Random Utility and Product Choice Probabilities

From Utility to Choice Probability

A demand model must connect what a consumer values to the product the consumer chooses. In a discrete-choice setting, each consumer selects one alternative from a finite set: a product, another product, or the option of purchasing none of the listed products. The choice is not necessarily deterministic from the analyst’s perspective. Consumers may respond differently to characteristics the analyst does not observe, and these unobserved influences can be represented as utility shocks.

The key distinction is between the systematic part of utility, which is described by observed product and consumer information and model parameters, and the unobserved part, which varies across alternatives. Given a probability model for those shocks, the utility model implies a probability for each individual consumer choosing each alternative. These are individual choice probabilities, conditional on the consumer’s observed situation and the model.

Definition: Let $J_i$ be the set of listed products available to consumer $i$, and let $0$ denote the outside alternative, such as buying none of those products. Consumer $i$’s utility from alternative $j\in J_i\cup\{0\}$ is $$ U_{ij}=v_{ij}+\varepsilon_{ij}, $$ where $v_{ij}$ is systematic indirect utility and $\varepsilon_{ij}$ is an unobserved utility shock. The consumer chooses an alternative that maximizes $U_{ij}$. If the maximizer is unique with probability one, the choice probability is $P_{ij}=P(U_{ij}>U_{ik}\text{ for every }k\ne j)$. If ties can occur with positive probability, a tie-breaking rule must instead be specified and included in the definition of the choice probabilities.

The utility $v_{ij}$ may depend on product characteristics $x_j$, consumer attributes $z_i$, prices $p_j$, and parameters $\theta$. For example, a model might include quality and price directly, as well as an interaction between quality and income. Product characteristics describe the available alternatives; consumer attributes describe the person making the choice. An interaction lets a product characteristic have different utility implications for consumers with different attributes. The outside alternative also has systematic utility $v_{i0}$ and a shock $\varepsilon_{i0}$; omitting it would force every consumer to choose a listed product.

Worked Example: Separating Product and Consumer Information

Suppose consumer $i$ considers products A and B, as well as an outside option. Product quality is $q_j$, price is $p_j$, and the consumer attribute $z_i$ measures income in units of ten thousand dollars. Consider the systematic utility specification $$ v_{ij}=0.8q_j-0.5p_j+0.3z_iq_j,\qquad v_{i0}=0. $$ For this hypothetical consumer, let $z_i=1$. Product A has $(q_A,p_A)=(2,3)$, and product B has $(q_B,p_B)=(1,1)$.

The quality and price values belong to the products, while $z_i=1$ belongs to this consumer. The interaction $z_iq_j$ combines them. Substitution gives $$ v_{iA}=0.8(2)-0.5(3)+0.3(1)(2)=0.7, \qquad v_{iB}=0.8(1)-0.5(1)+0.3(1)(1)=0.6. $$ The outside option has systematic utility $v_{i0}=0$. These values alone rank the systematic utilities as A, B, outside; they do not yet give choice probabilities. Probabilities also depend on the distribution of the shocks.

A General Choice Probability and a Useful Shock Model

For any specified joint distribution of the shocks, the probability of choosing $j$ is the probability that its total utility exceeds the total utility of every other available alternative, subject to the uniqueness or tie-breaking condition in the definition. This probability depends on the whole choice set. Even if the systematic utility of product A does not change, its probability can change when another alternative is added or removed.

A particularly tractable specification assumes that the shocks are independent and identically distributed type-I extreme-value random variables with scale $\tau>0$. Their common cumulative distribution function is $F(e)=\exp[-\exp(-e/\tau)]$, with density $f(e)=\tau^{-1}\exp(-e/\tau)\exp[-\exp(-e/\tau)]$. This continuous distribution makes exact ties have probability zero. Under these assumptions, the probability formula can be derived directly.

