Multinomial Logistic Regression Models

Dr. Vijay Victor

Indian Institute of Technology, Roorkee

February 4, 2026

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 1 / 30
Introduction to Multinomial Logit

What Is Multinomial Logit?

An extension of logistic regression to handle responses with more than
two categories.
Often called polytomous or multinomial logistic regression.
Nominal vs. Ordinal Categories
Nominal categories have no inherent order (e.g., Pass, Fail,
Withdraw).
Ordinal categories have a natural order (e.g., Low, Medium, High).
Different approaches are used for each type (baseline-category logit is
common for nominal outcomes).

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 2 / 30
Transitioning from Binary to Multinomial

Binary Logit Limitation: Only handles two outcomes (e.g., Pass vs. Fail).
Example: Exam outcomes can be:
Pass
Fail
Withdraw
A simple binary model cannot include a third category.
Solution: Use the baseline-category logit model, suitable for r ≥ 2
nominal categories.

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 3 / 30
Nominal (Unordered) Multinomial Regression Model

Definition:
Used when the outcome variable consists of multiple unordered
choices.
Examples include mode of transport (car, bus, train, bike).
The model estimates the probability of each outcome given a set of
independent variables.
Three Types of Nominal Multinomial Regression Models:
Case 1: Individual (Chooser)-Specific Data
Case 2: Choice-Specific Data
Case 3: Mixed Characteristics

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 4 / 30
The Intuition: Utility Maximization
The Behavioral Story:
Imagine every student assigns a ”Utility Score” (U) to each option
based on characteristics (e.g., tuition, prestige, distance).
We cannot see the score, but we see the choice.

The Decision Rule:

Choose Option j ⇐⇒ Uj > Uk for all k ̸= j
Example:
A student chooses a 4-Year College if:
U4-Year > U2-Year AND U4-Year > UNone

Key Takeaway
Multinomial Logit calculates the probability that the utility of one choice is
higher than the utility of all other choices.
Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 5 / 30
Case 1: Individual (Chooser)-Specific Data
Description:
The independent variables are related to individual characteristics
that influence their choices.
These variables do not depend on the choice alternatives.
Example: Education Choice
A person’s education level (secondary, diploma, undergraduate,
postgraduate, PhD) may depend on:
Age
Family income
Religion
Caste
Parents’ education
Estimation Method:
Typically estimated using Multinomial Logit (MNL) or Multinomial
Probit (MNP) models.
Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 6 / 30
Case 2: Choice-Specific Data (Conditional Logit)
Definition:
In this case, the independent variables describe the attributes of the
choices, not the person making the decision.
We assume that a person compares the features (e.g., price, speed) of
the available options.
Key Characteristic:
The variable is the same for all options (e.g., ”Price”), but the value
varies across options.
Example: Every car has a price, but Car A is $20k and Car B is $30k.
Example: Choosing a Mode of Transport A commuter’s choice depends on:
Travel Time (Bus takes 40 min, Train takes 20 min).
Cost (Bus is $2, Train is $5).
Comfort (Bus is crowded, Train is spacious).
Note: Individual characteristics (like the commuter’s Age or Income) are NOT included
in this specific model structure.
Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 7 / 30
The Long Format Data Structure
In Case 2 (Conditional Logit), we analyze options, not just people.
We cannot use the standard ”Wide Format” (one row per person).
We must use the ”Long Format” (one row per person-choice
combination).

Choosing Transport (Bus vs. Train) Notice how the variable Price
varies within the same person.

Person ID Alternative Price (Var X) Choice (Y)

1 (Alice) Bus $2.00 0
1 (Alice) Train $5.00 1
2 (Bob) Bus $2.00 1
2 (Bob) Train $5.00 0

Observation: Alice chooses the Train despite the higher price; Bob chooses the Bus.
Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 8 / 30
Data Structure: Wide vs. Long Format
The Problem: Standard regression uses ”Wide” data. Conditional Logit
requires ”Long” data to analyze attributes like Price.
1. Wide Format (Standard) One row per person

Person Choice Price Bus Price Train

Alice Train 2.00 5.00

↓ Reshape
2. Long Format (Required) One row per option

Person Option Price (Single Var) Choice

Alice Bus 2.00 0
Alice Train 5.00 1

Why? In Long Format, ”Price” is a single variable. We can estimate one βPrice that
applies to both buses and trains.
Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 9 / 30
Case 3: Mixed Characteristics

Description:
The independent variables include a mix of chooser-specific and
choice-specific characteristics.
Example: Car Choice with Individual and Choice Characteristics
A person’s car choice (Type A, Type B, Type C) may depend on:
Choice-specific factors: Price, fuel efficiency, maintenance cost.
Individual-specific factors: Age, income.
Estimation Method:
Estimated using Conditional Logit Model by incorporating interaction
terms or dummy variables for individual characteristics.

