University of Pennsylvania

E onomi s Department
November 1999

Consumption Smoothing in Island E onomies:

Can Publi Insuran e Redu e Welfare?

Orazio Attanasio

University College, London

Jos
e-V
 tor R
os-Rull

University of Pennsylvania, CEPR and NBER

ABSTRACT

In this paper we study the ee ts of ertain types of publi ompulsory insuran e arrangements for
aggregate sho ks on private allo ations in environments with limited ommitment. We show that
this type of insuran e an improve the wellbeing of private situations, but it an also deteriorate it.
We also des ribe how dierent hara teristi s of the environment ae t the role of publi insuran e.
Using data on the Mexi an PROGRESA program, we do ument the impa t that some government
programs have in rowding out private transfers.
Keywords: Consumption Smoothing, Publi Insuran e, In omplete Contra ts. JEL E60, H30,
D83.

 Ros-Rull thanks the National S ien e Foundation for Grant SBR-9309514 and the University of
Pennsylvania Resear h Foundation for their support. Thanks to Fabrizio Perri for his omments
and for his help with the ode and to Tim Besley, Ethan Ligon and Robert Holzmann. Finally,
we would like to thank Jose Gomez de Leon and Patri ia Muniz of PROGRESA for permission to
use the data and Ana Santiago and Gra iela Teruel for answering several questions about the data.
Corresponden e to Jose-V tor Ros-Rull, Department of E onomi s, 3718 Lo ust Walk, University
of Pennsylvania, Philadelphia, PA, 19104, USA. Phone 1-215-898-7767, Fax 1-215-573-4217, email
vr0je on.upenn.edu.
1 Introdu tion

In a re ent meeting between a World Bank oÆ ial and a nan e minister from a developing
ountry in whi h the provision of an in ome support s heme or safety net was being dis-
ussed the minister opposed strongly su h s heme. When questioned by the World Bank
oÆ ial about the reason for his opposition, the minister's reply indi ated the worry that
su h s hemes ould jeopardize the existen e of the support network provided by extended
families. In this paper, we onsider a model that justies these worries. In parti ular, we
onsider the impli ations of the provision of aggregate insuran e s hemes in situations where
agents annot fully insure idiosyn rati risk be ause of the presen e of enfor eability on-
straints. We show that the minister's worry was justied in that it is likely that the provision
of aggregate insuran e an rowd out private insuran e by more than the insuran e that is
publi ly provided. Furthermore, we onstru t examples in whi h the rowding out of private
insuran e auses overall welfare to de rease.
Obviously the results we provide annot be fully general, as they will depend on the
parti ular institutional features that we dis uss in our model. Our framework, however,
stresses that to evaluate the desirability and the design of various insuran e s hemes, one
has to pay attention to the way in whi h su h s hemes intera t with existing private (and
often informal) insuran e me hanisms.
We fo us on a situation in whi h individuals fa ing idiosyn rati risk an partly diversify
it by entering a ontra tual agreement with another individual. The arrangement we study,
however, annot fully a hieve the best possible allo ation within the pair be ause of the
absen e of enfor eability of ontra ts. Consumption an be smoothed between agents be ause
the agents that give up onsumption goods an expe t to get some extra onsumption in
the future, espe ially when their sho ks are bad. In other words, if agents do not omply
with the arrangement when times are relatively good by giving up some onsumption they
will be left alone in the future (autarky) to deal with the sho ks. It has been shown (
Thomas and Worrall (1988), Ligon, Thomas, and Worrall (1997), and Ko herlakota (1996)),
and we repli ate those ndings, that the amount of onsumption smoothing that an be
a hieved depends, among other things on the degree of agents' risk aversion and on the
varian e and persisten e of sho ks. Moreover, we show how a redu tion in the varian e of
agents' endowment due to sho ks that are ommon to both agents may redu e welfare. We
interpret the redu tion of this varian e omponent as the out ome of some a tuarially fair
insuran e poli y, or safety net, implemented by the government. We aggregate from pairs of
individuals into islands and from islands into a whole e onomy to ensure that it is feasible
for the government to provide su h an insuran e. The s heme we study ould be interpreted

1
as the government smoothing island level sho ks. We assume that the sho ks against whi h
insuran e is provided ae t all the individuals in the island and are fully observable by the
government and by the agents.
What we have in mind is the provision of insuran e against natural disasters, pri e
u tuations of agri ultural ommodities and so on.1 The phenomena linked to El Ni~no or
Hurri ane Mit h are obvious re ent examples of large sho ks against whi h the national
government or international organization might want to provide insuran e. However, other
phenomena and other insuran e s hemes (su h as support of agri ultural pri es) would also
fall within the type of s hemes we study in this paper. The important features we are
onsidering is that the type of sho ks the government is providing insuran e against are
aggregate and perfe tly observable.
The intuition behind our results is quite simple: while a redu tion of the varian e of the
endowment allows in prin iple more onsumption smoothing, but it also redu es the ost of
being in autarky. Sometimes the redu tion of the enfor ing power outweigh the redu tion
of the endowments' varian e. The overall ee t depends on the urvature of the utility
fun tion on the persisten e of idiosyn rati and aggregate sho ks and, on e again, on the
relative varian e of aggregate and idiosyn rati risks.
From what we have said so far, it is lear that the empiri al relevan e of our results
depends on a variety of hara teristi s that an potentially be measured. For this reason,
in the last part of the paper, we do ument the highly disruptive role of ertain government
poli ies on the realm of private transfers. We take this as eviden e of the important role
played by informal private arrangements and of its sensitivity to hanges in the environ-
ment. Hen eforth, we think that the type of welfare redu tions dis ussed goes well beyond
a theoreti al possibility.
While in our model, we onsider a relatively spe i insuran e s heme (in that is targeted
towards aggregate and observable sho ks), the sort of me hanism that lead to rowding out
of private insuran e arrangements is operative in a variety of situations and may be relevant
for many government programs. Therefore, the main lesson we learned from our exer ise,
is that government interventions do not o ur in a va uum. In addition to their dire t
ee ts there also are indire t ee ts as government programs an hange the in entives to
parti ipate into private arrangements and, more generally, the way in whi h these private
arrangements work. These indire t ee ts an be quite important as our example shows,
indu ing rowding out of private insuran e s hemes. In other words, government intervention
an break down the fragile so ial fabri that maintains some form of so ial insuran e among
related individuals. Even if one does not think that the so ial fabri is valuable by itself,
it an turn out that ertain types of government intervention, namely an attempt to insure

2
households, an have negative ee ts on the households overall insuran e possibilities. The
obvious poli y impli ation of this is that one should assess arefully the impa t of a proposed
intervention to see whether it may destroy private arrangements and the so ial fabri that
sustains it.2
A possible obje tion to our argument is that many 'safety nets', that is programs that
provide relief in ex eptional ir umstan es, su h as natural disasters, are nan ed by interna-
tional organizations it would not be ne essary to onsider the premium that an a tuarially
fair insuran e would imply in good times. Obviously, it should be stressed that a simple
transfer is very likely, espe ially if large in size, to in rease the welfare of the agents who
re eive it. However, by fo using on an a tuarially fair s heme, we want to onsider not
just the possible benets of a proposed s heme, but also its osts. Moreover, by stressing
the possible rowding out ee ts that an aggregate insuran e s heme might have, allows us
to fo us on its ineÆ ien ies and possible ways to improve it.3 That is, regardless of whether
a proposed insuran e s heme in reases or de reases welfare, the presen e of rowding out
ee ts stresses that su h a s heme might be sub-optimal and, subje t to some aveat, ould
be improved.
As a se ond possible obje tion, one might argue that the government ould devise more
ompli ated insuran e s hemes to avoid or minimize the interferen e that an aggregate
s heme might have with the fun tioning of private insuran e me hanisms. Our ndings all
for a more areful study of the details of how in ome support and other type of government
poli ies are implemented. We think that a me hanism design approa h is appropriate. Our
future work is geared towards answering the question of whether the government an de-
sign a poli y with that smoothes people's endowment, yet it does not redu e their ability
to sustain private arrangements. We are optimisti in this respe t and hope to ome up
with designs that will a tually improve beyond what the agents were doing without the
government provided insuran e.
Whether the me hanisms we stress are important or not is ultimately an empiri al issue.
Very few papers have performed empiri al exer ises to study the type of models we study.
In the last part of the paper we propose what we all 'suggestive empiri al eviden e' relating
to a large welfare program in rural Mexi o (PROGRESA) that suggests the presen e of
rowding out even if the program itself is not exa tly a publi insuran e program.
In a related but independent paper Krueger and Perri (1999), explore the properties of
progressive taxation. These taxes are not distortionary. However, they hange the relevant
individual endowment, and therefore, the value of autarky. They show that the perverse
ee t that we study in this paper may be very pervasive. Their model is dierent and they
study partial insuran e me hanisms that are private yet entralized sin e they involve all