Result: If the utility shocks are independent and identically distributed with the type-I extreme-value distribution of scale $\tau>0$, then, conditional on the systematic utilities, $$ P_{ij}=\frac{\exp(v_{ij}/\tau)}{\displaystyle\sum_{k\in J_i\cup\{0\}}\exp(v_{ik}/\tau)}. $$ The sum includes every available listed product and the outside alternative.

To see why, fix an alternative $j$. If its shock equals $e$, then it is chosen exactly when every other alternative’s shock is below the threshold that makes its total utility lower than $j$’s. Independence lets us multiply those conditional probabilities and integrate over $j$’s shock.

Condition on the candidate alternative’s shock.
For $j$ to win when $\varepsilon_{ij}=e$, each $k\ne j$ must satisfy $\varepsilon_{ik}<e+v_{ij}-v_{ik}$. The conditional probability of this event is the product of the corresponding CDF values.

Integrate over that shock.
Using the common density $f$ and CDF $F$, the probability is $\int_{-\infty}^{\infty}f(e)\prod_{k\ne j}F(e+v_{ij}-v_{ik})\,de$, where $k$ ranges over all available alternatives, including 0.

Evaluate the integral.
The extreme-value CDF turns the product into an exponential in $\exp(-e/\tau)$. Substituting $t=\exp(-e/\tau)$ reduces the expression to an exponential integral, yielding the stated probability.

More explicitly, insert the density and CDF into the integral. With $A=\sum_{k\in J_i\cup\{0\}}\exp((v_{ik}-v_{ij})/\tau)$, its integrand becomes $$ \frac{1}{\tau}\exp(-e/\tau)\exp\!\left[-A\exp(-e/\tau)\right]. $$ Under the substitution $t=\exp(-e/\tau)$, $dt=-(1/\tau)\exp(-e/\tau)\,de$. As $e$ runs from $-\infty$ to $+\infty$, $t$ runs from $+\infty$ to 0. Therefore, $$ P_{ij}=\int_0^\infty \exp(-At)\,dt=\frac{1}{A} =\frac{\exp(v_{ij}/\tau)}{\displaystyle\sum_{k\in J_i\cup\{0\}}\exp(v_{ik}/\tau)}. $$ The final equality follows by multiplying the numerator and denominator of $1/A$ by $\exp(v_{ij}/\tau)$.

Worked Example: Calculating Probabilities Including the Outside Option

Return to the systematic utilities $v_{iA}=0.7$, $v_{iB}=0.6$, and $v_{i0}=0$. Suppose the shocks follow the independent type-I extreme-value model with $\tau=1$. The choice set includes A, B, and the outside option, so the common denominator is $$ \exp(0.7)+\exp(0.6)+\exp(0) =2.0138+1.8221+1 =4.8359. $$

Thus, $$ P_{iA}=\frac{2.0138}{4.8359}\approx0.4164,\qquad P_{iB}=\frac{1.8221}{4.8359}\approx0.3768,\qquad P_{i0}=\frac{1}{4.8359}\approx0.2068. $$ The probabilities sum to one before rounding; their displayed values sum to 1.0000. Conditional on this consumer’s observed information and the specified model, the probabilities of choosing A, B, or neither listed product are approximately 41.64%, 37.68%, and 20.68%. They are not statements that the consumer makes a fraction of a choice.

Utility Differences, Scale, and Price Effects

The formula makes clear that choice depends on utility relative to the other available alternatives. Adding the same constant to every $v_{ik}$ leaves all probabilities unchanged, because it multiplies every exponential by the same factor. The scale $\tau$ also matters: it measures the spread of the unobserved shocks relative to differences in systematic utility. A larger $\tau$ makes those differences less decisive in the probability formula. Since multiplying all utilities and $\tau$ by the same positive constant leaves the probabilities unchanged, utility units require a scale normalization when parameters are estimated.

