Survey of Income and Program Participation (SIPP) Analytic Research Small Grants Competition Long-Run Earnings Volatility and Health Insurance Coverage Matthew S. Rutledge, Center for Retirement Research at Boston College

CONFERENCE DRAFT DO NOT CITE

Long-Run Earnings Volatility and Health Insurance Coverage: Evidence from the SIPP Gold Standard File1

Preliminary – Please Do Not Cite Matthew S. Rutledge2 Research Economist Center for Retirement Research at Boston College [email protected]

Abstract Despite the notable increase in earnings volatility and the attention paid to the growing ranks of the uninsured, the relationship between career earnings and short- and mediumrun health insurance status has been ignored due to a lack of data. I use a new dataset, the SIPP Gold Standard File, that merges health insurance status and demographics from the Survey of Income and Program Participation with career earnings records from the Social Security Administration (SSA) and the Internal Revenue Service (IRS) to examine the relationship between career earnings volatility and health insurance coverage. I find that more volatile career earnings are associated with an increased probability of experiencing an uninsured episode, especially for men. I also provide suggestive evidence that earnings volatility leads to gaps in health insurance coverage, but not vice versa. These findings are consistent with the “scarring” literature, and suggest the importance of safety-net measures for job losses and health insurance coverage.

1

This project was supported by the National Poverty Center (NPC) using funds received from the U.S. Census Bureau, Housing and Household Economics Statistics Division through contract number 50YABC266059/TO002. The opinions and conclusions expressed herein are solely those of the author and should not be construed as representing the opinions or policy of the NPC or of any agency of the Federal government. 2 The author thanks Luke Shaefer, Dan Leeds, Martha Stinson, Gary Benedetto, Jim Davis, Clint Carter, Shawn Pelak, Sheldon Danziger, Zoë McLaren, Brian Rowe, and George Lausch for their helpful comments and assistance in accessing and using the data.

2

Introduction Over the last few decades, the volatility of earnings have followed a distinct upward trend (Dahl, DeLeire, and Schwabish, 2008), contributing substantially to the rise in earnings inequality (Haider, 2001). Meanwhile, the growing ranks of the uninsured have received a great deal of attention both from researchers and policymakers. Despite the obvious link between labor income and health insurance, as the vast majority of nonelderly Americans receive health coverage through an employer (DeNavas-Walt, Proctor, and Smith, 2009), the relationship between increasing earnings instability and health insurance coverage has been largely unexplored. Until recently, there was no data source that provided long-term earnings histories and insurance status; for example, the most prominent long-run longitudinal dataset, the Panel Study of Income Dynamics (PSID), only began asking about insurance status in 1997. A recent joint effort led by the U.S. Census Bureau and the Social Security Administration (SSA) has resulted in a new data source that links health insurance information, among other variables, from the Survey of Income and Program Participation (SIPP) to administrative data on earnings from the SSA records. In this paper, I use this linked data, the SIPP Gold Standard File, to estimate the relationship between long-run earnings volatility and short- and medium-run health insurance coverage. I use a novel measure of earnings volatility that takes into account the natural growth of real earnings as one ages. I test the hypothesis that, for two individuals who have the same fitted long-run earnings trend, the one whose earnings deviate more from that trend is more likely to be uninsured no matter when the researcher

3

observes him or her. To my knowledge, this is the first paper to examine the relationship between career earnings, in either levels or volatility, and health insurance coverage. I find that there is a statistically significant, but small, positive relationship between career earnings volatility and the probability that someone ever experiences a period without health insurance coverage, though this effect is stronger for men. I also find, on the other hand, that individuals who have more volatile earnings are less likely to be continuously uninsured. I provide suggestive evidence that having more volatile earnings leads to an increased likelihood of being uninsured, and not the other way around, although data limitations prevent a definitive diagnosis of causality. In short, inconsistency in one’s career earnings profile translates into inconsistent health care coverage. Though the exact cause of earnings volatility is still subject to debate, a likely substantial contributor is the effect of job displacement (Stevens, 2001). Numerous researchers have found that job loss may have a “scarring” effect (Ellwood, 1982), making future losses more likely (Stevens, 1997), decreasing the quality of subsequent employer-employee matches (Fallick, 1996), and reducing earnings over the long run (Ruhm, 1991, Jacobsen, LaLonde, and Sullivan, 1993). If one job loss makes future displacement more likely, and leaves even re-employed workers with lower-quality jobs than other workers who were never displaced, then we should also observe that previously displaced workers are less likely to be covered by health insurance at any point-in-time. The “scarring” of job displacement may also lead to undesirable health insurance dynamics: insured workers who have a volatile earnings history may be more likely to lose insurance, and less likely to gain insurance once it is lost. These dynamics

4

suggest an increased need for safety-net insurance coverage in the event of job loss, and strengthening the motivation for measures that reduce the negative impacts of job displacement, such as job training, continuing education, and unemployment insurance.

Related Literature The literature suggests that earnings volatility has trended upward over the last few decades. Gottschalk and Moffitt (1994, 1995) decompose earnings variance into its permanent and transitory elements, and find that transitory earnings variance increased between the 1970s and 1980s. In a recent update (2008), they find that volatility remained at this higher level through 2004, though it did not increase further. Haider (2001) also finds that earnings instability increased substantially in the 1970’s, explaining about half of the growth in earnings inequality. Though Cameron and Tracy (1998) and Moffitt and Gottschalk (2002) find that earnings volatility declined in the late 1980’s and early 1990’s, recent evidence suggests that volatility has risen again in the 2000’s (Dynan, Elmendorf, and Sichel, 2007; Shin and Solon, 2008; Dahl, DeLeire, and Schwabish, 2008). Stevens (2001) attributes much, but not all, of this growth in the transitory earnings variance to the persistence of earnings losses after job displacement. Numerous papers have found that job loss leads to less stable future employment and permanently lower wages. Ruhm (1991), Jacobsen, LaLonde, and Sullivan (1993), and Stevens (1997) find evidence that workers who lose a job may be “scarred” for years even after

5

re-employment, with lower wages, increased probability of future job losses, and permanently increased earnings volatility. Job losses may also lead to the loss of health insurance, as employers are the source for about three-quarters of non-elderly insured Americans’ coverage (DeNavasWalt, Proctor, and Smith, 2009). Gruber and Madrian (1997) find that job loss is frequently associated with insurance loss, and Czajka and Olsen (2000) find a similar effect of parental job loss on children’s insurance coverage. While the research on the permanent effects of job displacement on direct labor market outcomes is rich, very little is known about the long-run effects of job losses and earnings volatility on health insurance. To my knowledge, the only previous study to examine this question is Simon and Schroder (2006). Using the 1996 and 2001 panels of the SIPP, they find that previously involuntarily displaced workers are significantly less likely to have insurance through their next job than workers who had not been displaced, though the displaced workers eventually catch up to the other workers after just more than a year. This study is the first to examine the relationship between long-run earnings volatility and point-in-time health insurance status. I use a new dataset constructed from administrative data that is likely to be more accurate than self-reported data. Eventually, this paper will also help to test the validity and usefulness of a public-use version of the same data, which the data administrators hope will make this data more accessible to researchers without drawing concerns about the disclosure of private information.

