Thus far, we have examined subgroup results one at a time. However, many of these subgroups are correlated with each other. For example, we have seen that less disadvantaged workers and those in higher-quality jobs tend to have more successful outcomes than other workers. However, better-off workers are more likely than those worse off to be in high-quality jobs. Thus, an important question is whether labor market success is due more to worker characteristics or initial job characteristics.
We isolated subgroup effects from others using multivariate regression methods. We estimated regression models for the four outcome measures used in the univariate subgroup analysis. In the main text, we present findings for the most important outcome measure: the percentage of months workers spent in medium- and high-wage jobs (Table IV.6). The results for the other three outcomes are presented in Table C.4 and are qualitatively similar to those presented in the text (although as discussed, in general, there was less variation in the total time workers spent employed than in the time workers spent in higher-wage jobs). We present "regression-adjusted" means for each subgroup level and indicate whether the difference between the regression-adjusted means for each subgroup and the "left-out" subgroup is statistically significant at the five percent significance level.(35)
We present estimates from three models for both males and females. The first model includes demographic variables only (that is, explanatory variables defined by individual, household, and area characteristics; model (1) in Table IV.6). The second model includes demographic variables as well as prepanel work experience variables from the wave 1 topical module [model (2)]. The third model includes demographic variables and initial job-related variables [model (3)]. In Table C.4, we present the model (3) results for the additional employment-related outcome measures only.
a. Models Including Demographic Variables Only
The regression-adjusted differences in labor market outcomes across subgroups defined by individual and household characteristics are similar to the univariate findings described above (Table IV.6). In particular, among our sample of low-wage workers, teenagers and older workers, African Americans and Hispanics, those with low levels of education, and those with health problems spent less time in medium- and high-wage jobs than their counterparts, and many of these differences are statistically significant at the ten percent level. There is also some evidence that those in higher-income households and males with children had better labor market outcomes than other workers, but these differences are not statistically significant. Thus, adjusting for the correlation among the demographic variables does not materially influence the subgroup findings.
|Explanatory Variables||Regression-Adjusted Means for Models with Demographic and Other Denoted Explanatory Variables|
|No Other Variables (1)||Pre-Panel Work History Variables (2)||Initial Job Variables (3)||No Other Variables (1)||Pre-Panel Work History Variables (2)||Initial Job Variables (3)|
|Younger than 20(a)||20||27||23||12||15||12|
|20 to 29||29**||30||30*||20**||20||20**|
|30 to 39||33***||32||33**||18*||17||18*|
|40 to 49||33**||29||30||19*||18||19*|
|50 to 59||22||17||16||18||17||18|
|60 or older||18||12*||12||17||16||14|
|White and other non-Hispanic(a)||31||31||31||20||19||19|
|Less than high school/GED(a)||23||24||26||12||12||14|
|College graduate or more||32*||33*||30||26***||25***||23***|
|Has a Health Limitation|
|Work Experience Prior to the Panel Period|
|Ever Worked for Six Straight Months|
|Number of Years Ever Worked Six Straight Months|
|Less than 5(a)||27||14|
|5 to 10||31||19*|
|10 to 20||26||22***|
|More than 20||32||20*|
|Usually Worked at Least 35 Hours Per Week When Working|
|Single adults with children(a)||30||31||30||18||18||19|
|Married couples with children||32||32||30||17||17||17|
|Married couples without children||26||25||27||16||16||16|
|Other adults without children||25||26||26||23*||24*||23|
|Household Income as a Percentage of the Poverty Level|
|100 percent or less(a)||26||24||29||17||17||20|
|101 to 200 percent||27||28||28||17||17||18|
|More than 200 percent||30||31||28||19||19||18|
|Received Public Assistance in the Past Year|
|Region of Residence|
|Lives in a Metropolitan Area|
|20th Percentile of the Hourly Wage Distribution in State|
|$250 or less(a)||27||27||27||16||16||17|
|$251 to $269||35*||34*||33||19||19||19|
|$270 or more||27||27||28||20||20||19|
|Percentage of State Population Residing in Metropolitan Areas|
|72 or less(a)||28||27||28||21||21||21|
|73 to 84||31||31||31||17*||17||17**|
