It is easy enough to calculate sample size and power in simple comparisons, but one frequently has to interpret results which originate from analyses, in which a small number of cases (10 - 20 per group) have been tortured by subjecting them to complex multivariate models (MACOVAs, regression models with multiple covariates etc.). Most of such exercises are evident nonsense, but is there a good rule of thumb for evaluating required number of cases in these complex situations?
Tanks Jan! You are absolutely correct. I formulated my question poorly. I'm thinking of non-simulated real life situations, like when one has to both visit the mother-in-law and approve or reject 100 congress abstracts during one weekend. Rules of thumb are welcome sometimes.
In aviation industry, where cost for samples can be exorbitant, for a given confidence in the test, we looked at the marginal increase in information for each extra test sample. This is the negative derivative of student’s t test. It seemed to work well.
I am unsure of the source for this, but it is frequently quoted that between 10-20 observations should be present for each variable added to a multivariate model ...I have found this works pretty well. As always some notion of the most important predictors (and a primary exposure) should be present before analysis. Finding the minimum AIC and BIC values can help identify the model which optimally balances goodness of fit to data with parsimony (i.e. explaining data with as few determinants as possible)....This is working the other way though i suppose to say how many variables can most efficiently explain a set of data.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size