3
$\begingroup$

I'm trying to implement survival analysis on conversion rates for a free trial product we provide. In general, I understand the applications on survival analysis along with dealing with right censored data (e.g., if a customer is currently on trial, we don't really know if they will or will not convert in the future).

My question is, how do you implement a survival analysis model in this context if you know the trial ends in 14 days? So you may have a handful of customers that simply leave the early (these would indicate a 'death'), and you have many who are currently in the trial (these are the censored customers). How would you go about modeling this knowing that the trial can't go on to some uncertain time in the future? it must end at some point in the future (14 days in my examples).

Also what sort of variables I can include based on your insight/knowledge/intuition.

Thanks in advance

$\endgroup$
1
  • $\begingroup$ You could treat the event of "leaving early" as a competing risk event which precludes the ocurrence of the event of interest (i.e., "conversion by the end of the trial period) and use a competing risks framework to analyze your data. Also, see ncbi.nlm.nih.gov/pmc/articles/PMC2811964. $\endgroup$ – Isabella Ghement Jun 20 '19 at 8:23
1
$\begingroup$

I'm doing a similar work related to patients' "death" in the remote patient monitoring setting. My thoughts based on your context are the following:

In your case, the study period or say observation time is 14 days since we only observe each customer's behavior during their free trial time, and then we want to understand whether a customer will "survival" (buy this product, start subscribing to the membership, or any other actions that represent this customer is willing to continue/extend/stay etc). The definition of event here could be "leaving early" as Isabella Ghement said, not using the product very frequently, e.g., if a customer still don't subscribe the product even after the 14 free trail days, then event=0 and survival_time=14+ where "+" means maybe the customer will subscribe in the future but we don't know the specific date of subscribing (right-censoring).

Since we just observe everyone's first 14 days of behavior, variables could include like # of log in time, # of click times, # of comment leaving times, # of response if comments are left, # of app opening times, the time of all these actions during the day(morning/afternoon/evening/midnight), etc... lots of variables you can include.

As for algorithms and metrics, Kaplan-Meier, CoxPH regression, Random Survival Forest, Conditional Survival Forest , XGBoost accelerated failure time, Deep Surv, Deep Hit, etc... C-index and cumulative dynamic AUC can be used for model performance, and libraries include like scikit-survival, pysurvival, lifelines. But first, the data must be of good quality.

Some extra thinkings, this type of problem can also switch to a binary classification problem. But, if you want to know like when a customer will start subscribing or stop using during this 14 free trial days, then go with survival analysis, though I think survival analysis is still not so mature compared to traditional regression/classification problem.

Hope this helps and happy to discuss more! :)

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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