Analytic Results
The analyses have used logistical regressions and ordinary least squares regressions to examine whether persons with alcohol and/or drug disorders differ from the general population in terms of the following:
- Employment rates; and
- Earnings per hour (monthly earnings divided by hours worked).
Analyses have been performed separately for males and females. It is well established in the labor market literature (and the findings reported here support) that males and females have fundamentally different experiences with respect to labor force participation, employment, wage rates, and earnings, which warrant separate analyses.
This study has also examined "reduced model" as well as "full model" effects. Reduced model estimates are those in which only basic demographic factors have been included as control variables to estimate the effect of alcohol and drug disorders as well as major depression. Full model estimates include additional variables as control factors that are generally considered measures of "human capital" (specifically educational attainment, choice of a white-collar profession, and marital status) and that strongly predict outcomes such as labor force participation, employment, and wage rates.
Estimated effects generated using both reduced models and full models are presented in order to demonstrate the apparent correlation between alcohol and drug disorders and the human capital variables. Effects estimated using reduced models are consistently greater than effects estimated with full models. This indicates that individuals with alcohol and drug disorders tend to have lower levels of educational attainment and/or selection into skilled professions. Because a key mechanism through which alcohol and drug disorders may achieve negative effects on labor market success is adverse effects on human capital (e.g., educational attainment and choice of occupation), it is justified to estimate alcohol and drug disorder effects using the reduced models (Mullahy and Sindelar 1989).
Also, regression analyses have been performed that include only a single measure of alcohol or drug disorders (summarized in table 5.10) and that include both major disorder variables (summarized in table 5.11). These separate sets of regressions have used the same observations. The single- and multiple-disorder regressions have been performed in order to identify effects on employment and the level of earnings among people with alcohol or drug dependence, ignoring the co-occurrence of the other disorders. In developing cost estimates - and in estimating disorder-specific effects - use is only made of the logistic and ordinary least squares regressions that simultaneously include alcohol and drug disorders as well as depression in order to separate their effects. Appendix B includes additional information regarding regressions that were completed.
Single-Disorder Impact Estimates
Results from 24 separate logistic and ordinary least squares (OLS) regressions are summarized in table 5.10. Each "cell" of coefficients in the table represents a separate regression. In the following sections "alcohol dependence" and "drug dependence" means dependence at any time in the respondents' lifetimes (i.e., "ever"). For style purposes, the rest of this chapter will drop the modifier "ever."
For males, alcohol dependence appears to not be a significant predictor of an individual's employment status (logistical regression, with employed = 1 and unemployed or not in the labor force = 0). The values on the left of table 5.10 are coefficients from logistical regressions. Statistically significant coefficients are indicated with asterisks.
However, when alcohol-dependent males are divided into "early drinkers" (first drink before reaching age 15) and "late drinkers" (first drink at age 15 or older), it is seen that alcohol dependence is strongly correlated with lower than expected employment for early-drinking males. Later drinking is not negatively associated and even may be positively associated. The reduced model yields larger negative impacts from dependence with early drinking than does the full model.
Earnings did appear to be negatively affected by alcohol dependence, both in association with early and later drinking initiation (the right of table 5.10). However, early initiation with dependence clearly has a stronger impact than does late initiation with dependence.
Drug dependence (ever in one's lifetime) is not significantly associated with employment status for males in either the reduced or the full model. Drug dependence does have a negative and significant association with reduced wage rates and earnings, and the estimated impact is lower in the full model than in the reduced model, as predicted by theory.
Results for females are equivocal. This is consistent with many prior studies of alcohol and drug disorders and female labor market success. Although alcohol dependence has a mild positive, but statistically insignificant, correlation with employment status and wage rates, females with dependence and early drinking are less likely to be employed than are their counterparts with no history of alcohol dependence. However, the estimate is of only marginal statistical significance. None of the estimates of the effects of alcohol dependence on females' wage rates approached statistical significance.
