Multinomial Logit Shares and Utility Inversion

From Choice Probabilities to Market Shares

In discrete-choice demand, the starting point is a consumer’s probability of choosing each available option. Market shares aggregate those probabilities across consumers. In the simple multinomial-logit model, this aggregation has a particularly direct form: when consumers share the same mean utilities and differ only through identically distributed random utility shocks, each product’s market share equals its choice probability.

The model has an inside option for each product \(j=1,\ldots,J\) in market \(t\), and an outside option indexed by \(0\). The outside option represents not choosing any of the inside products. The options are assumed to be mutually exclusive and exhaustive: each consumer chooses exactly one of them. This tutorial uses \(s_{jt}\) for product \(j\)’s market share and \(s_{0t}\) for the outside-good share.

Definition: In the simple multinomial-logit model, consumer \(i\)’s utility from inside product \(j\) in market \(t\) is \(u_{ijt}=\delta_{jt}+\varepsilon_{ijt}\), while utility from the outside option is \(u_{i0t}=\varepsilon_{i0t}\). The mean utility of the outside option is normalized to zero. The shocks \(\varepsilon_{i0t},\ldots,\varepsilon_{iJt}\) are independent and identically distributed type-I extreme-value random variables with scale normalized to one. The consumer chooses the option with the highest utility.

The scale normalization fixes the units in which utilities are measured. If the shocks instead had scale \(\sigma\), choice probabilities would depend on \(\delta_{jt}/\sigma\), so shares would identify utility relative to that scale rather than an unrestricted utility level. Here, recovered mean utilities are in normalized logit utility units.

Under the stated extreme-value assumption, the probability of choosing an option is proportional to the exponential of its mean utility. For the outside option, the normalized mean utility is zero, so its exponential term is \(\exp(0)=1\). Thus the choice probabilities are

P_{jt} = \frac{\exp(\delta_{jt})} {1+\sum_{k=1}^{J}\exp(\delta_{kt})}, \qquad P_{0t} = \frac{1} {1+\sum_{k=1}^{J}\exp(\delta_{kt})}.

For a population with the same mean utilities and the same shock distribution, the market share \(s_{jt}\) is the probability that a consumer in that population chooses product \(j\). Therefore \(s_{jt}=P_{jt}\), and \(s_{0t}=P_{0t}\). With a finite market, realized sample shares can differ from these probabilities because of the particular consumers and shocks observed; the model equations describe population probabilities or expected shares.

How the Outside Option Enters the Share Formula

The denominator includes the exponential terms for every inside product and the outside option. The outside option is not an extra product with an arbitrary denominator contribution: its contribution is one because its mean utility has been normalized to zero. Each inside product’s share is therefore determined both by its own mean utility and by the mean utilities of all competing options.

The options are mutually exclusive and exhaustive, so their shares sum to one:

s_{0t}+\sum_{j=1}^{J}s_{jt}=1.

Consequently, if the complete set of inside market shares is known, the outside share is the residual \(s_{0t}=1-\sum_{j=1}^{J}s_{jt}\). It is not an independent unknown in that case. However, conditional shares that describe only how choices are divided among inside products do not reveal the total share choosing an inside product, and therefore do not reveal the outside share.

Worked Example: Computing Shares From Mean Utilities

Consider a hypothetical market with two inside products. Suppose their normalized mean utilities are \(\delta_{1t}=\log(2)\) and \(\delta_{2t}=0\). Assume the simple multinomial-logit model above applies: the two inside products and the outside option are exhaustive alternatives, and the utility shocks are independent and identically distributed type-I extreme-value variables with scale one.

The exponentiated mean utilities are \(\exp(\delta_{1t})=2\) and \(\exp(\delta_{2t})=1\). Including the outside option’s term of one gives a denominator of \(1+2+1=4\). Thus

s_{1t}=\frac{2}{4}=0.5000,\qquad s_{2t}=\frac{1}{4}=0.2500,\qquad s_{0t}=\frac{1}{4}=0.2500.

