Journal of Health Economics 42 (2015) 81–89

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

The interaction of direct and indirect risk selection

Normann Lorenz ∗
Universität Trier, Universitätsring 15, 54286 Trier, Germany

a r t i c l e i n f o a b s t r a c t

Article history: This paper analyzes the interaction of direct and indirect risk selection in health insurance markets. It is
Received 15 April 2014 shown that direct risk selection – using measures unrelated to the benefit package like selective adver-
Received in revised form 6 September 2014 tising or ‘losing’ applications of high risk individuals – nevertheless has an influence on the distortions of
Accepted 9 December 2014
the benefit package caused by indirect risk selection. Direct risk selection (DRS) may either increase or
Available online 17 December 2014
decrease these distortions, depending on the type of equilibrium (pooling or separating), the type of DRS
(positive or negative) and the type of cost for DRS (individual-specific or not). Regulators who succeed in
JEL classification:
reducing DRS by, e.g., banning excessive advertising or implementing fines for ‘losing’ applications, may
I13
I18
therefore (unintendedly) mitigate or exacerbate the distortions of the benefit package caused by indirect
L13 risk selection. It is shown that the interaction of direct and indirect risk selection also alters the formula
for optimal risk adjustment.
Keywords: Risk © 2015 Elsevier B.V. All rights reserved.
selection Risk
adjustment
Health insurance
Discrete choice
Imperfect competition

1. Introduction advertising or ‘losing’ applications of high risk individuals.3 It has

Risk selection is considered to be one of the main problems in
regulated health insurance markets. If there is community rating,
so that insurers are not allowed to charge premiums according to
risk, they will make profits with some individuals and losses with
others. Insurers who act on these incentives to attract profitable and
repel unprofitable individuals are said to engage in risk selection.1
Two forms of risk selection can be distinguished: direct risk
selection (DRS) and indirect risk selection (IRS).2 With DRS, insur-
ers know that a particular individual or group of individuals is
characterized by non-average risk. DRS is therefore targeted at an
individual the insurer has identified to either be a high or a low
risk type (like, e.g., a hypochondriac) or at a group of individuals
the insurer knows to have non-average expected cost (like, e.g., a
certain age group or individuals living in a high cost area). Usu-
ally, the measures taken for DRS are not related to the benefit
package (i.e. the medical services) offered; examples are selective
been shown that potential profits associated with successful DRS
can be substantial.4
With IRS, insurers do not know which particular individual is
of high or low risk (or are prevented from using this information),
and so only act on their knowledge that there are different risk
types in the population. The measures taken to engage in IRS usually
consist of distorting the benefit package, so that it is attractive for
low risks, but not for high risks. Several studies have shown that
the incentives for IRS can be severe, and that insurers do indeed act
on these incentives. 5
A regulator can counteract the incentives for both DRS and IRS by
implementing a risk adjustment scheme, setting transfers to (and
from) insurers depending on signals which are informative about
individuals’ expected cost. In almost all risk adjustment schemes,
the formula used to calculate these transfers is based on a regres-
sion of actual health care expenditures on a set of explanatory
variables like age, gender and morbidity. Most of the literature on

3
van de Ven and van Vliet (1992) provide an extensive list of measures insur-
∗ Tel.: +49 651 2012624. ers may use for risk selection; for differential treatment of low and high risks’
E-mail address: lorenzn@uni-trier.de applications see Bauhoff (2012).
1 4 See Shen and Ellis (2002).
See van de Ven and Ellis (2000).
2 5
See Breyer et al. (2011). See Frank et al. (2000), Cao and McGuire (2003) and Ellis and McGuire (2007).

http://dx.doi.org/10.1016/j.jhealeco.2014.12.003
0167-6296/© 2015 Elsevier B.V. All rights reserved.
82 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 2

risk adjustment has been concerned with improving this underly- Table 1

Effect of DRS with individual-specific cost on the distortion of the benefit package.
ing regression by, e.g., including additional variables or altering the

grouping algorithm for diagnoses in morbidity based risk adjust- Type of equilibrium Positive DRS Negative DRS
ment, so that a larger part of the variance of actual expenditures is Pooling equilibrium Distortion decreases Distortion increases
explained. The larger the explained part of the variance, the closer Separating equilibrium Distortion decreases Distortion decreases
transfers are to actual cost, and the lower the incentives for risk
selection should be.

Initiated by the very influential study of Glazer and McGuire high risks is reduced. Negative DRS on the other hand creates a ‘sub-
(2000), there has developed a small literature that departs from this stitution effect’: If insurers are (somewhat) successful in repelling
statistical approach and instead explicitly models insurers’ incen- the high risks by DRS, they can reduce the degree of IRS. For an
tives for risk selection. One study in this literature has shown that overview of these results see, Table 1.
conventional, i.e. regression-based, risk adjustment may decrease A regulator who succeeds in reducing negative DRS (by, e.g.,
welfare if there is imperfect competition, another, that it may even charging a fine for ‘losing’ applications of high risk individuals) will
increase the extent of risk selection.6 These undesirable effects of therefore simultaneously reduce IRS in the pooling equilibrium, but
conventional risk adjustment exemplify the need for what Glazer unintendedly increase the distortion of the benefit package in the
and McGuire (2000) have termed optimal risk adjustment. They separating equilibrium. The distortions caused by IRS will also be
have shown that a regulator can increase the effectiveness of a risk increased if he succeeds in reducing positive DRS (by, e.g., banning
adjustment scheme by distorting the payments as calculated from excessive advertising), in this case for both the pooling and the
a regression: there has to be overpayment for signals which are
separating equilibrium.
correlated with high risks and underpayment for signals which are
In three of the four cases, optimal risk adjustment then becomes
correlated with low risks. If the over- and underpayments are cho-
even more important. We therefore derive the impact of DRS
sen optimally, incentives for IRS can be eliminated completely.7
on the formula for optimal risk adjustment developed by Glazer
Optimal risk adjustment has also been derived for a setting where
and McGuire (2000). We show that the overpayment for a signal
individuals differ in their elasticity to switch insurers or where
insurers are allowed to vary their premium in some dimension, that indicates a high risk has to be increased exactly by insurers’
expenditures on positive DRS; likewise, the underpayment has to
as is the case in the insurance exchanges in the US.8
be reduced by the expenditures on negative DRS. With this modi-
A concern, already raised by Glazer and McGuire (2000) them-
fication, their formula can eliminate the incentives for IRS even in
selves, is that such over- and underpayments create incentives
the presence of DRS.
for DRS regarding the signal, but so far it has not been analyzed
In the literature on optimal risk adjustment, some of the results
whether this has an influence on optimal risk adjustment. In fact,
regarding the distortions caused by IRS have been derived under
we are not aware of any theoretical study that explicitly models
perfect competition, but DRS seems incompatible with such a
the interaction of direct and indirect risk selection, even in the
setting where individuals are perfectly informed about all ben-
absence of risk adjustment.9 In this study we therefore develop
efit packages and premiums and always choose the insurer that
such a model and show that in general (the degree of) DRS has
offers the best benefit package-premium combination. We there-
an influence on the distortions of the benefit package caused by
fore derive our results within a discrete choice model, which can
IRS and that this alters the formula for optimal risk adjustment.
easily capture different levels of competition. To keep the model
DRS may either increase or decrease the distortions caused by
simple, we assume that the benefit package is one-dimensional, but
IRS, depending on whether insurers try to attract the low risks
the model can be extended to a multi-dimensional benefit package.
(positive DRS) or to repel the high risks (negative DRS), whether a
Also, to simplify the notation when deriving the results, we first
pooling or a separating equilibrium emerges, and whether the cost
consider the case of two risk types, but then show that the results
for DRS is individual-specific or not.
also hold for an arbitrary number of risk types.
If insurers’ expenditures for DRS are at least to some degree
The remainder of this paper is organized as follows: In Section
individual-specific (and not just a fixed cost), they affect risk-type-
2, we introduce the basic discrete choice model and show how DRS
specific cost: Positive DRS increases the cost per low risk, negative
can be incorporated in such a model. We analyze the pooling equi-
DRS the cost per high risk. In the first case, the cost difference
librium in Section 3 and the separating equilibrium in Section 4.
between the risk types is reduced, in the second, it is increased.
Section 5 concludes.
In the pooling equilibrium where both risk types pay the same pre-
mium, positive DRS therefore reduces the incentives for IRS, while
negative DRS increases it. 2. The model
In the separating equilibrium, insurers’ expenditures for posi-
tive DRS translate into a higher premium for the contract offered 2.1. Basic model without DRS
for the low risks; this makes this contract less attractive for the high
risks, so the distortion of the benefit package necessary to repel the Individual preferences regarding the benefit-premium bundle
are given by