3
agents. Moreover, they do not onsider the presen e of aggregate sho ks. As we dis uss
below, the presen e of aggregate sho ks an lead to parti ularly interesting dynami s.
The study of the substitution of private by publi insuran e is not new. Cutler and
Gruber (1996) estimate the extent of rowd-out arising from the expansions of Medi aid
to pregnant women and hildren over the 1987-1992 period and nd that approximately
50 per ent of the in rease in Medi aid overage was asso iated with a redu tion in private
insuran e overage. This type of rowding out is related to the one that we do ument using
Mexi an data, and it hints to the possible pervasiveness of the theoreti al problems that this
paper points to. S hoeni (1996) estimates whether in ome re eived from AFDC displa es
private familial assistan e in the form of ash and time help. Their ndings (displa ement is
pre isely estimated among bla ks but not whites) suggest that annual familial ash re eived is
redu ed by 17 ents per dollar in rease in AFDC benets, and time help re eived is redu ed
by 75 hours per year per $1,000 in rease in AFDC benets. Using a household survey
for Peru, Cox, Eser, and Jimenez (1998) nd that apital market imperfe tions are likely
to be an important ause of private transfers and that so ial se urity benets ' rowd out'
the in iden e of private transfers. Cox, Jimenez, and Okrasa (1997) found large transfers
among Polish families, partly targeted to those family members faring less well. They have
also found a weakening of the family networks after its transition to a market e onomy,
indi ating perhaps how this transfers substitute for other me hanisms.
The rest of the paper is organized as follows. In se tion 2 we present the model. We
onsider a parti ular form of institution: what we have labeled as the `extended family'.
Ee tively, our extended families are made of two individuals that enter a form of risk
sharing subje t to onstraint that allow for the la k of perfe t enfor eability. In this se tion
we also show how to build from a two agent e onomy to a multi-agent e onomy to be able to
interpret a redu tion in the varian e of the endowments due to aggregate sho ks as the result
of government provided insuran e. The model we present in Se tion 2 is not new: several
ontributions in the literature have onstru ted and used similar models. In Se tion 3, we
des ribe by means of an example how the model works: we dis uss the properties of the
equilibrium allo ation. In Se tion 4 we show an example where we ompare two e onomies
with dierent varian e in the endowment due to aggregate sho ks (the low varian e has
government provided insuran e) and show how the utility of the implementable allo ation is
lower in the insuran e e onomy even though the autarki and the rst best utilities are higher.
We also dis uss what are the hara teristi s of the e onomies that ae t these dierent
out omes. In Se tion 5, we dis uss some the empiri al eviden e that we have obtained from
the PROGRESA data set that do uments the rowding out of private transfers by publi
transfers. Finally, se tion 6 on ludes the paper.

4
2 The Model

While we will eventually onsider an endowment e onomy made of a large number of identi al
separate islands, we start onsidering two individual agents and a government that has
a ess to a tuarially fair insuran e. Endowments are sto hasti and there are no storage
possibilities. We think of this pair of agents as `an extended family'.
There are dierent kinds of sho ks in this e onomy. Let z denote a sho k with nite
support in Z . Furthermore, the sho k z is Markov with transition matrix z;z0 = Prob(zt+1 =
z 0 jzt = z ), and stationary distribution z .4 We label sho k z the aggregate or island sho k.
Let s 2 S denote another Markov sho k to ea h household that may be multi-valued, so
that it an in orporate both temporary and permanent elements. This sho k also has nite
support. We all sho k s the idiosyn rati or individual sho k. A ordingly, aggregate sho k
z is ommon to both agents, while this is not true for sho k idiosyn rati s. Conditional on
two onse utive realizations of the aggregate sho k,5 we write the sto hasti pro ess for s
as having transition s;z;z0;s0 = Prob(st+1 = s0 jzt+1 = z 0 ; zt = z; st = s), and un onditional
means z and s. In ea h state fz; sg agents get endowment e(z; s). We write ompa tly
  fz; sg and its transition ;;0 . We use the ompa t notation y = (z; s1 ; s2 ) and we refer
to its omponents as fz (y ); s1(y ); s2 (y )g, whi h are the aggregate sho k, and the idiosyn rati
sho k of agents 1 and 2 respe tively. We also write ompa tly the transition matrix of the
pair as y;y0 . We denote by  (y ) the stationary distribution of the sho ks.6 Moreover, the
history of sho ks up to t, is denoted by y t = fy0 ; y1;


; yt g. We use  (y t jy 1 ) to denote the
probability of history y t onditional on the initial state of the e onomy y 1 .
The government does not observe the idiosyn rati omponents of households sho ks, but
it does observe the aggregate sho k z . The government an raise taxes and make transfers.
We assume that the government has a ess to a fair insuran e s heme so there is an ex-
ante zero transfer ondition. We denote these taxes (net of transfer) by  (z ), or  () or
 (y ), depending on the ontext. However, it is understood that the tax only depends on
the aggregate state.
We assume that the agents of our model maximize the expe tations of a standard inter-
temporally separable, stri tly on ave and dierentiable utility fun tion. Future expe ted
utility is dis ounted at a rate  < 1.
(X )
E0  t u( t ) : (1)
t

If an agent were to be alone, we assume that there are no trading opportunities ex ept

5
for those involved by the government transfer, and we refer to this as autarky. Therefore the
onsumption of an agent in autarky is () = e() +  (). We write the value of the autarki
agent re ursively as
X 0

() = u [e() +  ()℄ + ;0
( ): (2)
0

If the two agents are not in autarky, they are ae ted by ea h other's idiosyn rati sho k
and their joint onsumption is restri ted by the pairwise feasibility onstraint whi h now
takes the form
1 (y ) + 2 (y ) = e1 (y ) + e2 (y ) + 2  (y ): (3)

In the absen e of enfor eability problems, agents would equate their marginal utilities in all
states of the world taking into a ount the transfers from the government.

u0 [ 1 (y )℄
= a onstant independent of y: (4)
u0 [ 2 (y )℄

While it is trivial to show7 that any optimal allo ation has to satisfy (4), theory is silent on
how the surplus is split. Repli ation arguments and equality between the agents imply that
a ompetitive equilibrium allo ation within the pair would be symmetri . In any ase, we
denote with W (y; 1) the value of the rst best that treats both types symmetri ally starting
from ea h of the possible states y .
Noti e that in su h a situation an a tuarially fair aggregate insuran e s heme would
substitute the random endowments e(z; s) by an endowment given by e(z ; s), with the same
mean and lower varian e. This insuran e s heme is welfare improving both if the agents are
able to share risk between them in the absen e of enfor eability problems and if the agents
are in autarky.
We assume that ontra ts are not enfor eable, and that the ost for any agent of breaking
an agreed arrangement is zero. Obviously, in this event the future ourse of the agreement
hanges but the agent that broke it suers no immediate ost as a result of his a tion.8
Sin e we are interested in the best allo ations a hievable, we further assume that if an
individual breaks a ontra t, he will not be able to enter any similar ontra t in the future
and will revert to autarky. We ould think of this as the most severe subgame-perfe t
punishment (see Abreu (1988)).9
Agents upon breaking the ontra t get utility
[(y )℄. This means, among other things,
that we assume that the government annot observe who broke a ontra t and who did not
in order to sele t the transfer. If this were possible, the government ould use the transfer

6
to enhan e the set of privately a hievable allo ations by redu ing the value of autarky.
The agents an engage in a mutually advantage relationship that may allow them to
smooth onsumption even without ommitment. To des ribe how this is done we draw from
Kehoe and Perri (1997) who in turn follow the re ursive approa h of Mar et and Marimon
(1992) and Mar et and Marimon (1995).10
Any allo ation for the pair should satisfy enfor ement onstraints. That means that at
ea h point in time and in every state of the world, y t, the members of the pair prefer the
allo ation they re eive to autarky. These enfor ement onstraints, therefore, take the form

1X
X
r t
 (y r jy t) u[ i (y r )℄ 
[(yt )℄: (5)
r=t y r

Let us onsider the problem of maximizing a weighted sum of utilities subje t to the resour e
onstraints and the enfor ement onstraints, that is, the problem of hoosing allo ations
f 1(yt); 2(yt)g for all yt to solve
1 X
X 1 X
X
max 1  t  (y t) u[ 1 (y t)℄ + 2  t  (y t ) u[ 2 (y t )℄ (6)
f (y )g
i
t
t=0 yt t=0 yt

subje t to (3) and (5) where 1 and 2 are non-negative initial weights. We an write the
Lagrangian as

1X
X (X
2
  (y )
t t
i u[ i (y t )℄+
t=0 yt i=1
21 39
X X X =
i ( y t ) 4  r t  (y r jy t )u[ i(y r )℄
i [(yt )℄5; (7)
i r=t yr

plus the standard terms that relate to the resour e onstraints.