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 10 / 30
Summary of Nominal Multinomial Regression Models
Comparison of the Three Cases:

Case Independent Variables Estimation Method

Individual-Specific Socioeconomic factors Multinomial Logit / Probit
Choice-Specific Choice attributes Conditional Logit / Probit
Mixed Both individual and choice attributes Conditional Logit with interaction terms

Key Takeaways:
Multinomial Logit (MNL) is used when choice-specific factors are not
included.
Conditional Logit (CL) is better suited when choice attributes vary
across alternatives.
Mixed models incorporate both types of variables and require careful
specification.
Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 11 / 30
Critical Assumption: Independence of Irrelevant
Alternatives (IIA)

The Intuition: The ”Copycat” Problem MNL assumes that if you

prefer Coffee over Tea, adding Hot Chocolate to the menu shouldn’t
change your relative preference for Coffee vs. Tea.
The Famous Counter-Example: Red Bus / Blue Bus Imagine a city
where commuters choose between a Car and a Red Bus.

Scenario 1: The Original Market Scenario 2: The Disruption

Options: Car, Red Bus. We introduce a Blue Bus.
Consumer Split: 50% drive It is identical to the Red Bus in
Car, 50% take Red Bus. speed, cost, and comfort. Only
50 the paint is different.
Odds (Car / Bus): 50 = 1.0.

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 12 / 30
Why MNL fails
What happens when the Blue Bus arrives?
1. The Logic (MNL Prediction)
2. The Human Logic (Reality)
MNL says: ”The odds of Car vs. Red
Bus MUST stay 1.0.” Car drivers keep driving (they don’t
care about bus colors).
It forces the probabilities to split
equally three ways: Bus riders simply split their votes.
33% Car, 33% Red Bus, 33% Blue 50% Car, 25% Red Bus, 25% Blue
Bus. Bus.
The Flaw: It predicts car usage drops The odds of Car vs. Red Bus changed
50
from 50% to 33% just because we to 25 = 2.0. IIA is violated.
painted some buses blue!

Nested Logit
If options are copycats (substitutes), we group them into a ”Nest”.
Step 1: Choose Private vs. Public Transport.
Step 2: If Public, choose Red vs. Blue.

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 13 / 30
Multinomial Logit Model

Step 1: Choose a Baseline Category

Suppose we select Withdraw as the reference (baseline).
In MNL, coefficients are meaningless without a reference point.

Intuition: The ”Height Analogy”

Imagine measuring the height of three students: Alice, Bob, and Charlie.
Absolute (Standard Regression): ”Alice is 170cm tall.”
Relative (Multinomial Logit): ”Alice is 10cm taller than Bob.”
In this model, we treat the Base Outcome (Withdraw) as ”Bob.”
We don’t ask: ”Does X increase the chance of Passing?”
We ask: ”Does X make Passing more likely than Withdrawing?”

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 14 / 30
Multinomial Logit Model

Step 2: Log-Odds (Vector Notation) Let x be the vector of predictors

(e.g., GPA, Income) and β j be the coefficients for outcome j.
 
πPass
ln = xβ Pass
πWithdraw
 
πFail
ln = xβ Fail
πWithdraw

Interpretation of xβ:
xβ Pass represents the relative utility (or score) of Passing compared
to Withdrawing.
Note: For the baseline itself, β Withdraw = 0, so xβ Withdraw = 0.

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 15 / 30
Recovering Probabilities
We convert the ”relative scores” (xβ) back into probabilities using the
logistic transformation.

