I'm aware of a variety of approaches to discovering “stand-alone” churn probabilities.

But I haven't been able to find by searching any info on “conditional” churn probabilities.

Use case: I'm a magazine publisher, and sell subscriptions to a number of magazines. I might be interested in knowing facts of the following form:

  1. If a customer cancels mag1 subscription, they're likely to cancel mag2 subscription.

  2. What “small” subset of my magazines are “core”, in the sense that canceling any of those subscriptions massively impacts cancellation of subscriptions in the larger group?

Are there specific routines for analyzing “conditional churn” for this kind of scenario?

Or do people just apply normal conditional probability estimators to stand-alone churn propensities?

If this is a well-explored scenario, assistance with useful search terms would be appreciated!

  • $\begingroup$ Survival analysis would be useful. You can use a Cox proportional hazards model to estimate the survival curve conditioned on any predictors you think are important. Survival analysis is likely better than examining risk of churn via logistic regression because of censoring (i.e. any customers which have not churned can be considered as yet to churn. Their churn time is thus censored from us, and the Cox model handles this with ease). $\endgroup$ Jan 16, 2021 at 20:19
  • $\begingroup$ You can also model the stand-alone churn probabilities of each product and use a copula to model the dependence between products. See e.g. Genest et al., Predicting dependent binary outcomes through logistic regressions and meta-elliptical copulas (2013) $\endgroup$ Jan 16, 2021 at 21:27
  • $\begingroup$ Thanks for the comments you two - a couple of directions to look into was all I was after! $\endgroup$ Jan 17, 2021 at 2:02


Your Answer

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