 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Services Research Outcomes Study (SROS)
|
|
|||||||||||
|
|
| |||||||||||
|
Services Research Outcomes Study (SROS) | ||||||||||||
ANALYTIC TECHNIQUES EMPLOYED: BEFORE/AFTER ANALYSIS AND REGRESSION
Before/After Comparisons
The SROS study compares events in the five years preceding admission ("Before Treatment") with the five years after discharge ("After Treatment"). The study describes and quantifies the changes between the periods for all clients and shows how the changes vary for the entire population and among subgroups of SROS clients. These subgroups are divided according to pretreatment characteristics, such as age, sex, race/ethnicity, and earlier treatment exposure; and aspects of the SROS treatment episode, such as the type of treatment, its duration, and whether the client completed or did not complete the treatment protocol. Comparisons are made between rates of occurrence on dimensions such as drug use, criminal activity, employment, living arrangements, and physical health.
The effects or outcomes of treatment are evaluated using two methods. The first is commonly known as a "before/after" or "pretest/post-test" design, and the second is regression analysis. The before/after design compares behaviors or other characteristics in a panel of the same research subjects, measured identically or comparably before and after the SROS sample treatment episode. The before/after design has some strengths over methods that compare two different samples of individuals to estimate the effects of an intervention, using a group never treated compared with a group after a round of treatment. In the before/after design, each subject serves as his/her own statistical control, keeping spurious or coincidental relationships from entering into the results. Specifically, behaviors or characteristics that tend to be permanent (e.g., gender, race/ethnicity) or slow to change (age, lifetime employment history, educational level) during the intervention exert a constant influence during both the before and after periods and may therefore be eliminated as competing explanations for a treatment’s effects. In addition, clients with similar types of premorbid behaviors or characteristics can be grouped, and treatment effects contingent on these factors can be examined by comparing subgroups.
All of the before/after changes reported here are net changes in a group statistic — for example, the percentage reduction in the proportion of all clients who used cocaine five or more times in the five years before the SROS sample treatment episode versus five years after treatment, or the percentage that committed burglary in the five years before admission to SROS compared with the five years after discharge from the SROS treatment episode. This analysis uses group change because for every behavior or characteristic measured as an outcome variable, there were individuals who changed in either direction, not only in the overall SROS client group but in every major subgroup of that population.
The "chronic relapsing disorder" view of drug treatment would expect little difference for the five-year "before/after" SROS differences. This lack of difference would be attributed to the regularly predicted relapse period, as the long outcome period used by SROS would show no treatment effect if relapse were a regularly occurring issue for a large proportion of the population discharged from drug treatment.
The SROS before/after analyses tested relapse behavior in two ways. First, the behavior (primarily drug use) was examined of those discharged from treatment along the basic dimension of use/no use of drugs. Any relapse during the five year period would result in findings that drug treatment is not effective, as the period of hazard is five or more years. Another more subtle test
of effectiveness is achieved by comparing the number of days per average month that individuals use drugs. Both measures are employed in this study.
Regression Techniques
While the before/after design focused on the entire sample and on subgroup level, regression analysis focused on the individual level. This section describes two predictive regression models for predicting outcomes associated with drug treatment.
It should be noted that regression analysis used in a study of this kind is a correlational technique. No claim is made for a direct causal relationship among the variables used in this analysis. Rather, the analysis seeks to explain covariation among the variables, that is, to assess whether certain behaviors or characteristics tend to coincide with the presence of a particular outcome to a greater or lesser degree than do others.
Also, the variables used in these regression analyses were not intended to be exhaustive of the data set nor of other potentially meaningful relationships that could be assessed from this rich data resource. The variables used in this regression analysis reflect the study’s interest in a model, explaining post-SROS treatment behavior through three (pre-treatment, treatment, and post-treatment) cumulative models.
Two types of regression analyses were employed. For continuous outcome variables (e.g., number of days per month used heroin), ordinary least-squares regression models were used. For dichotomous (binary) outcome variables (e.g., used heroin after discharge from SROS treatment episode), logistic regression models were used.
Continuous Variables
The principal model in the analyses of continuous variables is called the "conditional change" model:
YAFTER = _ + _YBEFORE + _ ßi Xi ++ _ ßii Xii + _ ßk Xk + e,
where YAFTER denotes the value of a continuous outcome variable measured for the after-SROS treatment period; YBEFORE denotes the value of the same outcome variable measured for the same individual for the before-SROS treatment period, the Xi s are other explanatory variables, Greek letters represent regression coefficients, and "e" is a random error assumed to have a mean of zero, have a constant variance, and be uncorrelated with the explanatory variables.
The distinguishing characteristic of the conditional change model is that the inclusion of the regression coefficient _ explicitly allows for causal dependence of an individual’s outcome after treatment on the individual’s status before treatment. This is a major strength of the conditional change model. For example, it makes sense to think that whether or not an individual used a particular drug during the before-treatment reference period affects the likelihood that one will use the same drug during the after-treatment reference period. Similarly, given the importance of previous work experience in obtaining and keeping a job, an individual’s employment status before treatment seems likely to affect employment status after treatment.
Dichotomous Variables
For dichotomous variables, logit analysis was employed. To illustrate the model, let D2 denote the after-treatment measurement of the dichotomous outcome (e.g., whether or not the respondent reporting using heroin after treatment). Then D2=0 if no (did not use heroin after treatment) and D2=1 if yes (did use heroin after treatment). Let D1 denote the before-treatment period of the same dichotomous outcome. In other words, D1=0 if did not use heroin before treatment and D1=1 if did use heroin before treatment. Let X and Z denote the other explanatory variables, which may be either dichotomous or continuous. (NOTE: two explanatory variables are used simply for illustrative purposes; additional explanatory variables would not change the form of the model.) Then the "unified model" for dichotomous outcomes is simply:
logit(D2) = b0 + b1*D1 + b2*X + b3*Z
|
|
|
 |
Click to Return to OAS Home Page |
|
This page was last updated on August 15, 2003.
|
|