6

Conceptual Framework It is well-established that individuals who have high-paying jobs are likely to have stable health insurance coverage (see, among many others, Kuttner and Rutledge, 2007). Because the employer is the dominant source of coverage in the United States, contemporaneous earnings volatility (especially when it is caused by a job loss) will be strongly associated with being uninsured point-in-time (Gruber and Madrian, 1997), or ever experiencing a gap in coverage (Short and Graefe, 2003). This paper examines a slightly different question: whether, all else (including the earnings level) equal, an individual who has a more volatile earnings history is more likely to be uninsured at any particular time. My hypothesis is that, controlling for the long-run level of earnings, an individual who has a history of more variable earnings is also more likely to have gaps in his or her health insurance coverage. The association between volatile earnings history and volatile insurance status may arise for any number of reasons. First, unstable earnings may be a signal of the unobserved ability of the individual. It is important to note that earnings volatility, at least across years, is likely to arise because of downside, not upside, risk. Unlike total income, which may have significant upside due to investment windfalls, property sales, inheritance, or lottery winnings, substantial year-to-year positive shocks to labor market income are unlikely. Substantial negative shocks, however, are quite common, due to job losses or cuts in hours. Volatile earnings, therefore, suggest a tenuous connection to the labor market, through frequent and/or persistent periods of non-employment, seasonal or temporary work, or inconsistent hours within a position. A history of volatile earnings,

7

then, could be evidence that the individual is of low ability, and therefore has trouble obtaining, and maintaining, high-quality jobs that are likely to offer stable health insurance coverage, even if earnings observed in the short-run, as it would be in many longitudinal studies, are stable. Low-ability individuals may also be less likely to seek coverage if it is not offered by an employer, as the individual insurance market has high transaction costs. Alternatively, there is some evidence in the literature that labor market shocks are persistent and may cause “scarring,” making individuals who suffer layoffs or cuts in hours more likely to repeat these negative results, while being unable to restore their previous earnings level.3 This suggests that previous earnings volatility makes current earnings more volatile, which may result in periods without health insurance coverage. Finally, an individual who has volatile earnings may have low risk aversion. She may therefore have little demand for protection against financial risk, and is unwilling to pay, or sacrifice wages, for health insurance at its market rate. Also, an individual who has been previously uninsured in the past without consequence may feel that she does not need continuous coverage. As a result of some or all of these factors, I expect individuals who have volatile earnings histories to be more likely to lack health insurance coverage in any period in which we observe them. I also expect past earnings volatility to be associated with an increased chance of losing coverage if one is insured, and a decreased probability of

3

Stevens (1997) finds that earnings remain below their peak level even six or more years after the initial job loss.

8

gaining coverage if one is uninsured, though the current version of this paper does not explicitly test this prediction. Differences between the genders in long-run stability in labor force participation are likely to affect this estimated relationship between earnings volatility and health insurance status. Men may leave the labor force, but for reasons for which I can account with the volatility measure. First, if men retire or pursue education, but these actions generally occurs at the beginning or end of careers, while my volatility measure focuses on mid-career gaps. Second, men may get discouraged by joblessness and stop searching, which would be coincident with the years of decreased earnings that lead to more volatile earnings, exactly the effect I am trying to capture. Women, however, often have years of decreased or zero earnings mid-career due not to the lack of success in a career, but for child raising or the care of an elderly relative. Goldin (2006) finds that though women’s labor force participation at any one time has risen, just under half of mothers who graduated from a “selective” college in 1976 have left the labor force within the first twenty years of their careers. Most women will be able to retain health insurance coverage during this period, either through a spouse or by becoming eligible for Medicaid. Women will therefore appear to have more volatile earnings due to gaps in their earnings histories, but this will not necessarily be associated with gaps in health insurance coverage. The conceptual framework outlined above is more appropriate for men, so it is important to estimate the genders separately.4

4

I expect that women who never have children should have the predicted positive correlation between earnings volatility and the probability of being uninsured, though Goldin (2006) finds that 23.5 percent of

9

Data The SIPP Gold Standard dataset is the result of a collaboration of the Census Bureau, the SSA, the Internal Revenue Service (IRS), and the Congressional Budget Office (Abowd, Stinson, Benedetto, 2006). This dataset links information collected in the 1990-1993, 1996, 2001, and 2004 panels of the SIPP to earnings and benefits histories from SSA records. The SIPP surveys households every four months during a two- to four-year period, about a variety of demographics, labor, welfare, and health topics in each intervening month. Each year between 1990 and 1993, and again in 1996, 2001, and 2004, SIPP began a new panel of households. The SIPP Gold Standard File contains a subset of variables available in the full SIPP data, including demographics and marital and fertility history, self-reported income and wealth, education, work status, industry and occupation, welfare receipt, and most importantly for this study, monthly health insurance coverage (any coverage, plus an indicator of employer-sponsored coverage). Survey questions featured in the Gold Standard subset are comparable across panels. An individual is included in the SIPP Gold Standard dataset if he or she is age 15 or older and has a valid Social Security number on file (approximately 88 percent of the age-eligible sample). Their earnings history is obtained from annual employer W-2 reporting to the IRS, which then goes into one’s Detailed Earnings Record (DER) in the

women without children leave the work force for at least six months at some point in their careers, compared to 14 percent of men. I will pursue this in the next draft of the paper.

