Workflow and counterfactual analysis

Suppose firms A and B propose a merger. How would prices change, and would a cost saving offset the increase? Fit the observed market once, then solve these scenarios under the same economic assumptions.

This guide assumes basic R and an installed antitrust package; see the installation instructions. Start here, read the economic models guide (source) to choose assumptions, and use the architecture guide (source) when writing an extension.

Fit a market and simulate the merger

Each of three firms initially sells one product. Quantity shares total 80%, leaving 20% for the outside option. Prices are in a common monetary unit; gross margins are proportions of price. insideSize = 100 sets total inside quantity. Under Logit demand and Bertrand pricing, calibrate() recovers demand parameters and marginal costs. A counterfactual() describes the proposed change; simulate() solves it.

library(antitrust)
prices <- c(2, 2.2, 2.5)
shares <- c(.35, .25, .20)
margins <- c(.40, .35, .30)
owners <- c("A", "B", "C")
fit <- calibrate(
  "logit", "bertrand", prices = prices, shares = shares,
  margins = margins, ownerPre = owners, insideSize = 100
)
merger <- counterfactual(ownership = c("A", "A", "C"))
merged <- simulate(fit, merger)
data.frame(before = merged@pricePre, after = merged@pricePost,
           proportional_change = merged@pricePost / merged@pricePre - 1)
#>       before    after proportional_change
#> Prod1    2.0 2.228681          0.11434039
#> Prod2    2.2 2.542324          0.15560183
#> Prod3    2.5 2.542632          0.01705298

All three prices rise. The proportional_change column uses the original price as denominator: .10 means a 10% increase, not an increase of .10 monetary units. A and B now internalize sales diverted between their products, which changes their pricing incentives; C also responds.

Check what calibration matched

The returned AntitrustFit stores observations, parameters and the fitted S4 model in fit@model. The merger result is a Logit object with the usual result methods. summary(fit) describes stored fitted state; it does not predict a new scenario.

data.frame(supplied_margin = margins,
           implied_margin = calcMargins(fit@model),
           difference = calcMargins(fit@model) - margins)
#>       supplied_margin implied_margin  difference
#> Prod1            0.40      0.4261622  0.02616222
#> Prod2            0.35      0.3357642 -0.01423583
#> Prod3            0.30      0.2770054 -0.02299456

The common Logit price coefficient cannot match all three supplied margins exactly. The implied margins are approximately .426, .336 and .277. These moment discrepancies describe the fit to inputs; they are different from a solver’s equilibrium residual and from sampling uncertainty. Calibration does not supply sampling standard errors. Check identifying moments before treating scenario results as evidence; ?calcDiagnostics explains model-specific checks.

Add a cost saving or change the products

Each call below starts from the original fitted market. A proportional cost shock of -.10 lowers product 2’s marginal cost by 10%.

efficient_merger <- simulate(fit, counterfactual(
  ownership = c("A", "A", "C"), costs = c(0, -.10, 0)
))
data.frame(merger = merged@pricePost,
           merger_with_efficiency = efficient_merger@pricePost)
#>         merger merger_with_efficiency
#> Prod1 2.228681               2.256325
#> Prod2 2.542324               2.423837
#> Prod3 2.542632               2.537745

The saving lowers product 2’s price relative to the merger alone, but product 1’s price rises slightly. A multiproduct firm’s responses need not move every price in the same direction. Neither calculation changes fit.

Exit and entry ask different questions about the active product set:

exit_result <- simulate(fit, counterfactual(exit = "Prod3"))
entrant_result <- simulate(fit, counterfactual(
  entry = entrant("New", meanval = 0, cost = 1, priceStart = 2)
))
exit_result@pricePost
#>   Prod1   Prod2   Prod3 
#> 2.09088 2.26371      NA
entrant_result@pricePost
#>    Prod1    Prod2    Prod3      New 
#> 1.995887 2.197210 2.497843 1.560605

Exit leaves an NA in product 3’s original price slot. The entrant supplies its own label, mean value, marginal cost and starting price, and shares the fitted Logit price coefficient. Entry requires a model with an entry implementation; ?entrant describes its inputs. A named quality = c(Prod1 = .05) shock would raise product 1’s choice weight by 5%, which is not a promise of 5% more sales.

