Mixed Logit Model
Also known as: Random Parameters Logit, Mixed Multinomial Logit, Error Components Logit, Karma Logit Modeli
The Mixed Logit model, introduced formally by McFadden and Train (2000) and elaborated in Train (2009), is a flexible discrete choice framework that allows preference parameters to vary randomly across decision-makers. By integrating standard logit probabilities over a mixing distribution of coefficients, it overcomes the restrictive independence of irrelevant alternatives (IIA) property and accommodates unobserved taste heterogeneity, panel data correlation, and complex substitution patterns across alternatives.
Key highlights
- Accommodates unobserved preference heterogeneity by allowing coefficients to vary randomly across individuals.
- Does not impose the IIA restriction, enabling realistic cross-alternative substitution patterns.
- Handles panel data naturally by correlating choices across repeated observations for the same decision-maker.
- The mixing distribution can take many parametric forms (normal, log-normal, triangular), offering modeling flexibility.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Mixed Logit when respondents face choices among three or more discrete alternatives and you expect unobserved taste heterogeneity across individuals. It is particularly valuable in stated-preference and revealed-preference surveys, panel choice data (repeated choices per individual), and settings where IIA is implausible. Key assumptions include additive random utility, a correctly specified mixing distribution, and sufficient simulation draws for reliable likelihood approximation. It is more computationally intensive than standard multinomial logit; if preference homogeneity is defensible, simpler models may suffice.
Strengths & limitations
- Accommodates unobserved preference heterogeneity by allowing coefficients to vary randomly across individuals.
- Does not impose the IIA restriction, enabling realistic cross-alternative substitution patterns.
- Handles panel data naturally by correlating choices across repeated observations for the same decision-maker.
- The mixing distribution can take many parametric forms (normal, log-normal, triangular), offering modeling flexibility.
- Simulated maximum likelihood estimation is computationally demanding, especially with many random parameters or large datasets.
- Results can be sensitive to the assumed form of the mixing distribution, and misspecification leads to biased estimates.
- Identification of the full covariance matrix of random coefficients requires large samples and careful specification.
- Welfare calculations (willingness-to-pay) derived from ratio of random coefficients can have undefined moments for some distributions.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does Mixed Logit differ from standard Multinomial Logit?
Standard multinomial logit assumes identical, fixed preference coefficients for all decision-makers and imposes IIA. Mixed logit relaxes both restrictions by allowing coefficients to be randomly distributed across individuals and by integrating over that distribution, producing flexible cross-alternative substitution patterns. The cost is computational: mixed logit requires simulation-based estimation rather than closed-form maximization.
Which mixing distribution should I choose for the random parameters?
The choice depends on the parameter's expected sign and behavior. Normal distributions are common for parameters that may be positive or negative. Log-normal is appropriate when the parameter must be strictly positive (e.g., a positive preference for a feature). Triangular or uniform distributions bound the support and can prevent extreme values. Goodness-of-fit tests and theoretical priors should guide the selection.
How many simulation draws are sufficient for reliable estimates?
The required number of draws depends on the number of random parameters and sample size. A common practical guideline starts with 500–1000 Halton or Sobol quasi-random draws and checks stability by doubling the draw count. If parameter estimates and standard errors change materially, more draws are needed. Halton sequences typically outperform pseudo-random draws by reducing simulation variance with fewer evaluations.
Sources
- 1.Train, K. E. (2009). Discrete Choice Methods with Simulation (2nd ed.). Cambridge University Press.ISBN 978-0-521-74738-7
- 2.McFadden, D., & Train, K. (2000). Mixed MNL models for discrete response. Journal of Applied Econometrics, 15(5), 447–470.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 2). Mixed Logit. ScholarGate. https://scholargate.app/econometrics/mixed-logit