Multinomial and Conditional Logit Models for Multiple Choices

Every decision we make, from which route to take to work to which brand of coffee to buy, is a discrete choice. For economists, marketers, and policy makers, understanding these choices is crucial for everything from demand forecasting to designing public transit systems. However, analyzing choices between three or more options is far more complex than a simple yes/no (binary) decision. Traditional linear regression models simply fall apart when the outcomes are categories, not continuous values. This is where specialized econometric tools-the Multinomial Logit (MNL), the Conditional Logit (CL), and the versatile Mixed Logit (MXL) models-step in, offering mathematically elegant frameworks for modeling human decision-making based on the powerful concept of Random Utility Maximization.

Decoding discrete choice: Why specialized models are essential 🔗

In economics, we assume individuals are rational actors who make choices based on maximizing their perceived utility. If you are choosing between Alternative A, Alternative B, and Alternative C, you will pick the one that gives you the highest utility, which is why these models are known as Random Utility Maximization (RUM) models. The “random” part acknowledges that researchers can never observe every single factor influencing a person’s decision (such as their mood, recent advertisements they saw, or personal biases). The model breaks down utility into two components:

• Systematic Utility (V): The part we can observe and measure (e.g., price, income, age).
• Random Utility (E): The unobserved component (the error term).

The total utility is U=V+E. To make the math solvable, we have to assume a specific distribution for that unobserved error term, E. When we assume the errors are independent and identically distributed (i.i.d) according to a Gumbel distribution (a type of extreme value distribution), the resulting choice model is the Logit model. This assumption of Gumbel-distributed errors is what gives the Logit models their closed-form, easy-to-calculate probability structure, but it also introduces their most significant-and often criticized-constraint: the Independence of Irrelevant Alternatives (IIA) assumption.

The workhorse: When to use multinomial logit 🔗

The Multinomial Logit (MNL) model is the most straightforward extension of the standard binary logit model (which handles only two choices, like ‘Buy’ or ‘Not Buy’). It is used when the dependent variable represents a choice between three or more alternatives that are distinctly unordered. Examples abound in policy and commerce:

• Consumer Marketing: Choosing between Brand X, Brand Y, or Brand Z coffee.
• Social Sciences: Deciding between three types of educational programs: Arts, Science, or Commerce.
• Labor Economics: Selecting an occupation among Manufacturing, Services, or Government.

The key characteristic of the MNL model is its focus on the attributes of the decision-maker. It tries to answer the question: “Which type of person is more likely to choose Alternative A over Alternative B?”

Unordered choices and decision-makers 🔗

In the MNL framework, the explanatory variables are case-specific or individual-specific. For instance, in the choice of educational programs in India, the predictors would be the student’s characteristics:

• Parental Income: A student from a high-income household might be more likely to pursue an expensive private medical education (a specific alternative).
• Gender: Historically, gender may have influenced the choice between engineering and humanities streams.
• Previous Academic Score: Students with higher marks might be more likely to enter highly competitive programs.

Crucially, variables like income or gender do not change across the choices for a single individual; the same student’s income influences the utility they derive from all available streams, leading to a set of coefficients, one for each alternative relative to a baseline option (often called the reference category). While powerful for segmenting populations and understanding demographic drivers of choice, this structure is what makes the MNL model highly susceptible to the IIA constraint.

The independence of irrelevant alternatives (IIA) assumption 🔗

The Independence of Irrelevant Alternatives (IIA) is the fundamental assumption that underpins the MNL model and is the source of both its mathematical convenience and its behavioral limitations. Simply put, IIA states that adding a new, irrelevant alternative to the choice set should not change the relative probability of choosing between any two existing alternatives. Mathematically, the ratio of the probability of choosing Alternative A over Alternative B is independent of whether Alternative C is also available.

Think of it this way: if I prefer a cup of tea (A) to a cup of coffee (B), the ratio of my choice probability P(A)/P(B) should theoretically remain unchanged whether the menu also offers a glass of milk (C) or not. In many real-world situations, this is a perfectly sound assumption. However, when alternatives are perceived as highly similar or are close substitutes, the IIA assumption can be catastrophically violated.

The red bus/blue bus paradox and its limitations 🔗

The most famous illustration of the IIA violation is the “Red Bus/Blue Bus” Paradox, famously articulated by economist Daniel McFadden, which led to his development of more flexible models. Imagine the following choice set for a commuter:

Scenario 1:Car: Fast, expensive, comfortable.Red Bus: Slow, cheap, slightly uncomfortable.

Assume, based on the MNL model estimation, the commuter chooses the Car 50% of the time and the Red Bus 50% of the time. The ratio P(Car)/P(Red Bus) is 50%/50%=1.0.

Scenario 2: Now, the city introduces a new service:

Blue Bus: Identical in every measurable way to the Red Bus (same route, same price, same speed), just a different colour.

In reality, the commuter should simply split their “Bus” preference between the Red Bus and the Blue Bus. The true choice probabilities should look like this:

• P(Car)=50% (unchanged)
• P(Red Bus)=25% (half of the original bus share)
• P(Blue Bus)=25%

But the MNL model, forced by the IIA assumption, requires that the ratio P(Car)/P(Red Bus) remains 1.0. The only way for this ratio to remain 1.0 is if the new probabilities become:

• P(Car)=33.3%
• P(Red Bus)=33.3%
• P(Blue Bus)=33.3%

The MNL model incorrectly predicts that the introduction of a new, nearly identical bus causes one-third of the original Car users to switch to the new Blue Bus, and one-third to switch to the original Red Bus. This is behaviorally nonsensical because the two bus options are perfect substitutes; they only cannibalize each other, not the fundamentally different Car option. This is why the IIA assumption severely restricts the MNL model when alternatives exhibit correlation or substitution patterns, leading to potential specification errors and invalid statistical inference. If alternatives are close substitutes, a model that violates IIA, such as the Nested Logit or the Mixed Logit, must be used.

