treatment outside the clinical trial; similarly, some fraction of treatment group may fail to attend treatment sessions. Statistically, this noncompliance adds “noise” to the design, and possibly a bias to the estimated treatment effect size. Random assignment of treatments is nonetheless of great value, even though compliance may be imperfect. Recent developments in analytical methodology simultaneously use both the random assignment of intended treatment and the treatment actually received. See Angrist et al. (1996), Balke and Pearl (1997), Manski (1990, 1995), Robins (1989), Rosenbaum (1996, 1999a), Sheiner and Rubin (1995), and Sommer and Zeger (1991) for various approaches.
When a pharmaceutical manufacturer makes claims about the efficacy or effectiveness of a new drug product, the U.S. Food and Drug Administration advisory committees looks to the evidence from randomized controlled trials in which eligible participants have been assigned at random to different conditions (e.g., new drug regimen versus usual and customary regimen, new drug regimen versus placebo regimen). In this context, over the long run, randomization is supposed to bring into balance all of the influences on the effects of interest, but for the randomly assigned intervention status. In consequence, randomized designs can provide an especially illuminating body of evidence about the efficacy and effectiveness of newly proposed interventions. It is when the new treatment regiment is compared with a no-treatment control condition that we can gain the most complete evidence of intervention impact. Of course, “no treatment” is rarely an absolute. In reality, new medications are compared against placebo treatment while both placebo and new medication groups receive standard evaluation and non-specific care.
Despite the broadly acknowledged superiority of evidence from randomized designs with no-treatment controls when the task is to assess treatment effects, some observers feel strongly that the benefits of randomization are overstated in studies about the effects of promising therapeutic or preventive interventions (see the next section). These observers argue that randomized trials create impediments to generalizable results of immediate public health significance. In addition, when individuals are seeking treatment for their drug problems, there might be ethical or logistical barriers to the no-treatment control condition that is required to gauge an intervention’s effects completely.
It is beyond the scope of this report to settle this issue. However, we do consider some possible situations in which it would be ethical and just to make a random assignment of different interventions, including the