2

I have a data set of people switching brands.
Say #people switching from Brand A to Brand B, C or D.

My data looks like this
Week | Brand | %Switchers | Price Week 1 | Brand B | 10% | 10 Week 1 | Brand C | 50% | 8 Week 1 | Brand D | 40% | 9

I have such data for 104 weeks. Note that sum of switcher to Brand B & C & D will always sum to 100. I want to model the influence of Brand & Price on switching behavior.

Since this is a proportion data, I am applying beta regression model on to it. My equation is

betareg(%Switchers ~ Brand + Price)

The model coefficients are significant & model fit is decent. However, the while I predict switching behavior across brands I am finding that sum(%Switchers) along Brand B Brand C Brand D are beyond 100% (which doesn't make logical sense).

While I can see where this is coming from (the model doesn't know that sum(%Switchers) is 100% in a given week), I am stumped as to how to interpret such scenarios? Also, is there some other model with will allow me to model proportions and limit their sum to 100% in a given week.

Edit: My problem is not that some proportions are zero or one. It is that sum of proportions in a given week are 1. Hence, IMHO, zero-one inflated beta regression doesn't particularly help.

Edit: I don't know why it is marked duplicate of another question. I am specifically mentioning that the linked questions doesn't help answers my key question viz What if switching probaility across brands come out to be greater than 100%

hardikudeshi
  • 121
  • 3
  • Use the logistic regression if you have # of switchers. – user158565 Oct 11 '18 at 21:16
  • 3
    I do not see how this question is a duplicate of the one indicated. That one deals with including probabilities of exactly 0 or 1 in a beta regression, this one deals with a Dirichlet-type problem where the estimated probabilities output by a series of beta regressions can and do sum to a number not equal to 1. – jbowman Oct 12 '18 at 16:58

0 Answers0