The Relationship Between Substance Use Patterns and Economic and Health Outcomes Among Low-Income Caregivers and Children

Like others who have studied the impact of substance use on economic outcomes, we draw on the theoretical framework proposed by Becker ( 20 ) and applied to health by Grossman ( 21 ). The key insight of these models is that utility is influenced by health, both directly and indirectly. Health influences utility indirectly because it contributes to productivity both at work, in the form of wages, and at home, by affecting the ability to raise children, for example. In the context presented here, heavy substance use can be viewed as one dimension of health and as a (negative) input to physical and mental health and thus productivity. Such a model predicts that heavy substance use would lower the probability of work, relative to public assistance or other options; worsen symptoms of mental health; and contribute to poor child behavior. Testing for such relationships is challenging because substance use and outcomes of interest may be related to some common factor other than substance use. To help address this confounding, we estimated how changes in outcomes are related to patterns of substance use. We also acknowledge the possibility that confounding factors could change substance use and our outcomes of interest simultaneously.

To assess how patterns of substance use are related to economic outcomes and well-being among low-income female caregivers and children, we estimated two sets of models using interview data from the Welfare of Children and Families study. Data were analyzed from baseline interviews that were conducted from March through December 1999 and follow-up interviews that were conducted 11 to 26 months after baseline (average, 16 months). We first modeled, at the time of follow-up, the likelihood of working, collecting income assistance (welfare or disability benefits), or neither working nor collecting income assistance (that is, being "detached") as a function of substance use patterns. Because we controlled for baseline work and income assistance, these follow-up models can be viewed as models of changes in work and public assistance status. Second, we modeled several measures of well-being as a function of substance use patterns: changes in caregiver mental health symptoms, changes in child behavior as reported by the caregiver, and changes in health insurance status of caregivers. Together, these analyses provide a picture of how substance use patterns and policy levers might influence work, public program use, and selected measures of well-being.

We used longitudinal data from the first two waves of the Welfare of Children and Families, or the Three-Cities Study, describing circumstances of low-income female caregivers living under 200% of the federal poverty level in low- and moderate-income neighborhoods in Boston, Chicago, and San Antonio ( 22 ). In March through December 1999, a total of 2,402 "focal" children aged zero to four or ten to 14 and their primary female caregivers were sampled. Information on work status, public program participation, mental health, and substance use of caregivers, as well as child behavior, caregiver demographic characteristics, and household structure were gathered through in-person interviews with caregivers conducted via a computerized instrument. The sample targeted children ages zero to four or ten to 14 because these ages mark major transitions for children and because for children living with female caregivers who work, these ages represent a time when preschool-age children "require high quality child care" and older children "require more intense supervision and scholastic guidance" ( 22 ). The survey oversampled blacks and Hispanics. Follow-up interviews were conducted with the same caregivers between September 2000 and June 2001.

Of the 2,402 women interviewed in 1999, a total of 384 women were not interviewed at follow-up and were excluded. There were no significant differences in moderate substance use, use of marijuana or other illicit drugs, and the probability of working at baseline between those who left the study and those who completed the follow-up interview. We further excluded 348 caregivers because they (N=152) or someone in their household (N=196) received Supplemental Security Income (SSI) at baseline; however, caregivers remained in the sample regardless of SSI receipt at follow-up. This allowed us to focus on transitions into SSI, which provides income assistance for poor individuals with a long-term disability. Terminations of SSI cases are rare, and once on SSI, recipients typically do not transition out of it. To focus on working-age caregivers, we excluded an additional 40 women who were younger than 18 or older than 64. We excluded seven women who were missing information on race, TANF status, work status, or Brief Symptom Inventory (BSI) scores, leaving 1,623 women in our sample.

Caregivers reported their substance use in the past 12 months. An ideal measure of substance use would capture symptoms of abuse and dependence or both frequency and quantity of substance use; however, this information was not available in the Three-Cities Study. However, several studies demonstrate significant correlation between even relatively crude measures of substance use with measures of treatment need, substance abuse, and substance dependence ( 23 24, 25 ), something we confirmed among women age 18 and older in the 2002 National Survey of Drug Use and Health. Substance use measures (based on either one or more occasions of having five or more drinks at a sitting or use of any illicit drug in the past 30 days) were significantly correlated with having either abuse or dependence ( ρ =.37, p<.01).

