PUB Diversion Analysis Hospital Mergers 0614
PUB Diversion Analysis Hospital Mergers 0614
By Subramaniam Introduction
Ramanarayanan, PhD1 Driven by a variety of factors, including changes in the reimbursement environment and
the need to have scale to successfully bear risk and support delivery models that emphasize
population-based health management, merger activity among hospitals has been robust over
the last couple of years. In fact, the annual number of transactions is nearly double what it was
five years ago. With continued debate over the benefits of hospital consolidation and whether
these benefits outweigh a possible harm to competition, hospital mergers have come under
renewed scrutiny from the Federal Trade Commission (FTC) and, on occasion, the Department
of Justice’s (DOJ’s) Antitrust Division.
Coincident with this increased scrutiny has been a change in the tools used by these agencies
in the analysis of hospital mergers. Many credit these new tools—such as hospital merger
simulation, the Willingness-to-Pay (WTP) framework, and diversion analysis—with explaining
the recent string of successes the FTC has had in litigating hospital merger cases.2 With the
use of these news tools, the analysis of hospital competition has moved away from patient
flow-based methods towards more structural models that aim to better capture the underlying
dynamics of the industry and come up with a prediction for the impact of the transaction
on prices.
Specifically, these tools are based on a “two-stage” model of competition, in which health
plans and hospitals bargain over prices and network composition in the first stage, and once
hospital networks have been formed, consumers choose from a set of hospitals in the second
stage based on a variety of factors under the assumption that they face the same out-of-pocket
costs across all hospitals within this set.3 This model of competition is used to construct a WTP
measure for each hospital (or system)—which is essentially the incremental value added by
that hospital (or system) to an insurer network. The model then estimates the change in WTP
resulting from the merged entity being part of the network as compared to a situation where
each hospital is in the network separately, and attempts to predict the change in price post-
merger based on this change in WTP.4
While potentially a useful tool in assessing the impact of a proposed transaction on prices,
implementation of the full merger simulation model can be a substantial undertaking owing to
the data requirements of the model. Specifically, one needs to construct price indices for each
hospital based on insurer claims data, which usually requires third-party discovery and can be a
time-intensive undertaking.5 Thus, a full merger simulation is better suited for later stages of the
merger review process where a challenge is likely.6
Given these limitations, the closely-related tool of diversion analysis is being increasingly used
in hospital merger reviews, particularly in the screening stage. The diversion analysis is built off
the same underlying theoretical principles as the WTP model but has the added benefit of being
estimable using publicly available data. This benefit comes with a tradeoff of not being able
to predict the magnitude of potential price increases resulting from a merger, but rather, the
analysis allows for the analyst to make an estimate of the upward pricing pressure incentive that
the combined entity will have to engage in anti-competitive pricing, as explained in Section IV.
Conceptual Framework
The notion of diversion is commonly used in antitrust economics to analyze mergers involving
differentiated products. The diversion ratio between two products, A and B, measures the
extent to which these products are close substitutes to each other. In particular, in the context
of hospital mergers, the diversion ratio aims to estimate the share of patients of each merging
hospital (or hospital system) that would switch to various competing hospitals in the market,
including the potential merger candidate, if the first hospital (or system) were no longer
available to patients as a possible option. That is, in evaluating a merger between hospitals A
and B, B would be considered a close substitute of A if a high share of patients would switch to
B if they were unable to use hospital A. Of course, the merger of firms that are close substitutes
presents a greater danger that post-merger price increases might be profitable.
A key feature to note here is that diversion is modeled in this case as arising from a hypothetical
scenario in which the merging hospital is excluded from the provider network. This is in contrast
to typical diversion analyses (in other differentiated products industries) which analyze customer
switching behavior in the wake of a hypothetical price increase. This is driven by the modeling
assumption that price differences do not matter to patients while making hospital choices
within their network, given that they would pay the same amount regardless of their choice, or
would likely exhaust their out-of-pocket co-payments under their health insurance plan.
