Is it valid to double or triple the size of a dataset in order to simulate a larger population, and would this increase the statistical significance of risk factors? I use logistic regression on an actual (but small) dataset to examine risk factors for an outcome of interest. Some of these risk factors have significance around p = .08. I want to project the findings onto a larger (simulated) dataset as a proxy for the larger population. Will this make the confidence intervals of candidate risk factors smaller, and their p-values become statistically significant?
Doubling the sample would underestimate the true population variance, because each sampling unit artificially has an identical twin. This would overestimate "significance". If you think that some of your "insignificant" risk factors are false negative results, you might consider the stability of your results by applying bootstrap techniques (given that assumptions for bootstrapping are fulfilled).
Doubling the sample would underestimate the true population variance, because each sampling unit artificially has an identical twin. This would overestimate "significance". If you think that some of your "insignificant" risk factors are false negative results, you might consider the stability of your results by applying bootstrap techniques (given that assumptions for bootstrapping are fulfilled).
Briefly no, this is not a good idea for the reasons Ulrich has identified. If you take a public health perspective it can be more important to look at the attributable risk, whihc combines effect size and prevalence, as well as the precision of the estimation of risk factors. On the whole the p-value tells you little of interest.
Sorry to say but answer is no! It will be considered adulteration in terms of statistical analysis. Bootstrapping is one method mentioned by Ulrich; however, if you want to capture the sampling variations and attributes of population in terms of clustering and population fraction, you may incorporate the probability weights and other survey design features during analysis.
The only exception I can think of is as a part of a power calculation, if you are able to justify a replication study. This would be using the existing data to make quite explicit what samples sizes would be needed to be adequately powered to find an association.
But in terms of reporting results, this is sadly a bad idea.
- Badges
- Science topic
- Similar topics
- Medicine
- Epidemiology