The shares sum to \(0.5000+0.2500+0.2500=1.0000\), as required. In this hypothetical population, half of consumers choose product 1, one quarter choose product 2, and one quarter choose the outside option.

Inverting Market Shares to Recover Mean Utilities

The share equation can be inverted by comparing an inside product’s share with the outside-good share. Both probabilities have the same denominator, so that denominator cancels in their ratio:

\frac{s_{jt}}{s_{0t}} = \frac{ \exp(\delta_{jt})/ \left(1+\sum_{k=1}^{J}\exp(\delta_{kt})\right) }{ 1/ \left(1+\sum_{k=1}^{J}\exp(\delta_{kt})\right) } = \exp(\delta_{jt}).

Taking logarithms gives the inversion. This is a direct algebraic result from the probability formulas; it recovers each inside product’s mean utility relative to the normalized outside option.

Share inversion: For a complete set of market shares generated by the normalized simple multinomial-logit model, with \(s_{0t}>0\) and \(s_{jt}>0\),

\delta_{jt}=\log\left(\frac{s_{jt}}{s_{0t}}\right), \qquad j=1,\ldots,J.

The ratio \(s_{jt}/s_{0t}\) is dimensionless, so its logarithm has no physical units. Its interpretation is in normalized utility units: a one-unit increase in \(\delta_{jt}\), holding the other mean utilities fixed, multiplies the product’s odds relative to the outside option by \(\exp(1)\). The inversion recovers the mean utility of each inside product relative to the outside option, not an absolute utility level independent of the normalization.

Worked Example: Recovering Utilities From Complete Shares

Suppose a hypothetical market has 1,000 consumers, and the observed choices are 400 for product 1, 300 for product 2, and 300 for the outside option. Assume these shares are being used as population shares for the simple multinomial-logit inversion. The options are exhaustive, and all three shares are positive.

The shares are \(s_{1t}=400/1000=0.4000\), \(s_{2t}=300/1000=0.3000\), and \(s_{0t}=300/1000=0.3000\). Apply the inversion separately to each inside product:

\delta_{1t} = \log\left(\frac{0.4000}{0.3000}\right) = \log(1.3333) \approx 0.2877,

\delta_{2t} = \log\left(\frac{0.3000}{0.3000}\right) = \log(1) = 0.

The recovered mean utility of product 1 is approximately \(0.2877\) normalized utility units relative to the outside option, and product 2’s recovered mean utility equals the normalized outside-option level. These values reproduce the shares under the model: the exponentiated utilities are approximately \(1.3333\) and \(1\), and the denominator is \(1+1.3333+1=3.3333\). The resulting shares are \(1.3333/3.3333=0.4000\), \(1/3.3333=0.3000\), and \(1/3.3333=0.3000\).

If these counts are realized observations rather than exact population probabilities, the calculated values are plug-in inversions of sample shares. They need not equal the underlying population mean utilities exactly.

Conditional Inside Shares Are Not Complete Market Shares

A conditional inside share is the fraction of inside-option choosers selecting a particular inside product. Write it as \(q_{jt}=s_{jt}/\sum_{k=1}^{J}s_{kt}\). These shares sum to one across inside products, but they omit the fraction of the market choosing the outside option. Consequently, conditional shares identify relative shares among inside products, not each inside product’s share relative to the outside option.

In fact, the ratio of two inside products’ shares eliminates the common denominator:

\frac{s_{jt}}{s_{\ell t}} = \exp(\delta_{jt}-\delta_{\ell t}), \qquad \log\left(\frac{s_{jt}}{s_{\ell t}}\right) = \delta_{jt}-\delta_{\ell t}.

Thus conditional shares can reveal differences in mean utility among inside products under the model. To recover each inside utility relative to the outside option, the total inside share—or, equivalently, the outside-good share—is also needed. Importantly, when all unconditional inside shares are known, the outside share is determined by their sum; the information gap applies when only conditional shares are supplied.