10

SSA’s files. The Gold Standard dataset includes annual earnings variables from the DER dating to 1978 that are not top-coded, unlike most other large-scale datasets. The Gold Standard File also includes (top-coded) earnings reports from the SSA’s Summary Earnings Record (SER) dating back to 1951. I retain only a single total earnings measure for each year of the matched SIPP respondent’s career; following the recommendation from the Gold Standard codebook, I use the SER total for most individuals, using the sum of the DER reports only if there are non-FICA earnings or the earnings have been topcoded. I restrict the sample to individuals 18 to 64 at some point during their time sampled by SIPP who have at least 10 years of positive earnings history, without which it would be difficult to measure the volatility of one’s earnings. There are approximately 130,000 women and 126,000 men in the restricted sample, or about 350,000 and 330,000 person-years, respectively.

Methodology This paper’s primary estimation model is a regression of short-run health insurance status on long-run earnings volatility, the long-run level of earnings, and other covariates. The first challenge is the measurement of long-run earnings volatility. In the data, real earnings trend upward over one’s career, likely due to returns to occupation- and firm-specific capital, peaking in one’s 50’s before reaching a plateau, or even declining, late in one’s career. An individual’s earnings variance is therefore a poor measure of volatility, as a worker who has a steep but consistent earnings profile (Figure 1, left

11

panel) may grade out as more volatile than one who has a narrow but inconsistent profile (right panel). Moreover, for two otherwise-identical individuals who have different years of experience, the one who has more experience will look more volatile, as there is more of a difference in earnings between the first few and last few years of his career than for someone who has a shorter, more truncated profile. An approach that better captures the consistency of year-to-year earnings, or lack thereof, while still taking into account the usual career earnings profile is to control for both predicted earnings throughout one’s career and the deviation from that fitted profile. For two workers who have identical fitted earnings profiles, as shown in Figure 2, the one whose earnings deviate more from that (flexible) trend line is clearly the more volatile one (right panel). As individuals will be heterogeneous, both in their capacity to earn in any given year and in their ability to grow those earnings, in a way that is not captured by observable characteristics such as age and education, I fit a cubic age-earnings profile to each matched respondent i’s earnings history:5  

   

   

.

The measure of volatility I use is the sum of the squared residuals (SSR) from this regression:6

5

As a robustness check, I replace the SSR calculated by a cubic regression of age on real earnings, delineated above, with an SSR calculated from a quadratic regression of age on real earnings in each regression. All results are nearly identical in both magnitude and significance. 6

Gottschalk and Moffitt (1994) use a similar methodology to calculate the permanent and transitory variances of earnings.

12

 

   

A more volatile earnings history results in a larger

   

.

[measured in real (2000) dollars

squared] as annual earnings more frequently deviate from the individual’s fitted trend.7 It is important to control for some measure of long-run earnings level, not just volatility; not only are higher earnings individuals less likely to be uninsured, but deviations that are small in percents are likely to add up to a much larger 

the higher

one’s lifetime level of earnings.8 To control for the long-run level of earnings, ideally I would include the matched respondent’s predicted value of real earnings,

, for

each year of age as a regressor in the main regression, to control for the earnings level. These regressors would be highly correlated, however, due to the serial correlation of earnings. Instead, I include the mean predicted real earnings level over four- or five-year periods (ages 18 to 21, 22 to 25, 26 to 30, and so on, up to ages 61 to 64).9

7

Using the individual’s own earnings history to predict the fitted trend, then summing up the squared deviations from the trend, may underestimate one’s earnings volatility. As a thought experiment, imagine replacing an individual’s highest earning year with zero earnings for the year. This increases the estimated SSR, because there is a point that is far from the trend line, but not by as much as if the fitted trend had been held constant, because the new $0 observation drags the fitted trend slightly downward. An alternative method would predict earnings based on observable characteristics, and then calculate the sum of squared residuals from that regression, but I believe that the SIPP Gold Standard File does not include enough variables to properly control for individual heterogeneity in earning potential and earnings growth potential. Also, those who take time off to pursue education or retire earlier than 65 will have additional years of low earnings that will make SSR larger, but these workers are more likely to have stable health insurance coverage. 8 9

Indeed,

is positively correlated with the mean of earnings over one’s career.

I had trouble achieving convergence in the logit model when using a continuous measure of average predicted earnings, so I instead use a categorical variable for each four- or five-year average, with each category representing either $5,000 or $10,000 ranges.

13

The dependent variable is a measure of the lack of health insurance coverage, so the regression model is ln

∑

.

The uninsured variable could take several different forms. Ultimately, I estimate the effect of earnings volatility, the level of earnings, and other covariates on six different coverage measures: 1. an indicator for whether individual i is ever uninsured during his/her time in the SIPP, 2. an indicator for whether individual i is always uninsured during his/her time in the SIPP, 3. the number of months individual i lacks health insurance coverage during his/her time in the SIPP, 4. an indicator for whether individual i is ever uninsured in a specific year (observations at the person-year level),10 5. an indicator for whether individual i is always uninsured in a specific year, 11 6. and the number of months individual i lacks health insurance coverage in a specific year. In specifications 1, 2, 4, and 5,

∙ is the logistic function. There is bunching in the

number of months uninsured, because many people are never uninsured, and others are 10

I also intend to estimate a similar model at the person-month level. The sample size is quite large, however, so convergence takes a long time. Preliminary estimates of the effect of Z on an indicator for whether the individual is uninsured in a given month, using only a random subsample of the data, are similar to the estimates for “ever uninsured” at the person-year level.

11

This measure of insurance status is comparable to the Current Population Survey’s health insurance variable, which is used to create the uninsured rate figure cited most often in policy discussions.

14

uninsured throughout their time in the SIPP sample. Specifications 3 and 6 are therefore estimated by Tobit regression, with the lower censoring limit set at zero months for both models; the upper censoring limit is 12 months for specification 6, and the maximum number of months one could be in the sample – 24 for most panels, except for 2004 (36) and 1996 (48) – for specification 3. The sample for the first three specifications includes one observation per person; for the latter three specifications, I include one observation per person-year. For each of these (nonlinear) regressions, I report the average marginal effect, or the mean of the derivative of the dependent variable with respect to ln

, and its

standard error (approximated by the Delta Method). In each case, my hypothesis is that the mean derivative, or marginal effect, of ln

on the measure of being uninsured

should be positive and statistically significant; i.e., greater earnings volatility over one’s career should be associated with one being more likely to lack health insurance coverage at any point that one is observed. The independent variables in each regression are measures of earnings volatility, the average level of predicted earnings over a period of four or five years (K periods), and other individual characteristics.

enters the regression as a natural logarithm because

of its long right tail. Average predicted earnings, dollars.