A high value for the diversion ratio between two merger candidates indicates that these
hospitals are perceived by consumers as being close substitutes. In a scenario where the
hospitals bargain with insurers on an all-or-nothing basis post-merger (as is typically assumed
in the merger simulation model), a merger between two such hospitals has the potential to
enhance the bargaining power of the now-merged hospitals vis-à-vis the insurer, given that
the merger removes whatever competitive constraint these hospitals imposed on each other.
Put differently, the close substitutability of the hospitals leads to the combined hospital entity
having a WTP that is larger than the sum of the WTPs of the individual hospitals.
www.nera.com 2
This analysis can be extended to a scenario where the merged hospitals bargain separately with
insurers post-merger. In such a case, if the hospitals are close substitutes, a merging hospital
may have a strong incentive to raise prices post-merger given that its merger partner will likely
recapture much of the lost patient volume in the event that the price increase leads to the
hospital being dropped from the network or “tiered”. Conditional on the level of gross margins
and the potential efficiencies generated by the merger (a point which we expand on in Section
IV), a price increase at one of the merging hospitals is then likelier to be profitable given that
the partner firm will recapture many of the patients of that hospital who seek care elsewhere.
As explained below, the diversion ratio becomes a crucial input to estimating the degree to
which there is upward pricing pressure following the merger. To the extent significant upward
pricing pressure is predicted, the merger is more likely to move to a second request and possibly
a challenge.
The diversion analysis thus acts as a useful merger screen given that the diversion ratios are
closely related to the likely changes in WTP, and thereby, to the magnitude of predicted price
changes stemming from the merger. Further, as discussed in the next subsection, the diversion
analysis can be implemented using publicly available data, which makes it a useful tool for
assessing the potential lessening of competition in a timely way during the initial 30-day period
of a merger review.7
The data used to estimate diversion ratios are typically inpatient discharge data organized at
the patient level. Such data are usually made available by various state agencies, such as the
Office of Statewide Health Planning and Development in California, the Agency for Healthcare
Research and Administration in Florida, and the Texas Health Care Information Collection in
Texas.8 These data contain information on a number of patient characteristics such as age,
gender, race, payer and 5-digit ZIP code as well as clinical covariates such as the principal
Diagnosis-Related Group (DRG) code, relevant diagnostic and procedure codes (if any), and
discharge status code (that identifies where the patient is at the end of the visit). The data
also identify the hospital the patient is being treated at, and can be merged to hospital-level
data (usually also available from the same state agencies) containing information on hospital
attributes like teaching status, ownership status (not-for-profit vs. for-profit), location, system
membership, and specialization by types of services offered (e.g., children’s hospital).
www.nera.com 3
An important consideration here relates to the identity of the hospitals in the “choice set” for
the patient. Because the model attempts to estimate the factors determining patient choice of
hospital, we need information on not just the hospital the patient ultimately chose, but also
on which other hospitals the patient could have chosen instead. In theory, while the model
estimation obviates the need for a strict definition of a relevant geographic market, identifying
the set of hospitals that could be potential alternative choices involves some analysis of the
service areas of the combining hospitals. This generally leads to considering many hospitals
over a broad geography. A key strength of the model, however, is that it is fairly robust to
the inclusion of hospitals (or geographic areas) that may not have much actual competitive
relevance. Computational limits do impose restrictions on the number of hospitals that can
feasibly be included in the choice set, especially in larger markets like Texas or California. In
such situations, a researcher can focus on the most competitively relevant set of hospitals and
systems and include the others as part of an outside option.9 The model is then estimated on
a reshaped version of the dataset that is structured so that each observation corresponds to a
patient-hospital dyad.10
Patient choices are then typically modeled and estimated using a standard discrete-choice
conditional logit specification.11 This model aims to capture the relationship between the actual
hospital choices made by patients (as recorded in the data) and the characteristics of patients
and hospitals in the sample that drive these choices. The price paid by the patient is assumed
to be the same across all hospitals (assuming they are all part of the patients’ managed-care