6
See Lorenz (2013) and Brown et al. (2011), respectively. In the empirical part of
their study, Brown et al. (2011) find such an increase in the extent of risk selection
for the Medicare Advantage program in the U.S.; however, there has been some
disagreement on this finding, see Newhouse et al. (2012).
7 See also Glazer and McGuire (2002) and Jack (2006).
8
For the first setting, see Bijlsma et al. (2011), and for the second, McGuire et al.
(2013) and Shi (2013).
9 Eggleston (2000) derives the optimal mix of supply and demand side cost shar-
ing for a setting with a single (semi-altruistic) HMO that can influence the level of
medical services (according to the outcome of a patient–provider bargaining pro-
cess) and can dump a share of the high risks at some cost; however, there is no
competition as there is only one provider.
u = pr v(m) − R, (1)
where R denotes the premium and m the level of medical services
(measured in monetary terms). pr is the probability of becoming
ill, and there are two risk types r = H, L, with pH > pL ; the share of
the low risks is . The utility of receiving medical treatment, v(m),
is increasing at a decreasing rate, i.e. v∗ (m) > 0 and v∗∗ (m) < 0. The
efficient level of medical services is implicitly defined by v∗ (mFB ) =
1.
There are n insurers j, each offering a benefit-premium bundle
{mj , Rj }. The individual’s decision of which insurer to choose may,
however, not only depend on these benefit-premium bundles, but
82 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 3

risk adjustment has been concerned with improving this underly- Table 1

also on some other factors, like perceived friendliness of personnel,
location, or which insurer was recommended by family or friends.
In a discrete choice model, these other factors are captured by aug-
menting the individual’s utility as given in (1) by an individual- and
j
insurer-specific utility component εi; the utility of an individual i
(being of risk type r) when choosing an insurer j therefore is
u i (mj , Rj ) = pr v(mj ) − Rj + εji .
If εj is assumed to be i.i.d. extreme value, the logit model with its
i g(aj ) is increasing and concave.13 With positive DRS, the (perceived)
analytically tractable choice probabilities arises. Denoting risk type
r’s utility of the benefit-premium bundle offered by insurer j by
Vrj = pr v(mj ) − Rj ,
and specifying the variance of ε as Var(ε ) =
j j
i i
ity of individual i (being of risk type r) choosing a particular insurer
k is10
Prob(i chooses k) = Prob(Vr k + εk > V l + εl ∀ l =
/ k)
i r
k
eVr / non-individual-specific and individual-specific cost. With non-
2.2. The model with positive and negative DRS
We consider positive DRS to be an activity each insurer is
engaged in which generates some cost and increases the proba-
bility of being chosen by the individual (or group of individuals)
the activity is targeted at. We model this increase in the probability
of being chosen to stem from an increase in the utility the individ-
(2) ual receives, which may either be real (as, e.g., with a discount for
a fitness club membership) or just perceived (as with advertising).
We denote the cost by aj and the increase in utility by g(aj ), where
utility of individual i choosing an insurer j therefore is
u i (mj , Rj ) = pr v(mj ) − Rj + g(aj ) + εji, (4)
so that insurer k’s market share among risk type r is given by14
2 2 /6, the probabil-
k k
e(V
r +g(a ))/
Prk = j
. (5)
e(Vr +g(aj ))/
j
i
Two cases regarding the cost aj can be distinguished:

= j
. (3)
n
e Vr /
j=1
Denote this probability by P kr ; it is also insurer k’s market share
among the individuals of risk type r. P k is increasing in V k : a higher
r r
share of individuals of risk type r will choose insurer k, if this insurer
offers a higher level of medical services or charges a lower premium.
The variance of the additional utility component, Var(εji ) =
2 2 /6, is a measure of the level of competition in this health insur-
individual-specific cost, total cost for DRS of an insurer j is
independent of the number of individuals choosing this insurer.
The prime example for this case is selective advertising, where
cost does not increase if an additional individual chooses insurer
j. With individual-specific cost, total cost for DRS of an insurer
10
See Train (2009, p. 40).
11
See Zweifel et al. (2009), chapter 7, for the Rothschild–Stiglitz equilibrium in
this setting.
12 We explain why this occurs in Section 4.1.

ance market. If is small, all the εij are very similar and therefore
only play a minor role in which insurer is chosen: Offering an only
somewhat higher utility level than all the other insurers will, in this
case, already attract a large share of all individuals; this implies a
high level of competition. If, on the other hand, is large, the other
factors besides the benefit level and the premium – captured by
j
large positive and large negative εi – are rather important, so that
insurers, when increasing their premium (or reducing their benefit
level), only lose a small share of their insured; a large level of
therefore corresponds to a low level of competition.
As shown by Lorenz (2013) who has analyzed this basic model
without DRS, the level of competition determines which type of
equilibrium emerges: If the level of competition is low (i.e., is
large), there will be a pooling equilibrium: all insurers offer the
same benefit-premium bundle, so each individual i chooses the
j
insurer j for which his εi is maximal. If the level of competition
is high (i.e., is small), a separating equilibrium similar (but not
identical) to the Rothschild–Stiglitz equilibrium under perfect com-
petition arises.11 Some of the n insurers offer a benefit-premium
bundle designated for the low risks, the remaining insurers a con-
tract designated for the high risks (with a larger benefit package
and a higher premium). All the low risks choose the insurer with
j
the highest εi among the first set of insurers, but only most of the
j
high risks choose the insurer with the highest εi among the second
set. Because a small share of the high risks chooses one of the insur-
ers offering the contract designated for the low risks, the separation
of risk types is not perfect.12
82 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 4

risk adjustment has been concerned with improving this underly- Table 1

j increases in the number of individuals choosing this insurer.

One example here are additional benefits which the regulator (or
society) considers not to be part of a ‘normal’ basic benefit package
insurers are supposed to provide, like discounts for fitness club
memberships or special counseling services. In this case, total
cost of DRS increases if an additional individual chooses insurer j.
It seems reasonable to assume that most risk selection activities
entail both non-individual-specific and individual-specific (i.e.
fixed and variable) cost.
Like positive DRS, we model negative DRS as an activity that
generates some cost, but decreases the probability of being chosen
by a particular individual (or group of individuals). We denote the
cost of negative DRS by bj and the utility decrease by f(bj ), where
f(bj ) is increasing and concave. With negative DRS, the (perceived)
utility of individual i and insurer k’s market share are as given in
(4) and (5), but with +g(a) replaced by −f(b).
Unlike with positive DRS, it is difficult to imagine some activity
where the cost an insurer incurs for negative DRS is independent of
the number of individuals choosing this insurer. ‘Negative adver-
tising’ might be an example, where an insurer informs about some
undesirable feature of its offer, like scrupulous utilization reviews,
but this and similar examples may seem rather far-fetched. We
think it is more realistic to consider negative DRS to be an activity
insurers are engaged in during the application process, so that cost
depends on the number of individuals applying at the insurer.15
Activities which fall into this category are that insurers require
additional (unnecessary) paper work or involve the high risk indi-
viduals in lengthy phone calls in which they try to convince (or

13
Since we are interested in a setting where insurers are engaged in DRS, we
assume lim g ∗ (a) → ∞ to guarantee an interior solution.
a→0+
14
To simplify the notation, we do not introduce different symbols for P kj for the
case of no, positive or negative DRS; we will, however, always make clear to which
case we refer.
15 In Section 5 we argue that the main effects should be similar if negative DRS

does not occur during the application process, but is targeted at individuals who
already hold a contract with the insurer.
84 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 84

urge) these individuals to choose a different insurer.16 In the work- Table 2

Shares rs of the four types; positive correlation of H and O for ı > 0.

ing paper version of this study we explicitly model the application
process and show that the main results are the same regardless of pL pH
whether the number of individuals applying at an insurer is equal Y LY = +ı HY = (1 − ) − ı
to or larger than the number of individuals eventually choosing O LO = (1 − ) − ı HO = (1 − )(1 − ) + ı 1−
the insurer.17 We will therefore refer to both cases as individual- 1−
specific cost.