Drug dependence is a strong and statistically significant predictor of lower employment among females. The magnitude of the impact is slightly less than for alcohol dependence with early drinking. However, it is more statistically significant, probably because there are almost twice as many women with drug dependence as with alcohol dependence combined with early drinking. Female wage rates do not appear to have a statistically significant association with drug dependence.
Multiple-Disorder Effect Estimates
Because of frequent co-occurrence of alcohol and drug use and depressive disorders, it is important to simultaneously estimate the respective efects. For example, estimates of the impact of drug dependence alone may be inflated because many of these individuals also have alcohol use disorders. For this reason, the single-disorder estimates presented immediately above may overstate the effects of these respective disorders. Table 5.11 presents the results of the multiple-disorder impact estimates, where variables representing alcohol and drug use disorders as well as depressive disorders are simultaneously included in the regressions along with the selected control variables (those in either the reduced model or the full model). As expected, the estimates in table 5.11 are generally smaller in magnitude and less likely to be statistically significant than the impacts reported in table 5.10.
The first finding (regression 1 in table 5.11) is that males with alcohol dependence and early drinking are significantly less likely to be employed than males otherwise having the same demographic characteristics - even the same drug use and depressive disorders. This effect loses statistical significance when human capital variables are added (the full model). Moreover, it appears that alcohol-dependent males with later drinking initiation have higher rates of employment than similar males with no history of alcohol dependence.
For males, drug dependence appears not to be a significant predictor of employment in the multiple-disorder analyses (regressions 1 and 2), in contrast to the findings of the single-disorder analyses. Tabulations not reported here found that about 70 percent of males with drug dependence also have an alcohol disorder. It would appear that much of the single-disorder impact for drug dependence detected in table 5.10 is a result of the fact that these individuals had co-occurring alcohol use disorders.
The multiple-disorder results in regressions 3 and 4 of table 5.11 ratify the earlier findings (table 5.10) that male wage rates are negatively associated with alcohol dependence, with greater effects for those with early initiation of drinking than for those with later drinking initiation. Once more, the full model (including human capital control variables) yields impact estimates that are materially smaller as well as statistically insignificant.
For females, the equivocal results of table 5.10 are nearly replicated in table 5.11. The only impact that is statistically significant for females is that of drug dependence (regression 5) on the probability of employment. Inclusion of human capital control variables (the full model) does not materially diminish the estimated impact in magnitude or significance. Although there is an apparently negative impact of alcohol dependence with early drinking initiation on expected employment that has nearly the same magnitude as the drug dependence effect, it is not statistically significant. As alluded to previously, there are relatively few females in the sample with alcohol dependence and early drinking initiation.
Productivity Losses Due to Alcohol and Drug Disorders
The estimation of lost productivity due to alcohol and drug disorders is described in this section and summarized in table 5.12. Specifically, it is estimated that the cost in terms of lost potential productivity due to alcohol disorders was $66.7 billion in 1992 and was $14.2 billion for drug disorders. These estimates are composed of separate estimates for males and females and separate estimates for the effects of alcohol and drug disorders on employment/nonemployment and on earnings. Estimates have been developed using the microsimulation techniques developed for and employed in many analyses of the RAND Health Insurance Experiment (Newhouse and the Health Insurance Group 1993; Manning et al. 1987; Duan 1983).
This is a materially different approach to the development of these estimates than has been employed in the prior studies of the costs of alcohol and drug dependence. The alternative, prior approach (which is also the approach that has been used in many studies of welfare and health care reform) develops estimates based on applying the impact estimates (proportional effects either on the probability of employment or on wage rates) reported in the prior section to general population averages for rates of employment or earnings (including the value of fringe benefits) by age and gender groups. This general approach was taken in Rice et al. (1990) and in the prior studies of the economic costs of alcohol and drug abuse that examined effects on employment and earnings. The RAND microsimulation methodology is an alternative approach to development of such estimates.
