5

I'm working with a survey that uses a rolling data collection format (i.e., there are multiple waves of sampling and initial contacts). I'm trying to develop a model to predict how likely a sample member is to respond to the survey within 7 weeks from today. Predicting whether a respondent who is first contacted today will respond within 7 weeks is fairly straightforward - just a basic propensity model predicting response within seven weeks of initial contact. However, predicting whether a case that's been in the field for several weeks already will respond in the next 7 weeks is more difficult.

My question is how do I take into account that the probability of response changes based on how long the case has been a nonrespondent (i.e., a respondent's initial probability of response may have been .75, but if they're still a nonrespondent after 4 weeks, they'll probably remain a nonrespondent)? I could just include the amount of time the case has been in the field as a variable in the propensity model, but I'm not sure if that's the appropriate way to handle it. It seems like this may be a situation for survival analysis, but my knowledge in that area is limited.

Any suggestions of an approach, model, or previous research I should consider would be much appreciated.

soupsandwich
  • 95
  • 4
  • 1
    This does sound like survival analysis might be the best approach. On a different note, I think [propensity-scores](http://en.wikipedia.org/wiki/Propensity_score) might be a little different from what you're thinking about. PS are used for observations within a *predictor variable* to aid in causal inference from observational data. – gung - Reinstate Monica Jul 26 '12 at 16:11
  • @gung Thanks for the insight. There's some overlap and informality in jargon the literature, this type of work is sometimes referred to as response propensity scoring. At any rate, my main question with survival modeling is how do I deal with the fact that the probability of response by a given time increases for a bit and then begins decreasing? Is this done via a transformation of the dep. var? Again, I'm not familiar with survival modeling, so any references would be appreciated. – soupsandwich Jul 26 '12 at 18:40
  • To (possibly) clarify my last comment, it's not that the probability of response **by a given time** increases then decreases, it's that probability of response **on a given day (or within a given rolling window) after first contact** increases then decreases. – soupsandwich Jul 26 '12 at 19:10
  • You could use a discrete-time logistic hazard model, with this model you could include the number of contacts that you already do. – Joe Apr 03 '16 at 07:37

0 Answers0