network) and is hence not included as a determinant of hospital choice.12 In particular,
Hospital attributes that typically drive patient choice include teaching status, size, or quality
rankings such as report cards.13 Because the model analyzes the choice for each patient
from among a set of hospital choices, patient characteristics (such as age, gender, DRG
weight, whether or not this was an emergency admission) are invariant across the hospital
alternatives and are hence not included in the model as separate variables. Interactions of
these characteristics with hospital attributes still matter in explaining choice and are included as
predictors.14 A key driver of patient choice that is included in the model is the driving distance
between the patient’s residence and each hospital in the choice set, typically calculated as
the distance between the residence zip-code centroid of the patient and the address of the
hospital.15 Not only is distance predicted to matter, the extent to which it drives choice of
hospital might well vary by patient characteristics such as age, or severity of illness (e.g.,
patients that are older might be less able or willing to travel greater distances to visit a hospital)
and such differences can be captured in the model by including the appropriate interaction
terms as predictors. Finally, there is also an element of “matching” between the patient’s
specific healthcare needs and the services offered by each hospital. For example, patients
with cardiac conditions may seek out a specialty heart hospital or a hospital that is particularly
recognized for handling chronic cardiac patients (or might have specialized equipment to
provide such care). Of course, such a patient would also not choose a hospital that did not have
a cardiology service. Such types of interactions can also be included as predictors in the model.
www.nera.com 4
The estimated parameters from the regression model provide a separate probability of each
patient choosing each choice set hospital as well as the “outside option”.16 The probabilities
are then used to calculate the predicted diversion ratios between relevant hospital pairs. The
diversion ratio between two hospitals, A and B, when hospital B is hypothetically excluded from
the network, is calculated as follows:
where the shares for each hospital are simply computed by adding the predicted probability
of choosing that hospital across all patients in the sample. A similar calculation can be used to
compute diversion ratios going in the other direction, i.e., from hospital A to hospital B, when
hospital A is hypothetically excluded from the network.
It is important to highlight a couple of key assumptions underlying this whole exercise. First, the
calculation of the diversion ratio assumes that all patients currently being treated at the hospital
being hypothetically excluded from the network (because of a price increase, for example)
leave and seek care elsewhere. This might not be entirely true in most cases as patients might
continue using a hospital even if it is out-of-network, but measuring the actual decrease in
patient volume because of exclusion can be difficult owing to data limitations. A second key
assumption, as discussed earlier, is that patients do not face any differences in prices (or out-of-
pocket expenses) while choosing across hospitals in their choice set. This assumption may well
be violated for patients with commercial insurance who may face larger out-of-pocket costs at
certain facilities that might be outside their network. An alternative is to estimate the model on
a sample of Medicare or Medicaid patients that typically face far fewer restrictions on provider
choice (and face no variation in out-of-pocket costs across these providers). However, the
preferences of such patients (including their propensity to travel) may not be representative of
patients with commercial insurance.
The UPP theory provides a measure of the combined firm’s incentives to increase price post-
merger and uses three key inputs: the diversion ratios, the pre-merger gross margins and an
estimate of or assumption about the likely efficiencies stemming from the merger. In examining
www.nera.com 5
a merger between two hospitals, A and B, the UPP theory states that there will be a positive
incentive for firm A to raise prices if
If the merging hospitals A and B are close substitutes, i.e., have high diversion ratios, that
increases the likelihood that a post-merger price increase at hospital A would be profitable given
that many of the patients that hospital A would lose as a result of the price increase, would in
turn be recaptured by hospital B.18 This incentive is also likely to be stronger if the recaptured
customers are profitable for hospital B, as reflected by B’s gross margin. However, it is possible
that the merger also helps the hospitals realize efficiencies that decrease the marginal cost for
the parties and hence exert a downward pressure on price. Hence, the net UPP accounts for
these efficiencies to arrive at a measure of the likely overall incentive for the merging firm to
increase prices post-merger.
