Suppose you want to give free trials to Amazon Prime but only want to give the trials to users who are most likely to subscribe since giving free trials to everyone would be expensive. What would be a good response variable for this problem? There are two flavors of Amazon: Amazon Prime and "Free Amazon". The goal is to target the most likely people to convert who use "Free Amazon". Would giving a survey and asking the respondents to rate their enjoyment of Amazon on a scale of 1-5 be a good proxy for a response variable? Would running an ordered logistic regression model be a good approach using this derived variable?

  • $\begingroup$ Sounds like you are asking the impossible. I suppose you can always construct a model but without data you have no way to check its value and you can't get sensible parameter estimates either. $\endgroup$ Jan 3, 2017 at 3:48
  • $\begingroup$ How do start-ups or companies with very little data target users who most likely to subscribe? I would assume administering a survey measuring purchase intent would be a good option. $\endgroup$
    – zorny
    Jan 3, 2017 at 6:16
  • $\begingroup$ First build a model for the latent variable effectively estimating it using other data (in this case: How often did a person actually use Amazon Prime and/or the questionaire you presented). Then secondly you can try to build a model for predicting the estimated latent variable based on customer data (age, location, etc.) $\endgroup$ Jan 3, 2017 at 9:03

2 Answers 2


Not sure I fully understand the question. If the question is how to choose the sample of people to offer the free trial, I would start by analyzing the properties of customers that paid the service and use it regularly (the best measure of customer satisfaction). Then you can build a classification algorithm that uses the same input variables, to predict how likely is a new customer to enroll in your paid service after a free trial.


Possible approach to this problem is to find some data about individuals that took free trials for something else (but similar to Amazon prime) and then make binary variable wheather they have subscribed after trial. Then you run probit/logit and get the model which assigns probability of subscription to individual based on some of his characteristics. Then apply this model to your potential trial recievers and give trials for those of them for whome estimated probability will be the highest.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.