This is the third post of a four-part series on simple statistics for clinical trials. Without too much technical detail, this series of posts is intended to offer some background in how statistical planning can influence study effectiveness and value.
The previous post in this series briefly discussed Type I and Type II errors, statistical significance, study power, and how they relate to sample size in general terms. Sample size is sometimes influenced by other considerations, as well, such as the need to avoid bias.
Bias
In discussing bias related to clinical study data collection, human factors need particular attention. Selection bias occurs when there are overall differences in the sample population that may make the sample not representative of the population of interest. This can occur in many ways but is strongly tied to recruiting methods, randomization, and how the sample for research is obtained. Expectancy bias occurs when researchers are more likely to observe the results they expect to see.
Bias can also be introduced by the study subject. Social desirability bias occurs when subjects provide responses that they think will be viewed favorably by researchers and peers. Response bias occurs when subjects respond with only information they want to share.
Confounding bias is one of several forms of study bias that does not arise from the human condition. Confounding bias occurs when study data is distorted by an extraneous, unintended variable leading to a spurious conclusion based on the variable’s association with exposure and outcome.
These are just a few of the many forms of bias that can improperly influence clinical study results. Qualified clinical statisticians can help the clinical study planning team to avoid bias from the outset.
Study populations
Clinical studies contain various “populations” (also called sets), which are essentially different ways of sorting data by the subjects’ relationship with the test product and the study procedure. We offer these terms so you’ll recognize them when you hear them mentioned in talks about your study.
The Safety Analysis Population includes all study subjects who received at least one exposure to the test product. The Per Protocol Population (PP) includes subjects who completed the study according to the instructions. The Clinically Evaluable Population includes those whose participation resulted in data that was able to be included in the study results.
The largest study set is the Intent-to-Treat Population (ITT), defined as all subjects involved in the study, regardless of randomized group assignment and level of study completion.
Randomization
You may have noticed above that sometimes not all subjects receive the same test product. Randomization is the process that objectively assigns clinical trial subjects to one group or another within the same trial in order to help avoid bias in study data. In a very simple example, a study may require that some subjects be assigned to use Shampoo A and some to use Shampoo B.
Simple randomization can be achieved by methods as commonplace as flipping a coin or drawing an odd- or even-numbered card from a deck, but the very simplicity of such methods can result in unbalanced groups within a study.
There are several other, more robust randomization techniques. Block randomization keeps subgroup sizes balanced. This is accomplished by recruiting, assigning a subject to a subgroup, for example, gender, and then performing simple random sampling within each group to assign treatment equally among the subgroups of men and women. Stratified randomization results in subgroups that are more representative of the population of interest in both size and member attributes. In this case, for example, subjects are recruited, researchers divide the sample into mutually exclusive groups (gender) and then randomly select a sample of men and women from each group based on simple random sampling so that the number of women selected and the number of men selected are representative of the overall population. If we have 70 men and 100 women, doing simple random sampling may be biased toward women, so we could, for example, select 7 men and 10 women randomly so the number in our sample is proportional. Covariate adaptive randomization assigns individuals to a treatment group for the purpose of balancing that group’s attributes based on potential covariates of interest, like age, for example.
Randomization software is available to quickly and effectively randomize subjects in a way that best supports study objectives. The professional statistician on your clinical study planning team can help if you want to know more about randomization.
Blinding (or Masking)
Even more important in avoiding bias in clinical studies is the process of blinding, also known as masking. Blinding is particularly important when study outcomes include comparisons of qualitative data (see Part I of this series for more on data types) such as comfort or pain relief.
In a blinded study, subjects and/or researchers don’t know specific details as to which group subjects have been assigned to. As mentioned in the previous example, a subject may be randomized to use either Shampoo A or Shampoo B. In a single-blinded study, the subject will not know which shampoo he or she is using. In a double-blinded study, neither the subject nor the researcher knows which group the subject is assigned to. Not all studies need to be blinded, and some cannot be blinded at all.
Reducing bias clarifies results
Because both clinical study subjects and research professionals are human, bias prevention must be built into every clinical study. There are many forms of bias that can influence the objectivity of study results at any stage of the trial. Some forms of bias can be reduced in the initial study design, and others during data analysis.
Talk with the qualified clinical statistician on your professional clinical research planning team about what methods should be used to reduce or eliminate bias in your next study.