The RAND microsimulation technique involves a series of procedures intended to improve the precision of estimated effects (whether for insurance design or health disorders). The application to alcohol and drug dependence involves comparing the expected value of earnings for an individual who is alcohol or drug dependent with his or her expected earnings in the absence of the alcohol or drug disorder. The difference between these two values is the estimated impact of the disorder on the individual, and the sum of these estimated impacts across all individuals with alcohol or drug disorders is the estimated national total impact.
The technique involves performing regressions similar to those reported in table 5.10 and table 5.11 to analyze the earnings of the alcohol- and drug-dependent populations, respectively, and of the non-alcohol- or non-drug-dependent population. Then, for a given alcohol- or drug-dependent individual, his or her expected value of earnings when alcohol- or drug-dependent is derived from the results of the regression on the alcohol- or drug-dependent population (regression packages will compute the "expected" or predicted value for an observation in a regression as well as the difference between the predicted/expected value and the actual value). In contrast, the expected earnings for each alcohol-dependent individual under the counterfactual assumption of never having been alcohol dependent is derived from the regression results from the non-alcohol- or non-drug-dependent population, respectively. This is done by applying the earnings regression coefficients from the non-alcohol-dependent population to the independent variable values for the individual alcohol-dependent observations. Summing these two expected values for the alcohol-dependent population (applying appropriate sampling weights) yields the national expected earnings for the alcohol-dependent population and their expected value if they had never been alcohol dependent. The difference in the two values is the estimated national impact.
An additional step in this RAND microsimulation technique is adjustment of predicted/expected values from the regressions for the "retransformation" bias (Duan 1983) that results from estimation of effects using log-linear models (which is both theoretically and empirically justified for the analysis of earnings). This step is necessary in order to address the retransformation problem as well as to protect against statistical issues such as heteroscedasticity in the data.
The costs for alcohol disorders were entirely for males. There were no statistically significant effects of alcohol disorders for females in the primary (multiple-disorder) analyses.
An important finding of this study is that early initiation of drinking (taking a first drink other than "sips" by age 15) combined with ever having met criteria for a diagnosis of alcohol dependence appears to have the strongest impact on employment success. In 1992, there were an estimated 2.66 million males who were not then full-time students and who met these criteria. An estimated 2.12 million of these were employed during the survey month, and on average they earned 13.1 percent less than would have been expected in the absence of their history of alcohol dependence. There were an additional 13.15 million males with a lifetime history of alcohol dependence (but with initiation of drinking after age 14), and an estimated 11.5 million of these were employed. This group experienced earnings that were estimated at 4.4 percent lower than they would have been in the absence of their history of alcohol dependence
The microsimulation techniques yielded estimates of expected monthly earnings of $2,514 for the 13.6 million employed males with dependence ever in their lifetimes. Their expected monthly earnings if healthy were $2,775, indicating a monthly loss of earnings of about $260 (9.4 percent). This set of estimates includes "smearing" adjustments of 1.348 and 1.398, respectively, for these two estimates, which has the impact of increasing the magnitude of the earnings loss estimate from about a 6-percent deficit (weighting the early and later drinkers together) to more than 9 percent. Note that this adjustment is necessary because the regression is done on "logged" dollars, which must be translated to "normal" dollars. The smearing factor can also be interpreted as adjusting for heteroscedasticity in the error term (Duan 1983). Heteroscedasticity can cause inefficient estimates of effects. Before the smearing adjustments, the expected monthly earnings of the alcohol-dependent males were $1,865 and $1,985 if never dependent (a 6.0-percent difference).
The total estimated productivity loss for employed, ever-alcohol-dependent males was $66.706 billion for 1992. Lost earnings for employed males ages 18 to 64 are equal to 13.6 million employed males x $260 loss per month x 12 months x 1.40 (to account for average fringe benefits beyond wages and salaries) x 1.1206 (to account for expected household productivity of employed males). The estimate of the contribution from fringe benefits is from studies on the cost of employee compensation (Braden and Hyland 1993), which find that employer-paid fringe benefits, such as paid leave, insurance, retirement and savings, and legally required contributions, are worth another 40 percent of the employee's wage or salary. The estimate of expected household productivity is based on data from Dorothy Rice (1997, personal communication) and the value represents the expected ratio of the value of employed males' household productivity contributions compared with market productivity (earnings plus fringe benefits), adjusted for the age distribution of males with lifetime alcohol dependence.
