How to Calculate Average Marginal Effects for a Continuous Variable
Use this premium calculator to compute the average marginal effect (AME) of a continuous predictor under a linear, logit, or probit model. Enter your intercept, coefficient for the continuous variable, and a list of observed x values. The tool calculates the observation-level marginal effects, averages them, and visualizes the pattern across your sample.
- Linear model: AME = coefficient on x
- Logit model: AME = average of beta × p × (1 – p)
- Probit model: AME = average of beta × phi(xb)
Expert Guide: How to Calculate Average Marginal Effects for a Continuous Variable
Average marginal effects, often abbreviated as AMEs, are one of the most practical ways to interpret regression results when your model is nonlinear. If you are estimating a logit or probit model and want to explain what a one-unit increase in a continuous variable does to the predicted probability of an outcome, the AME gives you a clean summary. Instead of reporting a raw coefficient that sits on a latent index scale, you translate the model into a more intuitive probability-based effect.
At a high level, the average marginal effect for a continuous variable answers this question: on average across the sample, how much does the predicted outcome change when the continuous predictor increases by one unit? In linear regression, this is trivial because the marginal effect is constant and simply equals the coefficient. In nonlinear probability models, however, the effect changes across observations because the slope depends on where each case sits on the probability curve. That is exactly why averaging individual marginal effects is so useful.
What an Average Marginal Effect Means
Suppose you estimate a model for whether a household adopts solar panels, whether a patient is readmitted, or whether a borrower defaults. Your dependent variable is binary, and one of your explanatory variables is continuous, such as income, age, debt ratio, or treatment intensity. In a logit or probit model, the coefficient on that variable does not directly equal the change in probability. Instead, the probability effect depends on both the coefficient and the current predicted probability or density level for each observation.
The AME solves this problem by computing the marginal effect observation by observation and then taking the arithmetic mean. This produces a single number that is often easier to communicate than a list of many observation-specific effects.
Core intuition
- In an OLS or linear probability model, the marginal effect is constant, so the AME equals the coefficient on the continuous variable.
- In a logit model, the slope depends on the logistic density and is largest around probabilities near 0.5.
- In a probit model, the slope depends on the standard normal density and also peaks near the center of the distribution.
- The AME is generally more representative than a marginal effect evaluated only at the mean of the covariates.
The Mathematical Formulas
Let the continuous variable be x and its coefficient be beta1. Let the linear index for observation i be:
eta_i = beta0 + beta1 x_i
Then the marginal effect differs by model type.
1. Linear model
If your outcome model is linear, then:
dE(y|x)/dx = beta1
Because the effect is constant for all observations, the AME is simply:
AME = beta1
2. Logit model
For logit, the predicted probability is:
p_i = 1 / (1 + exp(-eta_i))
The marginal effect of x for observation i is:
ME_i = beta1 × p_i × (1 – p_i)
Then the average marginal effect is:
AME = (1 / n) × sum(ME_i)
3. Probit model
For probit, the predicted probability is based on the standard normal cumulative distribution. The observation-level marginal effect is:
ME_i = beta1 × phi(eta_i)
where phi(eta_i) is the standard normal probability density function. Again, the AME is:
AME = (1 / n) × sum(ME_i)
Step-by-Step Process to Calculate the AME
- Estimate your model and identify the coefficient for the continuous variable.
- Compute the linear index for each observation using the intercept and the observed x value, and include any other covariates if your full model has them.
- Calculate the observation-level marginal effect using the correct derivative for your model type.
- Average all observation-level marginal effects across the sample.
- Interpret the result in the units of your dependent variable, usually a probability change per one-unit change in x.
This calculator simplifies the process by focusing on one continuous predictor plus an intercept. In a larger multivariable model, the same logic still applies, except the index would also include the contributions of the additional covariates.
Worked Example for a Logit Model
Assume a logit model with intercept beta0 = -1.2 and coefficient beta1 = 0.8. You observe 10 values of x. For each x value, compute the index eta_i, then the predicted probability p_i, then the marginal effect 0.8 × p_i × (1 – p_i). After doing that for all observations, average the 10 marginal effects. The result is your AME.