|85 or more||27||27||27||16||15*||16*|
|Poverty Rate in State|
|Less than 10 percent(a)||29||29||27||15||15||14|
|10 to 12 percent||31||30||30||19||19||20**|
|More than 12 percent||26||27||28||19||19||20*|
|Unemployment Rate in State|
|6 percent or less(a)||27||28||29||13||14||14|
|More than 6 percent||29||29||28||20*||20*||20*|
|Change in Unemployment Rate in State of Residence Between 1996 and 1999 (Percentage Points)|
|-2 percentage points or less(a)||28||28||27||18||18||19|
|-1 to -2||28||28||28||19||19||19|
|More than -1||30||30||30||16||16||16|
|Initial Job Characteristics|
|Less than $5.00(a)||19||11|
|$5.00 to $5.99||19||14|
|$6.00 to $6.99||37***||24***|
|$7.00 to $7.50||40***||34***|
|Usual Hours Worked per Week|
|1 to 19(a)||24||13|
|20 to 34||25||20***|
|35 to 40||29||19**|
|More than 40||32||18|
|Has More than One Job or Business|
|Owns Business (Self-Employed)|
|Health Insurance Coverage(b)|
|Regression R(2) Value||.12||.15||.27||.14||.15||.27|
|Source: 1996 SIPP longitudinal and wave 1 topical module files using the entry cohort sample of workers who started low-wage jobs within six months after the start of the panel period. All workers were followed for 42 months after job start.
Note: All figures are weighted using the 1996 calendar year weight, and standard errors account for design effects due to weighting and clustering.
a. Denotes the omitted explanatory variable in the regression model.
b. These figures pertain to health insurance coverage from all sources, including coverage through the employer as well as from other sources. We used this variable instead of the employer-based health insurance coverage variable, because data on overall health insurance coverage is available monthly, whereas the employer-based coverage variable pertains only to jobs in progress at the time of the interview. Thus, the employer-based health insurance variable could not always be linked to the job under investigation, which led to a significant number of missing values. However, the subsets of health insurance variables overlap considerably: the source of health insurance coverage was the employer for 80 percent of those with any coverage.
* Difference between the variable mean and the mean of the omitted explanatory variable is significantly different from zero at the .10 level, two-tailed test.
** Difference between the variable mean and the mean of the omitted explanatory variable is significantly different from zero at the .05 level, two-tailed test.
*** Difference between the variable mean and the mean of the omitted explanatory variable is significantly different from zero at the .01 level, two-tailed test
The explanatory variables measuring area characteristics have little predictive power in the regression models (Table IV.6). Those in metropolitan areas tended to have slightly better outcomes than those in other areas, and there is some evidence that females in the northeast region had more positive labor market experiences than females in other regions (although this result does not hold for males). However, in general, the state hourly wage and state unemployment measures are not statistically significant, and the parameter estimates are not in the expected direction. These weak results are somewhat surprising, because the area characteristics are intended to capture the economic conditions faced by sample members. Hence, we expected more positive labor market outcomes for those residing in areas with a higher demand for labor than those in other areas. A possible explanation for the weak findings is that the area characteristics are measured at the aggregated state level, so they might not accurately reflect demand conditions faced by the workers in their local areas.
The regression R(2) value from model (1) is about .13 for both males and females. Thus, although the demographic variables explain about 13 percent of the variance in the amount of time workers spent in the higher-wage labor market, substantial residual factors remain that account for differences across workers. Stated differently, there is substantial diversity in labor market outcomes among members within the subgroups under investigation.
b. Models Including Demographic and Prepanel Work Experience Measures
Work experience matters to some extent. All else equal, sample members with more than five years of labor market experience typically spent slightly more time in higher-wage jobs than those with less work experience, and this result holds for both men and women (Table IV.6). Furthermore, males who typically worked full-time while employed had more wage progression, on average, than part-time male workers, and these differences are statistically significant.