We categorized substance use patterns in two steps. Respondents were asked, "In the last 12 months, how often were you drunk?" Women were also asked, separately for marijuana and other illicit substances, "In the last 12 months, how often did you use?" Possible responses to all three questions included "never," "once or twice," "several times," or "often." In the first step, for each interview, we categorized a respondent's use as light or no use if respondents answered "never" or "once or twice" to all three questions (that is, frequency of drunkenness, marijuana use, and illicit substance use). Otherwise, use was categorized as moderate or heavy. We distinguished moderate or heavy use of alcohol and drugs from any use because previous research suggests that moderate to heavy substance use is most likely to coincide with substance-related harm and to reduce the probability of work and earnings ( 12, 26, 27 ). [A table showing substance use patterns by type of substance is available as an online supplement at ps.psychiatryonline.org.]

In the second step, we created four patterns of substance use based on substance use at baseline and follow-up: light or no use at both interviews (N=1,377); moderate or heavy use at both interviews (N=56); increased use during the study period (that is, light or no use at baseline and moderate or heavy use at follow-up) (N=99); and reduced use during the study period (that is, moderate or heavy use at baseline and light or no use at follow-up) (N=91). Because of the small sample of women with moderate or heavy use and because women with moderate or heavy use and women with increased use were expected to have worse trends in outcomes, we combined both of these categories into a new category called moderate or heavy substance use to improve precision.

We first assessed work and public program participation outcomes at follow-up; data from follow-up, rather than baseline, were analyzed to prevent the possibility of reverse causation or that a change in work or public program use "caused" women to change their substance use. We examined the likelihood of caregivers being in one of three mutually exclusive categories at follow-up: working now with or without receiving TANF or SSI, receiving TANF or SSI without working now, or neither working now nor collecting income assistance now (that is, being "detached"). Among this sample of predominantly unmarried, low-income female caregivers, we interpret the detached state as unfavorable, because it likely indicates a lack of resources.

Because Medicaid coverage and health insurance more generally offer important avenues to access substance abuse treatment services, we assessed health insurance status for the sample of women. We used caregiver reports of Medicaid coverage at follow-up and determined whether caregivers lost or gained health insurance coverage during the study period.

We analyzed two measures of mental well-being—caregiver mental health symptoms and child behavior problems reported by the caregiver. An increasing number of studies suggest that maternal mental health in general and maternal substance use in particular negatively influence a range of child outcomes ( 28, 29 ). To assess caregiver mental health, we used the BSI ( 30 ), in which female caregivers responded to 18 items relating to anxiety, depression, and somatization. These items were converted into an overall measure of poor mental health, with higher values indicating more symptoms. This measure is highly correlated with the existence of a diagnosable mental health disorder ( 30, 31 ). We created a binary measure of improved mental health—that is, a decrease of one standard deviation in the log-transformed BSI total mental health measure during the study period. Although this measure has not been used previously, the cutoff for poor mental health indicative of a likely diagnosis of a mental disorder is one standard deviation above the mean ( 20 ). Thus, if everyone in our sample had a decrease of one standard deviation in BSI scores, the share of women at baseline who would exceed the cutoff for poor mental health would fall from 7.6% to 0%. A rise of one standard deviation in BSI scores for our sample would result in 37% exceeding the cutoff for poor mental health.

To measure child behavior problems, we used caregiver responses to the multi-item Child Behavior Check List (CBCL) for children ages two to three and a separate CBCL for children ages four to 18 ( 32, 33 ), excluding 393 children younger than two at baseline (N=1,230). Similar to the BSI measure, higher scores indicate worse child behavior ( 32, 33 ), and we formed a variable equal to 1 if the focal child's CBCL score fell by one standard deviation between the two interviews.

We estimated two types of outcome models. First, to estimate the effect of substance use patterns on working, receiving income assistance (TANF or SSI), or being detached at follow-up, we estimated multinomial logit models of this three-category variable as a function of indicators for reduced substance use and moderate or heavy substance use, rendering light or no substance use as the reference category. Each model controlled for education, marital status and cohabitation, race and ethnicity, city, and age of respondent at follow-up and work status and use of TANF at baseline ( Table 1 ). In a second set of models of health insurance coverage, improvements in BSI scores, or improvements in CBCL scores, we estimated logistic regression models with the same controls as above.

For each result, we display the adjusted probability of the outcome, or the average marginal effect of substance use patterns based on the model coefficients. Given relatively modest sample sizes of caregivers who changed substance use patterns, we report results with borderline statistical significance, or p values under .10. [An appendix showing three tables with the full set of odds ratios (logit models) or relative risk ratios (multinomial logit models) is available as an online supplement at ps.psychiatryonline.org.] Analyses were conducted with Stata software (version 9.0). All estimates used sample weights provided in the Three-Cities Study data to make estimates representative of the target population of low-income caregivers and children in low- and moderate-income neighborhoods in Boston, Chicago, and San Antonio. As recommended by study investigators to prevent results from being skewed toward a single city, weights were scaled such that each city counted equally toward estimates. The Harvard Medical School Human Subjects committee deemed this study exempt from institutional review.