Discrete Choice Modeling: Latent Utility and Parametric Estimators
In econometrics, modeling how individuals make decisions among a set of alternatives is a fundamental challenge. This process is typically handled through discrete choice problems, where the goal is to determine why an agent selects one specific option over others. The core theory suggests that these choices are not random but are driven by an underlying, unobservable value known as latent utility.
To model this, we consider a population of agents (T) and a common set of choices (C). For any agent, the choice is represented as a binary indicator: it is 1 if a specific option is chosen and 0 otherwise. The model assumes that the latent utility is linear, consisting of observable factors and an additive response error.
[ไม่มีภาพประกอบ]
The Mechanics of Latent Utility
The decision-making process is defined by the comparison of utilities. An agent will choose option i if the utility derived from it is higher than the utility of any other available option j. This is expressed mathematically as the sum of observable covariates (x) multiplied by a coefficient (β), plus an error term (ϵ).
The observable covariates are highly flexible and can include both agent-specific characteristics and choice-specific attributes. For example, if the choice set consists of different coffee brands, the covariates might include:
- Agent characteristics: Age, gender, income, and ethnicity.
- Product characteristics: Price, taste, and whether the coffee is local or imported.
The error terms represent factors influencing the decision that the econometrician cannot observe. These are assumed to be independent and identically distributed (i.i.d.). The primary objective of the model is to estimate β, which reveals how different factors impact the final choice.
Parametric Estimators and Distribution Assumptions
To estimate the parameter β, researchers often impose specific distribution assumptions on the error term. These are known as parametric estimators. While these assumptions make computation more convenient, they introduce the risk of inconsistency if the distribution is misspecified.
Common Parametric Models
- Multinomial Probit Model: Used when the error term is assumed to follow a normal distribution.
- Multinomial Logit Model: Used when the error term is assumed to follow a Gumbel distribution.
Binary Response Models
A simplified version of this framework is the binary choice model, where the choice set (C) contains only two items. In this scenario, the agent chooses option 1 if its latent utility exceeds that of option 2.
The likelihood of this choice is calculated using a log likelihood function. If a normal distribution is assumed for the response error, the model becomes a probit model. This model utilizes the cumulative distribution function (CDF) of the standard normal distribution (Φ). Although the CDF itself lacks a closed-form representation, its derivative does, making the model computationally tractable.
[ไม่มีภาพประกอบ]
Distribution-Free Alternatives
To avoid the risks associated with misspecifying the error distribution, distribution-free models can be used. Instead of relying on a specific probability distribution in the log-likelihood function, these models replace probability terms with general weights (W), providing a more flexible approach to estimation.
Key Facts
- Latent Utility: The unobserved value that determines a discrete choice.
- Linearity: Latent utility is assumed to be linear in explanatory variables with an additive error.
- Probit vs. Logit: Probit models assume a normal distribution of errors, while Logit models assume a Gumbel distribution.
- Consistency Risk: Parametric models may produce inconsistent estimates if the error distribution is incorrectly specified.
- Binary Choice: A specific case of discrete choice where only two alternatives exist.
| Model Type | Error Distribution Assumption | Key Characteristic |
|---|---|---|
| Multinomial Probit | Normal | Based on standard normal CDF |
| Multinomial Logit | Gumbel | Computationally convenient |
| Binary Choice | Varies (e.g., Normal) | Limited to two alternatives |
| Distribution-Free | None | Uses weights instead of fixed probabilities |
Frequently Asked Questions
What is latent utility in discrete choice modeling?
Latent utility is the underlying, unobservable value an agent assigns to a choice. The agent is assumed to select the option that provides the highest latent utility.
What are observable covariates?
Observable covariates are the measurable variables used to explain a choice. These can include personal demographics (like age or income) or product attributes (like price or quality).
What is the difference between a Probit and a Logit model?
The primary difference lies in the assumed distribution of the response error: the Probit model assumes a normal distribution, while the Logit model assumes a Gumbel distribution.
Why is distribution misspecification a problem?
If the assumed distribution of the error term does not match the actual distribution in the data, the resulting parametric estimates for β may be inconsistent, leading to inaccurate conclusions.
How does a distribution-free model work?
A distribution-free model removes the reliance on a specific probability distribution by replacing the probability terms in the log-likelihood function with general weights.