2
$\begingroup$

In clinical trials, stratification may be used to ensure balance of treatments across covariates. According to guidance documents (e.g. EMA/CHMP/295050/2013) stratification variables should usually be included as covariates in the primary analysis.

What is the rationale of including the stratification variables in the model?

Improve this question
$\endgroup$

2 Answers 2

Reset to default
2
+50
$\begingroup$

There are several.

  1. Since stratification variables are chosen because of expected association with the outcome variable, including them can help control for confounding.
  2. Controlling for these variables allows for more precise estimation of the treatment effect.
  3. Relatedly, including the variables allows for controlling the variability that they have on the outcome variables, this is important for clinical trials where the populations might be heterogenous.
  4. including the variables helps ensure that the balance achieved by stratification is present in the model.
  5. Bias reduction caused by heterogeneity among the patient population's treatment groups.

In short, if I want to stratify by sex, I cannot determine how sex impacts the treatment effect without controlling for sex. Eg. Drug A has a treatment effect of 42 for women, of 35 overall, but in men it is only 21. Without including dummy variables for sex, you would not be able to determine that the drug is twice as potent in women than in men. This was actually a problem with ambien (although in that instance women were excluded from the trial completely). Because Ambien is more potent in women, it resulted in women driving under the influence the next morning and getting into accidents. Had women been included in the study and stratified, they would have been able to discern the difference in effect sizes and adjusted the dosages for women.

Improve this answer
$\endgroup$
5
  • $\begingroup$ Thanks for your input. But then, why should we care about stratifying the randomization (which may be difficult from a logistic perspective) if we adjust the analysis anyway? $\endgroup$ Commented May 31, 2024 at 9:15
  • 1
    $\begingroup$ Right, it's not always possible to stratify because you may not have the information (say, gender) you need before you start the RCT. Then you just adjust for gender in the analysis. However, if you stratify you will generally have less variance in your estimate than if you just adjust after the fact. $\endgroup$ Commented May 31, 2024 at 10:07
  • $\begingroup$ Short answer: it’s more precise and more accurate .(lower variance). Also, another reason to stratify is to ensure proper sample sizes for subgroup analysis. This proportionate sampling ensures representativeness. An undergrad class I TA’d back in the day did a simple sample of student veterans and gender analysis couldn’t be done because N was too small. 1. ncbi.nlm.nih.gov/pmc/articles/PMC3444136. 2. sciencedirect.com/science/article/pii/…. $\endgroup$ Commented Jun 1, 2024 at 1:01
  • $\begingroup$ Those links do a decent job of explaining. But that undergrad class could have stratified by gender and oversampled female student veterans and then applied weight adjustments to compensate (if there’s a sufficient sample size for each stratum weighting can simulate proportional stratification, albeit less precisely.) But a proper representative sample has a frequency distribution for important variables that matches the population distribution. Stratification helps ensure that the least represented level/subgroup of a variable has a proper sample size and the rest are calculated from that. $\endgroup$ Commented Jun 1, 2024 at 1:10
  • $\begingroup$ So which is the advantage of adjusting? Or when should you adjust over stratify if you have the choice and possibility? Because I usually see adjustments for sex but not that many studies with stratification $\endgroup$ Commented Dec 8, 2024 at 22:10
4
$\begingroup$

One way to do a test in a randomised trial is re-randomisation. That is, you generate the sampling distribution under the null by re-running the randomisation and dividing up the study participants according to the new imaginary assignments, and repeating this a lot of times.

For the usual statistical models, the re-randomisation test with no strata is close to the test (especially the score test) with no adjustment variables. Stratified randomisation, however, implies that between-stratum variation in the outcome doesn't contribute to the uncertainty in the treatment effect. If you do stratified re-randomisation, the between-stratum variation doesn't contribute to the variability of the null distribution, and the same is true if you adjust for the stratifying variables.

That is, an unadjusted test is conservative and an adjusted test has the correct level (in large samples, up to the usual assumptions).

Note that unlike most of the reasons in the other answer, this applies only to variables you actually stratified on, and applies whether or not they were good variables to stratify on. You might also want to adjust for other variables that are predictive of outcome, and that's a different question.

Improve this answer
$\endgroup$
1
  • 1
    $\begingroup$ Yeah, while absolutely valid tool and leads to greater precision, this doesn’t speak to why the stratum variables are included in the model in the the first place, it only offers a tool for improving hypothesis testing precision for the stratum variables after you decide to include them. Definitely a useful tool, just wanted to clarify the above point. $\endgroup$ Commented May 30, 2024 at 23:10

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.