Noting that  (y r jy t ) an be rewritten as  (y r ) =  (y r jy t )(y t ) we an rewrite the La-
grangian as

1 XX
X n h io
 t  (y t ) Mi (y t 1 )u[ i(y t )℄ + i (y t ) u[ i (y t )℄
i [(yt )℄ (8)
t=0 yt i

plus again the terms that refer to the feasibility onstraint. The newly introdu ed variable,

7
Mi (y t 1) is dened re ursively as Mi (y 1 ) = i and

Mi (y t ) = Mi (y t 1 ) + i (y t ) (9)

Note that at time t , the Mi (y t) 's are equal to the original weights plus the umulative sum
of the Lagrange multipliers on the enfor ement onstraint at all periods from 1 to t. The
rst order onditions that an be derived from this modied Lagrangian in lude

u0 [ 1 (y t)℄ M2 ( y t 1 ) +  (y t )
2
= 1 ) +  (y t ) ; (10)
u0 [ 2 (y t )℄ M1 ( y t 1

in addition to the omplementary sla kness onditions. The next step onsists in renormal-
izing the enfor eability multipliers by dening

i ( y t ) M2 ( y t )
'i ( y t ) = and x(y t ) = (11)
Mi ( y t ) M1 ( y t )

The virtue of this normalization is that it allows us to keep tra k only of the relative weight
x: Its transition law an be written as

[1 '1 (y t )℄
x(y t ) = x(y t 1 ) (12)
[1 '2 (y t )℄

by noting that [1 '1 (y t )℄ M (y t ) = M (y t 1 ).

We are now in a position to write this problem re ursively. To do so we dene a mapping
T from values into values, a xed point of whi h are the value fun tions that hara terize the
solution to our problem. To solve our model numeri ally, as we do in the next se tion, we
a tually follow this pro edure, that is, we iterate from a ertain initial set of value fun tions.
Su essive approximation have yielded in every ase the desired xed point. The state
variables are the urrent value of the sho k y (re all that, due to the fa t that the sho ks are
Markov, their urrent value is suÆ ient to evaluate onditional expe tations) and the urrent
value of the relative weights x. Let V = fV0 (y; x); V1(y; x); V2 (y; x)g be three fun tions one
for the planner and one for ea h of the agents, that satisfy the following property:

V0 (y; x) = V1 (y; x) + x V2 (y; x) (13)

8
The mapping T, whose xed point we are looking for, updates these three fun tions, and,
therefore, we write the updated fun tions as
T(V ) = fT0 ( V ); T1 (V ); T2 (V )g:

To dene T, we rst solve the following auxiliary problem where no in entive onstraints

are taken into a ount

X
(y; x; V) = max
1 ; 2
u( 1) + x u( 2 ) +  y;y0 V0 (y 0; x) (14)
y0

subje t to the feasibility onstraint (3), with solution i;V . Note that in this problem the
relative weight x is onstant. Next, we verify the enfor eability of the solution to (14). This
means verifying whether

X
u[ i ;V (y; x)℄ +  y;y0 Vi (y 0 ; x) 
[(y )℄ for i = 1; 2 (15)
y0

If (15) is satised, then T0 (V) = (y; x; V), and T1 (V) and T2 (V) are given by its left hand
side. It is easy to see that (15) annot be violated for both agents at the same time (just
note that autarky is a feasible allo ation). The only remaining problem is to update the
value fun tions when the onstraint is binding for one of the agents, say agent 1. In this
ase, we solve the following system of equations in f 1 ; 2 ; x0 g.

X

[(y )℄ = u( 1 ) +  y;y0 V1 (y 0; x0 ) (16)
y0

0 u0 ( 1 )
x = 0 (17)
u ( 2 )
1 + 2 = e1 (y ) + e2 (y ) + 2 (y ) (18)

With solution f 1 ; 2 ; x0 g.11 To update the value fun tions we let

X
T1 (V)(y; x) = u( 1) +  y;y0 V1 (y 0 ; x0 ) (19)
y0
X
T2 (V)(y; x) = u( 2) +  y;y0 V2 (y 0 ; x0 ) (20)
y0
T0 (V)(y; x) = T1 (V)(y; x) + x T2 (V)(y; x) (21)

9
A xed point of T, i.e. a V = T(V ), gives the value to the problem of maximizing a
weighted sum of utilities (see Mar et and Marimon (1992) and Mar et and Marimon (1995),
or Kehoe and Perri (1997) for an implementation of this approa h to international business
y les). Moreover, it also gives us a way to ompletely hara terize the properties of su h
a solution by numeri al methods. This means that for any parameterization we an tell
whether the enfor eable allo ation is autarky, the rst best or anything in between. We an
also study how the enfor eable allo ations are ae ted by hanges in the environment.
Note how dierent this type of problem is from a standard optimization problem. Note
that there is more than one relevant set of rst order onditions: binding states are repre-
sented by alternative Euler equations hara terized by the default onstraints.
One question that remains is how is this allo ation a tually implemented. Like in the
rst best, the theory is silent about how to split the surplus initially. We assume a symmetri
split. This means that the starting value for x is 1. From there a ontra t an be implemented
by a state ontingent transfer say  (y; x) that spe ies what agent 1 gives to agent 2 (it an
be negative) when the state is given by the pair fy; xg. This transfer is just the dieren e
between the endowment and the solution to the problem above. The law of motion for the
state variable x is also given by the pro edure des ribed above.
This ompletes our des ription of the model and its solution. The steps between equation
(13) and equation (19) re e t the steps of the simulation program we use below to hara -
terize the quantitative properties of some examples. In the next subse tion we des ribe how
to aggregate the model so that a redu tion of the varian e of the aggregate sho k an be
interpreted as government insuran e.

2.1 Aggregate Insuran e a ross Islands

Let the e onomy onsists of a large number (a ontinuum) of islands. Let ea h island itself
be populated by a large number (a ontinuum) of agents, all of whi h are paired with one
and only one other agent. Aggregate sho ks are ommon for all agents within an islands but
are independent a ross islands. Idiosyn rati sho ks are independent a ross agents. In this
e onomy a law of large numbers (Uhlig (1996)) an be applied twi e so that for ea h island
only the aggregate sho k determines aggregate output and so that there is no aggregate
un ertainty in the e onomy as a whole. Moreover, aggregate sho ks are observable by the
government.
In this e onomy, a small12 tax/transfer s heme  (z ) an be levied from all the agents. If

10
the transfer satises the following property

X
 (z ) z = 0 (22)
z

then the government an provide some a tuarially fair aggregate insuran e to all agents
without a ess to third parties. Moreover, if the varian e of the aggregate sho ks is not too
large relative to the endowment of agents in bad idiosyn rati states, the transfer an be made
large enough so that individual endowments net of the transfer are no longer dependent on z
and they have the same mean and lower varian e than the endowments before the transfer.
Formally,
e^(s) = e(z; s) +  (z ): (23)

3 The Model at Work

In this se tion we show, by means of an example and using numeri al solutions and simula-
tions, how the agents by themselves improve upon autarky. Next, we use a similar example
to show how the type of insuran e des ribed in the previous se tion an a tually de rease
the welfare of the two agents. We dis uss the features of the model that are more likely
to ause the provision of aggregate insuran e to intera t and interfere with the working of
private ontra ts.
Before delving into the details of the simulation it is worth dis ussing some features of
the model that an be des ribed even without the help of the numeri al omputations. First
of all, it is lear that the amount of risk sharing that an be a hieved in equilibrium depends
on the dieren e between the value fun tion within a given ontra t and that a hieved under
autarky. Therefore, whatever in reases the value of autarky de reases the amount of risk
sharing in our e onomy.
The most obvious example is the ee t of hanges in the dis ount fa tor. In reasing
the rate at whi h agents dis ount the future de reases the value of the punishment imposed
by autarky and therefore de reases the amount of risk sharing that an be a hieved in
equilibrium. Redu ing the dis ount fa tor, instead, in reases the amount of risk sharing
a hieved in equilibrium. For low enough levels of dis ounting one might be able to enfor e
rst best allo ations. In reasing the varian e of sho ks in reases the amount of risk sharing
as it makes `autarky' more painful. With CRRA preferen es, on the other hand, a shift in
the mean of the sho ks does not ae t the amount of risk sharing as it is equivalent to a
hange in units.

