum for each group in 2 or 22. There has been some decline since then: in 2, respondents, on average, said there was a 66.6 percent probability of surviving to age 75; by 214, this had decreased to 62.3 percent. There was also a temporary increase in wave 9, in 28, which may have reflected some optimism before the Great Recession hit later that year. The early increases in P75 paralleled those of P65 at the same time (compare Figures 8 and 12). The later decline in P75, however, occurred before the flattening of the trend in P65. Part of the decrease in P75 could be due to downward trends in self reported health and to other indicators of declining health. Figure 13 shows P75 results by self reported health for males across survey waves. For all males, mean subjective probability of survival to age 75 decreased from 63.6 in waves 2 4 to 59.8 in waves 1 12. Across all four waves, better self reported health is associated with 6
greater subjective probability of survival to age 75. At the same time, there is a decrease in P75 in almost every case. Among those reporting good health, for example, the mean level of P75 decreased from 6.9 to 56.4. From the perspective that subjective survival is a health measure, the decrease in overall self reported health for males was partly due to both reductions in P75 within each health group. For females, the trends were somewhat less consistent than for males. To smooth patterns and better discern them, we regressed P75 on age, wave number (linear), selfreported health, and a quadratic specification with wave number squared and an interaction term with self reported health. Table 3 presents the results of the resulting equations. The linear specification for males shows a very large difference in self reported health. The difference between those with excellent self reported health and those who reported poor health is 39 percentage points. That is, males in poor health report a mean level of P75 that is 39 percentage points lower than that for males in excellent health. P75 decreased slowly over time for those in the reference group of very good self reported health. P75 decreased more rapidly for those in good or poor health. Over the waves shown, P75 among those reporting good health decreased by 6.7 percentage points. The results are similar for females, although none of the health wave interactions is significant. In the quadratic specification, the coefficients on self reported health are about the same as those in the linear specification. The squared term on wave is negative, indicating a downturn in later waves, while few of the interaction terms are significant. To show the overall pattern, Figure 14 shows the fitted values of P75 by self reported health in each wave. For males reporting good or poor health, there is a monotonic decrease over time. For the other health categories, there is a small peak in wave 6 (22), but a decline since then. There is no indication of an upturn in P75. Females within each health category have higher P75 values than males do. For example, among females reporting very good health, P75 is, on average, about five percentage points higher than it is for males of similar health. There is more curvature in these trends among females, with a peak near wave 7 in 24. There is no indication of an upturn in P75 among these groups, with the exception of those in poor health in recent years. Altogether, the linear specification suggests subjective survival did not increase over this time, but the quadratic specification suggests it increased modestly in the 199s and early 2s. Thus, in the linear specification subjective survival not increasing but the quadratic specification suggests that in 199s and early 2s it was increasing modestly. Though the trends of P75, which peaked in the early 2s, and P65, which continued to rise before flattening in the late 2s, do not exactly match temporally, the lack of an increase in P65 is consistent with the lack of an increase in P75. This supports the view that perceptions of greater longevity contributed to the increases in P65 observed in earlier years. Conclusions 7
