18.8 Retention model

The modelling file contains one row per member and uses information that would be available during or at the end of the first membership period. The target is first-period renewal, with censored members removed from supervised modelling.

A defensible workflow would compare:

• a majority-class or overall-rate baseline;
• logistic regression for an interpretable probability model;
• a decision tree for transparent nonlinear rules; and
• a random forest if it produces a material validation improvement.

Data should be split by member, and preprocessing must be estimated using training data only. If the model will be used for future cohorts, a time-based validation is preferable to a purely random split.

Accuracy is not sufficient. NVRW has limited outreach capacity, so evaluation should emphasize calibration, precision among the highest-risk members, recall within the contactable group, and expected contribution margin preserved under alternative intervention effects.

The model may show that low first-90-day visit frequency is associated with non-renewal. That does not prove that increasing visits will cause renewal. The correct next step is a monitored pilot rather than a causal claim.