11
 An in rease in the persisten e of idiosyn rati sho ks makes 'lu ky' agents more relu tant
to share positive sho ks and therefore de reases the amount of risk sharing. In the limit, a
random walk sho k annot be shared at all. On the other hand, the persisten e of aggregate
sho ks does not have obvious impli ations. In the examples we study, with additive sho ks,
an in rease in the persisten e of aggregate sho ks improves risk sharing.
We next hoose a spe i example to illustrate the workings of the model. This implies
parameterizing the pro esses that generate the sto hasti endowments of our model as well
as individual preferen es. We ompute the solution of the model numeri ally. Welfare an
be omputed either dire tly from the solution or through the evaluation of the utility of long
simulations. We des ribe the properties of this solution. Next, we will show in the ontext
of another example how a redu tion of the varian e of the endowment, spe i ally, of the
part that is ommon to both agents sin e it is due to the aggregate sho ks, might redu e
welfare. Finally, we dis uss how spe i is the example.
We work with a parti ularly simple version of the model where aggregate and individual
sho ks an take only two values. The pro esses for the idiosyn rati sho ks are the same
for the two members of the extended family, although the realizations of these sho ks are
un orrelated. The example that we use has the endowment depend additively on the aggre-
gate sho k and the idiosyn rati sho k, e = z + s: Sin e there are two agents, there are three
sho ks and therefore 23 = 8 possible values of the state of the e onomy and in both ases,
the Markov pro ess that des ribes the joint pro ess fz; si g is the one des ribed above.
Utility is of the CRRA lass with risk aversion parameter  = 1:1, and the dis ount
fa tor is  = 0:85. Aggregate states are given by z 2 f1; 0:1g with persisten e of good
and bad aggregate states f0:9; 0:1g (diagonal elements of z;z0 ). Idiosyn rati individual
states are s 2 f1; 0:1g, with persisten e of good and bad individual states s;s0 = f0:7; 0:7g
(independent of the aggregate states). This e onomy, therefore, is hara terized by quite
severe u tuations both at the aggregate and at the individual level. Bad aggregate sho ks
are quite unlikely to o ur and when they do o ur are quite unlikely to persist. Bad
idiosyn rati sho ks, on the ontrary, are mu h more likely to o ur and are more persistent.
While the level of risk aversion (just above one) is quite standard, there is a fair amount of
dis ounting. Within this model, a lower degree of dis ounting would allow rst best to be
a hieved.
HERE GOES Figure 1

In Figure 1, we plot, for ea h value of the state (re all that the state onsists of the sho ks
and the relative weight that onsumer 2 gets into the problem), the onsumption fun tion
of agent 1 against (minus the log of) the state variable.13 In ea h panel, we also plot the
onsumption of agent 1 generated by a model where no ontra ts are enfor eable (`autarky')

12
and by a model where full risk sharing among the two agents is a hieved (`rst best').
The onsumption fun tion under rst best is obviously in reasing in the negative of
the state variable. It should be remembered that under rst best su h a variable is kept
onstant over time, at whatever level it happens to be. At the opposite extreme, under
autarky, onsumption of agent 1 oin ides with her endowment and is therefore independent
of the state variable.
The enfor eable onsumption fun tion oin ides over some intervals with the rst best or
with the autarky level and diers in others. When the sho ks of the two agents are the same,
the region over whi h the enfor eable allo ation oin ides with rst best is independent of
the aggregate state. However, when the idiosyn rati sho ks are dierent, that is when there
is s ope under rst best for private transfers, the interval of values of the state variable where
the enfor eable allo ation oin ides are dierent depending on the aggregate state. If the
aggregate state is bad, the enfor eable and rst best fun tion oin ide only for low values of
the state variable, when it is good, they oin ide for intermediate level. Finally, noti e that
it is possible to have situations where the relatively lu kier agent onsumes more than her
autarki level of onsumption. This happens for suÆ iently high levels of the state variable
when the aggregate state is good. SuÆ iently high level of the state variable means that the
lu kier agent has been relatively lu ky (and with a binding enfor eability onstraint) in the
past. If the aggregate state is good, the unlu ky agent is 'repaying' some of her debts even
though the realization of her idiosyn rati sho k is a bad one. In su h a situation not only
is the allo ation of resour es dierent from rst best, but the dire tion of transfers is the
opposite of that predi ted by rst best. Noti e that for this ee t to happen it is ne essary
to have an aggregate sho k, whi h has not re eived mu h attention in the literature.
HERE GOES Figure 2

On the same hord, Figure 2 shows the evolution of the negative of the log of the ratio of
marginal utilities x^ for ea h of the eight states of the e onomy under rst best and under the
enfor eable ontra t. Under rst best the state variable does not hange over time so that
the graph oin ides with the graph of the log. For the enfor eable equilibrium, this graph
des ribes the dynami of the system. When neither of the two onstraints is binding the
value fun tion is onstant and the graphs oin ides with that of the rst best. However, over
large regions of the state spa e, the behaviour of the enfor eable equilibrium diers from the
rst best. Typi ally, when the aggregate state is good, the enfor eability onstraints are not
binding only for intermediate levels of the state variable. When the aggregate state is bad,
the enfor eability onstraint is not binding for suÆ iently high levels of the state variable.
The ratio of marginal utility under autarky varies the most and re e ts the ratio of marginal
utilities in the dierent endowment points. We would represent this by a horizontal line

13
sin e it does not depend at all in the previous period ratio of marginal utilities.
The situation in whi h the relatively lu kier agent onsumes more than her endowment
and of the other agent, that is a situation in whi h not only the size but the sign of the transfer
diers from rst best, orresponds to regions where the state variable is tilted towards one
of the two agents. This represents the promises made in the past so that the favored agent
would give up some of her onsumption.
HERE GOES TABLE 1

The properties of the long run equilibrium of this example are summarized in Table 1.
The averages of the relevant variables are obtained letting the e onomy run for 20,000 periods
and averaging the resulting sample path. In the rst row, we report the average output per
apita in the e onomy. In rows 2 to 4 we report the average level of the value fun tion under
autarky, enfor eable and rst best equilibria. Obviously they are in reasing. In row 5 we
report the average absolute size of private transfers. This measures the extent of insuran e
provided in equilibrium. At about 0.1 they onstitute more than 5% of per apita output.
Finally, in row 6, we report the ratio of the varian e of individual onsumption (net of
aggregate onsumption) and the varian e of individual endowment (net of aggregate endow-
ment). This ratio is equal to zero under perfe t risk sharing (rst best) and to 1 under
autarky. In this example a number of 0.267 indi ates that there is a substantial amount of
risk sharing.

4 More Aggregate Insuran e May De rease Welfare

In this se tion we des ribe the ee ts of a simple aggregate insuran e s heme, of the kind
dis ussed above. The ee t that simple aggregate insuran e s hemes might have on welfare
and on the performan e of the e onomy des ribed by our model is omplex and depends on
a variety of fa tors. In parti ular, in addition to preferen es, the ee ts depend ru ially on
the properties of the aggregate sho ks and how they intera t with individual sho ks, on the
amount of individual risk that an be diversied in equilibrium and on the relative importan e
of aggregate and idiosyn rati sho ks. The ee ts of the introdu tion of a government
sponsored insuran e poli y an only be omputed with numeri al situations. Moreover,
mapping the features of the model into a pattern is not ompletely trivial as the various
omponents of the model intera t in a omplex way.
Introdu ing insuran e against aggregate sho ks has two ee ts. On the one hand it
redu es the varian e of aggregate sho ks and therefore in reases welfare be ause utilities are
on ave. On the other hand it redu es the amount of idiosyn rati risk that an be diversied
by enfor eable ontra ts. The reason for this is that in reasing all individual endowments in