It has become apparent in the last several years that the upward trend in labor force participation at older ages, strongly evident prior to 28, has declined or even halted. The cessation of an upward trend in P65 lends weight to the view that it has halted. A possible cause is worsening health both in self reported health but in actual health within self rated health categories. Accompanying the decline in health has been a differential change in P65 across health categories. Average value of P65 by those in the better health categories have increased and increased very strongly in some cases, but they have increased more slowly in the worse health categories. The widening of P65 across health categories suggests widening differentials in work life that would lead to increasing differences in economic preparation for retirement. 8
9 Figure 1. Labor Force Participation Source: CPS 1 2 3 4 5 6 199 1991 1992 1993 1994 1995 1996 1997 1998 1999 2 21 22 23 24 25 26 27 28 29 21 211 212 213 214 215 216 Men 62 64 65 69 5 1 15 2 25 3 35 4 45 5 199 1991 1992 1993 1994 1995 1996 1997 1998 1999 2 21 22 23 24 25 26 27 28 29 21 211 212 213 214 215 216 Women 62 64 65 69
Figure 2, P65, workers age 55 59 5 45 4 35 3 25 2 15 1 5 1992 1994 1996 1998 2 22 24 26 28 21 212 214 Male Female Figure 3, Percent with fair or poor self rated health age 55 59 6. 5. 4. 3. 2. 1. Working Not working all. 2 4 5 7 8 9 1 12 HRS wave 1
Figure 4, Change in male LFP, age 65 69, 3 year moving average 1.4 1.2 1.8.6.4.2.2 1991 1992 1993 1994 1995 1996 1997 1998 1999 2 21 22 23 24 25 26 27 28 29 21 211 212 213 214 215 216.4.6 Change in female LFP age 65 69 3 year moving average 1.2 1.8.6.4.2.2 1991 1992 1993 1994 1995 1996 1997 1998 1999 2 21 22 23 24 25 26 27 28 29 21 211 212 213 214 215 216.4 11
Figure 5, P65, workers 5 45 4 35 3 25 2 15 1 5 1992 1994 1996 1998 2 22 24 26 28 21 212 214 Male Female Figure 6, Males P65. Actual for workers and predicted for nonworkers, waves 2 7 4 35 3 25 2 15 1 5 51 52 53 54 55 56 57 58 59 Actual (workers) predicted (non workers) Females P65. Actual for workers and predicted for nonworkers, waves 2 7 3 25 2 15 1 5 51 52 53 54 55 56 57 58 59 Actual (workers) predicted (non workers) 12
Figure 7, P65, ages 55 59 45. 4. 35. 3. 25. 2. 15. 1. 5. 2 4 5 7 8 9 1 12 wave Not working Working All Figure 8, P65, workers and nonworkers across HRS waves 4. 35. 3. 25. 2. 15. Male Female All 1. 5.. 2 3 4 5 6 7 8 9 1 11 12 13
Figure 9, P65. gradient by self rated health, age adjusted, males 15 1 5 5 1994 1998 2 24 26 28 21 214 1 15 2 25 3 35 Excellent health Very good (ref) Good health Fair health Poor health P65. gradient by self rated health, age adjusted, females 5.. 5. 1994 1998 2 24 26 28 21 214 1. 15. 2. 25. 3. Excellent health Very good (ref) Good health Fair health Poor health 14
Figure 1, Fitted P65, males, workers and nonworkers, age 57 across HRS waves 45 4 35 3 25 2 15 1 excellent v good good fair poor 5 2 3 4 5 6 7 8 9 1 11 12 Fitted P65, females, workers and nonworkers, age 57 across HRS waves 4 35 3 25 2 15 1 excellent v good good fair poor 5 2 3 4 5 6 7 8 9 1 11 12 15
Figure 11. Health indicators, workers and nonworkers 55 59 Frequency of Diabetes.5.4.3.2.1..35.3.25 92 96.2 98 2.15.1 4 8.5 1 14. Frequency of BMI 35 or greater 92 96 98 2 4 8 1 14 Frequency of Health limits work Number of ADL limitations 1.8.6.4.2 1.4 1.2 1. 92 96.8 98 2.6.4 4 8.2 1 14. 92 96 98 2 4 8 1 14 Number of IADL limitations.5.4.3.2.1. Cognition score 2. 15. 92 96 1. 98 2 5. 4 8 1 14. 92 96 98 2 4 8 1 14 16
Index of limitations on mobility Index of large muscle limitations 3.5 3. 2.5 2. 1.5 1..5. 92 96 98 2 4 8 1 14 3.5 3. 2.5 2. 1.5 1..5. 92 96 98 2 4 8 1 14 Index of gross motor limitations Index of fine motor limitations 2..8 1.5 1..5. 92 96 98 2 4 8 1 14.6.4.2. 92 96 98 2 4 8 1 14 17