40% 25%
50% 20%
60% 16.7%
Another consideration that often arises in practice is whether the two merging hospitals are
each other’s closest substitutes as measured by the diversion ratios between them. Consider
a situation where the merging hospitals have high diversion ratios (going both ways, say) but
there are other non-merging competitors to which diversion is higher. Put differently, how does
one deal with a situation where the merging hospitals are close substitutes but are not each
other’s closest substitutes? Does the fact that a non-merging hospital is the closest substitute to
(one or both) of the merging parties resolve any anticompetitive concerns that might arise from
the merger?
www.nera.com 6
Recent Court rulings indicate that such an argument would likely not be a sufficient defense.
Commissioner Brill’s opinion in the ProMedica case states that “…the merging parties do not
need to be each other’s closest rival for a merger to have unilateral anticompetitive effects.”20
The 2010 Horizontal Merger Guidelines also state that “a merger may produce significant
unilateral effects for a given product even though many more sales are diverted to products
sold by non-merging firms than to products previously sold by the merger partner.” The upshot
of these rulings is that, even if the diversion analysis predicts greater diversion to non-merging
rivals than merger partners, defendants will have to prepare a case for showing why the merger
will not be anticompetitive especially if the values exceed the critical thresholds discussed above.
On the flip side, just because a transaction involves parties where diversion ratios between the
merging parties exceed critical values does not necessarily mean that it will get a protracted
review. The agencies clearly examine other factors in addition to the diversion ratios and in
cases where extensive repositioning by competitors and payers is likely, the weight placed on
diversion analysis might be minimized. The agencies generally require strong corroborating
evidence and do not go forward with a challenge based only on diversion analysis or even the
full merger simulation.
Discussion
In recent years, the FTC has increasingly adopted the WTP-based analysis of competition
in hospital markets, and has made extensive use of diversion analysis (a related tool) as an
important merger screen. There are several reasons underlying this shift towards using these
tools, especially in hospital markets. The WTP and diversion analyses have a stronger grounding
in rigorous economic theory as compared to past patient-flow based approaches. These
approaches also do not require identifying a relevant geographic market, thereby avoiding long-
drawn arguments with the merging hospitals on the appropriate size of the geographic market.
Instead, these methods rely on identifying a reasonable set of competitors that patients see as
alternatives to the merging hospitals. In addition, the estimates are largely unaffected by the
inclusion of hospitals that are competitively less relevant.
The diversion analysis in particular is relatively easy to apply and can be readily estimated using
publicly available data, thereby making it a good candidate for use as a merger screen. It is
a fairly flexible approach and has the ability to control for a number of different factors that
are not easy to incorporate into the standard geographic market tests that have been used
traditionally.21
While it has been used extensively in the analysis of hospital mergers, the approach used can
potentially be extended to the analysis of physician group mergers as well. A key deterrent in
the analysis of physician group mergers is the lack of computational power to estimate models
that include all providers in the market. While a model of patient choice of hospitals can be
estimated with up to 50 or so hospitals without much computational burden, this is insufficient
in physician services markets given the large number of physicians operating in each market.22
One possible approach that has been recently employed by the FTC is to divide patients into
various segments (for the sake of the analysis) and use the diversion ratios within each segment
to construct an aggregate diversion ratio among physicians.23
www.nera.com 7
While the diversion analysis has a number of strengths, one should not lose sight of the
limitations of using this approach. While the model controls for a number of factors, price is
not one of them. It is thus unclear how patients would respond if price were to increase and
patients had to pay a greater co-insurance and/or deductible amount, as is more likely in current
and future benefit designs.
The approach also does not account for repositioning—that is, the model is static and does not
currently capture how competing hospitals, physicians, insurers, and employers would respond
to a significant post-merger increase in price. Instead, it assumes that all of these parties will
continue to act in the same manner that they have acted in the past. In addition, the model is
not well-suited for handling markets that use limited provider networks. The model assumes
that patients are free to choose any hospital in their choice set. However, limited or narrow
networks are now coming back into play, driven in part by initiatives championed by the
Affordable Care Act, such as the formation of Health Insurance Marketplaces. If patients
are restricted in their choice of healthcare providers, the estimates from a choice model that
assume otherwise might well be unreliable.
