Bargaining in Health Care
Basic Idea
Solve the problem working backward, finding patient choice probabilities based on prices, and then solving for prices. Take as given the employer-insurer relationship, avoiding concerns regarding employers changing insurers or patients choosing different health care plans.
Setup
- Denote consumers by \(i\), insurance companies by \(m\), hospitals by \(j\), and diseases by \(d\)
- Denote by \(f_{mid}\) the probability that patient \(i\) enrolled in MCO \(m\) has disease \(d\) (\(d=0\) implies no disease)
- Denote by \(w_{d}\) the weights reflecting the relative intensity of resource use for illness \(d\), with \(w_{0}\) normalized to 0
- Insurer and hospital negotiate a base price, to which all other payments are adjusted using the relative weight
- Price paid for a given treatment is \(p_{mj} \times w_{d}\).
Patients
- Coinsurance rate, denoted \(c_{mit}\)
- Patient with disease \(d\) chooses the hospital \(j\) in their network (\(N_{m}\)) that maximizes their utility
- Patient utility is given by: \[u_{mijd} = \beta x_{mijd} - \alpha c_{mid} w_{d} p_{mj} + e_{mij}\]
- \(x_{mijd}\) denotes a vector of hospital and patient characteristics, and \(e_{mij}\) is an iid error term assumed to follow a Type I extreme value distribution
Patients
- Probability of choosing hospital \(j\) is given by \[s_{mijd}(N_{m},\vec{p}_{m}) = \frac{\exp (\beta x_{mijd} - \alpha c_{mid} w_{d} p_{mj})}{\sum_{k}\exp (\beta x_{mikd} - \alpha c_{mid} w_{d} p_{mk})},\] where \(\sum_{k}\) denotes the sum over all \(k\) hospitals in the patient’s network as well as their outside option.
- Expected number of patients admitted to hospital \(j\), weighted by the relative intensity of resource use, is \[q_{mj}(N_{m},\vec{p}_{m}) = \sum_{i=1}^{I_{m}} \sum_{d=1}^{D} f_{mid} w_{d} s_{mijd}(N_{m},\vec{p}_{m}).\]
Consumer surplus, revisited
- With one good and a demand curve, consumer surplus is what consumers would have paid minus what they did pay: the area under demand and above price
- Here each patient picks one hospital, so the patient’s surplus is the value of the hospital they chose net of what they pay for it, which in the model is the utility of the chosen hospital, \(u_{mijd}\)
- Utility is in “utils,” so we divide by \(\alpha\) to report surplus in dollars. \(\alpha\) is the utility lost per dollar paid out of pocket, so $1 \(\equiv \alpha\) utils, the same step as the certainty equivalent in the insurance module
Surplus for one patient
- We don’t see the patient’s \(e_{mij}\) draws, so we can’t say which hospital they pick or what surplus they get, only the probabilities \(s_{mijd}\)
- What we can compute is the expected value of their best option, \(\text{E}\left[\max_{k} u_{mikd}\right]\), averaging over the error draws
- With Type I extreme value errors this expectation has a closed form, the “log-sum”: \[\text{E}\left[\max_{k} u_{mikd}\right] = \ln \left( \sum_{k}\exp (\beta x_{mikd} - \alpha c_{mid} w_{d} p_{mk}) \right) + \text{constant}\]
- The constant is the same for every network and every set of prices, so it drops out of any comparison
Why the log-sum behaves like surplus
- Adding a hospital never lowers the patient’s expected surplus, and a hospital adds more when its \(\beta x_{mikd} - \alpha c_{mid} w_{d} p_{mk}\) is high relative to the others
- A $1 increase in \(p_{mj}\) lowers expected surplus by \(c_{mid} w_{d} s_{mijd}\) dollars, the patient’s expected out-of-pocket spending at hospital \(j\), the same quantity-times-price logic as the area under a demand curve
- Removing a hospital from the network lowers it by an amount that depends only on that hospital’s choice share (next slide)
Removing a hospital from the network