, are expressed in real (2000)

includes age and its square, categorical variables for race (with Hispanic as a

separate, mutually exclusive category) and education, and indicator variables for whether the individual is married or foreign-born.12

12

also includes a fixed effect for the year that

I also include an indicator for whether the information on nativity is missing.

15

the individual entered the SSA/IRS earnings sample, because a 45-year-old in 1990 who has worked 23 years will have faced a much different macroeconomy than a 45-year-old who has 23 years of experience in 2006. Each regression is run separately for men and women, to account for the difference in the measurement of volatility due to women’s increased prevalence for exiting the labor force for family reasons. I also use SIPP weights in each regression to account for oversampling of low-income individuals.13 I use Huber-White heteroskedasticity-robust standard errors, and in specifications 4 through 6, standard errors are clustered at the individual level. Table 1 presents summary statistics for men and women, both unweighted and weighted. The difference between the unweighted (almost two-thirds) and weighted (just more than 40 percent) proportions that are ever uninsured during their time in the sample are evidence of the importance of weighting to account for oversampling of low-income individuals, who are more likely to be uninsured. Overall, the uninsured rates seem a little high relative to numbers reported elsewhere, especially considering that the sample selection technique, particularly the requirement of 10 years of positive earnings, would suggest that the uninsured rate would be lower than in the overall population; for example, in the full (weighted) SIPP sample in 2002, only 9.1 percent of non-selfemployed workers were uninsured all year, compared to 15 percent for both genders

13

The coefficient on ln is generally larger (more positive) in weighted regressions compared to the same model unweighted, but there is no difference in the qualitative finding of statistical (in-)significance. A more complete weight would account for the possibly non-random sample selection that occurs because (1) individuals need to report Social Security numbers to be matched and (2) they need at least 10 years of positive earnings to be included in the regression; I plan to adjust the weights in the next draft of this paper.

16

(weighted) in this sample.14

is heavily skewed to the right, as evidenced by its large

standard deviation.

Results In Table 2, I report the results of logit regressions of indicators for whether the matched worker is ever, or always, uninsured during his/her time in the SIPP sample on earnings volatility, the level of earnings throughout one’s career, and personal characteristics. I do not report estimates for the level of earnings (almost universally negative, as expected) nor the SSA-IRS entry year fixed effects. The first two lines of Table 2 are the means of each individual’s marginal effect of ln

on the probability of being uninsured ever

or always and their Delta-method standard errors, while the rest of the table reports logit coefficients and standard errors. The first two columns of Table 2 contain the main result of this paper: as career earnings volatility increases, one is more likely to ever experience a period without health insurance coverage within a two- to four-year period of observation. This effect is highly statistically significant for both genders, though the magnitude of the effect is nearly four times larger for men, as expected. Interpreting the magnitude of the mean marginal effect is somewhat difficult, because most readers are likely not familiar with how real-life earnings variation

14

Full-sample SIPP estimates are from the Economic Research Initiative on the Uninsured’s Fast Facts Tables, found at http://www.rwjf-eriu.org/fastfacts/index.html.

17

translates into changes in the sum of squared residuals from a regression of real earnings on a cubic function of age, denominated in the natural log of dollars squared. To assist in the interpretation of these coefficients, I consider a counterfactual. I replace every matched worker’s highest earning year from the SSA-IRS record with zero earnings, as if the worker was laid off at the height of his or her career and spent the year unemployed, but then resumed his or her career without scarring.15 I then recalculate for a random subsample of 5,000 workers. This recalculated percent, or about 200 million dollars squared, larger than the actual percent (because

is a median of 6 . Multiplying 6

enters as a natural log) by the mean derivative suggests that such a

change would result in a man being on average 0.16 percentage points more likely to be uninsured at some point while he’s observed in the SIPP, while a woman is on average 0.04 percentage points more likely, so compared to a mean of 45 percent the magnitudes are small. The results in the third and fourth columns of Table 2 indicate that earnings volatility is not associated with being observed to always be uninsured, and may even make such an observation less likely. The mean marginal effect for men is negative but not statistically significant, but for women, the effect is negative and significant. The magnitudes, however, are not very large; replacing one’s best earnings year with zero earnings, as above, would increase the probability of always being uninsured by less than

15

This counterfactual is not completely correct, as this change from their actual earnings record would also reduce one’s average predicted earnings during that four- or five-year period in one’s career. A proper counterfactual would increase the earnings for a second year during the same four- or five-year period to maintain the average predicted earnings at its true level, so that the change in is indeed “all else equal.” I will adjust this exercise in the next draft of the paper.

18

one-hundredth of a percentage point for men and just more than 0.03 percentage points for women. At first glance this negative result may seem surprising, but high earnings volatility implies that at least some years are above one’s long-term trend. There is little reason, then, that someone who has inconsistent earnings should also have persistent (but poor) health insurance coverage.16 Table 3 presents logit regression results where the sample is, instead, person-year observations. Here, the two dependent variables are indicators for whether the matched worker was ever uninsured during that year (first two columns), or uninsured all year (third and fourth columns).17 The results for uninsured ever during the year are similar to ever uninsured during the full sample, though smaller in magnitude, as expected. For men, a 6-percent increase in the earnings volatility measure (in line with the counterfactual above) is associated with a 0.11 percentage-point increase, statistically significant but small relative to the mean of 33 percent. The increase for women is even smaller, and not statistically different from zero. The results in Table 3 also indicate that men who have highly volatile career earnings are more likely to be uninsured all year, unlike the finding for whether a man is uninsured for their entire time in the sample, but the effect is quite small, only 0.04 percentage points for the proposed change in earnings. As with being always uninsured

16

The coefficients on the other covariates are in line with existing research. One is more likely to be uninsured if one is non-white (especially Hispanic), unmarried, young, or less educated. 17

The SIPP Gold Standard File currently eliminates any years where the SIPP began or ended in the middle of the calendar year, so if someone is uninsured all year, that means for all 12 calendar months, not just for all the months they were present in the SIPP during that calendar year.