Probability of Pass:
e xβPass
πPass =
1 + e xβPass + e xβFail
Probability of Fail:
e xβFail
πFail =
1 + e xβPass + e xβFail
Probability of Withdraw (Base):
1
πWithdraw =
1+ e xβPass + e xβFail

Note: The 1 in the denominator represents the base category, since e xβBase = e 0 = 1.
Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 16 / 30
Recovering Probabilities: The Logic

Understanding the Denominator Notice that all three equations share

the same denominator D:
X
D =1+ e xβj
j̸=base

Key Property: Normalization

πPass + πFail + πWithdraw = 1

The denominator sums up the ”utility scores” of all possible options.

Dividing by this sum ensures the total probability equals 100%.
Intuition: This guarantees that every student ends up in exactly one
category (no one is left unaccounted for).

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 17 / 30
Relative Risk Ratios (RRR)
Definition: The coefficients (β) gives us log-odds. To get interpretable ratios, we
exponentiate them:
RRR = e β
This tells us how the relative risk (or odds) of an outcome changes for a
one-unit increase in x.

Example Calculation: Suppose βHours = 0.50 for the ”Pass” outcome.

RRRPass vs. Withdraw = e 0.50 ≈ 1.65

Interpretation:
An RRR of 1.65 means that for each additional hour of study, the relative
risk of Passing (relative to Withdrawing) increases by a factor of 1.65.
In percentage terms: It increases by 65% (1.65 − 1.00 = 0.65).

Note on Terminology
In Stata and many textbooks, this is called the Relative Risk Ratio (RRR), though it
functions similarly to an Odds Ratio in binary logit.
Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 18 / 30
Problem Statement
Objective: We aim to model what factors affect a student’s choice
regarding the type of college they attend.
Outcome (Y) Variable:
psechoice (Type of College)
Categories:
1 = No College
2 = 2-Year College
3 = 4-Year College
Predictor (X) Variables:
hscath: Type of School Graduate (Catholic school or not)
grades: Average grades in Maths, English, and Social Studies
faminc: Gross Family Income (in thousands)
famsiz: Number of family members
parcoll: Parents’ Educational Status (1 = College, 0 = Others)
female: Gender (1 = Female, 0 = Others)
black: Race/Ethnicity (1 = Black, 0 = Others)
Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 19 / 30
Data Set Overview

The dataset contains information on 6,649 students across 14

variables.
The dependent variable is psechoice, indicating the student’s college
choice.
There are 7 predictor (independent) variables influencing the
outcome.

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 20 / 30
Understanding Key Variables

grades: Captures average grades in Maths, English, and Social

Studies.
Measured on a 13-point scale (1 = highest score, 13 = lowest score).
faminc: Gross family income (in thousands).
famsiz: Number of family members.
parcoll: Parents’ educational status (1 = College, 0 = Others).
female: Gender (1 = Female, 0 = Others).
black: Race/Ethnicity (1 = Black, 0 = Others).
hscath: Catholic school graduate (1 = Yes, 0 = No).

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 21 / 30
Step 2: Labelling Categorical Response Variables

Defining Labels for Categorical Variables:

Code for Labeling Categorical Variables

label define Graduate School Type 1 ”Catholic school graduate” 0 ”Others”

label values hscath Graduate School Type
label define Parent Education 1 ”College” 0 ”Others”
label values parcoll Parent Education
label define Gender 1 ”Female” 0 ”Others”
label values female Gender
label define Race Ethnicity 1 ”Black” 0 ”Others”
label values black Race Ethnicity label define School Choice 1 "No College" 2 "2-Year College" 3 "4-Year
College"
label values psechoice School Choice
label define Graduate School Type 1 "Catholic school graduate" 0 "Others"
label values hscath Graduate School Type
label define Parent Education 1 "College" 0 "Others"
label values parcoll Parent Education
label define Gender 1 "Female" 0 "Others"
label values female Gender
label define Race Ethnicity 1 "Black" 0 "Others"
label values black Race Ethnicity

Purpose: Ensures that categorical variables are properly labeled for interpretation.

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 22 / 30
Step 3: Describing Data Set

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 23 / 30
Step 4: Tabulating Categorical Variables

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 24 / 30
Step 5: Tabulating Descriptive Statistics

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 25 / 30
Step 6: Model Description

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 26 / 30
Step 7: Estimating Multinomial Logit Model

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 27 / 30
Step 8: Results

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 28 / 30
Step 9: Interpretation of Model Coefficients

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 29 / 30
Step 10: Interpretation

Dr. Vijay Victor (IIT Roorkee) Multinomial and Ordinal Logit February 4, 2026 30 / 30