If the final AME were, for example, 0.145, the interpretation would be: on average, a one-unit increase in x is associated with a 14.5 percentage point increase in the predicted probability of the outcome, holding the model form fixed and averaging across the observed sample distribution.
Comparison Table: Marginal Effect Shape by Model Type
| Model | Observation-Level Marginal Effect | Maximum Slope Benchmark | Interpretation Note |
|---|---|---|---|
| Linear | beta1 | Constant at all x values | No curvature, so the AME equals the coefficient exactly. |
| Logit | beta1 × p × (1 – p) | p × (1 – p) peaks at 0.25 when p = 0.50 | Largest effects occur near the center of the probability range. |
| Probit | beta1 × phi(xb) | phi(0) ≈ 0.3989 | Effects are strongest around xb = 0 and taper in the tails. |
Reference Density Values Useful for Interpretation
Because nonlinear marginal effects depend on position along the curve, some benchmark values are worth remembering. The following table shows real mathematical reference points used constantly in applied econometrics and biostatistics.
| Index Value | Logit Probability p | Logit Slope p(1-p) | Standard Normal Density phi(x) |
|---|---|---|---|
| -2 | 0.1192 | 0.1050 | 0.0540 |
| -1 | 0.2689 | 0.1966 | 0.2420 |
| 0 | 0.5000 | 0.2500 | 0.3989 |
| 1 | 0.7311 | 0.1966 | 0.2420 |
| 2 | 0.8808 | 0.1050 | 0.0540 |
AME vs Marginal Effect at the Mean
A common question is whether to report the average marginal effect or the marginal effect at the mean. The marginal effect at the mean plugs the mean x value into the derivative formula and evaluates the effect there. This can be useful for a stylized benchmark, but it may describe a hypothetical person who does not actually exist in your data, especially when covariates are skewed or combined in ways that produce unrealistic averages. The AME is usually preferred because it respects the observed sample and averages actual observation-level derivatives.
- AME: average of each case’s marginal effect, usually more representative.
- MEM: marginal effect at mean covariate values, easier to calculate by hand but less robust to nonlinearity.
- MER: marginal effect at representative values, useful for profiles such as low, medium, and high x.
Common Mistakes When Calculating Average Marginal Effects
- Treating the nonlinear coefficient as the probability effect. In logit and probit, coefficients are not direct probability changes.
- Evaluating only one observation. A true AME requires averaging across all sample observations.
- Using the wrong derivative. Logit uses p(1-p), while probit uses the standard normal density.
- Ignoring units. If x is measured in thousands of dollars, one unit means one thousand dollars, not one dollar.
- Forgetting other regressors. In a multivariable model, the index should include all included variables.
How to Interpret the Sign and Magnitude
The sign of the AME follows the sign of the coefficient in most standard settings. A positive AME means larger x values are associated with larger predicted probabilities, on average. A negative AME means larger x values reduce the predicted probability, on average. Magnitude should be interpreted with care. For instance, an AME of 0.032 means a one-unit increase in x changes the predicted probability by 3.2 percentage points on average, not 32 percent.
Also remember that the AME is a sample-average derivative. It is not necessarily the effect for every individual. In nonlinear models, some observations can have much larger or smaller marginal effects depending on their position on the curve.
Why the Chart Matters
The chart in this calculator plots the observation-level marginal effects across your entered x values. This visual is valuable because it reveals whether the effect is nearly constant, strongly peaked, or tapering in the tails. In a logit or probit model, that shape helps you see whether your sample sits near the region of maximum responsiveness or mostly in low-slope zones where predicted probabilities are already near zero or one.
Authoritative Learning Resources
If you want a deeper treatment of marginal effects, interpretation of nonlinear models, and probability derivatives, review these authoritative resources:
- UCLA Statistical Methods and Data Analytics
- Penn State Eberly College of Science: Logistic Regression Materials
- U.S. Census Bureau Working Papers and Technical Research
Bottom Line
If you are trying to explain how to calculate average marginal effects for a continuous variable, the essential idea is straightforward: compute the derivative of the predicted outcome with respect to x for each observation, then average those derivatives. In a linear model the answer equals the coefficient. In a logit model use beta1 × p × (1-p). In a probit model use beta1 × phi(xb). The result is often the clearest way to express the average probability impact of a continuous predictor in a nonlinear regression.