Worked Example: Why Conditional Shares Need the Outside Share

Suppose a hypothetical report says that among consumers choosing an inside product, 60% choose product 1 and 40% choose product 2. These are conditional inside shares. Assume the simple multinomial-logit model and the same exhaustive choice set as above, but suppose the report does not state the fraction choosing any inside product.

If the total inside share is \(0.8000\), then the outside share is \(0.2000\), and the unconditional inside shares are \(s_{1t}=0.6000(0.8000)=0.4800\) and \(s_{2t}=0.4000(0.8000)=0.3200\). The inverted utilities are

\delta_{1t}=\log(0.4800/0.2000)=\log(2.4)\approx 0.8755, \qquad \delta_{2t}=\log(0.3200/0.2000)=\log(1.6)\approx 0.4700.

If instead the total inside share is \(0.5000\), the same conditional shares imply unconditional shares \(s_{1t}=0.3000\), \(s_{2t}=0.2000\), and \(s_{0t}=0.5000\). The corresponding utilities are

\delta_{1t}=\log(0.3000/0.5000)=\log(0.6)\approx -0.5108, \qquad \delta_{2t}=\log(0.2000/0.5000)=\log(0.4)\approx -0.9163.

Both cases have conditional inside shares of 60% and 40%, but the utilities relative to the outside option differ because the total inside share differs. The example also illustrates that the complete unconditional shares, including the outside share as their residual, provide the information needed for inversion.

Conditions, Boundaries, and Common Mistakes

The inversion applies to positive shares generated under the specified normalized simple multinomial-logit model. A zero share cannot be inserted into the logarithm: if \(s_{jt}=0\), the formula is undefined, and the data do not provide a finite plug-in utility through this expression. A zero outside share creates the same problem for every inside product. Such cases require attention to the data and modeling setup; replacing zero mechanically with a small positive number changes the calculation and does not follow from the inversion itself.

The model also imposes structure. The probabilities rely on the type-I extreme-value shocks, their common normalized scale, the outside-option normalization, and the stated choice set. The share inversion is not a general identity for arbitrary choice probabilities or for a model with unaccounted consumer heterogeneity. In particular, one should not interpret the recovered \(\delta_{jt}\) as a directly observed characteristic or as a causal effect.

Using conditional shares as if they were unconditional. A conditional split such as 60% and 40% among inside buyers does not supply the outside share. Obtain the total inside share or the outside share before applying \(\log(s_{jt}/s_{0t})\).
Claiming that complete inside shares leave the outside share unknown. With an exhaustive choice set, complete unconditional inside shares determine it as \(1-\sum_j s_{jt}\). The missing information problem concerns conditional inside shares alone.
Omitting the outside option’s denominator term. The denominator is \(1+\sum_k\exp(\delta_{kt})\), not just the sum over inside products. The one comes from the normalized outside mean utility.
Reversing the share ratio. The specified inversion is \(\log(s_{jt}/s_{0t})\). Reversing the ratio changes the sign and gives the outside-versus-inside comparison.
Treating sample shares as exact probabilities. Counts from a finite market produce empirical shares. Their plug-in inversion is a calculation from the observed data, not a guarantee that the underlying utilities have been recovered without sampling error.

The practical value of the formula is that it translates a complete set of market shares into normalized mean utilities with a simple calculation. It also makes clear why defining the outside option and the market-share denominator is essential. When product-level utility is later related to prices or observed characteristics, these recovered utilities can serve as demand-side objects; the inversion itself does not specify or estimate those relationships.

Check Your Understanding

Use the normalized simple multinomial-logit model and state any additional information needed for each calculation.

There is one inside product with share \(0.2500\). What is the outside share, and what mean utility does the share inversion give?
Two inside products have complete shares \(0.2000\) and \(0.5000\). Find the outside share and the two normalized mean utilities.
Explain why conditional inside shares of 70% and 30% do not, by themselves, determine utilities relative to the outside option.
If an inside product’s share equals the outside share, what is its normalized mean utility?
What does a zero observed outside share imply for the direct log-share inversion, and why should one not substitute zero into the formula?