19

over the full sample, women are less likely to be uninsured all year when career earnings are more volatile. The results are qualitatively similar for a Tobit regression of the number of months uninsured (Table 4). For a man, a 6-percent increase in earnings volatility is associated with a statistically significant increase in the number of months uninsured, but only by an average of 0.06 months, or about two extra days uninsured (if a “month uninsured” actually means that someone lacks coverage for the entire 31 days) over his two to four years in the sample. The mean marginal effect of earnings volatility on months uninsured in the full sample is positive but not statistically significant for women. For months uninsured in just the given year, the marginal effects are negative and significant for both genders, though the proposed increase in earnings leads only to three (men) or 4.5 (women) extra days uninsured all year. As a robustness check, I also re-estimated each model using ordinary least squares rather than logit or Tobit. The results were largely the same both in magnitude and significance. An exception was the regression of months uninsured in the given year for men (similar to Table 4, column 3); the mean marginal effect of earnings volatility was positive and significant, though very small. Attempting to discern the direction of causality. The results in Tables 2, 3, and 4 indicate that earnings volatility is associated with a statistically significant (but small in magnitude) increase in a male worker’s probability of experiencing an episode without health insurance. This does not imply causality, however; the results are consistent with

20

both career earnings volatility making one more likely to be uninsured and being uninsured making career earnings more volatile. It may be possible to use the data to get more information on the direction of causality. The SIPP Gold Standard File links earnings records from throughout one’s career, as late as 2006, regardless of when one was sampled by the SIPP, which means that some individuals were sampled relatively early in their careers, while others were sampled relatively late. It is therefore possible to calculate a new

for each person

just from the years of earnings before the individual entered the SIPP, and another just from the years after the individual entered the SIPP. If a history of volatile earnings over one’s career results in that person being uninsured, due to scarring from job losses leading to the inability to find a job offering health insurance, then there should be a positive correlation between earnings volatility in the years before one is observed in the SIPP and the probability that someone is observed as uninsured when he/she is sampled. If, on the other hand, being uninsured leads one to have a more volatile career, due to low unobserved ability or a low degree of risk aversion, then

calculated just from the post-SIPP years should be positively

correlated with being uninsured. It is also possible that both could be true, if the association between earnings volatility and being uninsured induces a negative feedback loop. The results in Table 5 suggest that, at most, only the former applies. The first two columns repeat the logit regressions from the first two columns of Table 2, where the dependent variable is an indicator for whether the matched individual is ever uninsured

21

during his/her time in the SIPP sample, but using a

calculated only from earnings

accumulated before entering the SIPP. The mean derivative of ln

) with respect to

the dependent variable is positive and statistically significant for men, though smaller in magnitude from the original regression, and negative and insignificant for women. This suggests that for men, volatile career earnings lead to an increased likelihood of being uninsured. Meanwhile, the regression that uses

calculated from just the post-SIPP years

indicates the exact opposite – being uninsured actually reduces earnings volatility later in one’s career. The coefficient for both genders is negative and statistically significant, but in both cases, the average marginal effect is extremely small. These results suggest that the arrow of causality points only from career earnings volatility to the increased probability of going with health insurance coverage at some point, but this conclusion is certainly too strong without more information. It is possible, and quite likely, that someone who is uninsured now is at greater risk of being uninsured in the future, and vice versa. To be able to assert confidently that earnings volatility causes future uninsurance, I would need to be able to observe that someone who has volatile earnings early in his/her career is no more likely during those years to be uninsured than someone with stable, on-trend earnings. Currently, no dataset allows for a long enough period of observation to make this case, though as time goes on, the Panel Survey of Income Dynamics, which began collecting health insurance coverage status in 1997, will have enough years of data to better answer this question.

22

Conclusion To my knowledge, this is the first paper to examine the relationship between long-run earnings volatility and short- and medium-run health insurance coverage. The results suggest that men who have more volatile career earnings are more likely to experience at least a short period without health insurance coverage. Women’s career earnings volatility is harder to measure without more information about the nature of periods away from the labor force, e.g., to take care of children or the elderly, but in some specifications of the model, there appears to be the same relationship between earnings stability and continuous health insurance coverage. These findings are subject to some important caveats, which need to be taken into account when drawing conclusions. While many of the estimates are strongly statistically significant, especially for men, the magnitudes of the effects are universally small; in a simple simulation exercise, I show that replacing one’s highest earning year with zero earnings increases the probability of ever being uninsured over a two- to four-year period by only 0.16 percentage points, a very modest increase over a mean of 45 percent. The estimates are also not meant to be interpreted as causal. The research question may imply that having volatile career earnings could lead to one becoming uninsured, but being uninsured, or being vulnerable to experiencing an uninsured episode, could, in addition (or instead), reduce one’s future earning potential. My estimates suggest that the former is more likely than the latter, as there is a positive correlation between one’s earnings volatility before entering the SIPP sample and later being uninsured, but no correlation (or even negative) between health insurance status in

23

the SIPP and subsequent earnings volatility. Still, without knowing one’s health insurance status in non-SIPP years, this relationship may arise for a different reason, such as serial correlation in health insurance status. Taking these results at face value, there appears to be a relationship between an inconsistent career earnings profile and the consistency of health insurance coverage, and though the pathway is not clear, early career earnings volatility is associated with lacking health insurance mid-career. These findings are consistent with other work showing that job losses lead to enduring “scars” that may affect future economic success. Though much has been done in recent years to help workers in down years, including COBRA health insurance coverage and unemployment insurance benefit extensions, policymakers should consider ways to strengthen the safety net, including increasing the funding for job training and adult education programs or encouraging temporary or contract work as a bridge, so that temporary setbacks in one’s career do not lead to lasting negative outcomes.

References Abowd, John M., Martha Stinson and Gary Benedetto, 2006. “Final Report to the Social Security Administration on the SIPP/SSA/IRS Public Use File Project,” http://www.census.gov/sipp/SSAfinal.pdf Cameron, Stephen and Joseph Tracy, 1998. “Earnings Instability in the United States: An Examination Using Matched-CPS Data,” Federal Reserve Bank of New York working paper, http://www.newyorkfed.org/research/economists/tracy/earnings_variability.pdf. Czajka, John L. and Cara Olsen. 2000. “The Effects of Trigger Events on Changes in Children’s Health Insurance Coverage.” Princeton, NJ: Mathematica Policy Research, Inc.

