3

I'm trying to understand confounders and I read the statement

'Age and sex are the most common confounders.'

Can someone explain why this is? I don't fully understand the concept of a confounder to be honest.

  • 4
    Can you say where you read that statement? – AdamO Apr 20 '18 at 15:21
  • Does the Wikipedia article help you? https://en.wikipedia.org/wiki/Confounding – mdewey Apr 20 '18 at 15:30
  • 1
    My personal impression is that age & sex are among the most common control variables in mainstream biomedical research. (Whether that's merited is a different issue.) This isn't the same thing as saying that they are confounders in the technical sense meant in biomedical research. A basic idea is that you should control for confounders; if we assume (rather naively, IMHO) that everyone is choosing their covariates mindfully & appropriately in accordance w/ that principle then that could be taken as evidence they are among the most common confounders. – gung - Reinstate Monica Apr 20 '18 at 15:32

1 Answers1

4

I don't think that statement makes any sense in isolation.

A "correcter" way of saying it: "Age and sex are the most commonly adjusted for confounders in multivariate analyses". I will admit that's true, but the sense of necessity or sufficiency to address confounding that way is unwarranted.

When we want to set up a causal model between an exposure and an outcome in observational research, it's important to consider the influence of confounders: these are things that are causal of the outcome and exposure such that, when ignoring their influence, the modeled association is different from the one that conditions on them.

The challenges in identifying confounders in observational research:

  1. They have to be easily measurable. Observational research is limited in the set of measures it can collect.
  2. You never actually know if something is a confounder.
  3. It's always a valid criticism that a proposed set of confounders don't satisfy the backdoor criterion.

For all these reasons, most causal analyses from observational data inevitably look back to age and sex as adjustments.

Two examples where age and sex don't confound anything:

  1. We might have genotype data on SNPs in a case-control study of cancer. Neither age nor sex can possibly confound genetics. They can possibly interact in epigenetic ways, but that's moot (you need do to GxE analyses).
  2. We might have a cohort of monozygotic twins who are discordant for a disease and measured for exposure status. The conditional analysis has 0 intracluster variability in sex and age: they are identical genetically, born within minutes of each other.

And yet, I have had reviewers suggest age and sex should be adjusted for in those scenarios. You have to laugh after a point...

AdamO
  • 62,637
  • 4
    Reviewers say the darnedest things. – Kodiologist Apr 20 '18 at 15:34
  • While the answer by AdamO is perfectly correct in the interpretation, I am not sure the examples are correct. Why wouldn't sex and age be a confounder in a cancer GWA study ? The outcome is not genetics but it is cancer, and why wouldn't sex and age be possible covariates for that ? The same goes for the second example. Yes, within a cluster there is no variability in sex and age, but the sex and age of the twins for sure can be confounders for the disease in consideration. Am I missing something ? – Santy.8128 Apr 25 '18 at 19:23
  • There's a difference between causes of the outcome and confounders. Confounders must cause variation in the outcome and selection into treatment/exposure (i.e., the variable whose causal effect is under study). The variables mentioned do indeed cause variation in the outcome, but they do not cause selection into treatment, so adjusting for them will not improve bias. They will, however, reduce the variability in the effect estimate by reducing the mean squared error, but they do nothing for bias. – Noah May 08 '18 at 17:49