Classical ML, validation, metrics, interpretation, debugging · insight-1174
Judge an interaction on the decision's scale: no log-odds interaction can still mean unequal probability gaps
- Evidence type
- practitioner heuristic
- Difficulty
- hard
- Roles
- Data scientist, Data analyst
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
A logistic model has two drivers, such as booking window and market, and the team asks whether they interact using the p-value of the product term.
Action
Compute predicted probabilities over the grid of both drivers, report the differences in probability with intervals next to the coefficient, and keep a product term only if it changes the probability gap by an amount the decision cares about.
Reason
Probability is a non-linear function of the linear predictor, so the effect of one driver depends on the baseline set by the other.
Boundary
Logistic, Poisson and other log- or logit-link models used for decisions in probability or count units.
Symptom or risk
A zero interaction on the log-odds scale still gives unequal effects in probability points, while a significant product term can be negligible where the decision is made.
Exceptions
Identity-link linear models, where the scale is the same.
Trade-offs
Probability-scale effects depend on the population mix.
Pitfalls
Reading a non-significant product term as proof of constant effects.
Example
Hypothetical log-odds: baseline -3, driver A +1, driver B +1, no interaction. Probabilities are 0.047, 0.119, 0.119 and 0.269, so A adds 7 points when B is off and 15 points when B is on.