presents descriptive statistics for the analysis sample segmented by gender and survey wave for all the variables. For males, the mean probability of any hospitalization in the past 12 months is 6.4 percent in Wave 1 and 9.2 percent in Wave 2. For females, the mean probability of hospitalization is 8.7 percent in Wave 1 and 10.7 percent in Wave 2. Thus, between Waves 1 and 2, the average probability of hospitalization increased by 44 percent for men (p
<.01) and 23 percent for women (p
<.01), an effect that corresponds primarily to the aging of our sample (respondents are five years older in Wave 2).2
Regarding ER utilization, 18.1 percent (20.3 percent) of male respondents visited the ER at least once in the past 12 months at Wave 1 (Wave 2). The percentage of female respondents who visited the ER in the past 12 months was slightly higher than that of male respondents in both waves. Yet, men were more likely than women to suffer an injury that required medical attention and/or disrupted normal activities. The proportions increased from Wave 1 to Wave 2 for both genders.
Mean Values for Analysis Variables
The drug use measures also displayed gender differences, with males consistently showing a greater participation rate and frequency of use than females. Four percent of men were casual drug users at Wave 1 and 4.5 percent were heavy drug users. These rates rose slightly to 4.8 percent and 5.2 percent, respectively, at Wave 2. The usage rates were considerably lower for women than for men at Wave 1, with casual users comprising 2.2 percent of the sample and heavy users coming in at 3.1 percent. These rates remained fairly stable for women at Wave 2. While not reported in , cannabis is the most common illicit substance used by either gender, a finding that has been well documented in other national surveys (SAMHSA 2009
). About 5 percent of men and 2 percent of women used cannabis exclusively in each wave. A relatively small fraction of respondents—2 percent of men and 1 percent of women—used multiple drugs in both waves.3
As noted earlier, we control for a long list of predisposing, enabling, and need variables in the logistic and negative binomial regression models. We report mean values for all these control variables in .
Identification in conditional fixed-effects models relies on the number of people changing the likelihood and/or intensity of health services utilization and shifting across the drug use categories from one wave to the other. displays changes in our health services utilization and drug use measures from Wave 1 to Wave 2. Number of hospitalizations changed between the two waves for 13 percent of men (N=1,477) and 16 percent of women (N=2,066); number of ER visits changed for 29 percent of men (N=3,198) and 32 percent of women (N=4,065); and number of serious injuries requiring medical attention changed for 32 percent of men (N=3,842) and 28 percent of women (N=3,548). In terms of drug use, 14 percent of men (1,505) and 9 percent of women (1,150) had a different number of drug-specific days of use in Waves 1 and 2. Moreover, 11 percent of men (1,221) and 8 percent of women (1,019) changed drug use categories between waves (i.e., no use, casual use, heavy use). Taken as a whole, these proportions are respectable (i.e., 8–32 percent of the sample for various measures of health services utilization and drug use) but identification would be further enhanced with more changers. We return to this issue later in the paper.
Changes in Health Services Utilization and Illicit Drug Use from Wave 1 to Wave 2
Logistic and Conditional Fixed-Effects Logit Models
presents selected estimation results by gender when binary measures of health care utilization are the dependent variables. The key explanatory variables are dummies for casual drug users and heavy drug users, with no drug use as the reference condition. The pooled panel estimation (logistic regression) in columns (1) and (3) shows that heavy drug use is positively and significantly (p<.05) related to all three health services utilization measures (i.e., odds ratio [OR] > 1) for both men and women. The estimated ORs for casual drug users are also greater than one in almost all cases, but only significant (p<.05) for any ER use for women and any serious injury for men. The effect sizes are relatively large with ORs in the range of 1.22 to 1.50. These pooled panel results suggest that illicit drug users, particularly heavy users, are about 25 to 50 percent more likely to consume these health services than non-drug users. However, this conclusion would be misguided since the fixed-effects estimates tell a somewhat different story.
Selected Estimation Results for Binary Measures of Health Services Utilization
Columns (2) and (4) of present estimates after controlling for unobserved individual heterogeneity using a fixed-effects estimator (conditional fixed-effects logit). For men (column (2)), both casual and heavy drug use continues to have a significant effect (p<.05) on the likelihood of experiencing a serious injury (similar effect sizes), but the estimates are no longer significant for any hospitalization and ER visit. None of the fixed-effects conditional logit estimates are statistically significant for women (column (4)), but the effect of heavy drug use on any ER visit is approaching significance (p<.10).
Negative Binomial and Conditional Fixed-Effects Negative Binomial Models
presents estimation results for count measures of health services utilization. The key statistic in our negative binomial models is the incident rate ratio (IRR), or the exponentiated coefficient.4
As with the ORs in , the majority of the IRRs in are significantly (p<.05) different from one when we apply negative binomial models to the pooled panel data (columns (1) and (3)). The effect sizes indicate that drug users have about a 30 percent higher rate of utilization of these services than non-drug users. Turning to the conditional fixed-effects negative binomial results (columns (2) and (4)), some of the IRRs are no longer significant. The four effects that remain significant are heavy drug use affecting the number of times hospitalized for both men and women and both casual and heavy drug use affecting the number of times seriously injured for men only. Except for heavy drug use’s effect on injuries, the other significant estimates are higher in magnitude than those resulting from the pooled sample estimation.
