Dear RG members.
I conducted cox regression analysis and unfortunately the HR of some of the variables turned out extremely large, like 1.17e+09, 1.31e+10, and extremely low for some others (1.87e-21).
FYI: 24 variables were included in the regression, and before running the regression, I have checked interaction and around three variables were excluded because of this. So, why I am encountering this problem and any solution please!
Thank you in advance.
24 is an awful lot of parameters to estimate; especially if the dataset is limited. Can you tell us more about data eg number of observations; how many are in the Dead/Alive category at the end of the measurement? How collinear are the predictor variables; are they categorical or continuous? Most importantly are they excellent but conceptually trivial variables eg the response is Dead/alive and the predictor is Extremely unwell yesterday. You would get a massive HR for that , but it may not mean much.
Have a look at my answer to this not unrelated question
https://www.researchgate.net/post/Is-it-weird-to-get-a-very-big-odds-ratio-in-logistic-regression
You may wan to look at approaches that 'shrink' unreliable estimates but I would start by answering the above questions first:
R Tibshirani (1997 ) The lasso method for variable selection in the Cox model doi: 10.1002/(sici)1097-0258(19970228)16:43.0.co;2-3.
https://pubmed.ncbi.nlm.nih.gov/9044528/
Dear Professor, thank you for your help. Here are some more informations.
-Number of observations=279
-Number of failure (readmission)=47
-On collinearity test, variance inflation factor (VIF) is less than 7 (maximum is 5.41) for all variables inserted in the regression.
-The predictors, some are categorical and some continuous.
-The predictors seems scientifically sound
You have lots of parameters to estimate with only 279 observations. I suspect you have categorical predictors where nearly everyone is either Dead or Alive thereby giving your extreme estimates.
I would start simpler - what is your primary exposure - the key predictor variable of interest and then add variables to that in a meaningful way. This may help with that process http://dagitty.net/
Dear prof, I think I got the answer. As you have said I have one categorical predictor in which nearly everyone has the outcome of interest (readmission). But, what is the best way to handle this variable?
If you have one predictor that is near perfect and it makes scientific sense, why do you need anymore?
Thank you professor #KelvynJones for the invaluable insight you gave!
- Badges
- Science topic