Longitudinal Analysis of Health and Behavioral Health Service Use: An Application of New Statistical Methods

Background and Purpose: Health service utilization can be a dynamic process, with consumers entering, exiting, and combining treatment over time. Yet few studies yoke the collection of longitudinal data and statistical analysis to formally model how service use changes over time. This paper uses longitudinal data on behavioral health service utilization by community corrections-involved adults to exemplify two strategies for analyzing longitudinal health service utilization data: multi-level regression modeling (MLM) and latent class growth modeling (LCGM).

Methods: Data were gathered from a recently-completed study of 152 rural adults currently or recently supervised by parole/probation officers in Lincoln County, Oregon. Quantitative data on medical, mental health, and substance use service utilization were gathered from adults every 60 days via telephone interviews for 8 waves. Information on demographic factors hypothesized to be associated with these three types of service use was gathered at study enrollment, including age, income, race/ethnicity, gender, education, and insurance status. Analytically, MLM and LCGM models were used to parameterize medical, mental health, and substance use service use.  MLM was employed to examine intercepts and slopes of service use across 8 waves of data in relation to individual demographic factors. Separate analyses employing LCGM examined whether latent groups of service use trajectories were present in the participant population; post-hoc descriptive analyses then examined whether demographic differences partially defined groups.

Results: Regarding MLM findings, gender and insurance status were significantly associated with initial levels of service utilization, with uninsured women having the largest odds of service utilization, followed by insured women. Over time, insured men showed lowest odds of health service utilization while insured women showed significantly higher rates of health and mental health service utilization. Regarding LCMG findings, 3 trajectories of service users emerged across all types of service utilization with classes of medical service utilization characterized as low-level users, moderate-level users and high-level users, with gender and employment status significantly associated with class membership. Mental health and substance use service utilization trajectory classes were characterized as non-users, users with increased intensity over time and users with decreased intensity over time, with gender, employment status and income significantly associated with class membership.

Conclusions and Implications: This paper applied two advanced statistical methods to longitudinal medical and behavioral health services data. Findings suggest slight differences by analytical strategy, although both methods were able to discern demographic correlates of service trajectories. While still rarely used in behavioral health services research, these analytical methods hold promise for social work research and practice. Regarding research, combining longitudinal methodology and statistical analysis moves service utilization research towards meeting the challenges of measuring dynamic, ongoing processes, which are arguably more reflective of individual experiences in behavioral health care settings. Regarding practice, longitudinal modeling strategies can help social workers differentiate usefully between high-versus-low level service users, and help direct service delivery to those who may have a high level of service need, but have historic low levels of utilization.