Shifting the focus: Introducing the conditional logit model 🔗

While the MNL model uses the chooser’s characteristics to explain the choice, the Conditional Logit (CL) model turns the entire framework on its head. It models the choice as a function of the attributes of the alternatives themselves. It asks: “How does a change in a choice’s features influence the utility of that choice?”

McFadden, who also developed the CL model, realized that in many contexts, what matters is the price, quality, or feature set of the option, not the demographics of the buyer. The CL model is exceptionally powerful in transportation economics, where the variables of interest are always related to the mode itself:

• Time: Travel time via train vs. air.
• Cost: Fare price of Option A vs. Option B.
• Comfort: The difference in utility between a sleeper bus and a private car.

Attributes of the alternatives vs. characteristics of the chooser 🔗

The structural difference is evident in the data setup and coefficients. In the standard CL model, the variables are alternative-specific, meaning their values change for each alternative within a choice set (e.g., the cost of a train ticket is different from the cost of a flight for the same journey). Crucially, the standard CL model estimates only one coefficient for each variable (e.g., a single coefficient for ‘Cost’ and a single coefficient for ‘Time’). This single coefficient indicates the marginal utility derived from that attribute, which is assumed to be constant across all alternatives.

For example, if the coefficient for ‘Price’ is estimated as βPrice​=−0.5, it means a ₹1 increase in price for any alternative (Bus, Train, or Flight) decreases the utility of that alternative by 0.5 units. The CL model naturally allows for the inclusion of alternative-specific constants (like a ‘Brand Loyalty’ constant for an Indian FMCG brand) to capture the intrinsic utility of that option that cannot be explained by the measured attributes. The CL model is a special case of MNL-in fact, they are mathematically identical but differ in how the independent variables are structured and interpreted, making CL a more appropriate model for behavioral interpretation.

However, the Conditional Logit model also suffers from the same restrictive IIA assumption because it still assumes the random error terms are i.i.d. Gumbel distributed. If we model the choice of vehicles (small, medium, or large SUV) based on features like mileage and price, the CL model will be tripped up if a new SUV is introduced that is highly similar to one of the existing options. This leads us to the most advanced model in the Logit family.

Overcoming IIA: The power of the mixed logit model 🔗

The Mixed Logit (MXL) model, also known as the Random Parameters Logit (RPL) or Mixed Multinomial Logit (MMNL) model, is often hailed as the “universal approximator” of discrete choice models. It was developed precisely to address the major shortcomings of the MNL and CL models, primarily the IIA assumption and the inability to capture genuine individual taste variation.

Allowing for heterogeneity with random parameters 🔗

The core innovation of the MXL model is the incorporation of preference heterogeneity. It recognizes that the utility derived from an attribute is not constant across all individuals-something the standard MNL and CL models fail to capture. For example, in a choice between a traditional market and a modern supermarket in an urban Indian setting:

• The Price Coefficient: A low-income customer (Person A) may have a highly negative coefficient for price (meaning they are extremely sensitive to price changes), while a high-income customer (Person B) may have a coefficient close to zero (meaning price is not a major factor).
• The Convenience Coefficient: A working professional may place a high positive utility on ‘Convenience’ (a store being close to work), while an elderly resident may prioritize ‘Local Community Feeling’.

MXL models these varying preferences by assuming that the coefficients (β) themselves are random variables that follow a specific distribution (such as Normal, Log-Normal, or Triangular) across the population. Instead of estimating a single βPrice​, the MXL estimates the mean (μ) and the standard deviation (σ) of the distribution of the Price coefficient. This flexibility provides two immediate, powerful advantages:

1. It Lifts the IIA Restriction: Because the utility function is different for every individual (due to the randomly drawn coefficients), the relative odds of choosing A over B are no longer independent of C. The MXL model can successfully distinguish between the substitution patterns of close alternatives (like the Red Bus/Blue Bus), making it a robust choice for modeling real-world behavior.
2. It Accounts for Unobserved Correlation in Utility: MXL can account for correlation between unobserved factors over time or across choices made by the same individual (e.g., brand loyalty in consumer panels), a crucial feature when working with panel data.

As McFadden and Train demonstrated, the mixed logit model is so flexible that it can approximate any discrete choice model derived from random utility theory to an arbitrary degree of accuracy. This adaptability makes it the gold standard in contemporary academic and commercial discrete choice modeling, particularly in complex scenarios like willingness-to-pay studies for new products or policy interventions. While its estimation is computationally intensive (relying on simulation methods rather than closed-form equations), the accuracy and robustness it offers far outweigh the complexity for sophisticated economic analysis.

What do you think? Given the prevalence of personalized marketing and heterogeneous preferences in the modern Indian market (e.g., choice of OTT platforms, mobile phones), do you think a researcher could rely solely on the MNL or CL model, or is the inherent complexity of the Mixed Logit model always necessary for accurate demand prediction?

References

1. https://zhanksun.github.io/files/Multinomial%20and%20Conditional%20Logit%20Models.pdf
2. https://www.cs.cornell.edu/~arb/papers/iia-www2016.pdf
3. https://ftp.iza.org/dp5826.pdf
4. https://thesis.eur.nl/pub/50097/Pfaff_433964.pdf
5. https://dspace.mit.edu/bitstream/handle/1721.1/150428/Hu_huyiqun_PhD_CEE_2022.pdf?sequence=1&isAllowed=y