Nova Acta Leopoldina Band 110 Nummer 377

association analysis is wrought with problems even in this environment with particularly large streams of relatively high quality data (BEKAERT et al. 2010, VANSTEELANDT et al. 2009). Still, under well understood assumptions the richer data structure and relevant practical needs justify addressing more ambitious questions, for instance, regarding dynamic treatment strategies. Optimal Dynamic Treatment Strategies Standard randomized clinical trials benefit from the perfect instrument (randomization) but have their limitations. They are typically designed to evaluate a single, continued prescription, which in routine implementation may require modification in response to events over time. Could experimental designs be exploited more directly to learn about the adaptive treatment strategies? While well designed trials tend to be very expensive and time consuming there is a growing concern that they may still fail. This happens not just by not yielding the positive answers hoped for, but sometimes miserably: without even adding useful information in one direction or another. This was dramatically illustrated by a number of HIV prevention trials seeking to confirm the protective potential of microbicide vaginal gels, a product which could be admin- istered by the woman for her own protection. Repeatedly inconclusive answers lead to a 2008 (US) Institute of Medicine report: “Methodological challenges in biomedical HIV prevention trials” (LAGAKOS and GABLES 2008). One big contributor to the trials failures was lack of com- pliance with assigned treatment, a feature which – more than missing data – hurts trial power. The p2 rule reflects for instance how a 20 % immediate treatment dropout for a random pop- ulation sample leaves just merely 68 % = (1–20 %)2 effective sample size. The microbicide studies required women to go off treatment upon becoming pregnant. Even when women had denied any intentions of pregnancy on recruitment, many pregnancies occurred over the course of the trial. As long as the experimental drug is not guaranteed to be safe for the fetus, it is common practice to withdraw from treatment at this point out of precaution for the baby. This phenomenon contributed to several large scale phase III studies ending with inconclusive re- sults due to lack of power. Since HIV prevention trials are conducted in sexually active populations, one might better acknowledge the possibly changing status in child wish, considering too that women are un- likely compliant with spermicidal gel use in periods when a child is desired. A design exem- plified in Figure 5 could provide explicitly for such situations and could allow to study what can be done for HIV prevention in sequence over the several stages of a woman’s sexual life. Figure 5 shows how the impact of a range of interventions, including E = Education reinforce- ment, P = Pre-exposure prophylaxis, and M = microbicide gel, may be evaluated when adher- ence to condom use has dropped and how one may move to a range of other treatments once the child wish is expressed and accounted for. Offering continued trial participation and a choice of workable treatments at such time should help avoid unwanted non-adherence with the trial products and the corresponding dropout and reduced power. Similar designs have been developed for prevention of relapse in drug addicted adolescents (LAVORI 2000 and MUR- PHy et al. 2001) and for improved drug adherence in diabetes patients (BUyzE et al. 2010). Nova Acta Leopoldina NF 110, Nr. 377, 47–64 (2011) Els Goetghebeur 60