|Source: Paul at FreeDigitalPhotos|
Randomized placebo controlled trials (controlled trials, for short) are arguably the best experimental design to understand human biology and behavior. Unlike observational epidemiology, such as cohort studies that can only observe causal associations, controlled trials are able to determine the causality between an exposure - think treatment - and an outcome. This is because controlled trials best approximate the "counter-factual," which is the impossible ideal where by the same subject is studied at the same time with and without an exposure. This is accomplished by randomizing sufficiently large numbers of people into a control and an experimental group(s) so that any confounding characteristics are equally distributed between the groups to remove their influence on the exposure-outcome relationship, thereby studying the "same people" at the same time. However, this assumes that both groups maintain randomization.
If a trial is sufficiently long enough, let's say between six months and a year or more, then some number of subjects will drop-out of the trial. As mentioned above, subjects are randomized into two or more groups, and then demonstrated to be identical at baseline, at least for any known pertinent characteristics. For example, groups will be shown to have the same distribution of body mass indexes, age, and exercise level. However, as the trial progresses, some subjects will leave the trial, which could result in 20% to 40% attrition. Every publication of a trial documents the similar baseline characteristics, the number of subjects in each group, and the attrition (See figure). Researchers must keep track of who leaves the trial and why.
A particular concern is that subjects will learn what group they are in, and if they are unsatisfied with their assignment, may leave the trial and create a selection bias since the groups are no longer identical. This is one of the reasons to blind the subjects, that is, give them a placebo treatment. But this is usually easier said than done, especially in dietary trials, which is a concern for this blog. One problem is that there is no sugar pill equivalent for a diet since we are all very aware of what are eating and have some idea, whether right or wrong, of the nutritional implications of our food. The A-to-Z trial* compared the effectiveness of several popular diets, including the Atkin's Diet and the Ornish Diet, on weight-loss and metabolic parameters. The Ornish Diet is essentially a whole-foods nearly-vegan diet, while the Atkin's Diet is, well, the Atkin's Diet. Unless you've been living under a rock, you would already have a preconceived notion of the Atkin's Diet and might even leave the experiment after watching an episode of Dr. Oz half-way through the trial. Fortunately, the investigators documented how many people left the trial and why, and there was no apparent bias. Although the degree of adherence was another matter.
The placebo is also necessary to ensure that the groups are exposed to the treatment in the desired manner - this is to avoid what is known as bias by differential misclassification. This was particularly evident in a primary prevention trial for children at high risk for diabetes. Previous non-trial research has shown a link between the age of dietary gluten introduction and the onset of diabetes, so that researchers hypothesized that delaying the introduction of gluten from 6 months of age to 12 months of age would prevent pancreatic islet auto-immunity and subsequent diabetes. While this was a preliminary study designed to test the feasability and safety of such an intervention, it was still disappointing that the results were null. But for our purposes, it demonstrated the placebo problem. Subjects, by and large, did not leave the trial, but many opted to ignore their assignment. At least 15% of subjects switched from the control group to the experimental group because they perceived a greater chance of success by delaying gluten exposure. The researchers did analyze the initial assignments, and also the "new" experimental and control groups to avoid the differential misclassification bias. However, this later analysis means that the groups were no longer randomized and that the experimental group now had a greater proportion of participants that are willing to switch groups - and any other characteristics (highly motivated, well-informed, aggressive?) that these types of people have. Thus, the actions of the subjects can make our initial assumptions, that both groups are random and appropriately exposed, invalid, or at least, less valid.
Experimental trials are the most sound means for determining the causal association between an exposure and an outcome, but they still have inherent flaws that we must anticipate. And his extends beyond mere methodological considerations and offers practical implications. First, it may help to interpret clinical trials based on how well you match the study population - are you the same sex, age, BMI classification, and would you yourself participate in such a trial (this may help gauge your behavioral similarities with the study). Second, the results of an experimental trial should not be regarded as dogma. One clinical trial with little grounding in basic science (e.g. homeopathic medicine) likely needs much more investigation; whereas several clinical trials that refute modest epidemiological evidence (e.g. dietary saturated fat and heart disease risk factors) probably clarify the true association. We cannot study ourselves as though we are lab rats, so let's not interpret our studies as though we are.
*In case you didn't look at the link, the Atkin's Diet surpassed all of the other diets, including the Ornish Diet, in terms of weight loss and improvements in cholesterol/metabolic markers after one year in pre-menopausal women. Perhaps this changes your preconceptions a bit?