Discrete-Time Event History (Hazard) Analysis

Description

Models the probability that an event such as churn happens in each week, session or step, given it has not happened yet, with logistic regression on person-period data.

Ordinary regression on whether an event occurred within a fixed window wastes information and mishandles censoring. Event history analysis models time to an event while respecting users still at risk when observation stops. When time is measured coarsely (weeks, billing cycles, sessions, funnel steps) or is intrinsically discrete, the natural quantity is the discrete-time hazard: the conditional probability of the event in period t given survival to t. Data are expanded to one row per user per period at risk, and a logistic regression with a separate intercept per period (or a smooth function of time) plus covariates is fitted with standard software; time-varying covariates enter naturally. If an underlying continuous proportional hazards process is assumed, a complementary log-log link is used instead. When per-period probabilities are below about .1, discrete and continuous results nearly coincide. Unobserved heterogeneity makes hazards appear to decline over time, so frailty or mixture extensions are worth checking.

Attributes

Loss type
Siloed
Stage
Post-launch
Effort
Days

Steps

  1. Define the state, the event (for example cancellation), the time origin, the period unit, the risk set and how censoring arises.
  2. Expand the data to a person-period file with one record per user per period at risk, ending at the event or censoring.
  3. Code the event indicator for each period and attach time-varying covariates (usage that week, price changes, support contacts).
  4. Fit a logistic regression with period dummies or a smooth time function plus covariates; use a complementary log-log link if a continuous-time proportional hazards interpretation is wanted.
  5. Interpret the baseline hazard shape and covariate odds or hazard ratios, and derive survival curves from the fitted hazards.
  6. Check for unobserved heterogeneity with a frailty or mixture term and test sensitivity to period definitions.

Sources

  • The SAGE Handbook of Quantitative Methods in Psychology (Millsap & Maydeu-Olivares, 2009)
  • Applied Bayesian Modelling (Congdon, 2003)
  • Bayesian Statistical Modelling (Congdon, 2006)

Also known as

  • Discrete-time survival analysis
  • Discrete-time logit hazard model
  • Complementary log-log hazard model
  • Person-period analysis
  • Piecewise exponential hazard model

Categories

  • Analytics & Behaviour
  • Statistics & Modelling