3. The pooling equilibrium

In this section we analyze the pooling equilibrium which occurs ∗

if the level of competition is low enough. We begin by briefly deriv- g ∗ (ak )( k k 19

L − a ) = n /(n − 1). The larger g (a), i.e., the more effec-

ing the equilibrium without DRS in Section 3.1. We consider the
equilibrium with positive DRS in Section 3.2 and analyze the impact
of positive DRS on risk adjustment in Section 3.3. We then show
that the results also hold for an arbitrary number of risk types in
Section 3.4. The effects of negative DRS, which are just the opposite
tive insurers’ expenditures on risk selection are, the larger, c.p.,
the equilibrium level of ak will be. These expenditures affect the
distortion of the benefit level m, which is now given by
2

of positive DRS, are summarized in Section 3.5. (1 − )(pH − pL ) k (1 − )(pH − pL ) k k

1− n m + a v∗ (m ) = 1.
p n
n−1 n−1 p
3.1. The pooling equilibrium without DRS (10)

Normalizing the mass of individuals to one and assuming profit Because the last term in the brackets [·] is positive, compared with
maximization, the objective of insurer k is
(8), the condition if there is no DRS, v∗ (mk ) has to decrease: mk
= P kL increases, so the distortion is reduced. More generally, the larger
L+ (1 − )P kH
k k k
max H, (6)
mk ,Rk the equilibrium level of ak , the larger mk .
where k = Rk − pr mk denotes insurer k’s profit per individual of
r Result 1. In the pooling equilibrium, the distortion of the benefit
risk type r. The solution to this objective yields the following two level decreases in the level of positive DRS if cost for DRS is individual-
conditions (see Appendix A.1): First, specific.
n
k
+ (1 − ) k
= . (7)
L H n−1 The incentive to distort the benefit level (with or without DRS)
arises because profit per high risk is lower than profit per low
Average profit per insured decreases in n and increases in : a risk, where the degree of the distortion depends on the difference
higher level of competition (a larger number of insurers n or a between these two profits, which, in the case without DRS, is given
smaller level of ) decreases profit. Secondly, the condition deter-
by20 − = (p − p )m .
k k H L k
mining the distortion of the benefit level is given by L H

With positive DRS, profit per low risk decreases because insurers
(1 − )(pH − pL )
2 waste part of this profit on their expenditures for DRS; this reduces
k k
1− n m v∗ (m ) = 1, (8) the difference between net profits (including ak ), and thereby the
p
n−1
incentive to distort the benefit level.
where p = pL + (1 − )pH . Because the fraction is positive, it is This is different for the case of non-individual-specific cost.21
immediately apparent that v∗ (mk ) > 1, so that mk is distorted below Although in this case, expenditures for risk selection decrease total
the efficient level mFB . As is to be expected, the distortion increases profit, they do not specifically decrease profit per low risk, so the
in the difference pH − pL and in the level of competition (captured difference between the risk-type-specific profits remains the same:
by n /(n − 1)). Therefore, positive DRS has no influence on the distortion of the
benefit level if cost is non-individual-specific. In the following, we
3.2. The pooling equilibrium with positive DRS will therefore only consider individual-specific cost.

3.2.1. Positive DRS of risk type

With positive DRS of the risk type itself and individual specific
cost, insurer k’s objective reads as 3.2.2. Positive DRS of a signal that is correlated with risk type
We now analyze the (probably more realistic) case that DRS is
k
= P kL( k
L− ak ) + (1 − )P kH k
H, (9) not targeted at the risk type itself, but at a signal that is (less than
perfectly) correlated with risk type. We assume that there are two
where now PLk contains g(ak ) and g(aj ) as given in (5).18 signal types s = Y, O, young and old. The shares of the four types of

The solution to this objective yields a positive equilibrium level of insurers’ expenditures on DRS, implicitly defined by
85 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 85

urge) these individuals to choose a different insurer.16 In the work- Table 2

individuals, rs , are given in Table 2, where ı > 0 captures the case

of a positive correlation of high risk and old age.
16
After a German sickness fund operating mainly in high cost areas went bankrupt
This formulation of a positive correlation has the advantage that
in 2011, members of this fund who then applied at other funds received phone calls increasing ı increases the level of correlation without altering the
in which some of the insurers told them that they could not continue their drug shares of the two risk types, and (1 − ), and the shares of the two
therapy or disease management program if they did not choose a different insurer; signal-types, and (1 − ).
see, e.g., Spiegel (2011).
17 See Lorenz (2014). There, we also illustrate all results with an example.
18
We assume Hk < 0, so that positive DRS is targeted only at the low risk. For a
very low level of competition (very high ), even the high risks would entail a profit
and positive DRS would be targeted at both risk types (albeit at different levels). 19
See Appendix A.2, where also condition (10) is derived.
20
See the second term in the brackets in condition (8), which contains (pH − pL )mk .
21 In this case, insurer k’s objective is k = (P kL kL − ak ) + (1 − )P kH kH.
86 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 86

urge) these individuals to choose a different insurer.16 In the work- Table 2

With positive DRS of the young and individual-specific cost, so that the distortion of the benefit level is eliminated for
insurer k’s objective reads as22 (1 − ) H
RAO = (p − pL )m. (15)
ı
k
= k k
rs P rs r − k k
rY P rYa , (11)
r s r
If there is perfect correlation, the share of the low risks equals the
share of the young, so = ; in addition, the mass of individuals
k contains g(ak ) and g(aj ). The distortion of the benefit in the lower left and the upper right corner in Table 2 has to be
where only PrY
level is now determined by23 zero, which requires ı = (1 − ) = (1 − ). Replacing in (15) shows
that with perfect correlation the efficient benefit level is therefore
(1 − )(pH − pL ) 2 (pH − pL ) implemented for RAO = (pH − pL )m, which is just the cost difference
1− n mk + n ıak v∗ (mk ) = 1. (12)

n−1 p n−1 p
between the two risk types. With less than perfect correlation,
ı < (1 − ), so there is overpayment, and the lower the level of

Because the last term in the brackets is positive, compared with no
DRS, v∗ (mk ) has to decrease, so the distortion is reduced. However,
for a given level of ak , the reduction of the distortion is of course not
as large as with DRS against the risk type itself, since ı < (1 − ).
If there is less than perfect correlation, expenditures on DRS of the
young not only increase the cost of the low risks, but also of the
high risks and therefore translate into a smaller reduction of the
cost difference between the risk types. As is to be expected, for a
given level of ak the reduction of the distortion increases in the level
of correlation (i.e. in ı).
Result 2. In the pooling equilibrium, the distortion of the benefit
level decreases in the level of positive DRS of a signal that is correlated
correlation (i.e., the lower ı), the larger this overpayment has to
be.
3.3.2. Optimal risk adjustment with positive DRS
With optimal risk adjustment there is overpayment for the old,
so this is the group positive DRS will be targeted at.25 Insurer k’s
objective in this case is given by (13) with RAO replaced by RAO − ak
rO
and with P k containing g(ak ) and g(aj ) as given in (5). The posi-
tive equilibrium level of ak then enters condition (14), where again
RAO has to be replaced by (RAO − ak ).26 The efficient benefit level is
therefore implemented with
(1 − ) H L k

with low risk if cost for DRS is individual-specific. A higher level of RAO = (p − p )m + a . (16)
ı
correlation (a higher ı) increases the effect of a given level of DRS on

the distortion of the benefit level. Because part of the overpayment for the old is spent on positive
DRS, the optimal overpayment for the old has to be raised by exactly
these expenditures, so that the net difference in payments (includ-
3.3. Implications of DRS on optimal risk adjustment in the ing ak ) equals the amount necessary to eliminate IRS as given in
pooling equilibrium (15).

