Description
Models whether users do something with a probit and how much they do with a regression whose errors may correlate with that decision, correcting for self-selection.
Many outcomes are two-stage: whether a user buys, donates or responds, then how much. The Type-2 Tobit has a probit selection equation and an amount equation observed only for participants, each with its own regressors and parameters, and correlated normal errors. Unlike the Type-1 Tobit it lets a variable push the two parts in opposite directions: older people may be more likely to donate but give less. The correlation captures self-selection: if unobserved factors that make people participate also make them spend more, a regression on participants alone is biased. Heckman's two-step estimator fits the probit, computes the inverse Mills ratio for each participant, and adds it as a regressor in the amount equation; its significance indicates selection bias. White heteroskedasticity-consistent standard errors are needed in the second step, or the model can be fitted by full maximum likelihood. Identification is fragile unless some variable affects participation but not amount.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Post-launch, Evaluation
- Data
- Quantitative
- Effort
- Days
Steps
- Specify a selection equation (participation) and an outcome equation (amount), each with its own regressors, ideally with at least one variable that affects participation but not amount.
- Step 1: estimate a probit for participation by maximum likelihood on all cases.
- Compute the inverse Mills ratio for each participant from the probit index.
- Step 2: regress the amount on its regressors plus the inverse Mills ratio for participants, with White standard errors, or fit both equations by full maximum likelihood.
- Test the Mills ratio coefficient for selection, and check the probit (pseudo-R-squared, likelihood ratio and heteroskedasticity tests) and the amount regression.
- Interpret effects on participation, conditional amount and unconditional expected amount separately, and validate by forecasting on a hold-out sample, for example ranking users by expected value.
Sources
- Quantitative Models in Marketing Research (Franses & Paap, 2001)
Also known as
- Heckman selection model
- Heckit
- Two-part participation and amount model
- Inverse Mills ratio correction
- Heckman correction
Categories
- Analytics & Behaviour
- Statistics & Modelling