24

Dahl, Molly, Thomas DeLeire, and Jonathan A. Schwabish, 2008. Recent Trends in the Variability of Individual Earnings and Household Income. Congressional Budget Office Paper, http://www.cbo.gov/ftpdocs/95xx/doc9507/06-30-Variability.pdf. DeNavas-Walt, Carmen, Bernadette D. Proctor, and Jessica C. Smith, 2009. “Income, Poverty, and Health Insurance Coverage in the United States: 2008,” U.S. Census Bureau Report P60-236. Dynan, Karen E., Douglas W. Elmendorf, and Daniel E. Sichel, 2008. “The Evolution of Household Income Volatility.” Federal Reserve Board, Finance and Economics Discussion Series, Working Paper 2007-61, http://www.federalreserve.gov/Pubs/feds/2007/200761/200761pap.pdf. Ellwood, David T., 1982. "Teenage Unemployment: Permanent Scars or Temporary Blem-ishes?" in R. B. Freeman and D. A. Wise, eds., The Youth Labor Market: Its Nature, Causes and Consequences. Chicago: University of Chicago Press, 34985. Fallick, Bruce C., 1996. “A Review of the Recent Empirical Literature on Displaced Workers,” Industrial and Labor Relations Review 50(1): 5-16. Goldin, Claudia, 2006. “The Quiet Revolution that Transformed Women’s Employment, Education, and Family,” The American Economic Review 96(2): 1-21. Gottschalk, Peter and Robert Moffitt, 1994. “The Growth of Earnings Instability in the U.S. Labor Market,” Brookings Papers on Economic Activity, 1994(2): 217-272. Gottschalk, Peter and Robert Moffitt, 2008. “Trends in the Transitory Variance of Male Earnings in the U.S., 1970-2004,” mimeo. Gruber, Jonathan and Brigitte Madrian, 1997. “Employment Separation and Health Insurance,” Journal of Public Economics, 66:349-382 Haider, Steven J., 2001. “Earnings Instability and Earnings Inequality of Males in the United States: 1967-1991,” Journal of Labor Economics 19(4): 799-836. Kuttner, Hanns and Matthew S. Rutledge, 2007. “Higher Income and Uninsured: Common or Rare?” Health Affairs 26(6): 1745-1752. Jacobson, Louis, Robert LaLonde, and Daniel Sullivan, 1993. The Costs of Worker Displacement. Kalamazoo, MI: The W.E. Upjohn Institute for Employment Research. Moffitt, Robert A. and Peter Gottschalk, 1995. “Trends in the Covariance Structure of Earnings in the U.S., 1969-1987,” mimeo.

25

Moffitt, Robert A., and Peter Gottschalk, 2002. “Trends in the Transitory Variance of Earnings in the United States,” Economic Journal 112: C68-73. Ruhm, Christopher, 1991. “Are Workers Permanently Scarred by Job Displacement?” The American Economic Review, 81(1):319-324. Shin, Donggyun and Gary Solon, 2008. “Trends in Men’s Earnings Volatility: What does the Panel Study of Income Dynamics Show?” NBER Working Paper 14075. Short, Pamela Farley and Deborah R. Graefe, 2003. “Battery-Powered Health Insurance? Stability in Coverage of the Uninsured,” Health Affairs 22(6): 244-255. Simon, Kosali, and Mathis Schroeder, 2006. “The Effect of Involuntary Job Displacement on Health Insurance,” draft presented at ERIU Conference, September 7-8, 2006. Stevens, Ann Huff, 1997. “Persistent Effects of Job Displacement: The Importance of Multiple Job Losses,” Journal of Labor Economics 15(1): 165-188. Stevens, Ann Huff, 2001. “Changes in Earnings Instability and Job Loss,” Industrial and Labor Relations Review, 55(1):60-78.

Observed values

Age

Higher Volatility, but Lower Variance

Real Earnings

Fitted values

Age

Lower Volatility, but Higher Variance

Real Earnings

Figure 1. Hypothetical Age-Earnings Profiles, by Variance

Lower Volatility (Lower SSR)

Real Earnings

Observed values

Fitted values

Age

Higher Volatility (Higher SSR)

Real Earnings

Figure 2. Hypothetical Age-Earnings Profiles, by Sum of Squared Residuals

Age

28 Table 1. Summary Statistics Male Ever Uninsured (0/1) Always Uninsured (0/1) Months Uninsured in Sample Months in Sample Months Uninsured in Year Uninsured Ever In Year (0/1) Uninsured All Year (0/1) SSR (Billions of dollars squared) ln(SSR) (Log of dollars squared) Age

Female Unwt Wt 0.66 0.43 (0.47) (0.50) 0.16 0.11 (0.36) (0.32) 12.0 7.4 (13.8) (11.4) 31.2 30.8 (9.9) (6.0) 4.6 2.7 (5.2) (4.6) 0.53 0.32 (0.50) (0.47) 0.27 0.15 (0.44) (0.36) 74.1 288.4 (23697) (49458) 19.98 19.97 (1.73) (1.70) 38.6 43.0 (13.0) (11.2)

Unwt 0.67 (0.47) 0.14 (0.35) 12.0 (13.7) 31.1 (9.9) 4.6 (5.2) 0.53 (0.50) 0.26 (0.44) 171.7 (26564) 20.84 (1.83) 38.3 (13.0)

Wt 0.45 (0.50) 0.11 (0.31) 7.5 (11.3) 30.8 (5.9) 2.8 (4.6) 0.33 (0.47) 0.15 (0.36) 99.9 (11389) 20.75 (1.81) 43.0 (11.1)

0.10 (0.30) 0.05 (0.21) 0.09 (0.28)

0.10 (0.30) 0.06 (0.24) 0.11 (0.32)

0.13 (0.33) 0.05 (0.21) 0.09 (0.28)

0.12 (0.33) 0.06 (0.23) 0.10 (0.30)

0.14 (0.35) 0.26 (0.44) 0.24 (0.43) 0.13 (0.34)

0.10 (0.30) 0.26 (0.44) 0.32 (0.47) 0.18 (0.39)

0.14 (0.34) 0.28 (0.45) 0.27 (0.44) 0.13 (0.34)

0.09 (0.28) 0.26 (0.44) 0.34 (0.48) 0.19 (0.39)

0.08 (0.28) 0.16 (0.36) 0.61 (0.49) 126,450 327,831

0.12 (0.32) 0.02 (0.15) 0.68 (0.47) 105,684 242,913

0.08 (0.27) 0.14 (0.34) 0.59 (0.49) 134,722 350,628

0.11 (0.31) 0.02 (0.14) 0.66 (0.47) 114,701 265,434

Race Black (0/1) Other Race (0/1) Hispanic (0/1) Education (0/1) Less than HS HS Grad Only Some College College Nativity Foreign-Born (0/1) N/A (0/1) Married (0/1) Unique Persons Person-Years Note: Standard Deviations in parentheses.