We now discuss the implications of the interaction of direct and Result 3. With individual-specific cost, if there is positive DRS
indirect risk selection for optimal risk adjustment. As shown by regarding a signal that is used for risk adjustment and that is correlated
Glazer and McGuire (2000), if a regulator does not observe indi- with high risk, the optimal overpayment of that signal to eliminate IRS
viduals’ risk type, but only a signal that is correlated with risk type has to be increased by the expenditures for DRS.
(like age), it is still feasible to eliminate the distortion of the benefit
This result shows that the formula for optimal risk adjustment
package by overpaying for a signal that indicates a high risk, and
derived by Glazer and McGuire (2000) is not invalidated by DRS:
underpaying for a signal that indicates a low risk. In Section 3.3.1,
there is still overpayment for a signal that is correlated with high
we first show how to implement optimal risk adjustment in our
risk, and underpayment for a signal correlated with low risk. Also,
setting with two risk types and two age groups if there is no DRS;
DRS does not invalidate their claim that optimal risk adjustment
we then determine how the optimal payments have to be modified
can implement the efficient benefit level. However, the formula to
if there is DRS in Section 3.3.2.
implement the efficient benefit level has to be modified and include
insurers’ expenditures on DRS if cost is individual specific. Whether
3.3.1. Optimal risk adjustment without DRS these expenditures are negligible or significant is an empirical mat-
With risk adjustment, each insurer receives a payment of RAO ter, but, e.g., the findings of Starc (2014), who reports that insurers
for each insured that is old; these payments are financed by a risk spend a large part of potential profits on marketing and insurance
adjustment fee RAF which each insurer has to pay for each insured brokers, indicate that these expenditures may be substantial.
(including the old). The balanced budget constraint requires this fee
to be RAF = (1 − )RAO . The insurer’s objective with risk adjustment 3.4. More than two risk types
is then given by
To keep the notation simple when deriving the results, so far
k
= rY P
k
( k
− RAF ) + rO P
k
( k
− RAF + RAO ). (13)
rY r rO r we have considered the case of only two risk types. The results,
r r however, also hold for an arbitrary number of risk types. Let the

number of risk types be r and denote the probability of risk type r

Simplifying the optimality conditions now yields24
by pr and its share by r .
2 In the conditions determining the distortion of the benefit level,
(1 − )(pH − pL ) (pH − pL )
1− n mk + n p
ıRAO v∗ (mk ) = 1, (14) 2
conditions (8), (12) and (14), the term (1 − )(pH − pL ) then has to
p
87 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 87

urge) these individuals to choose a different insurer.16 In the work- Table 2

n−1 n−1 r 2 (p
be replaced by p) > 0; see Appendix B.1. This is just
r=1 − r r
the variance of the illness probabilities, and the larger this variance,
the larger the distortion of m.
22
Like in Section 3.2.1, we assume that positive DRS is profitable for only one of
the two signal types; see footnote 18.
23 See Appendix A.3. 25
As before we assume that DRS is only targeted at one of the signal types.
24 26
See Appendix A.4. See Appendix A.5.
86 N. Lorenz
N. Lorenz
/ Journal
/ Journal
For this general case of an arbitrary number of risk types, we
define the shares of individuals (which were given in Table 2 for
the case of two risk types) by rY = r + ır and rO = r (1 − ) − ır ,
see Table B.1 in Appendix B.2. It is then straightforward to show that
there is a positive correlation of risk type and old age for r ır pr < 0;
ır has to be (mainly) positive for small illness probabilities and neg-
ative for large illness probabilities. The term (pH − pL )ı in the last
fraction of (12) then has to be replaced by − r ır pr , so positive DRS
again reduces the distortion caused by IRS (see Appendix B.3).
To eliminate the incentives for IRS with optimal risk adjust-
ment, as before there has to be overpayment (see Appendix B.4)
and this overpayment has to be increased by insurers’ expenditures
for positive DRS (see Appendix B.5).
3.5. The pooling equilibrium with negative DRS
The discussion of negative DRS can be very brief because for all
the settings we considered, negative DRS creates just the opposite
effects of positive DRS.27 Negative DRS increases the loss associated
with the high risks, which increases the cost difference between the
two risk types and thereby increases the incentives for IRS.
Negative DRS regarding a signal that is used for risk adjustment
and for which there is underpayment increases the loss for this sig-
nal, which allows keeping the underpayment less pronounced than
without DRS. We can therefore summarize the results for negative
DRS as follows:
Result 4. In the pooling equilibrium and with individual specific cost,
the distortion of the benefit level increases in the level of negative
DRS against the high risk type or against a signal that is correlated
with low risk. If there is negative DRS regarding a signal that is used
for risk adjustment and that is correlated with low risk, the optimal
underpayment of that signal to eliminate IRS has to be reduced by the
expenditures for DRS.
4. DRS in the separating equilibrium
4.1. The separating equilibrium without DRS
In Lorenz (2013), it is shown that the separating equilibrium
arises for a high level of competition (i.e. a low level of ) and that
it is similar, but not identical, to the Rothschild–Stiglitz equilibrium
under perfect competition.28 Both equilibria can be found in Fig. 1.
Under perfect competition, the separating equilibrium consists
of contract B, chosen by the high risks, and contract A1 , chosen by
the low risks, where A1 is located at the intersection of the low
risks’ iso-profit line pL and the high risks’ indifference curve
associated with contract B (so that the incentive compatibility con-
straint is satisfied); in addition, both iso-profit lines pass through
the origin and are in this case zero-profit lines.29 The number of
insurers offering the contract for the low risks (insurers of type A)
and for the high risks (insurers of type B), nA and nB respectively,
are indeterminate under perfect competition.30
This is different under imperfect competition, where nA and nB
are implicitly defined by the profit equality condition A = B : total
profit for an insurer of type A has to be the same as for an insurer of
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 86
Fig. 1. Separating equilibrium: contracts B and A3 are offered.
type B. If, e.g., B > A , it would be profitable for an insurer of type
A to become of type B; increasing nB decreases the market share of
each insurer of type B, which decreases B . In equilibrium, profit
per high risk equals nB /(nB − 1), so the iso-profit line associated
with contract B starts at that level, see Fig. 1. However, while B
is indeed one of the equilibrium contracts, A1 is not. To simplify
the exposition of which contract is offered for the low risks, in the
following we will simply speak of ‘insurer A’ and ‘insurer B’, i.e.,
present the explanation as if there was only one insurer of type A
and one insurer of type B.
B
Because B and A1 are both located on the indifference curve I VH ,
these two benefit-premium bundles provide the same utility to the
high risk, i.e., V HB = V HA1 ; therefore, all the high risks with εAi > εBi
will choose insurer A. From the perspective of insurer A who tries
to avoid being chosen by the high risks by offering a contract that
satisfies the incentive compatibility constraint, what is important
B
is not only the indifference curve I V H , but also the additional utility
components ε , in particular, the difference εA − εB . For some of the
j
i i i
high risks, this difference will be negative, for others somewhat,
or even considerably, larger than zero. There is therefore no single
incentive compatibility constraint (ICC), but an individual-specific
ICC for each of the high risks depending on this difference.
One way to capture these individual-specific ICCs is by a dis-
tribution function denoting the share of the high risks for which
the ICC is violated; formulated in this way, this distribution func-
tion coincides with insurer A’s market share among the high risks
A
as a function of VH given VHB, i.e., PHA = PHA(VHA|VHB). Graphically, this
B distribution function can be depicted by a shaded area around the
IV H
H
B
indifference curve I V , representing the density of this distribution
(with the darkness of the shaded area as a measure of this density,
which is bell-shaped). Above the shaded area, the ICC is satisfied for
all the high risks, so P A A
H = 0. Within the shaded area, PH increases
in V A A A A A
H according to the density ∂PH /∂VH = PH (1 − PH )/ . Below the
shaded area, PHA = 1, because there the ICC is violated for all the
27
This is straightforward to show by replacing ak by bk and +g(a) by −f(b) in all
the optimization problems and the derivations we derived in the appendix; see also
Lorenz (2014).
28 The separating equilibrium for a low level of exists under the same condition
as the Rothschild–Stiglitz equilibrium: the share of low risks must not be too large.
29
See Zweifel et al. (2009), chapter 7.
87 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 87

