The beauty (and truth) of randomization

One very important distinguishing feature of many studies is whether it is a randomized controlled trial, or just a regular controlled trial.

In a regular controlled trial (also referred to as a parallel cohort, or a study with a control group), subjects are either selected or self-selected to go into one group or another. You often see studies in which they simply refer to some control group. Sometimes (particularly in fitness studies), subjects are asked to bring a friend along who is going to act as their control.

The big downside to this sort of selection process is that there’s really no way to assure or even attempt to assure comparability between the two groups. Self-selection is notoriously known to bias one group from another group. And so-called “arbitrary” selection or assignment is similarly tricky. This includes assignments that are alternating (i.e. subject 1 goes into group 1, then subject 2 into group 2, subject 3 into group 1, subject 4 into group 2, and so on) as well as assignments according to birth date (e.g. all subjects born in odd years/months into one group) or ID numbers (e.g. drivers license numbers, hospital numbers, student numbers).

It is the non-randomzied study that has really given a lot of research a bad rap. And it also the reluctance on sport science researchers to embrace the randomized controlled trial that has caused so many lay-people reading articles and abstracts to give the obvious critique, “Well, this study isn’t useful because it’s impossible to track everyone’s diet/exercise/sleep/loading/unloading/rest/age/height/weight/hormones, and thus we can’t say that group A did better than group B because of the exercise program/diet/primary intervention.”

The big difference between a randomized controlled trial and non-randomized one is that a randomized trial DOES enable us to draw exactly that kind of conclusion.

The goal of randomization is not to control for every single variable that could possibly affect the outcome, but rather to create “equal”, or comparable groups without selection bias. With two groups that are essentially the “same”, with only the actual intervention being different, we are able to draw conclusions about the intervention’s effects (or lack thereof).

Randomization (particularly in large sample sizes) ensures that whatever confounding variables exist that might affect the outcome (known or unknown to the researchers), that they are equally distributed between the two groups. Since every subject has an equal chance of being in either group, any characteristics they might have also have an equal chance of landing in either group; and thus the distribution of any characteristic is going to be equal in one group compared to the other.

So, if we’re thinking about confounding variables like caloric intake, randomization basically ensures that there will be an equal number of high calorie consumers in both groups and equal number of low calorie consumes in both groups. And, if the study happens to show that exercise causes weight loss, compared to no exercise, that means that in spite of the calories consumed by individuals, exercise still causes weight loss.

In the case of the HIIT vs Steady State study out of Australia, there were some comments on JPFitness about tracking diets and caloric intake. Now, the randomization method used in that study was less than ideal, but if we assume for the sake of argument that it was passable, then what randomization does for that study is basically ensure that there were equal numbers of people who ate lots of calories in both groups. So, despite there being subjects who ate a lot being in _both_ groups, the groups still performed differently based on whether they were in the HIIT or steady state group (even if that difference was very small). One of the reasons why the study falls short, however, is that there is the possibility of some post-randomization bias (alas, a topic for another day).

Now, sometimes, randomization schemes don’t work. It’s more rare in large sample sizes and at higher risk in small sample sizes. After all, if you’re assigning people totally at random, those same random forces could theoretically put all the fat people into one group and all the thin people into the other group by chance alone. (Understanding why this is, or how likely it’s going to happen requires understanding binomial probability theory). In a case like this, as a research designer, if there’s a variable you’re not willing to leave entirely up to chance because it’s a really important variable, you can do what’s called stratified randomization, which basically is still random assignment to one group or another, but also ensures that you’ll have equal numbers of the variable you’re concerned with (in the case of fat and thin people, you’d be randomly assigning people to exercise or no exercise, but also making sure that there was an equal number of fat people in each group, and an equal number of thin people in each group).

Randomization is almost a field of study unto itself. There are a myraid of randomization schemes–almost as many as there are training methods! And just as there are poor training methods and good training methods, there are poor randomization methods and good ones as well. But the take home message here is that if you’re reading a study, or abstract, you should pay careful attention to whether subjects are randomly assigned to groups or not–because it makes a HUGE difference in a study’s interpretation.