PubMed2001
BACKGROUND: Health services researchers have increasingly used hazard functions to examine illness or treatment episode lengths and related treatment utilization and treatment costs. There has been little systematic hazard analysis, however, of mental health/substance abuse (MH/SA) treatment episodes. AIMS OF THE STUDY: This article uses proportional hazard functions to characterize multiple treatment episodes for a sample of insured clients with at least one alcohol or drug treatment diagnosis over a three-year period. It addresses the lengths and timing of treatment episodes, and the relationships of episode lengths to the types and locations of earlier episodes. It also identifies a problem that occurs when a portion of the sample observations is ǣpossibly censored. Failure to account for sample censoring will generate biased hazard function estimates, but treating all potentially censored observations as censored will overcompensate for the censoring bias. METHODS: Using insurance claims data, the analysis defines health care treatment episodes as all events that follow the initial event irrespective of diagnosis, so long as the events are not separated by more than 30 days. The distribution of observations ranges from 1 day to 3 years, and individuals have up to 10 episodes. Due to the data collection process, observations may be right censored if the episode is either ongoing at the time that data collection starts, or when the data collection effort ends. The Andersen-Gill (AG) and Wei-Lin-Weissfeld (WLW) estimation methods are used to address relationships among individuals multiple episodes. These methods are then augmented by a probit censoring model that estimates censoring probability and adjusts estimated behavioral coefficients and related treatment utilization and treatment costs. There has been little systematic hazard analysis, however, of mental health/substance abuse (MH/SA) treatment episodes. RESULTS: Five sets of variables explain episode duration: (i) individual; (ii) insurance; (iii) employer; (iv) binary, indicating episode diagnosis, location, and sequence; and (v) linkage, relating current diagnoses to previous diagnoses in a sequence. Sociodemographic variables such as age or gender have impacts at both the individual and at the firm level. Coinsurance rates and deductibles also have impacts at the individual and the firm levels. Binary variables indicate that surgical/outpatient episodes were the shortest, and psychiatric/outpatient episodes were the longest. Linkage variables reveal significant impacts of prior alcoholism, drug, and psychiatric episodes on the lengths of subsequent episodes. DISCUSSION: Health care treatment episodes are linked to each other both by diagnosis and by treatment location. Both the AG and the WLW models have merit for treating multiple episodes. The AG model permits more flexibility in estimating hazards, and allows researchers to model impacts of prior diagnoses on future episodes. The WLW model provides a convenient way to examine impacts of sociodemographic variables across episodes. It also provides efficient pooled estimates of coefficients and their standard errors. LIMITATIONS: The insurance claims data set covers 1989 through 1991, predating current managed care plans. It cannot identify untreated substance abusers, nor can it identify those with out-of-plan use. It provides treatment information only if services are covered by the insurance plan and are defined with a substance abuse diagnosis code. Like medical records, insurance claims will not specify substance abuse treatment received within the context of other health care (and thus identified by a non-substance abuse diagnosis code) or community services. IMPLICATIONS FOR POLICY AND RESEARCH: This article characterizes multiple health treatment episodes for a sample of insured clients with at least one alcohol or drug treatment diagnosis within a three-year period. We identify both individual and employer effects on episode length. We find that episode lengths vary by the diagnosis type, and that the lengths (and by inference cost and utilization) may depend on the treatments that occurred in previous episodes. We also recognize that health care or illness episodes may be ongoing at times of health care events prior to the ends of data collection periods, leading to uncertain episode lengths. Corresponding estimates of costs or utilization are also uncertain. We provide a method that adjusts the episode lengths according to the probability of censoring.