Figure 12. Subjective probability of survival to age 75, ages 55 59 across HRS waves 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 1 11 12 male female all Figure 13, Subjective probability of survival to age 75, males 55 59 by SRH and by wave group 9. 8. 7. 6. 5. 4. 3. 2. 2 4 5 7 8 9 1 12 1. excellent very good good fair poor all 18
Subjective probability of survival to age 75, females 55 59 by SRH and by wave group 9. 8. 7. 6. 5. 4. 3. 2. 2 4 5 7 8 9 1 12 1. excellent very good good fair poor all Figure 14. Fitted P75, males across HRS waves 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 1 11 12 excellent v good good fair poor 19
Fitted P75, females across HRS waves 9 8 7 6 5 4 3 2 1 2 3 4 5 6 7 8 9 1 11 12 excellent v good good fair poor 2
Table 1 Regression of P65, nonworkers waves 8 12 Males Females Prob. work in future, PWork.465.452 [.2]*** [.14]*** Age 51.433.171 [.296] [.196] High school 3.211.965 [1.962] [1.452] Some college.13 2.829 [2.] [1.459]* College.946.963 [2.238] [1.567] Black.433.867 [1.55] [1.74] Other 4.15.13 [2.3]** [1.542] constant.572 3.789 [2.554] [1.729]** R^2.282.288 N 157 264 *** p<.1, ** p<.5, * p<.1 Table 2, Regression of P65, population 55 59, HRS waves 2 12 Male Female Male Female Age -.27811 -.63147***.28125.63572*** (.19558) (.15227) (.19559) (.15227) SRH excellent 1.51237***.61824 1.98344*** 1.53645 (1.84289) (1.53167) (3.8797) (3.21676) SRH good -2.37876-1.1146.35598 3.9393 (1.58412) (1.27912) (3.34359) (2.6862) SRH fair -5.63695*** -5.99215*** 4.7879 2.645 (1.9637) (1.4847) (4.7425) (3.1372) SRH poor -14.59852*** -11.2157*** 16.7886*** 11.51437*** (2.67211) (1.9954) (5.595) (4.21551) Wave number (W#) 1.81137*** 1.67813*** 3.1418*** 2.246*** (.14288) (.11781) (.78423) (.62846) W#^2.9543*.425 (.5538) (.4421) SRH excellent*w# -.8319***.9597.94856.2352 (.25584) (.21111) (1.36492) (1.9325) SRH good*w# -.56156*** -.31745* 1.613.78727 (.229) (.16615) (1.13281) (.8947) 21
SRH fair*w# -1.3243*** -.8925*** 1.65641 2.3968** (.24358) (.18869) (1.36827) (1.3316) SRH poor*w# -1.55878*** -1.21853***.74719 1.325 (.33919) (.25828) (1.86722) (1.4558) SRH excellent*w#^2.468.2316 (.999) (.7855) SRH good*w#^2.7483.7868 (.7967) (.6284) SRH fair*w#^2.2487.8749 (.9577) (.7214) SRH poor*w#^2.5953.133 (.13275) (.9868) Constant 37.57442*** 5.8281*** 34.25216*** 49.558*** (11.2162) (8.73913) (11.37858) (8.89863) Observations 13,758 19,47 13,758 19,47 R squared.7524.6251.7569.6295 Standard errors in parentheses *** p<.1, ** p<.5, * p<.1 Table 3. Regression of P75, population ages 55 59, waves 2 12 Male Female Male Female Age.182.215.17725.1442 (.16869) (.14194) (.16871) (.14178) SRH excellent 7.8227*** 5.91892*** 6.77358** 5.35798* (1.58211) (1.4261) (3.3295) (2.9799) SRH good 6.57148*** 6.57744*** 5.2713* 3.24641 (1.36843) (1.19272) (2.88431) (2.49879) SRH fair 2.14361*** 18.38569*** 19.5965*** 25.38942*** (1.7145) (1.38938) (3.54995) (2.94531) SRH poor 31.1995*** 32.73541*** 28.69734*** 25.4677*** (2.3343) (1.89999) (4.79543) (4.153) Wave number (W#).588.2818.7394 1.788*** (.12293) (.1935) (.67345) (.58162) W#^2.5727.12948*** (.4755) (.49) SRH excellent*w#.7365.4256.38276.32566 (.2195) (.19561) (1.17194) (1.187) SRH good*w#.42168**.22391 1.842 1.44683* (.17486) (.15437) (.97593) (.83151) SRH fair*w#.11.16573.22629 2.36415** (.21113) (.17635) (1.18784) (.96787) SRH poor*w#.33181.268 1.27826 2.89978** 22
(.2942) (.24294) (1.62331) (1.3335) SRH excellent*w#^2.3668.2515 (.858) (.7257) SRH good*w#^2.4228.8775 (.686) (.5837) SRH fair*w#^2.1617.17627*** (.8296) (.6746) SRH poor*w#^2.684.19233** (.11532) (.938) Constant 58.7566*** 71.4251*** 56.8161*** 66.92972*** (9.67255) (8.14492) (9.8153) (8.28269) Observations 13,477 18,534 13,477 18,534 R squared.14363.14374.14387.1468 Standard errors in parentheses *** p<.1, ** p<.5, * p<.1 23