Males with a history of drug dependence are also estimated to have experienced productivity (and earnings) losses. The same method of estimation was used as for alcohol-dependent males. There were an estimated 2.85 million employed males with a history of drug dependence. The microsimulation estimated their expected monthly earnings at $2,356, against an expected value of $2,552 if they had never been drug dependent, yielding a loss of $196 per month (7.7 percent). As in the case of alcohol dependence, the smearing adjustment increased the estimated impact of drug dependence from 3.8 percent to 7.7 percent (the smearing factors were 1.334 and 1.390, respectively). The estimates of affected males and earnings losses must be adjusted from monthly to annual rates and incorporate the factors for fringe benefits and expected household productivity. This yields an estimated loss of $10.538 billion for ever-drug-dependent males.
Females have generally demonstrated weak or insignificant effects of alcohol and drug dependence on earnings, both in this study and in the prior reports. As discussed above, the only statistically significant effect for females is the impact of drug dependence on the probability of being employed. This loss has been estimated with the same microsimulation technique, with the additional calculation of expected probability of being employed for ever-drug-dependent women, based on the ever-drug-dependent and never-drug-dependent samples, respectively.
There were an estimated 2.1 million ever-drug-dependent females ages 18 to 64 not enrolled full time in school. There is no estimated negative impact on earnings (if employed) of ever-drug-dependent females (nor is there for alcohol): Expected monthly earnings for this population are $1,604 (including the smearing adjustment). However, women with a history of drug dependence have an expected employment rate of 62.7 percent, compared with a predicted rate of 69.0 percent if they had no history of drug dependence. This difference in expected level of employment translates into a loss of earnings of $102.50 per month averaged across the total 2.1 million women with and without employment. This translates into an estimate that about 136,000 ever-drug-dependent females were unemployed who would have been expected to have been employed if their experience had been like the non-drug-dependent population. The corresponding estimated loss of market productivity is $3.677 billion. No impact is estimated for loss of household productivity because no deficit was found in earnings among those employed.
It is highly noteworthy that among those males with a history of dependence on alcohol, the most severe effects were among individuals with early initiation of drinking (prior to age 15). The regression analyses (regressions 1 through 4) appear to indicate that a potential avenue through which early drinking operates on employment success is through negative effects on academic achievement. The analyses that omit the education variable consistently demonstrate effects of a larger magnitude than do those that control for education. This indicates that early drinking/ever dependent is correlated with lower educational attainment. This finding suggests a further reason to emphasize prevention efforts among adolescents.
Discussion
This analysis of the NLAES data has yielded new estimates of the nature and magnitude of effects of alcohol and drug disorders on employment. In several important respects this analysis represents an advance over prior analyses of this issue. The most notable improvements are that NLAES incorporates the most current diagnostic criteria (DSM-IV) for determination of alcohol and drug disorders and that the survey obtained data on earnings that excluded unearned income. Also critical is that this survey presents a recent (1992) picture of alcohol and drug use and disorder patterns. The survey gives equally rigorous attention to assessing both alcohol and drug disorders. Also, it assesses whether individuals had experienced episodes of depression that met DSM-IV criteria. Depression is a frequently co-occurring disorder with alcohol and drug disorders.
The major finding is that the strongest negative effects of alcohol dependence on earnings are experienced by individuals who began drinking before age 15. Although a history of alcohol dependence by itself is predictive of moderately reduced wage rates for employed males, this is merely an average of the stronger impact for early drinkers with the materially smaller impact for those who began drinking later. This suggests that alcohol prevention efforts that delay initiation of drinking may somewhat ameliorate the negative employment impacts that may be associated with alcohol dependence. Exploratory analyses also examined whether early initiation of drug use was associated with greater impacts; however, these analyses were inconclusive.