30
However, Olivella and Vera-Hernandez (2010) have shown that the smallest
number of insurers supporting this equilibrium is n = 3, with nB = 1 and nA = 2.
high risks.
Insurer A, when offering A2 , would therefore not be chosen by
any of the high risks. However, he can still increase his profit by
moving along the indifference curve of the low risks to the right,
which would have two effects31 : on the one hand, it increases profit
per low risk because the indifference curve has a larger slope than
pL ; on the other hand, it also increases the number of the high risks

31
We consider a movement along the indifference curve of the low risks because
with more than one insurer of type A, this would leave this insurer’s market share
among the low risks unaffected.
88 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
Fig. 2. Separating equilibrium with (a) positive or (b) negative DRS.
choosing this contract. Because at the boundary of the shaded area
this second effect is of second order (since the density is about zero
if PHA ≈ 0), the two effects balance somewhere inside the shaded
area, represented by contract A3 .32 Because a small share of the
high risks choose contract A3 , this contract is somewhat above the
iso-profit line pL (which would apply if only the low risks choose
insurer A).
Due to the ‘stochastic nature’ of the incentive compatibility con-
straint, the separating equilibrium under imperfect competition is
therefore not perfectly separating. Instead, a small share of the high
risks chooses the contract designated for the low risks, but none of
the low risks choose the contract designated for the high risks.
4.2. The separating equilibrium with DRS
4.2.1. The separating equilibrium with positive DRS
For positive DRS, we have to distinguish whether the utility
increase g(a) will only (or at least primarily) be experienced by the
low risks (as, e.g., might be the case with a discount for a fitness
club membership) or equally by both risk types (as, e.g., might be
the case with advertising). We begin with the case that the utility
increase can only be experienced by the low risks. Positive DRS will
than only be performed by insurers of type A.33
The objective of insurers of type A with positive DRS of the low
risks equals the objective as given in (9). The solution to this objec-
tive determines a positive equilibrium level of expenditures on
DRS, which increases the premium charged by these insurers.34
In Fig. 2(a), this increase in RA can be shown by an upward shift of
the corresponding iso-profit line. As is apparent, insurers of type
A can then increase their benefit level before attracting the same
share of the high risks as without DRS. Because a higher bene-
fit level (accompanied by the according increase in the premium,
pL mA ) increases the utility of the low risks, insurers will offer this
higher benefit level (if there is some competition), so that the new
equilibrium will be a contract like A4 . Since positive DRS reduces
the attractiveness of the contract offered by insurers of type A for
the high risks in the premium-dimension, they can increase the
attractiveness of their contract in the benefit level-dimension.
32
See Appendix C.1; for a more detailed derivation of this equilibrium, see Lorenz
(2013).
VA
33 Because contract B is far above the indifference curve I associated with con-
L
tract A3 , for the low risks there is a huge difference in utility between A3 and B. Any
moderate increase in the (perceived) utility of contract B due to positive DRS of insur-
ers of type B will reduce this difference only to a small degree and not induce any
42 (2015)
42 (2015)
81–89
81–89 88
Result 5. In the separating equilibrium, the distortion of the bene-
fit level is reduced if there is positive DRS of the low risk and cost is
individual-specific.
We now turn to the case that both risk types can experience the util-
ity increase g(a). In this case, DRS will be performed by all insurers:
insurers of type A engage in DRS of the low risks and insurers of type
B in DRS of the high risks (since high risks are profitable to insurers
of type B). Because the equilibrium level of expenditures on DRS in
this case is the same for both types of insurers,35 all insurers raise
their premium by the same amount. In Fig. 2(a), both iso-profit lines
are then shifted upwards by the same distance, so there is no effect
on the benefit levels mA and mB .
However, it seems reasonable to assume that positive DRS is
at least somewhat more effective when targeted at the low risks
than when targeted at the high risks. In equilibrium, insurers of
type A will spend more on DRS and raise their premium by a higher
amount than insurers of type B. The larger upward shift of their
iso-profit line then allows insurers of type A to increase their ben-
efit level (just as in the case where the iso-profit line of insurers
of type B is not shifted at all). Therefore, as long as positive DRS
reduces the attractiveness of the contract offered by insurers of
type A relative to the contract offered by insurers of type B in the
premium-dimension, insurers of type A can increase the attractive-
ness of their contract in the benefit level-dimension.
4.2.2. The separating equilibrium with negative DRS
Negative DRS will only be performed by insurers of type A,
because only these insurers try to avoid being chosen by the high
risks. As in Section 4.1, we explain the effect of negative DRS
as if there was only one insurer of type A and one insurer of
type B.
A − f (b A) +
With negative DRS, a high risk chooses insurer A if VH
εiA > VHB + εiB. Because negative DRS reduces the utility as perceived
by the high risks by f(b A), VHA can be increased by exactly that
amount without altering the number of the high risks choosing
insurer A. This allows insurer A to increase his benefit level, see
Fig. 2(b). There, AL5 denotes a contract offered by insurer A as
AH
perceived by the low risks, while 5 denotes the same contract as
perceived by the high risks. Compared to AL5, AH5 is shifted upwards
by f(bA ), the utility decrease of negative DRS (measured in mone-
tary terms). The larger bA , the larger f(bA ), and therefore the larger
mA can be without increasing the share of high risks choosing
of the low risks to choose this contract. Because insurers of type B cannot increase
their market share among the low risks (which remains at zero), they will refrain
from positive DRS.
89 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 89

insurer A. Because this effect occurs regardless of whether cost

for DRS is individual-specific or not, there exists one case where
35 This can be seen by replacing k
by ( k
− ak ) in (6) for insurers of type B and
H H
34 See Appendix C.2. comparing the respective optimality condition with the one for insurers of type A.
90 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 90

DRS influences the distortion caused by IRS even if cost is non- Appendix A.
individual-specific.
A.1. The pooling equilibrium without DRS
Result 6. In the separating equilibrium, negative DRS against the
high risks reduces the distortion of the benefit level of the low risks Using the property that the derivative of P k k

regardless of whether cost for negative DRS is individual-specific or ∂P k with respect to Vr

can be expressed in terms of P k r
P rk (1−P rk )
itself as r = , the FOCs to
r ∂ Vrk
not.
objective (6) are given by

The main mechanism in this last case is therefore different from all
the other cases we have considered. Here, IRS is substituted by DRS, ∂ k Pk k k k

L (1 − PL ) PH (1 − PH )
k
+ k k k

while in all the other cases, IRS affects risk type specific profits and = − PL + (1 − ) − + PH
∂ Rk L H
thereby the gains from IRS.

= 0 (17)

4.3. Optimal risk adjustment in the separating equilibrium

Pk k

∂ k
L L (1 − PL ) ∗ k k
= p v − PL
With optimal risk adjustment, profit per high risk equals ∂mk L

profit per low risk, so insurers’ incentives to distort the benefit

package are eliminated and the separating equilibrium turns into H PHk (1 − PHk ) ∗ k k
+ (1 − )p v H
− PH = 0. (18)
a pooling equilibrium where all insurers offer the same benefit
premium bundle. The impact of DRS on optimal risk adjustment
in the pooling equilibrium has, however, already been derived in Using the fact that in equilibrium P jr = (1/n) ∀ j, condition (17)
Section 3.3.2. yields (7). Solving (17) for Rk = (n /(n − 1)) + pmk and inserting
in (18) yields (8).

