Modeling fractions with stepwise selection

Viewed 20

We're trying to predict the fractions that people move from one postal code to another (which can also be the same postal code). Based on actual moves we have for every postal code the fraction to the other postal codes (e.g. postal_from: 1234, postal_to: 1234, fraction: 0.60 postal_from: 1234, postal_to: 1235, fraction: 0.30 postal_from: 1234, postal_to: 1236, fraction: 0.10 postal_from: 1234, postal_to: 1237, fraction: 0) We have quite a lot of predictors, so we would like to use forward selection to select which predictors to use in the model. Examples of predictors are number of households, kilometers between the postal codes, average purchasing power, etc.

We tried to fit a quasi-binomial model, but the problem is that this method doesn't give AIC, so we can't use stepAIC() in R. Which method would be suitable and what steps should we take when modeling the mover fractions?

1 Answers

you can avoid using stepAIC() by selecting your own variables (predictors) try to use correlation tests first , keep only significant varabiables. And then add your variables one by one so that your MAE and MAPE getting smaller ( i assume your have a train and test data parts)

Related