Competing Risks Survival Analysis

Description

Model time to an event that can end in several mutually exclusive ways, with a separate hazard and separate predictor effects for each way.

Many UX durations end in more than one way: a trial ends in upgrade, cancellation or silent lapse; a session ends in purchase, abandonment or hand-off to support. Fitting one hazard for 'any exit' blurs drivers that act in opposite directions. The competing risks approach gives each exit type its own hazard with its own covariate effects and duration dependence; the overall hazard is their sum. In the likelihood a spell that ends one way is treated as censored for every other way, and a spell with no exit is censored for all. An alternative, fitted by EM, models the probability of each cause with a logistic regression and a separate time-to-event curve for each cause, which avoids assuming the routes are independent. Congdon's job-mobility example shows education raising upward moves while experience protects against lateral and downward ones.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Days

Steps

  1. Define mutually exclusive and exhaustive exit types, plus censoring for spells still open.
  2. For each spell record its duration, how it ended (or that it is censored) and covariates.
  3. Specify a hazard for each exit type (for example Weibull or piecewise exponential) with its own coefficients and shape.
  4. Build the likelihood with each spell censored for every exit type other than the one observed, and fit by maximum likelihood or MCMC.
  5. Compare covariate effects across exit types and plot the predicted probability of each exit by time (cumulative incidence).
  6. Optionally fit the mixture form (logistic model for which exit, separate timing per exit) and compare conclusions; check proportionality of each cause-specific hazard.

Sources

  • Bayesian Statistical Modelling (Congdon, 2006)
  • The EM Algorithm and Extensions (McLachlan & Krishnan, 1997)
  • Statistical Methods for Reliability Data (Meeker & Escobar, 1998)
  • Statistical Methods for Survival Data Analysis (Lee & Wang, 2003)

Also known as

  • Competing risks
  • Competing risks model
  • Cause-specific hazards
  • Multiple-destination event history model
  • Multiple decrement model
  • Mixture model for failure times with competing causes

Categories

  • Analytics & Behaviour
  • Statistics & Modelling