Prof. Iven Van Mechelen, Dr Elise Dusseldorp & Dr Katrijn Van Deun
On September 30th, 2016,, Lisa Doove defenden her thesis entitled
For many medical and psychological problems, multiple treatment alternatives are available. An obvious question in such cases pertains to whether there is one globally best treatment alternative for the full population of clients under study, or whether the best treatment alternative varies over subgroups of clients that can be characterized in terms of pre-treatment characteristics. Formally speaking, the best treatment alternative varying over subgroups of clients may be referred to as a qualitative treatment-subgroup interaction, that is, in the case of two treatment alternatives A and B, an interaction that implies that for some subgroups of clients treatment A outperforms treatment B, whereas for other subgroups the reverse holds true. The detection of such interactions implies a clear need for the development of so-called treatment regimes, that is, decision rules that assign to each client a treatment alternative, out of the set of available treatment alternatives, based on his/her observed characteristics. The optimal treatment regime then is the one leading to the greatest expected outcome in the population under study.
For the detection of treatment-subgroup interactions and the estimation of optimal treatment regimes suitable statistical methods are needed. This is especially troublesome if the relevant subgroups of clients are unknown and are to be learned from the data. Recently, a promising class of tree-based methods has been proposed for this purpose. Unfortunately, however, this family of methods also goes with quite a few challenges. In this doctoral dissertation, we will address five of these in five consecutive chapters.
In Chapter 1, we focus on tree-based methods for the detection of subgroups involved in treatment-subgroup interactions. A major problem with regard to these methods reads that these have been developed almost independently, and that the relations between them are not yet fully understood. We clarify the conceptual basis of a selection of five tree-based methods, capture the relations between them, and give practical advice for end users with regard to the selection of a suitable method among them.
In Chapter 2, we focus on one tree-based method for the detection of subgroups involved in treatment-subgroup interactions, called QUINT (Dusseldorp & Van Mechelen, 2014), which is to be singled out because of its relevance for optimal treatment assignment (due to its exclusive focus on qualitative treatment-subgroup interactions instead of treatment-subgroup interactions in general). However, the paper in which QUINT has been introduced is not easily accessible for non-methodologists. We present a nontechnical review of the conceptual basis of QUINT and show its significance for psychological applications.
In Chapter 3, we address the problem of estimating optimal tree-based treatment regimes. Current methods for estimating such regimes either do not formally optimize an estimator of the expected outcome, or make various assumptions about models underlying the data (which may be prone to misspecification) and build the regimes in a greedy way. We present a novel approach for estimating optimal tree-based treatment regimes that directly maximizes an estimator of the overall expected outcome without making assumptions about models underlying the data, and with a look-ahead strategy to overcome greediness in the tree building.
Methodology for detecting treatment-subgroup interactions
For many medical and psychological problems, multiple treatment alternatives are available. A standard research question in such cases pertains to relative treatment effectiveness. A typical setting for the study of such a research question is that of randomized controlled trials (RCT’s), in which the persons under study are randomly assigned to different alternative treatment conditions. Beyond some treatment alternative being globally best, treatment effectiveness may vary over groups of persons that can be characterized in terms of pre-treatment characteristics. The latter results may have significant consequences for the development of optimal treatment assignment strategies. The cornerstone for the development of such strategies is the detection of subgroups that are involved in meaningful so-called qualitative treatment-subgroup interactions, that is, interactions that imply that for some groups of persons treatment A outperforms treatment B, whereas for other groups the reverse holds true.
First, we will develop and implement a methodology that, given data from simple RCT’s with a large number of background characteristics and one or more outcome variables, induces subgroups that are involved in sizeable qualitative treatment-subgroup interactions if these should be present in the data. Second, we will develop extensions of this methodology to more complex RCT’s that induce more than two treatment alternatives. Third, we will control the correctness and reliability of the inferences that result from the to be developed methodology. Throughout, the methodology will be applied on real and simulated benchmark data sets and evaluated in comparison with alternative methods for the detection of treatment-subgroup interactions.