Balanced groups with Logistic regressionDear members: I would like to know how critical are unbalanced treatment groups (e.g., T1= n40 vs T2= n14) when evaluating a binary response variable with logistic regression. Besides using the appropriate sample size, is there a way to determine what would be considered extreme imbalance? Is LR as demanding on balanced groups as OLS? Can JMP do a "weighted" LR to minimize the impact of imbalance? Many Thx.
AnalyticsDOERe: Balanced groups with Logistic regressionIf your sample is an honest representation of the population that can be described by a logistic regression, changing weights will distort your sample, then distort the model, then distort your view on the population. Unbalancing is not a problem for fitting a logistic regression. When one says unbalancing is a problem, there must be something else in play. Is the sample biased? Is logistic regression a right choice? Is mis-classification cost symmetric? ... Re-weighting is a patch to the first and the third situations. But the choice of weights should depend on the actual application.
