“I’m not going to the hospital, my father died in a hospital.”
In planning a study it’s a good practice to consider what confounding variables you may need to look out for.
Confounding variables are associated with the predictor (independent) and outcome (dependent) variables, but they are not in the causal chain. In the above example, disease is likely the predictor variable, death is the outcome variable, and going to the hospital is a confounder. Of course, this assumes the death was not iatrogenic; then the hospital would be in the causal chain.
Patients may be selected for interventions based on severity of disease, functional status, education level, and other factors, and these may be confounders.
Confounding can be addressed at the design stage, by:
Specification – excluding patients with the confounder (often not feasible)
Matching – selecting cases and controls matched by confounding variable levels
Randomization – randomly select patients for an intervention and hope confounding variables will balance out
It can be addressed in the analysis stage by:
Stratification – analyzing data in strata defined by confounding variable levels
Adjustment – mathematically adjusting for the confounding variable (usually by regression)
References
Adapted from Hulley SB, Cummings SR. Designing clinical research, 4th edition, Lippincott, Williams and Wilkins, 2013.
