SAS® Macros for Estimating the Attributable Benefit of an Optimal Treatment Regime

It is sometimes the case there is no general consensus on the best way to treat patients suffering from an illness or disorder. Consider a scenario where there are two competing treatments and it has been shown that one treatment works for some patients while another treatment works well for others. In such cases we may want to define an algorithm (or treatment regime) that dictates treatment based on individual characteristics. The optimal treatment regime would be the so-call “best” algorithm and minimizes the number of poor outcomes. It can be very difficult to assess the overall public health impact of such an algorithm. Attributable benefit (AB) of a treatment regime is a useful metric for assessing such algorithms by looking at the proportion of poor outcomes that could have been prevented had the algorithm or regime of interest been implemented. Here we will give an overview of the assumptions for using the attributable benefit measure and discuss two SAS macros for estimating AB for the optimal treatment regime. These macros are designed for the scenario where there is binary treatment, binary outcome, and possibly many different covariates. The first macro uses an estimator based on a logistic regression model on the outcome of interest while the second macro augments this estimator with a propensity score model to provide doubly robust protection from model misspecification. INTRODUCTION In many medical research scenarios, observational data is used to determine the effectiveness of different treatment strategies on patient health. In some cases, a treatment that works well for some patients may not work as well for others. When treatments compete like this, it may be that the best way to deal with individual patients is to develop a strategy or policy that assigns treatment to the individual based on his or her risk factors. The hope is that overall health will be improved if each patient receives the treatment that is right for them. While the terminology varies across disciplines (treatment strategy, policy, algorithm, and regime are just a few terms), the overall goal of this type of research is to identify which treatment (or combination of treatments) has the most impact on overall public health. Thus the mission here is two-fold: first to identify the best strategy for dealing with individual patients and second to assess the public health impact of such a strategy. The focus of this paper is later, to measure and quantify the public health impact of treatment policies which may or may not treat to the individual patient. Define a treatment regime as an algorithm or set of rules for treating patients. In addition, we will also call the “best” algorithm for dealing with patients to be the optimal treatment regime. We will assume that the optimal treatment regime is “best” because it minimizes the number of poor outcomes. Generally speaking, there is no limit on the possible number of different treatment regimes to handle patients who have some disease or disorder. So instead of clinical testing, observational databases have become a key component for such research, helping to identify potential subgroups of individuals who seem to respond differently to certain treatments than others. This paper will discuss one method for estimating the optimal treatment regime and two ways of estimating the public health impact. The current discussion is restricted to the scenario where the outcome is binary (i.e. disease/no disease or death/no death), treatment is binary, and there are any number of patient covariates or factors. SAS Macros have been developed to perform these analyses given certain assumptions about the data are met. CAUSAL INFERENCE The underlying theory behind estimating the effects of interest is grounded in causal inference, more specifically counter-factual data analysis. As such it may be important to review some of the necessary background information. Suppose we have data of the form {Y,T,X} where Y is a 0/1 binary outcome variable (Y = 1 is an indicator of poor outcome). T is a 0/1 binary treatment variable (0 is the standard treatment and 1 is the alternate treatment), and X is one or more confounding variables. Let’s say that the primary interest is in determining the causal effect of treatment on the chance of a poor outcome, going well beyond the idea that treatment and outcome are possibly associated. This is certainly much easier in the clinical setting where a proper experiment can be performed, but the need to make such conclusions from observational data has become increasingly more important. So consider two hypothetical worlds, one where all patients receive the standard treatment and another where all patients receive the