A.2. Positive DRS of risk type in the pooling equilibrium

5. Conclusion

In this paper, the interaction of direct and indirect risk selec- The FOC of (9) with respect to ak is
tion (DRS and IRS) has been analyzed. It has been shown that DRS,
Pk k

using measures unrelated to the benefit package, nevertheless has ∂ k (1 − PL ) ∗

L (ak )( k
− ak ) − P k = 0. (19)
= g L L
an influence on the distortions of the benefit package caused by IRS. ∂ ak
If cost for DRS is (at least to some degree) individual-specific, DRS
selectively reduces the profit per individual of the risk type it is tar- With P j = (1/n) ∀ j, this can be simplified to g ∗ (ak )( k − ak ) =
r L
geted at. Positive DRS therefore reduces and negative DRS increases n /(n − 1), which implicitly defines a positive level of ak , because
the difference in profits between the low and the high risks. Because lim g ∗ (a) → ∞. Replacing k k k

L by ( L − a ) in (17) and (18), solving

in the pooling equilibrium the degree of the distortion depends a→0+

on the difference between these profits, positive DRS reduces the (17) for Rk and substituting in (18) then yields (10).
distortion of the benefit level, while negative DRS increases it. In
the separating equilibrium, DRS can act as a substitute for IRS and A.3. Positive DRS of a signal that is correlated with risk type
thereby reduce the distortion of the benefit level. In addition, it
has been shown that the effects of DRS on type- and signal-specific The FOC of (11) with respect to ak defines a positive equilibrium
profits also have an influence on the formula for optimal risk adjust- level of ak . Solving the FOC with respect to Rk for Rk = (n /(n −
ment: The over- and underpayments have to be inflated by insurers’ 1)) + pmk + ak and inserting in the FOC with respect to mk yields
expenditures on positive DRS and reduced by their expenditures on (12).
negative DRS.
We have derived these results for a setting where negative DRS
occurs during the application process, but the main mechanisms A.4. Optimal risk adjustment without DRS
should also hold if it is targeted at the high risks who already hold

a contract with the insurer. If an insurer attracts a larger share of Solving the FOC of (13) with respect to Rk for Rk = (n /(n − 1)) +
91 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
Economics
42 (2015)
42 (2015)
81–89
81–89 91

the high risks, this will increase the cost of negative DRS, even if the
activity of risk selection and the cost associated with it occur only
later (when the insurer tries to induce the high risks to switch to
another insurer). With negative DRS, high risks are more expensive
than without DRS; anticipating these additional cost, insurers will
therefore not make their contract as attractive for this risk group
as they would without DRS. Therefore, negative DRS will increase
the distortion in such a setting as well.
pmk + RAF − (1 − )RAO and substituting in the FOC with respect to
mk (using the definitions of rs as given in Table 2) yields (14).
A.5. Optimal risk adjustment with DRS
The derivation of condition (16) is identical to the derivation of
condition (14) described in Appendix A.4; the only difference is that
RAO has to be replaced by (RAO − ak ).

Appendix B and C. Supplementary data

Acknowledgements
Supplementary data associated with this article can be found, in
I thank Friedrich Breyer and Esther Schuch for helpful comments the online version, at http://dx.doi.org/10.1016/j.jhealeco.2014.12.
and suggestions. 003
92 N. Lorenz
N. Lorenz
/ Journal
/ Journal
of Health
of Health
Economics
References
Bauhoff, S., 2012. Do health plans risk-select? An audit study on Germany’s
Social Health Insurance. Journal of Public Economics 96 (9–10),
750–759.
Bijlsma, M., Boone, J., Zwart, G., 2011. Competition leverage: how the demand side
affects optimal risk adjustment. TILEC Discussion Paper, 2011-039.
Breyer, F., Bundorf, M.K., Pauly, M.V., 2011. Health care spending risk,
health insurance, and payment to health plans. In: Pauly, M.V., McGuire,
T.G., Barros, P.P. (Eds.), Handbook of Health Economics, vol. 2. Elsevier,
Amsterdam, pp. 691–762.
Brown, J., Duggan, M., Kuziemko, I., Woolston, W., 2011. How does risk selection
respond to risk adjustment? Evidence from the Medicare Advantage Program.
NBER Working Paper 16977.
Cao, Z., McGuire, T.G., 2003. Service-level selection by HMOs in Medicare. Journal of
Health Economics 22 (6), 915–931.
Eggleston, K., 2000. Risk selection and optimal health insurance-provider payment
systems. Journal of Risk and Insurance 67 (2), 173–196.
Ellis, R.P., McGuire, T.G., 2007. Predictability and predictiveness of health care spend-
ing. Journal of Health Economics 26 (1), 25–48.
Frank, R.G., Glazer, J., McGuire, T.G., 2000. Measuring adverse selection in managed
health care. Journal of Health Economics 19 (6), 829–854.
Glazer, J., McGuire, T.G., 2000. Optimal risk adjustment in markets with adverse
selection: an application to managed care. American Economic Review 90 (4),
1055–1071.
Glazer, J., McGuire, T.G., 2002. Setting health plan premiums to ensure efficient qual-
ity in health care: minimum variance optimal risk adjustment. Journal of Public
Economics 84 (2), 153–173.
Jack, W., 2006. Optimal risk adjustment with adverse selection and spatial compe-
tition. Journal of Health Economics 25 (5), 908–926.
Lorenz, N., 2013. Adverse selection and risk adjustment under imperfect competi-
tion. Trier University Working Paper 5/13. http://ideas.repec.org/p/trr/wpaper/
201305.html
Economics
42 (2015)
42 (2015)
81–89
81–89 92
Lorenz, N., 2014. The interaction of direct and indirect risk selection. Trier University
Working Paper 12/14. http://ideas.repec.org/p/trr/wpaper/201412.html
McGuire, T.G., Glazer, J., Newhouse, J.P., Normand, S.-L., Shi, J., 2013. Integrating risk
adjustment and enrollee premiums in health plan payment. Journal of Health
Economics 32 (6), 1263–1277.
Newhouse, J.P., Price, M., Huang, J., McWilliams, J.M., Hsu, J., 2012. Steps to reduce
favorable risk selection in Medicare Advantage largely succeeded, boding well
for Health Insurance Exchanges. Health Affairs 31 (12), 2618–2628.
Olivella, P., Vera-Hernandez, M., 2010. How complex are the contracts offered by
health plans? SERIEs 1 (3), 305–323.
Rothschild, M., Stiglitz, J., 1976. Equilibrium in competitive insurance markets: an
essay on the economics of imperfect information. Quarterly Journal of Eco-
nomics 90 (4), 629–649.
Shen, Y., Ellis, R.P., 2002. How profitable is risk selection? A comparison of four risk
adjustment models. Health Economics 11 (2), 165–174.
Shi, J., 2013. Efficiency in plan choice with risk adjustment and premium discrimi-
nation in Health Insurance Exchanges. Unpublished.
Spiegel, 2011. Fragwürdige Beratung: Krankenversicherer wimmelt Senioren ab.
Spiegel, May 9, 2011. http://www.spiegel.de/wirtschaft/soziales/fragwuerdige-
beratung-krankenversicherer-wimmelt-senioren-ab-a-761384.html
Starc, A., 2014. Insurer pricing and consumer welfare: evidence from Medigap. RAND
Journal of Economics 45 (1), 198–220.
Train, K.E., 2009. Discrete Choice Methods with Simulation, 2nd ed. Cambridge Uni-
versity Press, New York.
van de Ven, W.P., Ellis, R., 2000. Risk adjustment in competitive health plan markets.
In: Culyer, A.J., Newhouse, J.P. (Eds.), Handbook of Health Economics. Elsevier,
Amsterdam, pp. 755–845.
van de Ven, W.P., van Vliet, R.C., 1992. How can we prevent cream skimming in
a competitive health insurance market? The great challenge for the 90’s. In:
Zweifel, P., Frech, H. (Eds.), Healtheconomics Worldwide. Kluwer Academic Pub-
lishers, Dordrecht, Boston, London, pp. 23–46.
Zweifel, P., Breyer, F., Kifmann, M., 2009. Health Economics, 2nd ed. Springer, Berlin,
Heidelberg.
 Journal of Health Economics 42 (2015) 90–103

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbas e

Environmental regulations on air pollution in China and their impact

on infant mortality
Shinsuke Tanaka ∗
Tufts University, United States

a r t i c l e i n f o a b s t r a c t

Article history: This study explores the impact of environmental regulations in China on infant mortality. In 1998, the
Received 27 January 2014 Chinese government imposed stringent air pollution regulations, in one of the first large-scale regulatory
Received in revised form 24 February 2015 attempts in a developing country. We find that the infant mortality rate fell by 20 percent in the treatment
Accepted 26 February 2015
cities designated as “Two Control Zones.” The greatest reduction in mortality occurred during the neonatal
Available online 6 March 2015
period, highlighting an important pathophysiologic mechanism, and was largest among infants born
to mothers with low levels of education. The finding is robust to various alternative hypotheses and
JEL classification:
specifications. Further, a falsification test using deaths from causes unrelated to air pollution supports
Q56
I18
these findings.
Q53 © 2015 Elsevier B.V. All rights reserved.
I12
O13