Finally, the diversion model is based only on patient hospital choice data. It does not really
model how hospitals, particularly regional or statewide hospital systems, negotiate price with
managed care. If prices over a wide set of geographies are negotiated all at once, then the
structural conditions in one specific geographic area may not have much effect on the prices set
for that area during negotiations.
All of these concerns suggest that the model may not always be a reliable predictor of the likely
diversion or the actual degree of upward pricing pressure resulting from hospital mergers. It
is a useful tool, but one of possibly many indicia that ought to be used while assessing likely
anti-competitive impacts in such transactions. Further research especially looking at empirical
validation of this method is required in order to better understand the set of conditions under
which the model generates reliable predictions of pricing post-merger.
www.nera.com 8
Notes
1 The author is a Senior Consultant at NERA Economic Consulting and 12 The price to the patient could be an important missing variable to the
would like to thank Thomas McCarthy and Scott Thomas for helpful extent that hospitals are increasingly “tiered” with materially different
comments and discussion. Email: [email protected] copayments owed by the patients depending on which hospital is
2 A prominent recent example was the proposed acquisition of Rockford chosen. Similarly, narrow networks create a similar measurement
Health System by OSF Healthcare System in Illinois, which was problem if patients really cannot choose from any and all of the
subsequently abandoned by the parties after the FTC won a preliminary hospitals in the choice set. However, network configurations can
injunction in federal district court. change by the patient switching health plans and may not be as rigid as
implied by initially being part of a narrow network.
3 The WTP model was developed and applied to insurer-hospital
bargaining in a couple of influential academic articles: Town, R. and 13 An alternative to including various hospital attributes is to include an
Vistnes, G. “Hospital Competition in HMO Networks,” Journal of Health indicator variable (sometimes referred to as a “dummy” variable) for
Economics, Vol. 20 (2001), pp. 733 – 753 and Capps, C., Dranove, each hospital as a predictor. These indicators account for all hospital-
D.D., and Satterthwaite M.A, “Competition and Market Power in Option specific attributes that are time-invariant.
Demand Markets,” The RAND Journal of Economics, Vol. 34 (Winter 14 An interaction term in a regression model is used when the effect that a
2003), pp. 737 – 763. particular predictor has on the final outcome depends on the values of
4 For a non-technical write up of the theory and implementation another predictor. For example, one might introduce an interaction term
underlying hospital merger simulation, see Argue, D.A and Shin, between age and driving distance as a predictor in the hospital choice
R. T, “An Innovative Approach to an Old Problem: Hospital Merger model above (this is done by simply introducing a product between
Simulation,” Antitrust (Fall 2009), pp. 49 - 54 and Brand, K.J and the two variables as a predictor) in order to account for the possibility
Garmon, C. “Hospital Merger Simulation,” AHLA Member Briefing that the extent to which driving distance influences a patient’s hospital
(2014). choice is dependent on how old the patient is.
5 In place of insurer claims data, it is possible to use publicly available 15 To the extent that the influence of driving distance on hospital choice
financial data from state agencies or Medicare Cost Reports but one may be non-linear, a researcher might include non-linear functions of
might have to sacrifice some accuracy in such cases. In any event, when driving distance (such as polynomial terms) as predictors to capture
compared to a diversion analysis, the data and analysis requirements this effect.
for a merger simulation are much more substantial. Also, the merging 16 These predicted probabilities are also used for computing WTP in
parties may be able to gain access to the insurer claims data only if the hospital merger simulation model discussed earlier. As such, the
the merger is litigated, since these data are typically obtained by the diversion analysis is closely related to the WTP model with diversion
agencies through third-party subpoenas. Otherwise, the merging parties ratios being strongly correlated with changes in WTP. In fact, the
have to rely upon the publicly available data if they want to get a feel diversion analysis is often thought of as the first step in the hospital
for what the agencies are likely to find. merger simulation since the same conditional logit model is used in
6 Another related data limitation here relates to the lack of data on both analyses.
consummated mergers that would allow for a “before-and after” 17 See Farrell, J. and Shapiro, C., “Antitrust Evaluation of Horizontal
comparison of WTP and prices driven by the merger. Given this lack of Mergers: An Economic Alternative to Market Definition,” The B.E.