Interestingly, differences in mean outcomes across age groups diminish somewhat when the prepanel work experience variables are included in the models. Thus, our initial findings across age groups can be explained by the higher levels of work experience among older workers, which gave them more job-related skills and made it easier for them to find higher-paying jobs.
c. Models Including Demographic and Initial Job-Related Variables
In general, the inclusion of the job-related variables leads to slightly smaller differences across the demographic subgroups than those presented above (model (3) in Table IV.6). (36) For example, when the initial job characteristics are included in the model, the Hispanic and education effects for males become statistically insignificant. The effects become slightly smaller due to the fact that less disadvantaged workers tend to get better jobs, even in the low-wage worker population.
The multivariate findings support our conclusions from the univariate analysis that job quality matters (Table IV.6 and Table C.4). Low-wage workers who had higher starting wages, worked more hours, and had available health benefits spent more time, on average, in higher-wage jobs than those in lower-quality jobs. Most of these differences are statistically significant at the 5 percent significance level. However, the regression-adjusted means across the job-related subgroups are slightly smaller than the univariate means because of the correlation between the demographic and job-related variables and the correlation among the job-related variables. For example, the regression results no longer suggest that males in professional and sales occupations and females in professional and clerical occupations experienced more wage progression than other workers. The occupational effects, however, more closely resemble those from the univariate analysis if the demographic variables are excluded from the models, or if the demographic variables are included but other job-related variables are excluded (not shown).
Interestingly, those who had more than one job at the start of the low-wage job spell had slightly better outcomes than those who did not, perhaps capturing differences in the motivation to work and succeed across the two groups of workers (Table IV.6 and Table C.4). In addition, self-employed workers typically spent substantially more time than jobholders in the medium- and high-wage labor market sectors, and these differences are statistically significant for males.
Finally, the inclusion of both the job and demographic characteristics yields a model R(2) value of .27 for both males and females (Table IV.6). Thus, we find again that there remain substantial residual factors that account for differences in labor market success across low-wage workers, even after controlling for a large number of demographic and job-related factors. In sum, although we have identified some important differences in medium-term labor market outcomes across key subgroups of the low-wage worker population, there are clearly other important factors that we could not identify using the SIPP data.
(27) In the previous chapter, we focused our discussion on the comparison of the characteristics of low-wage workers to those of all workers. However, in this chapter, we focus our discussion on the comparison of low-wage workers to medium- and high-wage workers in order to assess the extent to which the labor market experiences (such as total time employed) of low-wage workers differ from those of higher-wage workers.
(28) An individual was defined to have been employed in a month if he or she was employed for at least one week during the month.
(29) We find similar results for the percentage weeks worked (Table IV.3), because most individuals were employed for all weeks during the month. Thus, for simplicity, in this chapter, we focus on the months measure.
(30) The hours figures for medium-wage jobs include the zero hours worked by those who never held medium-wage jobs.
(31) We did not examine subgroup differences across the three male and three female low-worker typologies presented in the previous chapter, because the much smaller sample size used in the overall employment analysis yielded unstable clusters that were difficult to interpret.
(32) The standard errors of the estimates account also for design effects in the SIPP data due to clustering.
(33) We measured these indicators using information on the state in which the worker lived at the beginning and end of the follow-up period.
(34) These figures pertain to health insurance coverage from all sources, including coverage through the employer as well as from other sources. We used this variable instead of the employer-based health insurance coverage variable, because data on overall health insurance coverage is available monthly, whereas the employer-based coverage variable pertains only to jobs in progress at the time of the interview. Thus, the employer-based health insurance variable could not always be linked to the job under investigation, which led to a significant number of missing values. However, the subsets of health insurance variables overlap considerably: the source of health insurance coverage was the employer for 80 percent of those with any coverage.
(35) The regression-adjusted mean for Hispanics, for example, was the average predicted value from the regression model, where the value of 1 was inserted for the Hispanic dummy variable for all individuals but where the other explanatory variables were calculated at their actual values. The regression-adjusted means for other explanatory variables were constructed in an analogous way.
(36) We are aware that the job variables are likely to be correlated with the error term in the regression models (that is, that the job variables are likely to be endogenous), which could lead to biased coefficient estimates on all the explanatory variables. Thus, we do not view our parameter estimates as "structural" relationships between the explanatory and dependent variables. Rather, our goal is to identify broad associations between subgroup variables and labor market outcomes.