0
$\begingroup$

I am using logistic regression to calculate the OR of falling using certain predictors (e.g. strength), age is also a confounder. The effect of age appears linear - risk of falling incrementally increases with age. Age also has a weak/negative correlation with strength.

When age is added as a covariate to the full group (n=2000), strength is significant. However, I have been advised to consider splitting my group into different age groups i.e. 2-3 age group strata. If I do this, firstly groups are uneven (e.g. old n=1500, very old n=500) and secondly strength loses significance in the old group; age also loses significance in both groups.

I struggle to understand the difference between age as a covariate and age group strata. I have seen the following post - Cox regression: age covariate within age group strata - but I am struggling to apply this to my analysis as it is a retrospective observational cohort study and the effect of age is linear.

I would be really grateful if anyone could explain (simply) how the analysis of age as a covariate and age group strata differs? And also why I might be losing significance for strength when using age strata?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

The idea of losing information about continuous age is in general absurd and will result in meaningless analyses with unaccounted-for outcome heterogeneity. The only reason to ever stratify on age in a Cox model setting is if you have serious non-proportional hazards for age that you want to partially correct for. On top of stratification you would include smooth nonlinear flexible effects for age (e.g., using a regression spline). But there are perhaps better solution. You might for example find that an accelerated failure time model fits better in some situations, and you might include an age by time interaction as a special time-dependent covariate instead of stratifying.

Improve this answer
$\endgroup$
3
  • $\begingroup$ Frank, thank you so much for taking the time to respond to my question. Apologies for being unclear in my post, I have now edited it, but my study is a large, observational cross-sectional cohort study and I am using logistic rather than Cox regression. The suggestion to stratify is based on theoretical/biological reasons i.e. a 60y old and an 85y old have very different body composition characteristics etc. However, the risk increase for every extra 5y of age seems roughly proportionate (30-40% increase with every 5 years). $\endgroup$ Commented Oct 27, 2021 at 15:17
  • $\begingroup$ I would be so grateful if you could explain how adding age as a covariate and analysing the full group (n=2000) differs from analyses stratified by age e.g. old (n=1500) and very old (n=500). Thank you again $\endgroup$ Commented Oct 27, 2021 at 15:17
  • $\begingroup$ Splitting by age pretends that age is a binary variable. That is not the case. Such splitting is invalid as it assumes a piecewise flat relationship with outcome with a sharp discontinuity at the cut point. Adjust for age flexibly and continuously (using e.g. a regression spline) and forget about stratification. Only stratifying by age will result in residual outcome heterogeneity that is left unexplained. $\endgroup$ Commented Sep 10, 2022 at 17:22

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.