ECON31031 Chap.3 Binary Response Models: LPM, Logit and Probit
Binary Response Models: LPM, Logit and Probit
Binary Response Models treat a yes or no outcome as a Bernoulli variable whose success probability depends on individual characteristics through a distribution function of the linear index.
Choosing the uniform distribution gives the linear probability model, which OLS can estimate but which can predict probabilities outside the unit interval and has heteroscedasticity built in, so White's robust covariance matrix is needed for inference. The logistic distribution gives logit, whose log odds are linear in the regressors and whose coefficients exponentiate into odds ratios.
The normal distribution gives probit, which also arises from a latent utility crossing a threshold. Because only the side of the threshold is observed, the error variance must be fixed, which explains why logit and probit coefficients differ in scale. Binary data are only one kind of discrete response: ordered responses rank answers such as survey scales, and multinomial responses choose one of several exclusive alternatives.
The generalised linear model view ties the three binary models together through their link functions: identity for the linear probability model, logit as the canonical link for binary data, and the inverse normal for probit.
What this chapter covers
- 01
Binary, ordinal and multinomial responses
- 02
Bernoulli probabilities that vary across individuals
- 03
The linear probability model and its two problems
- 04
White's heteroscedasticity-consistent covariance matrix
- 05
Logit, log odds and odds ratios
- 06
Generalised linear models and link functions
- 07
The latent variable approach, probit and limited dependent variables
Worked example · free
Update a logit probability through the odds
- 2Convert the probability to odds: 0.8/0.2 = 4, or odds of 4 to 1 on.
- 2Multiply by the odds ratio e^0.5 = 1.6487 to get new odds of 6.595.
- 2Convert back: 6.595/(1 + 6.595) = 0.868, a rise of about 6.8 percentage points.
Key terms
- Link function
- The function that connects the expected value of the response to the linear index in a generalised linear model.
- Latent variable
- An unobserved continuous outcome whose crossing of a threshold determines the observed binary response.
- Heteroscedasticity
- Error variance that differs across observations, which binary data have by construction.
Binary Response Models: LPM, Logit and Probit FAQ
Why not just use OLS for a binary outcome?
OLS on a binary outcome is the linear probability model. It is easy to estimate, but it can produce fitted probabilities above one or below zero, forces each characteristic to have a constant effect, and has heteroscedastic errors by construction, so its usual standard errors are biased.
Why are logit coefficients larger than probit ones?
The logistic distribution has variance π²/3 while the standard normal has variance one. Because the scale of the latent utility cannot be identified, each model fixes its own variance, so logit coefficients come out roughly 1.6 to 1.8 times the probit values in the module's benefits example.
What does an odds ratio mean in a logit model?
It is the exponential of a coefficient, the factor by which the odds of success are multiplied when that regressor rises by one unit. For a dummy it compares the odds with and without the characteristic, holding the other regressors fixed.
How is a probit probability calculated by hand?
Compute the linear index from the coefficients and the person's characteristics, then read the standard normal distribution function at that value. For a negative index use symmetry: the probability equals one minus the tabulated value for the matching positive index.
What are truncated and censored models?
They are limited dependent variable models defined by how a latent outcome is observed. A truncated model keeps only observations above a threshold, a censored model records the threshold value whenever the latent outcome falls below it, and a binary model records only which side is reached.
Exam move
Learn the three models as one family that differs only in the distribution function F. Draw the three curves from memory and mark where the linear probability model leaves the unit interval. Practise converting between probabilities, odds and log odds in both directions, and compute probit probabilities from table values, remembering symmetry for negative indices.
For theory questions, be ready to explain the latent variable story and why the error variance has to be fixed. Use R alongside the theory: fitting the same specification with lm() and glm() shows which statistics each function reports, and a scatterplot of a binary outcome against a regressor shows why a smoother is needed to see any relationship at all.
Finish by explaining in two sentences why the latent variable story requires a fixed error variance.
Working through Binary Response Models: LPM, Logit and Probit in ECON31031? Sia is AskSia’s AI Economics tutor — ask any ECON31031 Binary Response Models: LPM, Logit and Probit question and get a clear, step-by-step explanation grounded in how ECON31031 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.