For a direct utility effect, differentiate the probability formula. If $v_{ij}$ changes while the other alternatives’ systematic utilities stay fixed, then $$ \frac{\partial P_{ij}}{\partial v_{ij}}=\frac{1}{\tau}P_{ij}(1-P_{ij}). $$ For a different alternative $k\ne j$, changing its systematic utility has effect $$ \frac{\partial P_{ij}}{\partial v_{ik}}=-\frac{1}{\tau}P_{ij}P_{ik}. $$ These derivatives show that a higher utility for one option raises its own probability and lowers the probabilities of the others. A price effect follows through the utility specification, not from the probability formula alone.

Worked Example: A Price Change and an Individual Probability

Suppose a consumer faces products A and B and the outside option, with systematic utilities $v_{iA}=0.5$, $v_{iB}=0.2$, and $v_{i0}=0$, and $\tau=1$. Assume A’s price enters its systematic utility as $v_{iA}=a-0.4p_A$, where $a$ and the other alternatives’ utilities are fixed. The choice probability for A at the stated utilities is $$ P_{iA}=\frac{\exp(0.5)}{\exp(0.5)+\exp(0.2)+1} =\frac{1.6487}{1.6487+1.2214+1} \approx0.4260. $$

Because $\partial v_{iA}/\partial p_A=-0.4$, the chain rule gives $$ \frac{\partial P_{iA}}{\partial p_A} =\frac{\partial P_{iA}}{\partial v_{iA}}\frac{\partial v_{iA}}{\partial p_A} =-0.4(0.4260)(1-0.4260) \approx-0.0978. $$ At these values, a one-unit increase in A’s price changes its individual choice probability locally by about $-0.0978$, or $-9.78$ percentage points per price unit. This is a local derivative under the stated utility model, not an exact prediction for a large price change. It also describes a model-based response, not necessarily a causal effect unless the model’s assumptions support that interpretation.

Common Pitfalls and Interpretation

The outside option is a genuine alternative in the maximization, not a residual category that can be ignored until after probabilities for products have been calculated. Its utility and shock belong in the same model, and its exponential term belongs in the denominator. Omitting it generally inflates the probabilities assigned to listed products and changes the model’s implications.

Another mistake is to treat the ordering of systematic utilities as if it determined the choice with certainty. Under the continuous-shock model, the highest $v_{ij}$ gives the highest choice probability, but other alternatives can still be chosen. A third error is to define each probability as the probability that an alternative weakly beats all the others without addressing ties. If utilities can tie with positive probability, those events may overlap; they need not define probabilities that sum to one. Establish uniqueness almost surely or state a tie-breaking rule.

Finally, the probability formula is conditional on the specified systematic utilities, available alternatives, and shock distribution. It does not by itself establish that the utility specification is empirically adequate, that parameters are identified, or that a price coefficient has a causal interpretation. Those conclusions require additional assumptions and analysis. The individual probability formula is the model-level link between utility and choice; market-level aggregation is a separate step.

Key takeaway: A random-utility model assigns each alternative systematic utility plus an unobserved shock. Under independent, identically distributed type-I extreme-value shocks, an individual’s probability of choosing an alternative is its exponentiated systematic utility divided by the sum over all available alternatives, including the outside option.

Check Your Understanding

Answer each question using the choice model and assumptions stated in the tutorial.

In $v_{ij}=0.8q_j-0.5p_j+0.3z_iq_j$, which quantities describe the product, which describe the consumer, and which term combines both?
Why must the outside alternative appear in the probability denominator when the consumer can choose not to buy a listed product?
For two listed products and an outside option, explain why probabilities defined using weak utility inequalities may fail to sum to one if ties have positive probability.
Under the extreme-value model, what happens to an alternative’s choice probability if its systematic utility increases while all other utilities remain fixed?
If $\tau=1$, $v_{iA}=0$, and $v_{i0}=0$, what is the probability of choosing A when these are the only two alternatives?

Random Utility and Product Choice Probabilities

What You'll Learn