data, most merger simulations are run on cross-sectional data, i.e., the Journal of Theoretical Economics (2010), Vol. 10(1).
effect of a change in WTP on price is estimated using information on
18 Note that the diversion ratios are estimated under the assumption of
variation in WTP and prices across hospitals (and not over time within
hospital A being completely excluded from the network, and thus losing
a hospital). For more information on this and other limitations of the
all of its patient volume to competing hospitals. While the UPP analysis
model, see Brand and Garmon (2014) supra note 4.
is carried out in the context of a price increase, one can reconcile these
7 Note that, in settings outside of healthcare, diversion ratios are by considering a situation where the price increase instituted by hospital
computed using survey data or win/loss reports which may not be A leads to it being dropped from the network. Note that applying the
publicly available and might therefore be applicable only in later stages analysis in this way also assumes that the two hospitals, A and B, are
of the review process. In the absence of alternate information, diversion not bargaining with the payer on an all-or-nothing basis (as they would
ratios may also be computed using data on market shares, but this then both be excluded from the network).
would necessitate defining a relevant geographic market, a factor this
19 In their article outlining the UPP theory, Farrell and Shapiro suggest
approach was designed to avoid.
the use of a 10 percent efficiency credit for the purpose of screening
8 More information on the types of variables included, the ordering transactions that might be presumptively anticompetitive.
process, and the cost of procuring these data can be found online at See Farrell, J. and Shapiro, C. supra note 18.
the websites of these agencies: OSHPD (http://www.oshpd.ca.gov),
20 See In the Matter of ProMedica Health System, Inc. No. 9346,
THCIC (https://www.dshs.state.tx.us/thcic), and AHCA (http://ahca.
Opinion of the Commission by Commissioner Brill at 46-47.
myflorida.com).
21 It should be noted that, while the model includes a number of other
9 The outside option includes hospitals that are typically on the periphery
factors impacting patient choice, they often end up having lower
of the area being considered (or across state borders) as well as
predictive power in comparison to the effect of the driving distance
hospitals that have very little competitive relevance.
between the patient and the hospital.
10 As an example, if the data contain 30 choices in the hospital choice set
22 In addition, many physicians split time across multiple hospitals or
(including the outside option), and these hospital choices had a total
clinics, making it difficult to assign them to a particular location and
of 50,000 patient discharges in that year, the sample on which the
increasing the number of possible alternatives patients face while
model is estimated would include a total of 1.5 million observations
choosing physicians.
corresponding to each patient-hospital combination. Once the choice
set includes more than 50-60 hospitals, the computational burden 23 Of course, one of the key advantages of the diversion ratio approach
increases greatly, at which point the researcher might consider in hospital markets—the availability of public data—is lost while
expanding the definition of the outside option to include a greater examining physician mergers with this approach. While a full discussion
number of hospitals. of this model is beyond the scope of this paper, see Carlson, J.A. et
al., “Physician Acquisitions, Standard Essential Patents, and Accuracy of
11 A discrete-choice model is used in situations where the dependent
Credit Reporting,” Economics at the FTC (2013) for more details.
variable takes on a handful of discrete values (e.g., 0 or 1). In this case,
the dependent variable takes the value 1 for the observation where
the hospital is the patient’s actual choice and 0 for all other hospital
observations in the patient’s choice set.
www.nera.com 9
About NERA
NERA’s clients value our ability to apply and communicate state-of-the-art approaches clearly
and convincingly, our commitment to deliver unbiased findings, and our reputation for quality
and independence. Our clients rely on the integrity and skills of our unparalleled team of
economists and other experts backed by the resources and reliability of one of the world’s
largest economic consultancies. With its main office in New York City, NERA serves clients
from more than 25 offices across North America, Europe, and Asia Pacific.
Contact
For further information and questions, please contact the author:
The opinions expressed herein do not necessarily represent the views of NERA Economic Consulting or any other NERA
consultant. Please do not cite without explicit permission from the author.