14
`bad' states of the world makes `autarky' less unappealing and therefore the enfor eability
onstraints more likely to be binding.
The net ee t depends on whi h of these two ee ts prevails. Obviously, if we start
from a situation in whi h no risk sharing ontra t is enfor eable (`autarky' equilibrium), the
introdu tion of the aggregate insuran e s heme annot make things worse as it annot rowd
out any private insuran e. On the other hand, if in the initial situation private arrangements
are able to diversify part of idiosyn rati risk, there is potential for a substantial amount of
rowding out that might lead to a welfare redu tion.
This result is not entirely surprising. Ligon, Thomas, and Worrall (1998), for instan e,
show that in a model similar to that onsidered here, the introdu tion of storage possibilities
an lead to a redu tion in welfare. The reason for their result is the same as that onsidered
here. Giving the individual households the possibility of self insure via storage, makes
`autarky` less unappealing and therefore, via the enfor eability onstraints, rowds out some
of the private insuran e. Also, Krueger and Perri (1999) nd that the private se tor's ability
to partially ensure against sho ks diminishes with ertain lass of government sponsored
redistributive poli ies (in their ase progressive taxation).
In the rest of this sub-se tion we show by means of an example how simple aggregate
insuran e redu es welfare. Furthermore, in addition to the ee ts on welfare, we try to
quantify the amount of rowding out indu ed by the introdu tion of a simple aggregate
insuran e s heme. This is important to assess the amount of ineÆ ien y implied by a ertain
s heme, regardless of its overall welfare ee t.
In this se tion we onsider two simulations that dier somewhat from that presented
in Se tion 3. While we use the same dis ount fa tor, we de rease the oeÆ ient of risk
aversion to 0.8. The most dramati hange relative to the previous simulations, however, is
in the pro esses that generate aggregate and idiosyn rati sho ks. Aggregate states are now
z 2 f1; 0:05g with persisten e of good and bad aggregate states f0:95; 0:8g (diagonal elements
of z;z0 ), idiosyn rati individual s 2 f1; 0:015g with persisten e of good and bad individual
states s;s0 = f0:95; 0:75g (independent of the aggregate states). That is we onsider a world
in whi h bad sho ks are mu h more extreme and mu h more persistent. The ombination of
a bad aggregate and a bad idiosyn rati sho k implies a level of the endowment of 0.055 to
be ompared to a value of 2, relevant for good states.
HERE GOES Figure 3

HERE GOES Figure 4

Figures 3 and 4 are equivalent to Figures 1 and 2. Figure 3, for instan e, plots the
onsumption fun tion against the negative of the log of the state variable, the ratio of the
weights in the surrogate so ial planner's problem, x^, in ea h of the eight states of the world.

15
As before, as rst best is not enfor eable, the equilibrium allo ation of our model oin ides
with the rst best only in ertain regions of the relative weights. Outside those regions, the
rst best is no longer enfor eable. Agents an, however, still do better than autarky in some
of the states and get a small transfer bounding away their allo ations from the autarki ones.
HERE GOES TABLE 2

The rst olumn of Table 2 summarizes the properties of this version of the model.
Noti e that the enfor eable equilibrium there is relatively little risk sharing. The ratio of the
varian e of onsumption to the varian e of endowment is now 0.88, mu h loser to 1 than in
Table 1. The average size of private transfers is also quite small: only about 1% of output
per apita. The reason for this lies mainly in the persisten e of the individual sho ks. The
fa t that sho ks are so extreme would, other things being equal, allow for more risk sharing
than in the environment summarized by Table 1. However, the mu h higher persisten e of
the idiosyn rati omponent results in the impossibility of sharing this in reased risk.
In the se ond olumn of Table 2, we introdu e the mandatory insuran e s heme dis ussed
above. As the s heme is a tuarially fair, average output per apita does not hange. Average
welfare under autarky and rst best both in rease as we de rease the varian e of aggregate
sho ks. However, average welfare a hieved in the enfor eable equilibrium de reases relative
to that in olumn 1. This is aused by the rowding out of the already small private sharing.
Average private transfers go from 0.016 to 0.013. Furthermore, the ratio of the varian e of
onsumption to the varian e of endowment in reases to almost 0.9: the model gets mu h
loser to autarky.
HERE GOES Figure 5

HERE GOES Figure 6

If we provided some aggregate insuran e to the e onomy in Table 1, we would obtain a

larger amount of rowding out, but a welfare in rease rather than a de rease. This might
seem a little surprising: as in Table 1 private transfers is mu h larger and the equilibrium
mu h ' loser' to rst best, one would imagine there is mu h more s ope for rowding out.
To understand what is happening, it is useful to look at Figures 5 and 6 that plot sample
paths for the onsumption of agent 1 and for the ratio of marginal utilities for our se ond
model, whose summary statisti s are reported in Table 2. Noti e that most of the times the
enfor eable equilibrium is very lose to autarky and onsiderably dierent from the rst best.
However, there are situations in whi h there is a substantial amount of risk sharing. This
happens, for example, when both the aggregate state and the idiosyn rati state of agent 1 are
bad, while the idiosyn rati state of agent 2 is good. Su h a situation is parti ularly evident
in Figure 6, where, in su h a situation, the marginal utility of onsumption is onsiderably
smaller than that that would obtain under autarky. Even a small amount of rowding out,

16
su h as that do umented in Table 2, an be extremely detrimental in su h a situation.

4.1 How Restri tive is the Example

The model we sket hed in Se tion 2 an be extended in several dire tions. The rst, obvious
extension is to allow for the possibility of storage. As dis ussed in Ligon et al. (1997) and
in Alvarez and Jermann (1998), the problem be omes numeri ally mu h more omplex.
However, the main result obtained in this paper, that simple aggregate insuran e s hemes
are ineÆ ient and an potentially de rease welfare should go through. The only aveat one
has to bear in mind is that the presen e of storage in a model with enfor eability onstraints
might put severe limitations to the amount of idiosyn rati risk that is diversied. This is
be ause the possibility of self insuran e makes autarky mu h less unappealing than without
storage.14 When onsidering the introdu tion of aggregate insuran e, therefore, we start
from a situation in whi h there is potentially very little private insuran e to rowd out.
However, as we saw in Table 2, one an get a de rease in welfare in situations in whi h there
is very little risk sharing to start with. The important point is that risk sharing might be
happening at ru ial moments.
So far we have onsidered symmetri idiosyn rati sho ks. It might be interesting to
onsider situations in whi h individuals are hara terized by very dierent pro esses. It
would be interesting to establish both whether one gets more or less risk sharing and whether
the introdu tion of aggregate insuran e is more or less likely to result in welfare de reases
in su h a situation.

5 A suggestive empiri al analysis

While there is now a onsiderable theoreti al literature that uses models similar to the one
above, very little empiri al eviden e on their ability to des ribe a tual e onomies exists.
Moreover, as far as we know, no eviden e exists on the introdu tion of aggregate insuran e
s hemes on the fun tion of private informal arrangements. And yet the empiri al relevan e
of the sort of me hanisms we have des ribed is bound to be of ru ial importan e for the
design and operation of a large variety of government programs.
In addition to the papers of Cox and o-authors, Cutler and Gruber (1996) and S hoeni
(1996) that we mentioned in the introdu tion, there are three pie es of empiri al eviden e we
are aware of that are relevant for the models we have used. Ligon et al. (1998) using a model
similar to ours, try to t the data from Indian villages in semi-arid tropi s that have been
used in a variety of studies of risk sharing, in luding Townsend (1994). Ligon et al. (1998)
estimate the model by simulating numeri ally the solution of the enfor eable ontra t. There

17
are two problems with their approa h. First, instead of onsidering a multi-agent framework
they solve the problem of ea h household vs. the rest of the village. Se ond, they ignore
ompletely storage. Both simpli ation were done to simplify the numeri al omputations.
While the rst might not be too important in pra ti e, the se ond is mu h more serious.
Nonetheless, Ligon et al. (1998) report that the version of the model they estimate ts the
data better than a model that assumes rst best allo ations.
Foster and Rosenzweig (1999), instead, take an opposite approa h and test some simple
impli ations of models with imperfe t enfor eability. In parti ular, they test the hypothesis
that, onditional on urrent sho ks, urrent transfers are negatively related to the umulate
of past transfers. Furthermore, to take into a ount altruism, they also test the hypothesis
that this negative orrelation is stronger for transfers from non-relatives than from relatives.
This approa h onstitutes a rst important step in establishing the plausibility and empiri al
relevan e of enfor eability problems.
Finally, there exist a relatively large literature in anthropology that do uments the exis-
ten e of phenomena, sometimes dened as quasi- redit, that ould be on eivably explained
by model with limited risk sharing. A large part of this literature is summarized in some
re ent papers by Platteau (1997). Platteau himself has studied shing ommunities in South
India and Senegal (see Platteau and Abraham (1987)).
One of the reasons why the study of models with endogenously limited risk sharing is
diÆ ult is be ause of the stringent data requirements that both a stru tural and a des rip-
tive approa h imply. In parti ular, one would like to have information about the nature and
importan e of private insuran e s hemes, about the dynami environment where the house-
holds live, about the degree of prevailing risk sharing and, possibly, about the importan e
of enfor eability problems. Ideally, one would like to ompare dierent small e onomies that
dier, for instan e, in the varian e and persisten e of aggregate and idiosyn rati sho ks and
ompare the degree of risk sharing among these e onomies. The main problems to be fa ed
when attempting to gather information on these issues are the fa t that surveys are often not
targeted towards a small village or island but are instead nationally representative surveys
and the fa t that they are typi ally la k a long longitudinal dimension. These onsiderations
are even more important if one would like to evaluate spe i government programs, in that
one would want to have information on the dierent islands that dier in the a ess to this
type of s hemes.
However, more and more surveys in whi h the sampling is done at the village or island
level are be oming available. Furthermore, su h surveys are starting to in lude retrospe tive
information and ould therefore be used to measure not only the mean but possibly the
amount of variability fa ed by individual households and by ea h island.

