We estimated the sum of the COMBINE treatment costs plus the social costs of health care use, arrests, and motor vehicle accidents over three years for the 9 alcohol dependence treatments in the COMBINE study.4
If alcohol treatment interventions generate future social cost savings, then these treatment interventions are more attractive from the social perspective because they generate social benefits that are not limited to alcohol-dependent patients. Furthermore, adding social costs to treatment costs may change the relative attractiveness of costly interventions—an initially costly treatment intervention may result in lower future social costs, increasing its attractiveness relative to other alternatives.
Our multivariate results show no significant differences in mean costs between any of the treatment arms as compared to MM + placebo for the 3-year post-randomization sample. We hypothesize that this lack of statistical significance was largely caused by the skewness of the data, which in turn was caused by the occurrence of rare but expensive inpatient hospital stays, arrests, and motor vehicle accidents. These high cost outliers increased both the mean and variance of costs and typically a larger sample size is required to show significant cost differences between treatment interventions.
Because our sample size was predetermined by the design of COMBINE and could not be increased, we addressed the skewness by estimating median regression models. In contrast to mean costs, the medians are less affected by high cost outliers. At 3 years post-randomization, the median
costs of MM + acamprosate, MM + naltrexone, MM + acamprosate + naltrexone, and MM + acamprosate + CBI were significantly lower than the median cost for MM + placebo. Median cost differences ranged from $2,500 to $3,800 less than the median costs of MM + placebo. These results show that social cost savings are generated relative to MM + placebo by 3 years post-randomization and, importantly, the magnitude of these cost savings is greater than the costs of the COMBINE treatment received 3 years prior. Qualitatively, this result is similar to the UKATT9
results, although it is difficult to make a definitive comparison because UKATT did not have a placebo condition, as was the case here.
To explore our results in more detail, we re-ran our mean and median regression analyses with just the initial treatment costs plus health care expenses (i.e., excluding arrests and motor vehicle accidents). We found that the 3-year post-randomization median results were very different from : no treatment intervention was individually significant, only 4 of the coefficients were negative in sign, and we could not reject the null that the coefficients were jointly equal to zero (P = .37). Thus, we concluded that there are no offsetting health care costs 3 years post-randomization and that including arrests and motor vehicle accident outcomes had a substantial effect on our estimated social cost differences.
As an alternative to our OLS specification, we transformed our model to account for the skewness of our cost data. We followed the methodology described by Manning and Mullahy15
to determine the appropriate data transformation. This process suggested a generalized linear model with a log-link and gamma distribution. With this model, the differences in costs represented by the treatment arm coefficients and the costs of MM + placebo were the same in sign (with one exception) and generally of a similar magnitude as the differences in means shown in for the OLS models. These results highlight the robustness of the OLS model to our skewed data distribution, and we report the OLS specification along with the median regression model.
A related literature examines the potential cost-offset of alcohol treatment by evaluating whether alcohol treatment as compared to no treatment reduces subsequent medical utilization and health care costs (e.g., 16–21
). The results of this literature vary widely due in part to differences in study population, study design, and comparison group. For example, Parthasarathty et al.18
found that inpatient, ER, and total medical costs declined 18 months after intake into the outpatient chemical dependency program at Sacramento Kaiser Permanente; non-ER outpatient costs were unchanged. Booth et al.19
found differences in the use of inpatient and outpatient care as alcohol treatment varied in intensity from short detox to extended detox to incomplete treatment and completed treatment. In their study of male alcoholics in the VA, Booth et al. found that inpatient days and outpatient visits increased post-alcohol treatment for all treatment groups; however, inpatient medical care decreased and substance abuse inpatient care increased for most treatment groups; overall, the use of inpatient services increased for these groups. In a study of adults receiving benefits from a behavioral managed care company and its parent medical care insurance company, Kane et al.16
found that the pattern of outpatient and inpatient medical utilization before and after alcoholism treatment was symmetric: utilization increased gradually in the year before treatment and then decreased post-treatment. Their results suggested that alcoholism treatment did not reduce subsequent health care utilization.
In contrast to previous cost offset papers of which we are aware, our paper takes advantage of the COMBINE trial’s RCT design. Related papers typically create an untreated alcohol-dependent group from health care records. Weaknesses of this design include the potential regression to a lower mean utilization post-treatment and potential selection bias caused by the inability to control for unobserved differences between alcohol-dependent patients who do and do not go to alcohol treatment. Our RCT design does not have these weaknesses and thus our conclusions have more internal validity.
Our paper has several limitations. First, our analysis examines only a subset of possible social costs: health care utilization, arrests, and motor vehicle accidents. We did not examine labor market outcomes because of the difficulty of placing a dollar value on all the outcomes, such as the value of being “unemployed.” But because a majority of the COMBINE patients were not unemployed (approximately 75% were employed at intake and at weeks 52 and 156), excluding this variable is unlikely to change our conclusions. Second, although we drew on peer-reviewed literature for our unit cost estimates, we exercised some judgment in selecting the unit cost components to include. For example, in selecting the unit costs of arrests, we used the monetary subtotal of arrest costs, but we adopted a conservative approach and did not include accident-related medical costs for those involved in accidents who were not patients in the COMBINE trial or nonmonetary cost estimates of pain and suffering, which would have increased our unit cost estimates substantially. Although the size of the cost differences would increase with larger unit cost estimates, there is no reason to believe that our conservative approach affects our comparisons across the alternative treatment interventions because the unit costs were applied consistently to all treatment arms. Third, all treatment interventions are compared to MM + placebo, which is not a “no treatment” group as MM plays an important treatment role. If the comparison group did not include MM, an active treatment, the estimated cost differences may have been even larger than estimated here. Fourth, we recognize that expensive rare events such as lengthy hospital stays, arrests, and motor vehicle accidents may greatly affect mean costs especially given the size of our sample. Our analysis of median costs, which are less affected by outliers, is one way to address this concern. In the future, additional RCT studies with larger sample sizes are needed to confirm the joint and individual impact of alcohol treatment on mean and median costs.
Finally, while it is reasonable to compare the clinical outcomes with the cost results reported here, we are reluctant to make causal statements about the relationship between specific clinical therapies and the impact on costs 3 years post-randomization because COMBINE was not designed for that analysis. We do note, however, that MM + acamprosate + naltrexone had the highest mean effectiveness6
for all 3 clinical outcomes (percent days absent, proportion of patients who avoid heavy drinking, and proportion of patients with good clinical outcomes) and also the largest estimated decrease in median costs 3 years post-randomization. Similarly, MM + acamprosate + CBI and MM + naltrexone had the next largest estimated median cost reductions and the second or third largest mean effectiveness estimates across the clinical outcomes. Additional analyses with larger clinical and economic samples and/or analyses that develop structural relationships between clinical outcomes and cost would be required to be more definitive about the relationship. In the meantime, our study suggests that several alcohol treatments—MM + acamprosate, MM + naltrexone, MM + acamprosate + naltrexone, and MM + acamprosate + CBI—may indeed lead to reduced median social costs associated with health care, arrests, and motor vehicle accidents.