Keywords:
Infant mortality Air
pollution Environmental
regulation China

1. Introduction themselves to a legal framework for reducing pollution emissions.

There is little disagreement that air pollution poses a major envi-
ronmental risk to human health. Improved air quality worldwide
is correlated with the amelioration of numerous health problems,
including respiratory infections, cardiovascular diseases, and lung
cancer.1 Developing countries rank highest in air pollution world-
wide, and children under age five in these countries are considered
to be the most vulnerable population. Elevated air pollution is
generally beyond the scope of individual control and falls to the
public sector. However, environmental regulations on air pollution
are extremely contentious. This was evident at the 2009 United
Nations Climate Change Conference, also known as COP15, held in
Copenhagen, where many developing countries refused to commit
∗ Correspondence to: The Fletcher School, Tufts University, 160 Packard Avenue,
Medford, MA 02155, United States. Tel.: +1 6176274619.
E-mail address: Shinsuke.Tanaka@tufts.edu
1
A sizable literature documents health risks associated with air pollution expo-
sure. See, for example, Schwartz et al. (1996), Levy et al. (2000), Samet et al. (2000),
Chay and Greenstone (2003a), Currie and Neidell (2005), and Currie et al. (2009),
and Arceo et al. (2012).
Their opposition is largely due to the prevailing concern that the
economic costs associated with pollution abatement may outweigh
the health benefits. Accordingly, air pollution regulations are still
rare in developing countries, and whether, and to what extent,
environmental regulations on air pollution lead to health benefits
remains an important question yet to be answered.
In this paper, we examine the effect of environmental regula-
tions pertaining to air pollution on infant mortality in China. As
China’s economy continued to grow at unprecedented rates for the
last several decades, ambient air quality deteriorated to one of the
worst levels in the world due to its heavy reliance on coal-fired
energy generation. In 1998, the Chinese government imposed strin-
gent regulations on pollutant emissions from power plants, in one
of the first attempts on such a large scale in developing countries.
This so-called the Two Control Zone (TCZ) policy designated nearly
175 prefectures exceeding the nationally mandated pollution
standards as the TCZ. In these areas, the power industry, which con-
tributed more than 90 percent of air pollution, was forced to reduce
emissions and install new pollution control technologies, while also
shutting down a massive number of small, inefficient power plants.
For our purpose, the TCZ policy provides a quasi-experimental