This study has not searched exhaustively for other patterns of alcohol or drug disorders that might be more strongly and negatively associated with employment and earnings. The finding that early drinkers with alcohol dependence disorders have stronger negative effects suggests that further analysis of the very rich NLAES data base may yield further insights. This is only logical, since the dependence and abuse measures were developed to identify individuals that might experience problems across a broad range of life experiences, rather than solely on employment impacts.
It is important to note that beyond alcohol dependence, other patterns of alcohol consumption and problems (e.g., abuse) were not negatively and significantly associated with effects on employment and earnings. Virtually all of these other variables were insignificant when they were tested in reduced or full models, and many of these variables were even positively correlated with employment and earnings. This was true for measures of current consumption and even for measures of alcohol abuse (either current or lifetime) without concurrent dependence.
These estimates are only modestly greater (by about one-third) than the estimates in Rice et al. (1990), which used the ECA survey from the early 1980's. That study estimated 1985 alcohol costs at $27.2 billion and drug costs at $5.9 billion, which would equal about $50 billion and $11 billion, respectively, if projected to 1992, versus the estimates in this study of $66.7 billion and $14.2 billion, respectively. Among the notable differences is that the Rice et al. (1990) study included both alcohol abuse and dependence in the prevalence estimates. However, NLAES appears to yield higher estimates of the prevalence of lifetime alcohol dependence than does the ECA survey. This analysis has not found significant negative employment effects associated with alcohol abuse, and in certain analyses, alcohol abuse is positively correlated with employment outcomes.
Most important, the Rice et al. (1990) study used a quite different approach to measure alcohol and drug effects, which directly lead to lower estimates of effects than would otherwise be expected. Rice et al. (1990) applied both their "timing" technique and the "indicator" method used in this and most other studies of this issue (see section 5.3).
An important area for further exploration is the fact that ECA assessed a wide range of mental disorders, whereas NLAES assessed only depression. This analysis did control for depression, finding consistently negative and often statistically significant effects on earnings. This is quite in contrast with the findings in the ECA analysis, which actually detected strong positive correlations between affective disorders and personal income.
As reported in earlier studies, females appear to have few significant employment-related effects from alcohol and drug disorders. It is likely that more complex models will be necessary in order to understand whether and how alcohol and drug disorders among women may correlate with employment outcomes. More sophisticated modeling of female labor market decisions requires incorporation of elements for marital status combined with fecundity. These elements are not nearly as significant in modeling the labor market experiences of males.
- Table 5.10: Regression Results for Single-Disorder Impact Estimates
-
Table 5.10: Regression Results for Single-Disorder Impact Estimates (each cell represents a separate regression) Alcohol or Drug Use Disorder Logistic Regression:
Earnings Past Month > 0OLS Regression:
ln (Earnings/Hr Past Month)Reduced Model Full Model Reduced Model Full Model Results for Males Alcohol dependence -0.029
(0.069)0.096
(0.075)-0.082
(0.018)**-0.050
(0.017)**ØAlcohol dependence and early drinking -0.440
(0.156)**-0.306
(0.167)*-0.178
(0.048)**-0.099
(0.045)**and later drinking 0.109
(0.067)0.161
(0.073)**-0.065
(0.025)*0.041
(0.022)*Drug dependence -0.149
(0.166-0.034
(0.169)-0.153
(0.036)**-0.098
(0.034)*ØResults for Females Alcohol dependence 0.035
(0.073)0.010
(0.071)-0.042
(0.025)-0.043
(0.023)ØAlcohol dependence and early drinking -0.346
(0.177)*-0.323
(0.180)*-0.046
(0.083)-0.011
(0.079)and later drinking 0.083
(0.075)0.046
(0.075)-0.041
(0.040)0.047
(0.037)Drug dependence -0.283
<0.106)**-0.255
(0.107)**-0.054
(0.046)0.008
(0.043)ØSource: Analysis of National Longitudinal Alcohol Epidemiologic Survey.
