1
$\begingroup$

My question has three levels of complexity, I will try to explain one by one. Please accept I am no professional in statistics and still trying to broaden my understanding.

1) I would like to investigate the effect of continuous variable (X1) on incidence of a disease. I have complete data on positive cases (time and age of diagnosis, and value of X). However, for the negative cases (without the disease) I have only epidemiological data - exact numbers of people from statistical office and the distribution of X (mean and SD) taken from populational study performed a few years ago. However, I lack the knowledge on specific value of X for each of the people without the disease. Can I create an "artificial population" with particular N, mean and SD and then use it in regression analysis?

2) If the disease I am modelling is rare (0-heavy) should I use Poisson regression or negative binominal model?

3) If I have more data I would like to include in the analysis (sex and city/rural residence) can I do this in the same way as in 1) i.e. distribute males and females and city and rural residence randomly (but summing to the numbers I have from statistical office) among this "artificial population"?

4) Can I really use the statistical data on X from a study performed a few years ago? It assessed the value of X in a representative sample of population and created percentile charts which are still used to assess X in patients. However, I am afraid that with years the population may have changed and I may have some false positive findings when comparing the patients (in whom, the X change over time, if one exists, is reflected in gathered data) to fixed-in-time population distribution.

$\endgroup$
1
$\begingroup$

A Bayesian approach to include your historical data in the analysis is to use power priors. A power prior can be specified setting a parameter $a_{0}$, which will weight your historical data. The power prior distribution of your parameter of interest so called $\theta$ is given by:

$\pi(\theta|D_{0}, a_{0}) \propto L(\theta| D_{0})^{a_{o}} \pi_{0}(\theta|c_{0}) $

Where $L(\theta| D_{0})$ is the likelihood of the current study, $c_{0}$ is the hyperprior parameter that controls the impact of the $ \pi_{0}(\theta|c_{0}) $ and $a_{o}$ is a scalar that controls the influence of the historical data.

I think that the frequentist version is to pool all your data and compare the results with a stratified analysis by class.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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