- Write \(v_{mikd} = \beta x_{mikd} - \alpha c_{mid} w_{d} p_{mk}\). With hospital \(j\) in the network, the patient’s expected surplus is \(\ln \sum_{k \in N_{m}} e^{v_{mikd}}\), and without it, \(\ln \sum_{k \in N_{m,-j}} e^{v_{mikd}}\)
- The difference is \[\ln \sum_{k \in N_{m}} e^{v_{mikd}} - \ln \sum_{k \in N_{m,-j}} e^{v_{mikd}} = -\ln \left( \frac{\sum_{k \in N_{m,-j}} e^{v_{mikd}}}{\sum_{k \in N_{m}} e^{v_{mikd}}} \right) = -\ln \left( 1 - s_{mijd} \right) = \ln \left( \frac{1}{1 - s_{mijd}} \right)\]
- The ratio inside the log is the combined choice share of every hospital other than \(j\), so the loss from dropping \(j\) is pinned down by \(j\)’s own share. A hospital the patient picks 10% of the time is worth \(\ln (1/0.9) \approx 0.11\) utils, one picked half the time \(\ln 2 \approx 0.69\). Dividing by \(\alpha\) gives dollars, which is the willingness-to-pay measure of Capps, Dranove, and Satterthwaite (2003)
Patients
- Summing the log-sum over enrollees and diseases, weighted by the chance of each disease, gives the closed form for consumer surplus, \[W_{m}(N_{m},\vec{p}_{m})=\frac{1}{\alpha} \sum_{i=1}^{I_{m}} \sum_{d=1}^{D} f_{mid} \ln \left( \sum_{k}\exp (\beta x_{mikd} - \alpha c_{mid} w_{d} p_{mk}) \right)\]
- Difference between \(W_{m}\) with and without hospital \(j\) in the patient’s network, \(W_{m}(N_{m},\vec{p}_{m}) - W_{m}(N_{m,-j},\vec{p}_{m})\), then serves as an estimate of the patient’s willingness to pay for hospital \(j\)
Insurers
- Expected cost to the MCO for a given hospital and associated prices, \[TC_{m}(N_{m},\vec{p}_{m})=\sum_{i=1}^{I_{m}} \sum_{d=1}^{D} (1 - c_{mid}) \times \sum_{j\in N_{m}} p_{mj} f_{mid} s_{mijd}(N_{m},\vec{p}_{m})\]
- Value for the MCO is \[V_{m}(N_{m},\vec{p}_{m}) = \tau W_{m}(N_{m},\vec{p}_{m}) - TC_{m}(N_{m},\vec{p}_{m}),\] where \(\tau\) is the relative weight placed on employee welfare
- Net value that MCO \(m\) receives from including system \(s\) in its network is then \(V_{m}(N_{m},\vec{p}_{m})-V_{m}(N_{m,-j},\vec{p}_{m})\).
Hospitals
- Hospital \(j\)’s marginal cost for services provided to patients in MCO \(m\) is given by \[mc_{mj}=\gamma \nu_{mj} + \varepsilon_{mj},\] where \(\nu_{mj}\) denotes a set of cost shifters, \(\gamma\) are parameters to estimate, and \(\varepsilon\) is an error term
- Profit for hospital \(j\) for a given set of MCO contracts (denoted \(M_{s}\)), is \[\pi_{s}\left(M_{s},\{\vec{p}_{m}\}_{m\in M_{s}},\{N_{m} \}_{m\in M_{s}} \right)=\sum_{m\in M_{s}} \sum_{j \in J_{s}} q_{mj}(N_{m},\vec{p}_{m}) \left[p_{mj} - mc_{mj} \right]\]
- Net value that system \(s\) receives from including MCO \(m\) in its network is \[\sum_{j \in J_{s}} q_{mj}(N_{m},\vec{p}_{m}) \left[p_{mj} - mc_{mj} \right].\]
Nash Bargaining Solution
Nash bargaining solution is the choice of prices maximizing exponentiated product of the net value from agreement:
\[\begin{align*}
NB^{m,s} \left(p_{mj, j\in J_{s}} | \vec{p}_{m,-s}\right) &= \left(\sum_{j\in J_{s}}q_{mj}(N_{m},\vec{p}_{m}) \left[p_{mj} - mc_{mj} \right]\right)^{b_{s(m)}} \\
& \times \left(V_{m}(N_{m},\vec{p}_{m})-V_{m}(N_{m,-j},\vec{p}_{m})\right)^{b_{m(s)}},
\end{align*}\] where \(b_{s(m)}\) is the bargaining weight of system \(s\) when facing MCO \(m\), \(b_{m(s)}\) is the bargaining weight of MCO \(m\) when facing system \(s\), and \(\vec{p}_{m,-s}\) is the vector of prices for MCO \(m\) and hospitals in systems other than \(s\). We can normalize bargaining weight such that \(b_{s(m)} + b_{m(s)} = 1\).
Nash Bargaining Solution
Note that taking the natural log of the objective does not change the maximum, since the natural log is a “monotonic transformation.” In other words, the log will change the value of some function \(f(x)\), but it will not change the order, so that if \(f(x_{1})>f(x_{2})\), it follows that \(\ln (f(x_{1})) > \ln (f(x_{2}))\).
