I am working on a project that involves a population-based data set, to study the impact of a clinical variable on the mortality of patients with certain cancer. In all similar studies, age, which is normally a continuous variable, was categorized into "age groups" according to each study's methodological reasons. I have also reviewed a lot of studies that show that "discretization" of continuous variables when creating predictive models is better than using them as continuous variables because it decreases noise and thus increases the accuracy of prediction. I was told by one of my senior co-authors that I should redo the analysis after going back and using "age" as a continuous variable. Was anybody confronted with a similar issue in his research? If so, what is the most solid evidence that you've got to defend categorizing age, or any other continuous variable in clinical or epidemiological data?
Thanks Paul for your timely feedback. Your answer is definitely with going for the continuous variable. Yet I noticed that the model you chose was a univariate model. Does your answer also apply to multivariate models? In other words, if I changed from categorized age to continuous age variable within my model, I would expect a significant change in the effect, hazard ratio, or odds ratio of my predictor of interest?
I agree with Paul's feedback to you, Amr.
I sometimes adjust for age in cancer spatial studies by entering data on the age distribution in age intervals 0-4, 5-9, 10-14, .... etc.. If the software uses a Poisson model, then such age adjustment will identify clusters of cancers after the age effect has been removed. If the response variable is continuous, I can use the continuous age variable in a linear regression mode, and then I can output the residuals for predicting (say) some continuous index.
A key reference against dichotomisation has as a title that it is a 'bad idea'; many reasons listed there in favour of keeping continuous variables as such.
https://www.ncbi.nlm.nih.gov/pubmed/16217841
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis