To predict the effect of a merger,
we need to know where sales go when a price changes and how firms take
those sales into account. Prices and shares alone rarely answer either
question. antitrust combines observed market information
with a demand system and an assumption about firm behavior to recover
the parameters needed for a simulation.
The answer is conditional on those inputs and assumptions. A different margin, diversion estimate, or conduct model can change recovered demand and costs, and therefore the predicted effect. Calibration does not provide sampling standard errors; checking plausible inputs and model choices is part of an economic analysis. The workflow vignette (source) shows how to use a fitted market. Read it first for an executable example; here we explain what its fit means. The architecture guide (source) is for extension authors.
Most calibrations start with product prices \(p\), quantities or shares \(q\), pre-change ownership \(\Omega\), and at least some gross margins \(m_i=(p_i-c_i)/p_i\). A model may also require diversion ratios, elasticities, capacities, bargaining powers, nest assignments, or auction moments. Check what a “share” means before supplying one: Logit uses quantity shares, while CES and AIDS use revenue shares. Utilities and valuations may also need an observed outside share or a normalization to fix their common level.
Consider Bertrand pricing. Firm \(f\) chooses the prices of its products to maximize \(\sum_{j\in J_f}(p_j-c_j)q_j(p)\). For one of those products, \(i\), the first-order condition is
\[ q_i(p)+\sum_{j\in J_f}(p_j-c_j)\frac{\partial q_j(p)}{\partial p_i}=0. \]
The first term is the gain from charging a slightly higher price on units of \(i\) still sold. The sum captures lost sales of \(i\) and any change in profits on the firm’s other products. A merger adds terms to that sum: sales diverted from one merging product to another now matter to the same owner. Partial ownership is handled with ownership weights.
Observed shares and margins restrict the demand derivatives in this condition. The chosen demand system fills in the remaining structure and allows marginal costs to be recovered. A common parameter can leave discrepancies between supplied and implied margins, as the workflow example shows. These moment discrepancies are distinct from equilibrium solver error. A counterfactual then changes ownership or another market condition while holding the calibrated demand primitives fixed. Two demand systems can agree at the observed prices but predict different outcomes if their derivatives change differently away from that point.
One useful observation about substitution is quantity diversion from
\(i\) to \(j\): \(D^q_{ij}=-(\partial q_j/\partial p_i)/(\partial
q_i/\partial p_i)\). It can help identify derivatives that shares
and margins leave undetermined. Under the package’s diversion-matrix
convention, the diagonal is \(-1\).
With an outside option, the other modeled products need not absorb all
sales lost by \(i\). See
?diversion for each model’s quantity or revenue
convention.
Demand assumptions determine both who substitutes for whom and how those substitution patterns change as prices move. Linear and Log-linear models put restrictions directly on slopes or elasticities. Logit and CES give substitution a more structured form, while nests and random coefficients allow more variation. The table gives the main calibration tradeoffs; the model help pages give exact arguments and solver controls.
| Family | Substitution and recovered parameters | What the calibration needs or restricts |
|---|---|---|
| Linear | \(q=a+Bp\) has constant slopes in quantity levels. Calibration recovers intercepts and a slope matrix. | Prices, quantities, margins, and diversion or symmetry restrictions are needed to pin down slopes. There is no curvature in quantity levels. |
Log-linear (LogLin) |
\(\log q_i=a_i+\sum_j B_{ij}\log p_j\) has constant elasticities, collected in \(B\). | Prices, quantities, margins, and diversion or other restrictions identify the elasticities. |
LA-AIDS (AIDS) |
Budget shares respond to log prices and an expenditure/price index. Adding up, homogeneity, and symmetry restrict cross-price effects. | Revenue shares, margins, and diversion or curvature restrictions identify the local system. Prices can be optional at calibration but matter for level outputs. |
| PCAIDS and nested PCAIDS | Share proportionality restricts substitution; nests allow stronger within-group substitution. | knownElast and mktElast
provide identifying information. Nested calibration also needs nests and
enough margins. |
| Logit | \(s_i=\exp(\delta_i+\alpha p_i)/(1+\sum_j\exp(\delta_j+\alpha p_j))\). A common \(\alpha<0\) governs price response; substitution is proportional to shares. | Prices, quantity shares, ownership, and enough margins identify \(\alpha\) and mean values. Without an outside option, one mean value is normalized. |
| Nested Logit | Correlation within nests directs more diversion toward products in the same nest. | Nest assignments and additional margins identify nesting parameters. A singleton nest cannot identify its own parameter. |
| CES | Quality weights and price powers determine revenue
shares. gamma controls substitution; alpha and
the inside share govern expenditure allocation. |
Prices, revenue shares, margins, and an outside share
or normalization identify primitives for the chosen construction.
Ordinary output-market translations require
gamma > 1. |
| Nested CES | Substitution can differ within and across nests. | Nest assignments and extra restrictions or margins
identify the curvatures. In ordinary output markets, within-nest
sigma exceeds between-nest gamma > 1. |
| BLP/random coefficients | Integrating heterogeneous price tastes over consumer draws allows flexible diversion. | Prices, shares, known outside share s0,
and conduct-specific margins identify price-taste parameters by
contraction and margin matching. Calibration does not estimate
s0 or bargaining power. |
Capacity Logit (LogitCap) |
Logit demand is subject to product quantity ceilings. | Market size and capacities locate binding constraints. Margins at exactly binding products matter because unconstrained FOCs do not identify their costs. |
In Linear, Log-linear, and AIDS models, restrictions on derivatives and diversion help fill gaps in observed market data (Haefen 2002; LaFrance 2004; Epstein and Rubinfeld 2004). Logit, CES, and capacity models instead put more structure on substitution or supply constraints (Werden and Froeb 1994; Sheu 2011; Froeb et al. 2003).
Some models have an "alm" variant for an unobserved
outside share. It uses additional margins to identify that share along
with demand curvature. The variant is a separate implemented model, even
when its demand name is the same, and it may lack a
specify() route.
Observed synthetic experiments start from known costs, shares,
ownership and one reference margin, then solve prices. That anchor does
not identify every model: nests, relative slopes, capacities or
bargaining state can require additional inputs. See
?synthetic_market for the supported inversions and
?supportedModels for the observed_synthetic
capability. This is a different experiment from calibration using
observed prices.
Demand tells us how buyers respond to prices; conduct tells us which choices firms make and which profits they internalize. The distinction matters for cost recovery as well as merger predictions. A Bertrand margin, for example, cannot simply be fed into a Cournot first-order condition. Each conduct model below uses its own equilibrium argument.
| Conduct | How firms compete and what a merger changes | What calibration requires |
|---|---|---|
| Bertrand | Firms choose prices and internalize the margins on their own products. Common ownership adds recaptured sales to that calculation. | Margins help identify demand derivatives; fitted prices and margins yield costs. |
| Cournot | Firms or plants choose quantities, taking account of their effect on inverse demand. Common ownership changes those quantity incentives. | General Linear/LogLin implementations use plant quantities and production or cost information. Logit and CES Cournot have separate complete implementations. |
| Stackelberg | Quantity leaders move before followers, whose responses they anticipate. A merger or a capacity/leadership change may alter that sequence. | Implemented Linear/LogLin games require plant and leadership information. |
Monopolistic competition (moncom) |
Each product is priced against its perceived own demand. CES holds the aggregate index fixed for that derivative. | Prices, shares, and at least one margin identify supported Logit/CES/BLP variants. Ownership does not enter the cross-product FOC. |
| Differentiated second-score auction | A seller offers the portfolio product with the highest surplus and bids cost; payment reflects the next-best score. A merger changes portfolio alternatives. | Shares and margins recover offer-value and cost primitives. A valuation may need normalization; price data help separate value from cost under cost changes (Brannman and Froeb 2000; Miller 2014). |
| Nash bargaining | Buyer and seller negotiate against disagreement payoffs. A merger changes recapture within the seller’s portfolio and the bargaining alternatives. | Bargaining power is supplied. Supported Logit, CES, and BLP models use their own FOCs rather than Bertrand margins alone. |
| Second-score bargaining | Bargaining occurs with a second-score auction disagreement option. | The Logit and CES implementations each have their own demand, bargaining, and auction state. |
| Capacity-constrained second-price auction | A buyer sets a reserve, while seller costs are drawn from a distribution; capacity changes the minimum-cost distribution. | Capacities, prices/margins, and sometimes reserve or self-supply moments identify the cost distribution. More flexible distributions need more moments (Waehrer and Perry 2003). |
The demand and conduct descriptions above are a guide to choosing
assumptions, not a menu of components that can be paired at will. The
table below shows four standard Logit/CES routes selected from
supportedModels() when this vignette is built. It is an
illustrative selection, not the complete registry. Each row is a
complete implemented model. TRUE means the
listed lifecycle operation is available; simulate() still
checks whether a particular scenario field is supported by that
model.
| demand | conduct | class | calibrate | specify | simulate |
|---|---|---|---|---|---|
| logit | bertrand | Logit | TRUE | TRUE | TRUE |
| logit | cournot | LogitCournot | TRUE | TRUE | TRUE |
| ces | bertrand | CES | TRUE | TRUE | TRUE |
| ces | cournot | CESCournot | TRUE | TRUE | TRUE |
Run antitrust::supportedModels() for the complete
implementation list. The model registry identifies what can be fitted or
simulated. A separate, narrower transition registry determines which
fitted models can be translated with respecify(); an
implemented row does not imply a transition from every other row. The workflow vignette (source)
shows such comparisons, and the architecture
vignette (source)
explains the registry’s role in dispatch and extensions.