18
 In this paper, we present some redu ed form eviden e from a re ent so ial program in-
trodu ed in Mexi o. The PROGRESA program is a large welfare initiative targeted towards
rural ommunities. The program aims at providing poor rural households with help in three
dimensions: nutrition, edu ation and health. Started at the end of 1997, it now overs about
8 million individuals in about fortyl thousands villages. The program is implemented by rst
targeting villages on the basis of a well spe ied statisti al algorithm. Su h an algorithm
onsiders the so- alled index and `degree of marginalization' of ea h ommunity as well as
the availability of ertain stru tures, su h as s hools, hospitals et in the region. For a de-
s ription of the targeting pro edure and more generally of the program see Gomez de Leon
(1998).
At the start of the program, the agen y that runs it de ided to start the olle tion of a
panel data set to evaluate its ee ts and impa t. For this purpose, the program's oÆ ials
hose 506 villages that qualied for the program in whi h a very ri h questionnaire was
administered to about 25,000 households living in these villages. Interestingly, in 186 of the
506 villages hose for the data olle tion, the implementation of the program was delayed
until November 1999. As far as we know, the 186 villages were hosen in a random fashion
and the delay in the program implementation was done only with the aim of introdu ing an
experimental feature in the evaluation of the program. Therefore, we an work with a set of
`treated' villages (in the sense that the program was implemented from the beginning) that
an be ompared to a set of otherwise `identi al' ' ontrol' villages.
As mentioned above, the program has three omponents: nutrition, health and edu ation.
The rst onsists in the provision of some vou hers, delivered to the females in the household,
that an be used to pur hase food items. This aspe t of the program is linked to the health
omponent, in the sense that the parti ipant households are entitled to the food subsidy only
if they take their hildren to health enters and hospitals for some va inations and visits.
Finally, the program oers s holarships, dierentiated by age and gender, for kids to attend
s hool. They are onditioned on s hool attendan e.
As it is lear from this brief des ription, PROGRESA in not an aggregate insuran e
s heme of the type we dis ussed in the rst part of the paper. However, the program is likely
to have the ee t of limiting the impa t of bad sho ks and, therefore, if the me hanisms
we study are operative, rowd out private insuran e for the same reasons our aggregate
insuran e s heme does. Moreover, what we want to stress is that government programs
involving transfers to households may have an ee t on the intera tions among households.
What we he k below is whether the existen e and size private transfers is ae ted by the
introdu tion of the program.
Before going into the details of the empiri al analysis, it is important to provide a few

19
words of aution. Partly be ause the program at hand is very dierent from an `aggregate
insuran e program', its welfare impli ations ould be very dierent from those we onsidered
in the theoreti al model dis ussed above. This is parti ularly so be ause the main goals of
the program are long run obje tive (human apital a umulation in parti ular) so that the
program annot be evaluated on the basis of the ee t it might have on the prevalen e and
importan e of private insuran e. The exer ise we perform is only indi ative of the importan e
that the me hanism we study might have.
Given the nature of the data, and in parti ular be ause of the absen e of a long panel
of observations on individual households, we do not estimate a stru tural model. Instead,
we provide some eviden e on the intera tion of the program itself with the existen e and
importan e of private transfers. In parti ular, we use expli itly the ' ontrol' and 'treatment'
samples to test the hypothesis that private transfers are somehow rowded out by publi
programmes. In addition, we also perform the same exer ise by looking at whether the
village re eived some form of support from the outside.
The exer ise we perform is redu ed form in nature. In parti ular, we relate the existen e
and magnitude of private transfers re eived by individual households to a number of ontrols,
in luding variables that measure the o urren e of various negative sho ks that might have
ae ted the household in the re ent past. In addition to these variables, we also onsider a
dummy for the existen e of the program. As the ontrol villages in whi h the implementation
of the program was delayed were hosen randomly, this variable is exogenous.

5.1 The data

The data available to us was olle ted in O tober 1998 by PROGRESA, the agen y that
runs the program, in ollaboration with IFPRI. In total we have data from 506 villages. Of
these, in 320 the program was started in July 1997, about one year before the interviews
we use took pla e. In the other 186 ` ontrol' villages the program will be introdu ed in
November 1999. Supposedly, these 186 villages were hosen at random. We present some
eviden e below that substantiates this laim.
Our sample in ludes a total of 23511 households, of whi h 14672 are living in `treated'
and 8839 in ` ontrol' villages. The villages are from 7 states in Mexi o: Guerrero, Hidalgo,
Mi hoa an, Queretaro, Puebla, San Luis de Potosi and Vera ruz. Some of these households
are not used in the analysis that follows be ause of missing or in onsistent information on
some of the key variables.
The data set in lude information both on re ent sho ks re eived by the households and
the nature and prevalen e of private transfers. Furthermore we have information on a variety

20
of other family and village hara teristi s. In parti ular, we use information on the following
variables.

5.1.1 Welfare indi ators and sho ks

The data set ontains additional and more detailed information on expenditure share that we
have not used. We use a omprehensive measure of onsumption obtained from the question:
`How mu h money does your household allo ate to expenditure'? We also know whether the
household is above or below the poverty line.
In terms of sho ks experien ed re ently by the household we know the number of days ea h
household member was si k in the last month and the number of work days lost be ause of
si kness by ea h earner in the last six months. Furthermore we know whether the household
was ae ted by some large sho ks. These in lude things like draughts, oods, hurri anes et .
A omplete list is given in the Data Appendix.

5.1.2 Demographi s

We have information on family size and omposition, in luding the age of ea h household
member. Furthermore we have information on o upation and labour supply behaviour of
ea h member. The survey also ontains information on household members that have left
the family, usually for migration.

5.1.3 Transfers and loans

We know the nature of the transfers re eived by ea h individual household member in the
last month. The transfers an be in money or in kind (food, lothes et .). For the monetary
transfers we also know the amount re eived by ea h household member for ea h transfer.
Moreover, we know for how long these transfers have been re eived and by whom they
were given. We also have information on all the loans re eived by ea h household member,
in luding their amount, their sour es and the reason for the loan.
There are relatively few households where more than one transfer is re eived. In what
follows, however, we onsider the total of all transfers, ex ept for remittan es re eived by
immigrants. Furthermore, we also onsider the sum of individual transfers and loans re eived
by family or friends (that is not re eived by nan ial institutions or money lenders). The
results were not greatly ae ted by the denition of loans we used.

21
5.1.4 Village information

We know whether the village has been ae ted by one of several large sho ks. The sho k
typology used in the question is the same as that used in the questions asked to the individual
households. The information is given by a village representative, not by the individual
households. We also know whether the village re eived some external support to ope with
the above mentioned sho ks. Finally we know the village index of marginalization and degree
of marginalization as well as the so- alled `deepness of poverty' in the village.

