Used in the context of clinical trials, stratification refers to dividing up study participants and/or outcomes into subgroups (also called strata or blocks) based on a characteristic besides just the treatment or exposure. Stratification can be random (such as selecting a certain percentage of participants from different groups to represent a real-life population) or non-random, in which a predetermined number of participants are selected from different subgroups (such as a quota for different races/ethnicities in a trial).
Deeper dive
The idea behind stratification is that since an intervention (or an exposure in observational studies) may affect different people in different ways, it’s important to look at the outcomes within subgroups to find out if such differences exist. In a trial investigating Drug A versus placebo, for example, researchers might stratify participants in terms of their symptoms or demographics to look at whether the drug affects one group (such as those with more severe symptoms or those of a specific ethnicity) more or less than another. Or, in an observational study looking at lead exposure and effects, participants might be stratified according to their age of exposure or their duration of exposure.
Although stratification can provide a more precise understanding of possible outcomes from an intervention or exposure, too much stratification can lead to falsely inflated significance or p-hacking: the more subgroups you have, the more statistically likely it is that something significant will emerge from one of them. In covering studies that stratified participants, then, it would be helpful for journalists to ask why the researchers stratified the patients the way they did, whether the population was stratified too much or not enough, and whether stratification substantially affected the findings. If the study involves a lot of subgroups, it might be wise to ask if the researcher used some kind of statistical correction method to reduce the likelihood of a randomly occurring significant finding, such as a Bonferroni correction (though many other types of correction exist).