Separate simultaneous and staged changes

Combining changes answers one equilibrium question. This gives the same result as the efficient merger above; conflicting values for the same field are rejected.

combined <- combine_counterfactuals(
  merger, counterfactual(costs = c(0, -.10, 0))
)
combined_result <- simulate(fit, combined)
combined_result@pricePost
#>    Prod1    Prod2    Prod3 
#> 2.256325 2.423837 2.537745

For a merger followed by a cost saving, add_step() requests a second solve starting from the first result. The CounterfactualPath retains both equilibria:

path <- simulate(fit, add_step(merger, costs = c(0, -.10, 0)))
data.frame(after_merger = result_at(path, 1)@pricePost,
           after_efficiency = final_result(path)@pricePost)
#>       after_merger after_efficiency
#> Prod1     2.228681         2.256325
#> Prod2     2.542324         2.423837
#> Prod3     2.542632         2.537745

Costs, ownership and product state carry forward according to the fitted model’s rules. simulate(path, counterfactual(...)) can continue a path; ?add_step and ?simulate describe that state persistence.

Choose the correct lifecycle operation

New observations call for recalibration. Proposed changes call for simulation. Known parameters and translations have their own routes:

Operation Use it when… Result
calibrate() You have observations and a model. Recovers demand parameters and costs using its calibration equations.
specify() You already know the structural parameters. Builds a fit without recalibrating demand.
simulate() You want the equilibrium under a proposed change. Applies a scenario and solves under the fitted model.
update() An observed input or calibration option changed. Calibrates again under the same model specification.
respecify() You want a registered model translation. Retains and reconstructs only the information that transition permits.
revised <- update(fit, margins = c(.38, .35, .30))
known <- specify("logit", "bertrand", prices = prices,
                 ownerPre = owners, parameters = fit@parameters,
                 insideSize = 100)
c(revised_alpha = revised@parameters$alpha,
  supplied_alpha = known@parameters$alpha)
#>  revised_alpha.alpha supplied_alpha.alpha 
#>            -1.847460            -1.805019

update() cannot change the fit’s demand, conduct or variant. To identify a different model from observations, call calibrate() for that model. A supplied fit has no observed-data calibration call to repeat.

A Logit-to-CES translation requires a target curvature gamma:

ces_fit <- respecify(fit, demand = "ces", gamma = 2)
#> Warning in get(".sim_legacy", envir = asNamespace("antitrust"), inherits =
#> FALSE)(...): 'shares' argument is only used for BLP demand (for BLP contraction
#> mapping). For 'CES', shares will be ignored.
#> Warning in validityMethod(as(object, superClass)): 'ownerPost' and 'ownerPre'
#> are the same
#> Warning in validityMethod(as(object, superClass)): 'ownerPost' and 'ownerPre'
#> are the same
ces_fit@spec$id
#> [1] "ces::bertrand"

This transition preserves baseline prices, shares or quantities, ownership and active products, while deriving CES mean values and recovering costs under the target model. It discards the Logit coefficient and costs. The comparison is local: demand curvature away from the baseline, and merger predictions, can differ. ces_fit@diagnostics$translation records the translation; the fit has no CES calibration call, so update(ces_fit) cannot silently recalibrate it from Logit margins. Inspect supported transitions with ?model_transition.

Construction or translation can warn that supplied shares are ignored by a non-BLP specify() route, or that scenario-neutral pre/post ownership is equal. Those messages describe construction, not failed estimation; inspect unexpected warnings and model-specific diagnostics before proceeding.

Optional: explore an observed synthetic market

When structural parameters are hard to choose, an observed synthetic experiment starts with shares, ownership, known costs and one reference margin. Here two inside firms own two products each; an additional active reference firm owns product 5. The five active-product shares sum to one. This is a separate intuition/calibration exercise, with a different price scale from our merger.

intuition <- synthetic_market(
  demand = "logit", supply = "bertrand", n_firms = 2, n_products = 2,
  shares = c(.2, .1, .15, .25, .3), costs = c(40, 50, 60, 70, 80),
  reference_margin = .2, seed = 7
)
intuition@diagnostics$synthetic$solved_prices[5]
#> [1] 100
intuition@diagnostics$synthetic[c("share_residual", "foc_residual")]
#> $share_residual
#> [1] 2.775558e-17
#> 
#> $foc_residual
#> [1] 7.105427e-15

The reference price is 80/(1-.2) = 100. Other prices satisfy the ownership-adjusted Bertrand first-order conditions. Small share and FOC residuals check this inversion, not empirical recovery or statistical coverage. ?synthetic_market explains identification inputs and the separate passive outside option in ALM models.

Not every model supports every scenario. Use ?counterfactual, ?supportedModels and the models guide (source) for field-specific limits. Tariffs and quotas belong to the sibling trade package.