29 Table 2. Logit Regression Results, One Observation Per Person Dependent Variable

Mean Derivative ln(SSR)

Logit Coefficients ln(SSR) Age Age Squared Black Other Race Hispanic Less Than HS HS Grad Only Some College College Foreign-Born N/A Nativity Married Constant

Number of Persons R2

Ever Uninsured Men Women (1) (2)

Always Uninsured Men Women (3) (4)

0.0273 *** (0.0029)

0.0073 *** (0.0028)

-0.0014 (0.0019)

-0.0061 *** (0.0015)

0.1515 (0.0165) -0.0716 (0.0397) 0.0022 (0.0005) 0.3896 (0.0682) 0.0645 (0.0860) 0.5276 (0.0739) 0.6846 (0.0939) 0.0427 (0.0727) -0.2230 (0.0700) -0.5832 (0.0741) 0.0253 (0.0788) 2.4734 (0.1439) -0.6462 (0.0445) -5.2877 (1.0499)

0.0391 (0.0148) -0.0228 (0.0366) 0.0014 (0.0004) 0.7781 (0.0578) 0.2846 (0.0829) 0.6104 (0.0690) 0.6989 (0.0911) -0.1462 (0.0691) -0.3024 (0.0661) -0.7741 (0.0724) -0.0034 (0.0693) 2.2399 (0.1379) -0.9850 (0.0388) -4.2813 (0.9185)

-0.0203 (0.0277) -0.1234 (0.0618) 0.0016 (0.0007) -0.0185 (0.0909) -0.0757 (0.1311) 0.4021 (0.0931) 0.9668 (0.1398) 0.4689 (0.1323) -0.0319 (0.1324) -0.7145 (0.1697) 0.2127 (0.1070) 0.3321 (0.1571) -0.7497 (0.0654) 3.4335 (1.5246)

-0.0833 (0.0205) 0.0268 (0.0546) -0.0001 (0.0006) 0.3590 (0.0771) 0.0171 (0.1243) 0.5631 (0.0832) 0.9285 (0.1332) 0.2472 (0.1227) -0.0175 (0.1227) -1.0252 (0.1705) -0.0474 (0.0920) 0.4966 (0.1522) -1.5123 (0.0599) 0.0556 (1.3417)

82445 0.2193

*** * *** ***

*** ***

*** ***

*** *** ***

89790 0.1913

***

*** *** *** *** *** ** *** ***

*** *** ***

82394 0.3184

** **

*** *** ***

*** ** ** *** **

***

***

*** *** **

***

*** ***

89733 0.3224

Note: ln(SSR) is the natural log of the sum of squared residuals from the individual-byindividual OLS regression of log earnings on a cubic function of age. Sample restricted to individuals with 10 or more years of positive earnings. Regressions reported above use SIPP weights and include average predicted earnings over a four- to five-year age groups and SSA-IRS entry year fixed effects. Heteroskedasticity-robust standard errors in parentheses. *** - Significantly different from zero at the 99 percent confidence level ** - 95 percent confidence level * - 90 percent confidence level

30 Table 3. Logit Regression Results, One Observation Per Person-Year Dependent Variable

Mean Derivative ln(SSR)

Logit Coefficients ln(SSR) Age Age Squared Black Other Race Hispanic Less Than HS HS Grad Only Some College College Foreign-Born N/A Nativity Married Constant

Number of Person-Years Number of Persons R2

Uninsured Ever in Year Men Women (1) (2)

Uninsured All Year Men Women (3) (4)

0.0194 *** (0.0019)

0.0023 (0.0017)

0.0063 *** (0.0016)

-0.0049 *** (0.0012)

0.1229 (0.0120) -0.1687 (0.0302) 0.0023 (0.0004) 0.2561 (0.0450) 0.0419 (0.0594) 0.4675 (0.0498) 0.6824 (0.0637) 0.0726 (0.0500) -0.2271 (0.0475) -0.6951 (0.0527) -0.0030 (0.0548) 1.5241 (0.0861) -0.7064 (0.0302) 1.1552 (0.7674)

0.0143 (0.0107) -0.1318 (0.0283) 0.0018 (0.0003) 0.6068 (0.0391) 0.2511 (0.0577) 0.5759 (0.0476) 0.7842 (0.0628) -0.0494 (0.0469) -0.2603 (0.0441) -0.7826 (0.0500) -0.0214 (0.0492) 1.5210 (0.0876) -1.1867 (0.0278) 1.2574 (0.6911)

0.0701 (0.0173) -0.0148 (0.0401) 0.0008 (0.0005) -0.0209 (0.0576) -0.0170 (0.0808) 0.4203 (0.0598) 1.1654 (0.0851) 0.6134 (0.0762) 0.1651 (0.0752) -0.5733 (0.0963) 0.0268 (0.0715) 0.4740 (0.1065) -0.8443 (0.0409) -1.9939 (0.9808)

-0.0536 (0.0137) 0.0689 (0.0364) -0.0001 (0.0004) 0.4184 (0.0497) 0.1215 (0.0753) 0.5983 (0.0574) 1.1348 (0.0820) 0.4486 (0.0718) 0.1702 (0.0701) -0.7609 (0.0940) -0.0409 (0.0622) 0.4512 (0.1100) -1.5680 (0.0382) -3.0668 (0.8815)

242913 112712 0.2344

*** *** *** ***

*** ***

*** ***

*** ***

265434 121536 0.2155

*** *** *** *** *** ***

*** ***

*** *** *

242913 112712 0.3163

***

*

*** *** *** ** ***

*** *** **

*** *

***

*** *** *** ** ***

*** *** ***

265434 121536 0.3193

Note: ln(SSR) is the natural log of the sum of squared residuals from the individual-byindividual OLS regression of log earnings on a cubic function of age. Sample restricted to individuals with 10 or more years of positive earnings. Regressions reported above use SIPP weights and include average predicted earnings over a four- to five-year age groups and SSA-IRS entry year fixed effects. Heteroskedasticity-robust, clustered (by individual) standard errors in parentheses. *** - Significantly different from zero at the 99 percent confidence level ** - 95 percent confidence level * - 90 percent confidence level

31 Table 4. Tobit Regression Results Dependent Variable

Mean Derivative ln(SSR)

Logit Coefficients ln(SSR) Age Age Squared Black Other Race Hispanic Less Than HS HS Grad Only Some College College Foreign-Born N/A Nativity Married Constant