http://dx.doi.org/10.1016/j.jhealeco.2015.02.004
0167-6296/© 2015 Elsevier B.V. All rights reserved.
 S. Tanaka / Journal of Health Economics 42 (2015) 90–103
environment, wherein the intensity of exposure to the regulations
can be defined by the TCZ regulatory status, and we are able to com-
pare changes in infant mortality rate (IMR) before and after the pol-
icy reform, between the cities assigned and not assigned as the TCZ.
To implement the analysis, we draw IMR data from the Chi-
nese Disease Surveillance Points (DSP) system that collected birth
and death registrations for 145 nationally representative sites from
1991 through 2000. IMR, defined as the number of infant deaths
under age one per 1000 live births in a given year, is available for
each DSP site by year level, linked with detailed information on
birth characteristics and parental attributes. We match this dataset
to the TCZ regulatory status assigned to individual cities, based on
the governmental report, and thereby estimate the treatment effect
of the regulations.
We find that the air pollution regulations led to significant
reductions in infant mortality. The difference-in-differences esti-
mates suggest that the regulations have led to 3.29 fewer infant
deaths per 1000 live births than would have occurred in the absence
of the regulations. This corresponds to a 20 percent reduction
in IMR. 63 percent of the reduction in infant mortality occurred
during the neonatal period, highlighting an important pathophysi-
ologic mechanism, and the greatest reduction of mortality occurred
among children born to mothers with low educational levels.
A major methodological challenge, however, is that the TCZ des-
ignation rule may not be orthogonal to unobserved characteristics
that contribute to reductions in air pollution and infant mortal-
ity. The present study conducts a number of robustness checks
and a falsification test to address this issue. First, we confirm that
the TCZ status has little association with changes in observable
covariates, assuring that there is no systematic difference in con-
current trends in observable characteristics between the TCZ and
non-TCZ cities. Although this is not a direct test of the exclusion
restriction, since it requires that TCZ status not be correlated with
trends in unobservable factors, this result leads us to believe that
the treatment effect is less likely to be confounded by differential
trends in unobservable factors as well (Altonji et al., 2005). Second,
the regression is also directly adjusted for differential pre-existing
trends in mortality, yet the estimates are essentially unaffected.
Lastly, the policy had no impact on infant deaths due to acciden-
tal causes. The absence of a causal mechanism linking air pollution
to these causes of death serves as falsification evidence, suggest-
ing that differences in access to or quality of medical services and
technologies cannot be the sources of bias. Overall, there is no evi-
dence that the estimates are driven by inappropriate identification
assumptions, leading us to believe that the treatment effect based
on the TCZ status is indeed not spurious but causal.
This study makes three major contributions to the existing liter-
ature. First, by exploiting regulation-induced changes in air quality,
it addresses a policy-relevant question: to what extent do envi-
ronmental regulations in developing countries lead to reductions
in infant mortality? Several prior studies have focused on vari-
ation in air quality induced by recession (Chay and Greenstone,
2003a), weekly fluctuations (Currie and Neidell, 2005), wildfires
(Jayachandran, 2009), or wind directions (Luechinger, 2014). Chay
and Greenstone (2003b) provide compelling evidence for the link-
age between the Clean Air Act of 1970 and infant mortality in the
U.S. It remains to be determined, however, whether, and how effec-
tively, environmental regulations can improve human health in
developing countries.2 A recent study by Greenstone and Hanna
2
Extrapolating evidence in developed countries to developing countries is dif-
ficult for a number of reasons. First and foremost importantly, we do not know
whether environmental regulations in developing countries had any impact in
reducing pollution due to weaker implemental mechanisms and an enforcement
91
(2011) examines regulations on air pollution and water pollution
in India since 1987. They find these regulations efficacious in reduc-
ing air pollution, but such reductions led to modest and statistically
insignificant reductions in infant mortality. Our study provides
an interesting contrast, finding that infant mortality significantly
responded to the environmental regulation. Further, the regulation
we focus on targeted coal for energy generation, which is the major
contributor to air pollution in many other developing countries as
well, whereas Greenstone and Hanna (2011) focus on vehicular
pollution.3 , 4 The findings in this study accordingly present relevant
estimates for the effect of environmental regulations in develop-
ing countries implementing similar policies on coal in the power
industry.
Second, the present study contributes to our understanding
of the relationship between air pollution and infant mortality at
greater concentration levels. Previous evidence is predominantly
derived from the United States or other developed countries, where
pollution is relatively low.5 Since we know little about the shape of
the dose–response relationship, it is consequently difficult to pre-
dict the marginal impact of pollution reduction in the presence of
non-linear relationship.6 Air pollution in China is one of the highest
in the world. Its total suspended particulates (TSP) level in 1995 was
four times higher than the WHO standards, and four times higher
than the level in the United States in 1970, when the Clean Air
Act was amended, as examined in Chay and Greenstone (2003b).
Thus, estimates in China provide compelling evidence applicable to
the distinctive context of developing countries where air pollution
levels are relatively high.
Third, there is extensive literature showing differential patterns
according to socioeconomic status, yet it is still an open question as
to whether air pollution also exhibits differential impact on infant
mortality (Currie and Hyson, 1999; Case et al., 2002; Jayachandran,
2009). While infants in poor countries are considered to be the
most susceptible to the effect of pollution, not only because of
high pollution levels but also because families lack the resources
or knowledge necessary to avoid exposure, the impacts may be
small if air pollution does not have first-order effect on them.7 Thus,
the present study helps identify vulnerable population in designing
policies.
The current research design has several advantages over previ-
ous studies. First, this study focuses on infants, not only because
system. Second, even if pollution is successfully reduced to some extent, infant mor-
tality may not fall if a concave relationship between mortality and pollution level
leads to low marginal pollution effect at high concentration levels. Third, magni-
tudes of impacts in reducing air pollution should be greater if people have limited
access to medical services, initially have lower health status, and/or have limited
knowledge in avoiding pollution.
3 Since the most heavily affected industry is the power industry, which was a
driving force behind China’s rapid economic growth, the findings in this study
accordingly highlight the important tradeoffs among economic growth, environ-
mental quality, and human health. See Tanaka et al. (2014) for the impact of the
environmental regulation on industrial performance.
4 A relatively smaller-scale air quality regulatory regime, targeting a different
industry, in another developing country, can be found in the Indian transportation
sector, which was mandated to use compressed national gas vehicles in Delhi during
working hours. Kumar and Foster (2007) show its effect on respiratory health.
5 Examining the pollution effect at low levels, especially levels lower than what
is often considered to be the standard, is also an interesting question in itself.
Currie and Neidell (2005) find that CO has significant impact on infant mortality
in California over the 1990s at relatively low levels.
6 For example, Arceo et al. (2012) show a non-linear relationship between CO
and health in Mexico. Evidence in developed countries may understate the impact
of pollution reduction in developing countries if the marginal impact of pollution is
higher at greater concentration levels.
7 For example, people with poor health tend to stay indoor with little exposure
to pollution. Children from rich households, who tend to have better health, may be
exposed more to pollution if they are more likely to have outside activities.
92 S. Tanaka / Journal of Health Economics 42 (2015) 90–103
they are particularly vulnerable to air pollution due to their weak
respiratory system, but because focusing on infant mortality mit-
igates complicating factors associated with adult mortality. For
example, adult deaths correlate more closely to chronic disease
conditions than to acute changes in air quality. In addition, adults
may migrate into less polluted areas. Addressing infant mortality
circumvents these issues, if not completely, because it is relatively
less difficult to identify causes of death during the first year of
birth, and because migration rates are low for pregnant women and
infants. Lastly, China is not only one of the most polluted countries
but also one of the first developing countries to regulate air pollu-
tion on such a large-scale. It is evident that China serves as a rare
research environment in which to assess the impact of environ-
mental regulations at greater concentration levels.
The rest of the paper is organized as follows. Section 2 provides
the historical background on air pollution and national air pollution
regulations in China. Section 3 describes the data and the descrip-
tive statistics. Section 4 presents the econometric framework and
its validity, and Section 5 presents empirical results. Section 6 con-
cludes.
2. Background on air pollution and regulations in China
2.1. Brief history
China is infamous for its air pollution, due to emissions from a
power sector that relies heavily on coal to generate electric power.
As the world’s largest coal producer, China possesses abundant
and relatively cheap coal, which constitutes the country’s primary
energy resource endowment, accounting for 75.5 percent of total
energy production in 1995 (National Bureau of Statistics of China,
2006). However, coal generally emits more pollutants than other
fossil fuels. As China underwent rapid economic growth, total SO2
emissions increased from 18.4 million tons in 1990 to 23.7 million
tons in 1995, and the ambient air pollution rose to levels detri-
mental to human health (State Environmental Protection Agency
[SEPA], 1996).
Fig. 1 illustrates the world distribution of TSP (Panel A) and SO2
(Panel B) concentration levels in 1995. The TSP level in Beijing, the
capital city of China, was 377 g/m3 , almost four times higher than
the WHO guideline of 90 g/m3 , and its SO2 concentration level was
90 g/m3 , almost double the WHO guideline of 50 g/m3 (WHO,
2002). SO2 is also an important precursor of acid rain. From the
1980s to the mid-1990s, the area of China experiencing acid rain
expanded by more than 1 million km2 (Yang and Schreifels, 2003).
During that decade, elevated air pollution gave rise to increas-
ing public concern about adverse impacts on human health.8
In response, the Chinese government formulated a series of
8
It is generally known that the smaller a particulate, the more detrimental it is
to health. For example, PM10 or PM2.5 , the particles with a diameter of 10 or 2.5
micrometers or less, respectively, or toxic gas, such as SO2 , are considered to be
the most hazardous because, when inhaled, these particulate matters or gas can
penetrate deep into the lungs and interfere with internal gas exchange. Further,
Laden et al. (2000) find that fine particles emitted from combustion sources (i.e.,
motor vehicles or coal combustion) have a stronger association with mortality than
those from non-combustion sources. Alternatively, SO2 becomes sulfuric acid when
it interacts with water, which is the main component of acid rain that may have a
direct or indirect impact on health. Yet, epidemiological evidence of the impact of
SO2 on mortality in developed countries is somewhat mixed. While Mendelsohn and
Orcutt (1979) show close associations between the two, SO2 is also considered a less
important determinant of mortality (Schwartz and Marcus, 1990; Nielsen and Ho,
2007). Hedley et al. (2002) is one of the few intervention studies that investigates
changes in SO2 caused by an overnight restriction on all power plants and road
vehicles in Hong Kong using fuel oil with a sulfur content of more than 0.5 percent.
They found that the intervention resulted in an immediate reduction in ambient
SO2 concentrations and a reduction in death rates particularly due to respiratory
environmental regulatory policies, the first version of which
was enacted in 1987, known as the Air Pollution Prevention and
Control Law (APPCL). This original law, however, failed to reduce
air pollution, mainly because it excluded the power sector, the
major contributor of SO2 emissions (Qian and Zhang, 1998). Even
worse, SO2 emissions continued to surge, and areas affected by
acid rain expanded.
APPCL was consequently amended in 1995. The major part of
the amendment was to include a section to regulate pollutant emis-
sions and coal combustion, particularly regarding the usage of high
sulfur-content coal, at power plants (Hao et al., 2007). Although the
1995 APPCL still had a weak enforcement mechanism and limited
efficacy, a prominent feature of the amendment was to propose a
future regional strategy, which would identify priority regions to
improve air quality and prevent the spread of acid rain.9 , 10
This was officially approved and implemented as the Two Con-
trol Zone (TCZ) policy in January 1998 (State Council, 1998). This
legislation designated prefectures exceeding nationally mandated
thresholds as either acid rain control zone or SO2 pollution control
zone.11 Based on the records in preceding years12 , prefectures were
designated as SO2 pollution control zone if;
• Average annual ambient SO2 concentrations exceeded the Class
II standard,13
• Daily average concentrations exceeded the Class III standard, or
• High SO2 emissions were recorded.14
Alternatively, prefectures were designated as acid rain control
zone if
• Average annual pH values for precipitation were less than or
equal to 4.5,
• Sulfate deposition was greater than the critical load, or
• High SO2 emissions were recorded.
In total, 175 prefectures out of 333 prefectures across 27
provinces were designated as TCZs. They accounted for 40.6 per-
cent of its population, 62.4 percent of GDP, and 58.9 percent of
the total SO2 emissions in 1995 (Hao et al., 2001). The SO2 pollu-
tion control zone was concentrated in the north due to high SO2
emissions for heating,15 whereas the acid rain control zones were
primarily in the south, where heat, humidity, and solar radiation
combine to create high atmospheric acidity. Hence, acid rain in the
and cardiovascular reasons. See also Aunan and Pan (2004) and Matus et al. (2012)
for health effects of air pollution in China.
9
Article 27 of the 1995 APPCL stipulates: “The environmental protection depart-
ment under the State Council together with relevant departments under the State
Council may, in light of the meteorological, topographical, soil and other natural
conditions, delimit the areas where acid rain has occurred or will probably occur
and areas that are seriously polluted by sulfur dioxide as acid rain control areas and
sulfur dioxide pollution control areas, subject to approval by the State Council.”
10 It is a standard practice of policy experimentation in China to implement strate-
gies in a particular region or for a set of time period, attempting to demonstrate their
effectiveness before expanding their implementation to the entire nation.
11 In this sense, the legislation can be considered to be parallel to the attainment
and nonattainment county designation by the Clean Air Acts in the United States.
12
The original document does not specify exactly which years of records they refer
to.
13 According to the Chinese National Ambient Air Quality Standards (CNAAQS)
for SO2 , Class I standard designates an annual average concentration level not
exceeding 20 g/m 3 , Class II ranges 20 g/m3 < SO2 < 60 g/m 3 , and Class III ranges
60 g/m3 < SO2 < 100 g/m 3 . Cities should meet Class II, which is considered to be
less harmful.
14
The original document does not specify the levels of SO2 emissions that are
considered to be “high.”
15
See Almond et al. (2009) for the impact of heating policy, which created a
discrepancy in air quality north and south of the Huai River.