ECON31031 Chap.4 Marginal Effects and Odds Ratios
Marginal Effects and Odds Ratios
Marginal Effects and Odds Ratios answer the question every applied binary model raises: how much does a characteristic change the probability? In the linear probability model the coefficient is the effect. In logit and probit the effect of a continuous regressor is the density at the index times the coefficient, so it is largest near a probability of one half and shrinks towards the extremes.
Dummies use the difference between two fitted probabilities. Effects can be summarised at the sample average of the regressors or averaged over individuals, and their standard errors come from the delta method. In the module's benefits data the three models disagree on coefficients but agree closely on effects. A useful check on any reported effect is its sign and scale.
The sign always matches the coefficient, because densities are positive, while the size depends on where the individual sits on the curve. Odds ratios offer a second, logit-specific reading: they are constant across individuals, whereas changes in probability are not.
What this chapter covers
- 01
Why coefficients are not effects in nonlinear models
- 02
Marginal effects for continuous regressors
- 03
Discrete changes for dummy variables
- 04
Effects at the average versus average marginal effects
- 05
Delta-method standard errors
- 06
Odds ratios and percentage-point interpretation
Worked example · free
Compare a probit derivative with an exact dummy effect
- 2Continuous effect: 0.3989 × 0.25 = 0.0997, almost 10 percentage points per unit.
- 2Exact dummy effect: Φ(0.5) - Φ(0) = 0.6915 - 0.5 = 0.1915.
- 2Derivative approximation: 0.3989 × 0.5 = 0.1995, slightly too large because the curve flattens between 0 and 0.5.
Key terms
- Delta method
- A technique that turns the covariance matrix of estimated coefficients into standard errors for functions of them.
- Average marginal effect
- The mean of individual marginal effects computed for every observation in the sample.
- Discrete change
- The difference between fitted probabilities with a dummy set to one and set to zero.
Marginal Effects and Odds Ratios FAQ
Where is a logit marginal effect largest?
It is largest where the fitted probability is one half, because the logistic density equals p(1 - p) and peaks there at 0.25. The same coefficient therefore moves probabilities less for individuals whose probabilities are already close to zero or one.
Should I report effects at the average or average effects?
Both are used in the notes. The effect at the average evaluates the derivative at mean characteristics, while the average effect computes it for each person and averages. In the benefits example they differ only in the third decimal place, but say which one you report.
How do I interpret a negative odds ratio?
An odds ratio is never negative, because it is the exponential of a coefficient. A negative coefficient gives an odds ratio below one, meaning the characteristic lowers the odds; a value of 0.786, for example, means odds about 21% lower.
Why do logit and probit give similar effects but different coefficients?
The two models fix different error variances, so their coefficients are on different scales, roughly a factor of 1.6 to 1.8 apart in the module's benefits example. Marginal effects convert both back to probability units, which is why the reported effects agree closely across logit, probit and the linear probability model.
How do I scale a marginal effect?
A marginal effect is the change in probability per one-unit change in the regressor. Multiply it by the change the question actually describes; an effect of -0.065 per unit of the replacement rate becomes about -0.0065, or two thirds of a percentage point, for a rise of 0.1.
Exam move
Practise each effect formula with a calculator until the steps are automatic: density times coefficient for continuous regressors, a difference of probabilities for dummies. Always name the individual you evaluate. Convert effects into percentage points and scale them to a realistic change in the regressor.
Then practise reading tables like the benefits output in the notes, explaining why coefficients differ across models while effects agree. Build a small table for one fitted probit and one fitted logit: index, probability, density, effect of a continuous regressor and exact effect of a dummy. Repeat it at three different indices so you can see the effect shrink towards the tails.
Practise stating each result in percentage points and each odds ratio as a percentage change in the odds. Before the exam, rehearse one complete answer that reports a coefficient, its marginal effect at a stated profile and its odds ratio, each with a one-line interpretation.
Working through Marginal Effects and Odds Ratios in ECON31031? Sia is AskSia’s AI Economics tutor — ask any ECON31031 Marginal Effects and Odds Ratios 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.