Months Uninsured in Sample Men Women (1) (2)

Months Uninsured in Year Men Women (3) (4)

1.0563 *** (0.1321)

0.0170 (0.1202)

-1.6673 *** (0.0976)

-2.4051 *** (0.1063)

1.0563 (0.1321) 0.5868 (0.2926) 0.0129 (0.0034) 2.1923 (0.4799) 0.1677 (0.6362) 4.4014 (0.5093) 5.5386 (0.6717) 0.7661 (0.5797) -2.6439 (0.5577) -6.6054 (0.5966) 0.2307 (0.5825) 11.5441 (0.6311) -6.3788 (0.3417) -86.5842 (7.5992)

0.0170 (0.1202) 1.0509 (0.2958) 0.0075 (0.0034) 6.0828 (0.4374) 2.5158 (0.6660) 5.5964 (0.5187) 6.8941 (0.6963) -0.5009 (0.5880) -2.4206 (0.5606) -8.1392 (0.6243) -0.0812 (0.5527) 11.4435 (0.6609) -10.9017 (0.3223) -87.2023 (7.3125)

-1.6673 (0.0976) -0.6983 (0.2770) 0.0131 (0.0031) 4.4801 (0.4289) 2.7145 (0.5837) 5.4189 (0.4944) 13.5910 (0.5749) 3.9983 (0.4280) -0.6100 (0.4033) -6.1243 (0.4633) 2.7184 (0.4924) 12.7106 (0.6451) -10.6960 (0.3090) 22.1136 (6.7423)

-2.4051 (0.1063) -0.5518 (0.2725) 0.0116 (0.0031) 5.8392 (0.3798) 3.0245 (0.5650) 6.2417 (0.4852) 15.6429 (0.6026) 4.3352 (0.4252) 1.0011 (0.3942) -6.2664 (0.4556) 1.7136 (0.4700) 12.9941 (0.6694) -11.2918 (0.2903) 27.6405 (6.6033)

*** ** *** ***

*** ***

*** ***

*** *** ***

Number of Persons Number of Person-Years R2

82445

89790

0.0929

Censored Observations Zero Months Uncensored Maximum (or 12) Months

41788 40657 0

*** ** *** *** *** ***

*** ***

*** *** ***

*** ** *** *** *** *** *** ***

*** *** *** *** ***

0.0808

112712 242913 0.0796

121536 265434 0.095

45661 44129 0

152638 52786 37489

166315 53278 45841

*** ** *** *** *** *** *** *** ** *** *** *** *** ***

Note: ln(SSR) is the natural log of the sum of squared residuals from the individual-byindividual OLS regression of log earnings on a cubic function of age. Sample restricted to individuals with 10 or more years of positive earnings. Regressions reported above use SIPP weights and include average predicted earnings over a four- to five-year age groups and SSA-IRS entry year fixed effects. Heteroskedasticity-robust, clustered (by individual) standard errors in parentheses. *** - Significantly different from zero at the 99 percent confidence level ** - 95 percent confidence level * - 90 percent confidence level

32 Table 5. Logit Regression Results, Pre-SIPP vs. Post-SIPP Dependent Variable

Mean Derivative ln(SSR) - Pre-SIPP Only

Ever Uninsured Men Women (1) (2) 0.0174 *** (0.0031)

Ever Uninsured Men Women (3) (4)

-0.0016 (0.0029)

ln(SSR) - Post-SIPP Only

Logit Coefficients ln(SSR) - Pre-SIPP Only

0.0950 *** (0.0169)

Age Squared Black Other Race Hispanic Less Than HS HS Grad Only Some College College Foreign-Born N/A Nativity Married Constant

Number of Persons R2

0.0252 (0.0531) 0.0018 (0.0006) 0.4146 (0.0696) 0.0859 (0.0876) 0.5370 (0.0765) 0.7996 (0.0927) 0.0995 (0.0720) -0.1723 (0.0695) -0.5523 (0.0744) -0.0686 (0.0767) 2.4772 (0.1479) -0.6731 (0.0453) -13.3008 (1.3765) 65150 0.2035

*** ***

*** ***

** ***

*** *** ***

-0.0068 *** (0.0012)

-0.0604 (0.0068) 0.4417 (0.0362) -0.0047 (0.0006) 0.5665 (0.0452) 0.3829 (0.0678) 0.6422 (0.0499) 0.0615 (0.0520) -0.3848 (0.0476) -0.7458 (0.0483) -1.1704 (0.0544) 0.0844 (0.0573) 0.9195 (0.0485) -0.6058 (0.0313) -10.9290 (0.4957)

-0.0358 (0.0065) 0.5255 (0.0348) -0.0062 (0.0005) 0.7613 (0.0382) 0.3394 (0.0643) 0.7042 (0.0458) 0.1157 (0.0539) -0.4894 (0.0485) -0.8147 (0.0494) -1.3453 (0.0571) 0.1221 (0.0524) 0.9157 (0.0514) -0.7396 (0.0286) -10.9173 (0.4819)

-0.0084 (0.0152)

ln(SSR) - Post-SIPP Only Age

-0.0114 *** (0.0013)

0.0535 (0.0454) 0.0009 (0.0005) 0.7878 (0.0592) 0.3042 (0.0844) 0.5948 (0.0720) 0.6902 (0.0922) -0.1630 (0.0693) -0.2898 (0.0659) -0.7751 (0.0731) 0.0384 (0.0686) 2.3050 (0.1436) -0.9661 (0.0398) -8.0552 (1.9598) 71192 0.1814

* *** *** *** *** ** *** ***

*** *** ***

41475 0.1836

*** *** *** *** *** ***

*** *** ***

*** *** ***

*** *** *** *** *** *** ** *** *** *** ** *** *** ***

43465 0.1757

Note: In columns 1 and 2, ln(SSR) is the natural log of the sum of squared residuals from the individual-by-individual OLS regression of log earnings in the years before the individual enters the SIPP on a cubic function of age; in columns 3 and 4, the dependent variable in the regression used to calculate ln(SSR) is log earnings in the years after leaving the SIPP sample. Sample limited to individuals with ten years of positive earnings pre- or post-SIPP, respectively. Regressions reported above use SIPP weights and include average predicted earnings over a four- to five-year age groups and SSA-IRS entry year fixed effects. Heteroskedasticity-robust standard errors in parentheses. *** - Significantly different from zero at the 99 percent confidence level ** - 95 percent confidence level * - 90 percent confidence level