The resulting first order condition yields: \[\begin{align*}
\frac{d \ln (NB^{m,s})}{d p_{mj}} &= b_{s(m)} \frac{q_{mj} + \sum_{k\in J_{s}} \frac{d q_{mk}}{d p_{mj}} \left[p_{mk}-mc_{mk}\right]}{\sum_{k\in J_{s}} q_{mk}\left[p_{mk}-mc_{mk}\right]} + b_{m(s)} \frac{\frac{d V_{m}}{d p_{mj}}}{\triangle V_{m}} \\
&= 0, \end{align*}\] where \(\triangle V_{m} = V_{m}(N_{m},\vec{p}_{m})-V_{m}(N_{m,-j},\vec{p}_{m})\) is the insurer’s net value from the agreement.
Nash Bargaining Solution
Rearrange this equation to write: \[q_{mj} + \sum_{k\in J_{s}} \frac{d q_{mk}}{d p_{mj}} \left[p_{mk}-mc_{mk}\right] = -\frac{b_{m(s)}}{b_{s(m)}} \frac{\frac{d V_{m}}{d p_{mj}} \sum_{k\in J_{s}} q_{mk}\left[p_{mk}-mc_{mk}\right]}{\triangle V_{m}}\]
By assumption, the first order conditions are separable across insurers. Combining all first order conditions therefore yields the following system of equations: \[q + \omega (p-mc) = -\lambda (p-mc),\] where \(\omega\) and \(\lambda\) are each \(J_{s} \times J_{s}\) matrices, while \(q\) and \(p-mc\) are \(J_{s} \times 1\) vectors
We can then solve for prices: \[p = mc - (\omega + \lambda)^{-1} q\]
Insurer Steering
“Reasonable assumptions” such that \[\frac{d V_{m}}{d p_{mj}}=-q_{mj}-\alpha \sum_{i}\sum_{d}\gamma_{id}c_{id}(1-c_{id}) \left(\sum_{k\in N_{m}} p_{mk}s_{ikd} - p_{mj}\right),\] where \(\gamma_{id}\) includes several terms including disease weights and probability of disease
- \(c_{id}\times (1-c_{id})\) captures role of insurer coinsurance rates in steering patients
- If \(p_{mj}\) is high relative to the weighted average price of other hospitals, then \(\frac{d V_{m}}{d p_{mj}}\) is less negative due to the term in parentheses, and thus it is optimal (in the sense of maximizing joint surplus) to allow \(p_{mj}\) to increase so as to channel customers to the low price hospital.
- In the extreme cases, at \(c_{id}=1\) the insurer wouldn’t care about negotiating price with physicians since all of the extra cost is borne by patients. Conversely, at \(c_{id}=0\) the term in parenthesis is again irrelevant as the patients do not switch to lower price hospitals.
Prices without bargaining
Consider a single-hospital system. If the hospital set \(p_{mj}\) on its own, with no say for the insurer (\(b_{m(j)} = 0\)), it would pick the price that maximizes \(q_{mj}(p_{mj}) \left[p_{mj} - mc_{mj}\right]\), the same problem as the monopolist earlier in this module: \[q_{mj} + \frac{d q_{mj}}{d p_{mj}} \left[p_{mj} - mc_{mj}\right] = 0 \quad \Longrightarrow \quad p_{mj} - mc_{mj} = -q_{mj} \left(\frac{d q_{mj}}{d p_{mj}} \right)^{-1}\]
- The markup over marginal cost depends only on how fast the hospital’s own demand falls as its price rises. Patients leaving for other hospitals is the only thing holding the price down.
Prices with bargaining
Keep the insurer at the table (\(b_{m(j)} > 0\)) and the single-hospital case of the first-order condition becomes \[p_{mj} - mc_{mj} = -q_{mj} \left(\frac{d q_{mj}}{d p_{mj}} + \underbrace{q_{mj} \times \frac{b_{m(j)}}{b_{j(m)}} \times \frac{\frac{d V_{m}}{d p_{mj}}}{\triangle V_{m}}}_{\text{new term}} \right)^{-1}\]
- The new term is the insurer’s side of the negotiation, namely how much a higher price hurts the insurer, \(\frac{d V_{m}}{d p_{mj}}\), relative to what hospital \(j\) is worth to it, \(\triangle V_{m}\), scaled by the insurer’s bargaining weight. At \(b_{m(j)} = 0\) it vanishes.
- \(\triangle V_{m} > 0\) by construction, and with some work \(\frac{d V_{m}}{d p_{mj}}<0\) under most conditions, so bargaining raises the “effective” price sensitivity and lowers hospital margins relative to the previous slide.
- This does not always hold. When one hospital is priced far above the others, the insurer may let a given hospital raise its margin so as to steer patients away from the very high priced hospital.