5.2 Results

In Table 3, we report the mean of some village-level variables in the ontrol and treatment
samples. All means are weighted by the sample size in ea h village. In the third olumn,
we report the p-value of the test that the two means are equal in the treatment and ontrol
villages. In parti ular, we onsider four variables from the village data set and a number
of village averages omputed from the household data set. The variables from the village
data set in lude three indexes of poverty and marginalization that are omputed by the
PROGRESA oÆ ials on the basis of well dened algorithms ( alled the marginalization
index, the degree of marginalization, the degree of poverty) and a dummy that indi ates
whether a village has been ae ted by one or more 'large sho k'. Moreover, we report
the per entage of households in a village under the poverty line, the means of the log of
onsumption expenditure, family size, age of the referen e person, the number of days of
si kness of all household members, the number of days of work lost be ause of si kness, the
value of monetary private transfers (un onditional and onditional on a positive transfer) and
the per entage of households in the village re eiving some kind of private transfer (monetary
or in kind). Finally, we report the average duration of the existing transfers.
HERE GOES TABLE 3

Most of the averages are almost identi al in the treatment and ontrol villages. In par-
ti ular, we an never reje t the hypothesis that these means are dierent in the ontrol and
treatment villages. Testing for the statisti al signi an e of the means in the two sets of
villages, we fail to reje t the null for most of the variables in the table. When we try to use
any of these (and a variety of other village averages) to 'predi t' whether a village belongs to
the treatment or ontrol group, we failed to nd any signi ant regressor. This is omforting
about the random nature of the allo ation of villages to the two groups.
The majority of these villages were ae ted by some sort of large negative sho k: in
parti ular, 83% and 82% of the treatment and ontrol village representatives reported that
the village had suered be ause of one of these sho ks. The most ommon of these sho ks

22
were draughts that ae ted about 70% of the villages onsidered.
Of the variables in Table 3, the only noti eable dieren e among treatment and ontrol
villages is in the average level of transfers ( onditional on positive transfers). In the ontrol
villages, individual transfers seem to be larger, on average, than in the 'treatment' villages.
This dieren e, although not signi ant at standard levels, attra ts the smallest p-value and
is suggestive of the results to ome.
Not very many households re eive private transfers (5.8% and 6.2% in the treatment
and ontrol villages respe tively). However, these transfers, when positive, an be quite
important in the household budget: for the households with a positive transfer and positive
food expenditure, the median ratio of the (monthly) transfer to the monthly food budget
is 0.7. The existing transfers seem to be the out ome of relatively long relationships: the
average and median duration is around 5 years, both in the ontrol and in the treatment
villages.
As anti ipated above, we onsider both some Probit models (for the indi ator that at
least a household member re eives a transfer -either monetary or in kind) and some Tobit
models (for the monetary transfers). The results for the Probit models are reported in Table
4, while those for the Tobits are reported in Table 5. Similar results were obtained for slightly
dierent denitions of transfers (for instan e in luding a variable indi ating loans re eived
by relatives and friends, or onsidering the transfers of the referen e person only).
HERE GOES TABLE 4

In both tables we start with the simplest spe i ation that in ludes only the dummy
for the ontrol villages. In interpreting these oeÆ ients, it should be remembered that the
treatment villages in lude households that do not benet dire tly from the program. This
is the right way to pro eed both be ause the program is targeted to households that are
parti ularly needy, so that it is not surprising that they re eive more transfers, and be ause
we are interested in the rowding out ee ts at the village level.
Given the results in Table 3, it is not surprising that, while the sign of this variable
is positive, it is not signi antly dierent from zero. The size of the oeÆ ient, however,
espe ially for the monetary transfers, is e onomi ally signi ant. It is therefore worthwhile
to he k whether we an obtain more pre ise estimates by in luding in the spe i ation
variables that are likely to be important determinants of transfers and are exogenous for the
individual households.
In the se ond olumn of Tables 4 and 5, we in lude a village level variable (the marginal-
ization index) and several dummies to ontrol for the sho ks that a household re eives. The
oeÆ ient on the marginalization index is onsistently negative, both in the Probit and in the
Tobit models. The sho k dummies are strongly signi ant and their oeÆ ient take, for the

23
most part, the expe ted sign, in that a sho k is usually asso iated with a larger transfer. In
this spe i ation, the signi an e of the oeÆ ient on the ontrol village's dummy in reases
in size both in the Probit and in the Tobit models. In the rst ase, the p-value asso iated to
the hypothesis that this oeÆ ient is zero is 0.21, while in the se ond is 0.075. As far as the
amount of monetary transfers is on erned, therefore, on e we ontrol for dieren es a ross
villages and for sho ks experien ed by the sample households, the oeÆ ient is marginally
signi ant, indi ating the presen e of some rowding out, at least in terms of the size of the
transfers re eived by some households.
HERE GOES TABLE 5

In olumn (3) we add to our models the number of days of si kness experien ed by the
household members and the number of days of work lost be ause of si kness by the household
earners. The oeÆ ient on the day si k is positive, while that on the days of work lost is
negative. In interpreting this oeÆ ient it should be remembered that the net ee t of a
day of work lost be ause of si kness is the sum of the two oeÆ ients, so that the net ee t
is lose to zero. This result, that is that transfers re eived are ae ted only by the days of
si kness of the non-earners, is somewhat surprising. The signi an e of the oeÆ ient on the
ontrol villages dummy marginally in reases again in both the Probit and Tobit models. The
p-value on the hypothesis of no dieren e in the inter ept between treatment and ontrol
villages is now 0.19 in the Probit and 0.064 for the Tobit model.
The introdu tion of these two variables ould be questioned if one thinks that the program
has an important health omponent. However, it is unlikely that the program, whose fo us
is on va inations and preventive medi ine, ould have an ee t on these out omes in the
short period over whi h it was implemented. Furthermore, as we saw in Table 3, there is no
indi ation that these variables were dierent in the ontrol and treatment villages.
In olumn (4) of Tables 4 and 5, we add the (log) onsumption expenditure, in an
attempt to ontrol for the size of transfers. This variable takes a onsistently negative and
signi ant oeÆ ient. As the phrasing and timing of the question is not ompletely lear, the
interpretation of this oeÆ ient is not ompletely straightforward. It might be interpreted as
indi ating that households with a relatively low average level of onsumption are targeted by
relatives with higher transfers. Or it might be interpreted as indi ating that households with
a low level of re ent onsumption expenditure have been ae ted by some kind of sho ks
and therefore re eive some transfers. On e again, and with more plausibility than for the
days of si kness, it may be argued that su h a variable is endogenous to the determination
of transfers so that its introdu tion ould introdu e some important biases.
The oeÆ ient in olumn (4) on the ontrol villages dummy is now on the verge of being
signi ant for the Probit (p-value of 0.061) and strongly signi ant for the Tobit (p-value of

24
0.025), indi ating the presen e of some rowding out indu ed by the program.
To summarize the eviden e in Tables 4 and 5, we nd some eviden e of a negative ee t
of the program on the size of monetary transfers. To un over su h eviden e, however, it is
ne essary to ondition on various types of sho ks re eived by the sample households. The
eviden e of the ee t on the existen e of any transfers (that is monetary or in kind) is mu h
weaker, even though, when one onditions on log onsumption, the oeÆ ient on the ontrol
villages is marginally signi ant.

6 Con lusions

In this paper we have analyzed the ee ts that the introdu tion of aggregate insuran e
might have in a situation in whi h individuals in a small e onomy fa e both idiosyn rati
and aggregate risk. While we allow risk sharing among members of what we all extended
families, we are interested in situations in whi h idiosyn rati risk is less than fully insured
be ause of the presen e of enfor eability problems. We show that in su h a situation, the
provision of simple insuran e s hemes against aggregate sho ks is almost surely ineÆ ient,
in that it rowds out private insuran e against idiosyn rati sho ks. While the net ee t
of su h a s heme is ambiguous, we have shown that it is possible to onstru t examples in
whi h the rowding out ee t leads to a welfare de rease.
We solve and simulate our model in the ase in whi h individual households do not have
a ess to storage te hnologies. We onje ture, however, that the zest of our results arries
through to the ase of e onomies with storage.
The important message of this paper is that government interventions do not o ur
in a va uum. They o ur in the ontext of, and intera t with, private se tor me hanism
whose nature an, in all likelihood, be altered. Simplisti government interventions might
have perni ious ee ts, resulting in a worsening of people's welfare. And even if this does
not happen, our results indi ate that in designing aggregate insuran e, one should both be
aware of its rowding out ee t and possibly think of ways to avoid them.
Two important questions arise from our resear h. First, how relevant are these types of
issues in reality? Se ond, if they are important, an one design aggregate insuran e s hemes,
or more generally a wide variety of government programs, so to avoid the ineÆ ien ies implied
by our arguments?
The answer to the rst question must be based on empiri al eviden e. The empiri al study
of these models is still at the beginning. However, some early ndings, that we mentioned
in Se tion 4, and our own eviden e indi ate that these types of models might be important.

25
The eviden e we have presented in Se tion 4 is derived from the analysis of a new data set on
a Mexi an welfare program. While the program is not an aggregate insuran e s heme of the
kind dis ussed in the theoreti al model, it shows that publi programs targeted at a variety
of goals in general intera t with private arrangements among households. The evaluation of
su h programs, along with their dire t ee ts, should also take into a ount these indire t
ee ts.
As far as the se ond question is on erned, in this paper we have only onsidered simplisti
insuran e me hanisms. However, one an think of implementing alternative s hemes that,
by indu ing the right type of intera tions between the members of the extended family, avoid
the rowding out of private arrangements that we des ribed in Se tion 3. We have left the
study of these me hanisms for future resear h.

