Confounders in observational studies Flashcards
how to control for confounders
IN STUDY DESIGN:
Restriction
Matching
IN ANALYSIS
Stratification
Multivariate models
Propensity Scores
RESTRICTION
IN STUDY DESIGN.
It means performing the study including only subjects with the same category of the confounding variable, for instance, subjects with similar age (only young or only old).
CONS: This model reduces external generalizability and reduces sample size and power!
MATCHING
by matching subjects in both exposure and control groups with one or more category of confounding variable, for instance, subjects in the exposure and control groups are paired together based on their similar age (individual matching model) or the exposure and control groups are matched due frequency of confounding variables (frequency matching – e.g., same distribution of category age in both groups). This strategy is often limited to 2 or 3 relevant confounding, also it increases internal validity, but compromises external generalizability (1).
STRATIFICATION
it occurs in the analysis phase by dividing groups into strata of suspected confounding factors; the outcome is assessed in each stratum. Here the model works well with a limited number of confounding
MULTIVARIATE MODELS
IN THE ANALYSIS PHASE.
In the analysis phase by using logistic, linear or Cox regression, to adjust multiple confounding simultaneously
PROPENSITY SCORE
by generating a score for each subject based on the probability of being exposed, then the groups are balanced by matching subjects with similar scores. The model can be used in matching, stratification, and regression analysis. Likewise multiple regression models, it can be used to adjust for multiple confounding, but here a single score is obtained before the outcome assessments (pre treatment variables). Nevertheless, PS is more adequate for a maximum of 7 events per confounding, otherwise, multiple regression models are desirable
ADVANTAGES OF PROPENSITY SCORES
Easier method to explain to a nontechnical audience as groups with similar baseline characteristics (covariates) are created and compared.
The diagnostics for the efficacy of propensity scores modeling requires
just checking for balance of covariates in both comparison groups.
It is much more straightforward than regression and allows determination of the range over which comparisons can be made.
DISADVANTAGES OF PROPENSITY SCORES
Propensity scores obscure identification of interactions between treatment and confounders.
If matching by propensity scores is the method chosen, not all patients are used in analysis, only those who could be matched in both groups. That means we will lose power in the analysis.
ADVANTAGES OF REGRESSION MODELS
Estimate the effect size of each confounder and also to identify interaction effects between treatment and confounders.
Diagnostics for regression are not so straightforward as for propensity scores (i.e., just checking for balance in baseline characteristics between comparison groups).
DISADVANTAGES OF REGRESSION MODELS
Outcome regression models do not allow separation of modeling and outcome analysis as propensity scores do.
Modeling may influence the choice of covariates in the model and how they are used (squares, interaction…). Manipulating covariates, in turn, may change the strength or even the direction of the intervention on the outcome.
It is not so straightforward to explain to a nontechnical audience
how regression controls for confounders.
What models do we have for propensity scores?
There are several PS models: stratification PS, covariate adjustment PS, PS matching etc., and it is important to choose the model with caution because they can generate different results. The analysis is very complex. If there are more than seven events per confounder, it is more appropriate to use the multivariate model, otherwise PS is more accurate.
