To gain access to individual-level, de-identified STTR study data, we submitted a concept proposal to the STTR Data Coordination Center at the University of Washington, which was reviewed and approved in 2016. Individuals interested in collaborating or working with these data should contact the STTR Data Coordination Center at [email protected]. For this study, we initially selected six STTR studies based on similarity across baseline questionnaires and relative completeness of de-identified baseline data. Three of these studies were later excluded because the baseline questionnaires collected healthcare resource utilization data without specific recall periods and could not be used to meaningfully calculate costs.
Collectively, 868 people living with HIV with or at high-risk for SUD are represented in three studies: (1) PACTo: Enhanced Access to HIV Care for Drug Users in San Juan, Puerto Rico, which implemented and evaluated a community-level, structured approach to 409 people living with HIV who use substances in five communities in San Juan from 2014 to 2017 (18); (2) Project RETAIN: Providing Integrated Care for HIV-Infected Crack Cocaine Users, which evaluated the efficacy of an integrated HIV and primary care “retention clinic” in achieving virologic suppression compared to treatment as usual in 360 people living with HIV who used cocaine in Miami, FL and Atlanta, GA from 2013 to 2017 (19); and (3) BRIGHT 2: Baltimore-Rhode Island Get HIV Tested, which evaluated the effectiveness of HIV linkage to care comparing intensive case management to treatment as usual in community corrections offices in 99 people living with HIV who were on probation or parole in Baltimore, MD from 2011 to 2015 (20, 21). Healthcare resource utilization data were self-reported by study participants at baseline.
We identified healthcare resource utilization measures common to at least two of the selected studies with comparable recall periods, and categorized them into three domains: general medical care (e.g., hospital-based ED visits), SUD treatment (e.g., times treated for alcohol use disorder (AUD)), and medications (e.g., prescribed medication for AUD) (Table 1). We also included participant spending on substances, a measure shared by PACTo and RETAIN. Baseline healthcare resource utilization was reported across varying recall timeframes ranging from past 30 days to lifetime.
We reviewed common healthcare resource utilization measures to identify outcomes that are comparable across studies and could potentially be used for economic analyses. A prerequisite was that variables must represent units (e.g., number of hospital-based ED visits) over a specific recall period (e.g., last 30 days). Dichotomous measures such as “ever been treated for substance use disorder” or measures over lifetime cannot be meaningfully monetized for use in economic evaluations. We identified 10 measures that met this criteria and were representative of the healthcare sector perspective. These measures captured data on ED, inpatient hospital and residential facility, and outpatient encounters. Additional measures informed a broader, societal perspective by capturing reported number of days experiencing alcohol- or drug-related problems, and participant spending on alcohol or drugs in a given recall period. Some of these measures evaluated utilization during a specified time frame using response from a single question (e.g., number of hospital-based ED visits in a specific recall period) whereas other measures captured utilization using a combination of questions (e.g., number of hospitalizations in a specific recall period and number of days spent in the hospital per reported hospitalization) to calculate the total number of hospital days in the recall period.
We constructed descriptive statistics for each measure across all three studies (Table 2). To normalize different baseline assessment time-frames, we considered extrapolating data to the longest recall period (12 months). For instance, responses to measures reported “per 30 days” can be multiplied by 12 to represent a “per 12 month” measure. However, we instead created measures of the “average healthcare resource utilization per 30 days” by dividing 6-month and 12-month data by the represented number of months. While both adjustments rely on a limiting assumption that the rate of healthcare resource utilization remains constant over time, creating an average with real data points vs. adding data points through extrapolation was deemed a more conservative and preferred approach. Inpatient hospital days, residential facility days, and outpatient visits were calculated using a combination of the number of these events reported and the corresponding number of days per event. All studies collected data on the number of hospitalizations as well as the number of days per hospitalization (up to five most recent hospitalizations at baseline). In these instances, we created a measure of average event frequency (number of hospitalizations) per 30 days. If any individually-reported event-length (number of days spent in the hospital per hospitalization) exceeded 30 days, we adjusted to a 30-day maximum. We then multiplied the adjusted event frequency by the average event length in order to calculate a utilization measure that was translatable to dollars (e.g., number of days spent in the hospital per 30-day period).
Missingness was low across the three studies. We categorized missing data as: 1) an absence of information or 2) invalid responses (22). In our studies, absence of information included responses left blank and responses of “I don’t know,” “Refuse to respond,” or “N/A.” Invalid responses included out-of-range responses (e.g., 50 ED visits in 30 days), and incompatible compound responses. For questions assessing frequency and duration separately, if either question was left blank or if one of the two questions was answered with a positive response and the other with a zero (e.g., zero hospitalizations, 2 days each), we considered the response to that measure to be missing due to incompatability of combined responses. Invalid compound missingness was only applicable to the combined measures of utilization, whereas missingness due to absence of information or out-of-range responses was applicable to all measures. For the purpose of this study, which was limited to baseline data, we removed missing responses to individual measures from our analysis through case deletion, rather than create a complete data set through imputation.
As dictated by the healthcare sector and societal perspectives, we attempted to find MCFs designed to capture the value of the resources utilized, without accounting for other characterstics, such as profit (23). We used the U.S. nationally-representative Medical Expenditure Panel Survey (24), to capture Medicare payments for hospital-based ED visits, hospitalization days, hospital clinic or outpatient department visits, community clinic or neighborhood health center visits, and physician visits; the Alcohol and Drug Services Study, from the Substance Abuse and Mental Health Services Administration (25), to value drug or alcohol residential treatment, detoxification hospital stays, and treatment provider visits; the Medicare physician fee schedule (26–28), to value mental healthcare provider visits; and data pooled by McCollister et al. (17), to value days experiencing alchohol or drug problems. Mean resource utilization figures were then multiplied by corresponding MCFs, converted to 2017 USD, in order to generate mean costs per 30 days at baseline for each measure across all three studies. Participant spending on substances was reported in dollar units, thus no MCFs were applied.
This study does not represent direct human subjects research and was a secondary analysis of de-identified data from STTR studies; each original study had IRB approval. This study was completed under a data sharing agreement with the STTR Data Coordination Center in which all authors agreed to respect and protect the privacy of the original participants.