Case control studies are used to identify potential risk factors usually through logistic regression. One possible extension of this procedure is to use logistic regression coefficients as weights and build a risk prediction model. However, in some literature, case control studies are not recommended for risk prediction models and the reason given was case control studies have some bias in data collection. Why this argument does not apply to identification of risk factors (may be through logistic regression approach)?
The problem is that case comparisons do not give you prevalence without doing something more. You have selected on the outcome so that you cannot study prevalence as you have chosen the number of controls and cases and not nature! Eg if it is 1 case to 1 comparison, the prevalence is 50% by design it cannot be anything else.
The risk from an exposure is given by the estimated slope in the logistic model - but the intercept is the result of the design and cannot simply be applied to the population to get the general risk.
In a case control study you over-sample cases (selection bias). Risk ratios can not be determined only approximated if prevalence of disease is low. Only odds ratios could be determined.
That argument applies to identification of risk factors too. Due to different biases that may be confronted in these studies like recall, selection and interviewer biases, risk factors can not be determined truely even if logistic regression is used. I attached a link.
http://www.healthknowledge.org.uk/e-learning/epidemiology/practitioners/introduction-study-design-ccs
Article Design matters in patient-level prediction: evaluation of a ...
- Badges
- Science topic
- Similar topics
- Methodology
- Research Methodology
- Research Papers