Notes: Other variables included in the reduced model are age, age squared, ethnicity, rural/urban residence, and the number of children living in the individual's household. Other variables in the full model include educational attainment, marital status, and whether the stated occupation was a skilled profession.
"Early drinking" means initiation of drinking (other than "sips") before reaching age 15. "Later drinking" means initiation of drinking at age 15 or older.
Values in parentheses are standard errors of the coefficients.
SUDAAN software was used to correct standard errors for sampling design.
* indicates significance at 0.10 level.
** indicates significance at 0.05 level.
Ø indicates that the estimated standard error was not corrected for the sampling design (estimated with SAS). Such correction leads to increases in estimated standard errors of up to 20 percent.
- Table 5.11: Regression Results for Multiple-Disorder Impact Estimates
-
Table 5.11: Regression Results for Multiple-Disorder Impact Estimates (each cell represents a separate regression) Alcohol or Drug Use Disorder Logistic Regression:
Earnings Past Month > 0OLS Regression:
ln (Earnings/Hr Past Month)Reduced Model Full Model Reduced Model Full Model Results for Males Regression 1 Regression 2 Regression 3 Regression 4 Alcohol dependence and early drinking -0.390
(0.171)**-0.274
(0.180)-0.140
(0.049)**-0.070
(0.046)and later drinking 0.137
(0.072)*0.180
(0.077)*-0.045
(0.025)*0.026
(0.023)Drug dependence -0.133
(0.189)-0.062
(0.191-0.099
(0.048)**-0.061
(0.046)Results for Females Regression 5 Regression 6 Regression 7 Regression 8 Alcohol dependence and early drinking -0.258
(0.184)-0.225
(0.188)-0.025
(0.086)-0.001
(0.082)and later drinking 0.119
(0.077)0.091
(0.077)-0.031
(0.041)-0.040
(0.039)Drug dependence -0.298
(0.111)**-0.250
(0.113)**-0.028
(0.046)-0.042
(0.043)Source: Analysis of National Longitudinal Alcohol Epidemiologic Survey.
Notes: Other variables included in the reduced model are age, age squared, ethnicity, rural/urban residence, the number of children living in the individual's household, and history of major depression. Other variables in the full model include educational attainment, marital status, and whether the stated occupation was a skilled profession.
"Early drinking" means initiation of drinking (other than "sips") before reaching age 15. "Later drinking" means initiation of drinking at age 15 or older.
Values in parentheses are standard errors of the coefficients.
SUDAAN software used to correct standard errors for sampling design.
* indicates significance at 0.10 level.
** indicates significance at 0.05 level.
- Table 5.12: Summary of Estimated Productivity Losses in the Workforce Due to Alcohol and Drug Disorders, 1992
-
Table 5.12: Summary of Estimated Productivity Losses in the Workforce Due to Alcohol and Drug Disorders, 1992 Gender Group Nature of Impact Basis for Estimate Labor Market Status/ Diagnosis No. of Persons With Loss
(in millions)<Size of Impact
(percent)Cost to Society
(billions of dollars)ALCOHOL DISORDERS Males Lower wages/ productivity Regression 5 Employed, ever alcohol dependent - 1st drink by age 15 2.117 13.1* 23.631 Males Lower wages/ productivity Regression 5 Employed, ever alcohol dependent - 1st drink after age 14 11.489 4.4* 43.075 Females No statistically significant impacts Subtotal, Alcohol 13.569 $66.706 DRUG DISORDERS Males Lower wages/ productivity Regression 5 Employed, ever drug dependent 2.855 9.4* $10.538 Females Excess unemployment Regression 7 Unemployed, ever drug dependent .136 7.1* $3.677 Subtotal, Drug Abuse 3.006 $14.205 TOTAL 16.575 $80.911 *These estimates are from the regression analysis above. They were not directly used in development of the cost estimates on this table. A microsimulation technique was employed that estimated expected earnings and/or employment for the alcohol- and drug-dependent populations, respectively, when dependent and not affected by alcohol or drug dependence.