2
$\begingroup$

Usually the condition of the validity of a logistic regression is to have 10 events per predictor.

In our model the binary outcome variable (1 if Healthy aging ; 0 otherwise) has a frequency of healthy aging for around 1000 observations (37% of the sample).

Also there are 13 predictor variables, so the assumption of having 10 events per variable is widely satisfied.

But the question is the frequency of 37%.

Edited:

In this logistic regression model, we can't interpret odds-ratios as relative risks due to high event rate. What other regression methods can we use to model a binary outcome?

Improve this question
$\endgroup$
7
  • 3
    $\begingroup$ I often find that 20 events per predictor are required. But this is oversimplified because you need at least 96 observations just to estimate the intercept. I don't understand why you think a proportion of 0.37 for $Y=1$ could be a problem. And isn't healthy aging a continuous concept? If you used any dichotomization to arrive at $Y$ you'll get an inappropriate analysis. $\endgroup$
    – Frank Harrell
    Commented Apr 14, 2016 at 11:49
  • 1
    $\begingroup$ Poisson models are for count data for $Y$. Your $Y$ dichotomized several continuous variables. Although it's challenging to put together multiple outcomes into one variable, it's worth doing because binary $Y$ has minimal information/minimum power. $\endgroup$
    – Frank Harrell
    Commented Apr 14, 2016 at 12:24
  • 1
    $\begingroup$ Did all the subjects die already or do you have aome right- and/or left-censoring here? $\endgroup$
    – Björn
    Commented Apr 15, 2016 at 6:08
  • 1
    $\begingroup$ My point was the selection bias of looking at it that way, when some may have becone unhealthy but you do not know (losses to follow-up) and some may yet do so (i.e. after the end of follow-up). Why only look at those alive? Presumably all will die eventually, what is the effect on your inference of looking at only those alive? $\endgroup$
    – Björn
    Commented Apr 15, 2016 at 9:27
  • 1
    $\begingroup$ What do you mean by using Poisson regression? Is your outcome not just a yes/no, but rather a 0,1,2,3,4...? Or do you mean approximating the binomial distribution for extremely low rates and large risk sets by a Poisson distribution? Or do you mean using a Poisson regression likelihood with a follow-up offset to fit an exponential time-to-event model? $\endgroup$
    – Björn
    Commented Apr 15, 2016 at 12:40

1 Answer 1

Reset to default
1
$\begingroup$

In epidemiology, sometimes we are interested in calculating RRs than ORs. When the rare disease assumption is not valid (i.e. event rate is more than 10%) we cannot interprete odds-ratios as relative risks in logistic regression.

In order to calculate risks in this case, we used modified Poisson regression (cf. reference below) with robust error variance.

References:

  1. Guangyong Zou, A Modified Poisson Regression Approach to Prospective Studies with Binary Data, Am. J. Epidemiol. (2004) 159 (7): 702-706 doi:10.1093/aje/kwh090

  2. How can I estimate relative risk in SAS using proc genmod for common outcomes in cohort studies? Introduction to SAS. UCLA: Statistical Consulting Group. from http://www.ats.ucla.edu/stat/sas/notes2/.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.