Selected Estimation Results for Count Measures of Health Services Utilization
Our results indicate that unobserved individual heterogeneity is not an important source of bias when estimating the effects of illicit drug use on serious injuries for men. This is also the case when estimating the effects of heavy drug use on the number of times hospitalized (intensive margin) for both genders. However, all of the other estimated ORs and IRRs are either non-significant throughout (p<.05) or become non-significant after we control for unobserved individual heterogeneity.
Sensitivity Analysis 1: Excluding Health Status Controls
The core results discussed above include a set of health status measures (i.e., SF-12 physical health score; SF-12 mental health score; dummy variables for arthritis, gastritis, heart disease, and hypertension) and other substance use variables (tobacco and alcohol) as controls. One potential problem with this approach is that the health status measures and substance use variables are contemporaneous and thus potentially endogenous to the drug use and health care utilization variables. Specifically, the effects of drug use on health care utilization could work through certain chronic health conditions or be diluted by the co-morbid use of alcohol or tobacco. If our measures of health status and substance use behaviors do not predate the decision to use illicit drugs, we may over-control when adjusting for these measures.
To analyze how sensitive our results are to the inclusion of these controls, we reran all models excluding the six health status and two substance use variables from the core models presented in and . For both genders, ORs and IRRs from the fixed-effects models are generally larger (i.e., further away from one) and more significant than those in the core models. In particular, female heavy drug users now have higher odds of being hospitalized (p<0.05). These results suggest that the use of other substances and poor health status could mediate the effects of drug use on health care utilization.
Sensitivity Analysis 2: Analysis by Health Insurance Status
The marginal cost of care (i.e., price) for the consumer is one of the most important variables in a healthcare demand relationship. In the healthcare market, most patients do not directly respond to the true marginal cost of care because they are insured. Thus, insurance type, coverage parameters, and plan flexibility are often more important for consumers than healthcare prices per se. The NESARC does not include comprehensive measures of insurance features, but it does contain information on whether a person is covered by public (i.e. Medicaid, Medicare, VA) or private insurance or is uninsured. With these distinctions, we can explore how drug use’s effects on healthcare utilization vary by type of insurance. Conditional fixed-effects results (available upon request) show a positive and statistically significant association (p<.05) between drug use and the number of visits to the ER for uninsured men, but not for men with private or public insurance. In addition, drug use is associated with a higher likelihood of serious injuries for privately insured men and a higher number of injuries for uninsured men. For women, drug use is associated with a higher number of times in the hospital (p<.05), but only among those with public health insurance. These subgroup analyses should be viewed with caution, however. Because fixed-effects estimates are identified by those individuals who change health care utilization and drug consumption over time, the sample sizes become quite small when running separate models by insurance type. Thus, some non-significant estimates could be influenced more by a lack of statistical power than by a weak relationship.
Sensitivity Analysis 3: Analysis by Age
To investigate the presence of age-specific effects, we divided the male and female samples into two groups (age 35 and younger, and older than age 35). As with the insurance type analysis reported above, identification and statistical power is a concern for some of the specifications. With this caveat in mind, findings indicate that older females (> age 35) who are heavy drug users have a higher probability of entering a hospital and a greater number of hospital visits. None of the estimates are significant for men at the 5 percent level or better. These results are available upon request.
Sensitivity Analysis 4: Total Number of Drug-Specific Days of Use
In our core analyses, we divide drug users into two groups (casual and heavy users) to minimize the influence of extreme outliers. To explore the effects of a continuous measure of drug use on health care utilization, we replaced our two categories with the total number of drug-specific days of use. Findings from the fixed-effects models confirm a positive and significant effect of drug use on the number of hospitalizations for men. In contrast to the core analysis, the fixed-effects models demonstrate no evidence of an effect of drug use on serious injuries. However, they do show a positive and significant effect of drug use on the likelihood and number of ER visits for men, which were not detected in the categorical analysis. For women, the measure of drug-specific days of use is not statistically significant in any of the fixed-effects specifications.
Sensitivity Analysis 5: Interactions between Illicit Drug Use and Alcohol Use
The literature has firmly established that illicit drug users, especially heavy users, are often alcohol users as well. Thus, illicit drug users who also consume alcohol may disproportionately use health care services relative to those who use alcohol or illicit drugs only. To explore this relationship while also preserving statistical power, we re-estimated models with five substance use variables (casual drug use, heavy drug use, ounces of ethanol consumed, and the two drug/alcohol interactions) and the standard set of controls. Given that all of our fixed-effects specifications are nonlinear (i.e., conditional fixed-effects logit and negative binomial), interpreting interaction terms is more complicated than simply assessing the statistical significance (Ai and Norton 2003
). Following the advice of Norton (2004)
, we employed linear fixed-effects regression (e.g., linear probability models instead of logit) to the conditional samples because computations for interactions in nonlinear fixed-effects models with more than one interaction term are currently not available in Stata or other statistical packages (Norton, Wang, and Ai 2004
). Although linear estimation of binary and count variables is not preferable, the advantage here is that the effect sizes and associated standard errors for interaction terms are easy to interpret. These estimates are available upon request.
With three health services utilization measures, two margins (intensive and extensive), and two genders, we estimated a total of 12 specifications with drug/alcohol interactions. For males, the interactions of both drug use variables (casual and heavy) with alcohol use are positive and statistically significant (p<.05) in the specification for the number of times hospitalized. This suggests that drug-using males who also consume alcohol are being admitted to a hospital more often than their drug-using counterparts who do not use alcohol. None of the other drug/alcohol interactions in the remaining 11 specifications are statistically significant at the 5% level or better.