26
 Table 1: Computed statisti s from the Model E onomy

Average output per apita

P 
y (y ) [e1 (y ) + e2 (y )℄ 1.46000
P (y)enfor eable
Average
Vi (y; 1)
value fun tion
-10.705
y

Average symmetri value fun tion (rst best)

P
y  (y ) W (y; 1) -10.698

Average
P (y)value fun tion under autarky
y f
[1 (y)℄ +
[2(y)℄g -10.739

Sample average private transfers

P
y  (y ) je (y ) +  (y ) (y; :)j
1 1 0.1024

Ratio of the varian e of individual onsumption

(net of aggregate) and the varian e of individual
endowment (halso net of aggregate)
i
P  (y) 1 (y;:)
P (y) (y;:) 2
1
P
y
n P (y)[e (y)+ (y)℄o2
y
0.2674
y
 (y) e1 (y)+ (y) y 1

Notes:  = 1:1,  = :85, z 2 f1:0; 0:1g, z1 ;z1 = :9, z2 ;z2 = :1, z2 ;z2 = :1 s 2 f1:0; 0:1g,
s1 ;s1 = :7, s2 ;s2 = :7:

27
 Table 2: Computed statisti s from the Model E onomies
No insuran e 0.001 insuran e
 (1) = 0  (1) = :001
Average output per apita
P  (y ) [e1 (y ) + e2 (y )℄ 1.64583 1.64583
y

Average enfor eable value fun tion

P
y  (y ) V (y; 1) i 4.44575 4.44558

Average sym. value fun tion (rst best)

P
y  (y ) W (y; 1) 4.47292 4.47350

Average
P (y)value fun tion under autarky
y f
[1 (y)℄ +
[2(y)℄g 4.43563 4.43692

Sample average private transfers

P
y  (y ) [e (y ) +  (y ) (y; 1)℄
1 1 0.01626 0.013780

Ratio of the varian e of individual onsumption

(net of aggregate) and the varian e of individual
endowment ( also
h net of aggregate)i
P  (y) 1 (y;1)
P (y) (y;1) 2
1
P
y
n P (y)[e (y;1)+ (y)℄o2
y
0.88376 0.89550
y
 (y) e1 (y;1)+ (y) y 1

Notes:  = :8,  = :85, z 2 f1; :05g, z1 ;z1 = :9, z2 ;z2 = :8, s 2 f1:; :015g, s1 ;s1 = :95,
s1 ;s1 = :75,

28
 Table 3: Des riptive statisti s in ontrol and treatment villages
treatment villages ontrol villages p-value
of dieren e

degree of poverty 2.550 2.474 0.345

marginalization index 0.405 0.447 0.536
degree of marginalization 4.643 4.638 0.917
% of poor households 0.535 0.511 0.224
average log onsumption 4.834 4.848 0.613
family size 5.496 5.554 0.350
days of si kness 3.856 3.850 0.979
days of work lost 0.959 1.057 0.293
frequen y of village sho ks 0.835 0.845 0.762
age of referen e person 47.19 47.59 0.989
value of monetary 27.32 36.18 0.167
private transfers
value of monetary private 636.40 772.67 0.277
transfers ( ondit. on pos.)
per entage of h.h. re eiving 0.058 0.062 0.592
a transfer (monet. or in kind)
time sin e transfer started 58.44 54.95 0.707
(in months)
Number of villages 320 186 -

29
Table 4: Probit models for individual monetary and in-kind transfers

(1) (2) (3) (4)

onstant -1.567 -1.611 -1.665 -0.152

(0.017) (0.072) (0.073) (0.123)
ontrol villages 0.025 0.034 0.036 0.052
(0.027) (0.027) (0.027) (0.028)
marg. index - -0.249 -0.234 -0.300
- (0.021) (0.021) (0.022)

days si k - - 0.013 0.013

- (0.001) (0.001)

days of work lost - - -0.009 -0.009

- - (0.003) (0.003)

log( onsumption) - - - -0.319

- - - (0.021)
sho ks dummies no yes yes yes

Standard errors in parentheses.

Number of observations: 23306

30
Table 5: Tobit models for individual monetary transfers

(1) (2) (3) (4)

onstant -4487.6 -4604.7 -4716.6 -2193.7

(125.8) (241.9) (243.4) (356.5)

ontrol villages 119.0 135.0 141.0 167.5

(75.54) (75.88) (75.98) 76.6

marg. index - -556.4 -522.2 -628.3

(59.13) (59.11) (61.75)

days si k - - 29.84 29.51

(3.66) (3.69)
days of work lost - - -29.81 -28.33
(8.81) (8.72)

log( onsumption) - - - -528.78

(61.03)
sho ks dummies no yes yes yes

Standard errors in parentheses.

Number of observations: 23306
of whi h 1038 un ensored

31
Referen es

Abreu, D. 1988. On the Theory of Innite Repeated Games with Dis ounting. E onometri a,
56, 383{96.

Alvarez, F. and Jermann, U. 1998. Asset Pri ing when Risk Sharing is Limited by Default.
Unpublished Manus ript, University of Pennsylvania.
Casta~neda, A., Daz-Gimenez, J., and Ros-Rull, J.-V. 1998. Exploring the In ome Distri-
bution Business Cy le Dynami s. Journal of Monetary E onomi s, 42 (1).
Cox, D., Eser, Z., and Jimenez, E. 1998. Motives for Private Transfers Over the Life Cy le:
An Analyti al Framework and Eviden e for Peru. Journal of Development E onomi s,
55 (1), 57{80.

Cox, D., Jimenez, E., and Okrasa, W. 1997. Family Safety Nets and E onomi Transition: A
Study of Worker Households in Poland.. Review of In ome and Wealth, 43 (2), 121{209.
Cutler, D. M. and Gruber, J. 1996. Does Publi Insuran e Crowd Out Private Insuran e?.
Quarterly Journal of E onomi s, 111 (2), 391{430.

Foster, A. and Rosenzweig, M. 1999. Imperfe t Commitment, Altruism, and the Family;
Eviden e from Transfer Behavior in Low-In ome Rural Areas. Mimeo, University of
Pennsylvania.
Gomez de Leon, J. 1998. Applying Dis riminant Analysis for the Sele tion of Bene iaries:
the Case of Progresa in Mexi o. Mimeo. Paper presented at the First Annual Meeting
of the LACEA/IDB/WorldBank Network on Inequality and Poverty, Buenos Aires.
Kehoe, P. and Perri, F. 1997. International Business Cy les with Endogenous In omplete
Marjets. Working Paper, Federal Reserve bank of Minnepolis.
Ko herlakota, N. R. 1996. Impli ations of EÆ ient Risk Sharing without Commitment.
Review of E onomi Studies, 63 (4), 595{609.

Krueger, D. and Perri, F. 1999. Risk Sharing: Private Insuran e Markets or Redistributive
Taxes?. Working Paper, Federal Reserve bank of Minnepolis.
Ligon, E., Thomas, J. P., and Worrall, T. 1997. Informal Insuran e Arrangements in Village
E onomies. Mimeo.
Ligon, E., Thomas, J. P., and Worrall, T. 1998. Mutual Insuran e, Individual Savings and
Limited Commitment. Mimeo.
Mar et, A. and Marimon, R. 1992. Communi ation, Commitment and Growth. Journal of
E onomi Theory, 58 (2), 219{49.

Mar et, A. and Marimon, R. 1995. Re ursive Contra ts. Unpublished manus ript, Univer-
sitat Pompeu Fabra.

32
Platteau, J. P. 1997. Mutual Insuran e as an Elusive Con ept in Traditional Rural Commu-
nities. Journal of Development Studies, 33, 764{96.
Platteau, J. P. and Abraham, A. 1987. An Inquiry into Quasi-Credit Contra ts: The Role of
Re ipro al Credit and Interlinked Deals in Small-s ale Fishing Communities. Journal
of Development Studies, 23, 461{90.

S hoeni, R. F. 1996. Does Aid to Families with Dependent Children Displa e Familial
Assistan e?. Rand Working Paper, DRU-1453-RC.
Thomas, J. and Worrall, T. 1988. Self-Enfor ing Wage Contra ts. Review of E onomi

Studies, 55, 541{54.

Townsend, R. M. 1994. Risk and Insuran e in Village India. ,

E onometri a 62 (3), 539{91.
Uhlig, H. 1996. A Law of Large Numbers for Large E onomies. E onomi Theory , 8 (1),
41{50.

33
Notes

1 In the Mexi an data set that we study, about 82% of the 506 villages in luded in the
sample are ae ted, a ording to the village authorities, by some sort of aggregate sho ks,
su h as draught (the most ommon), res, hurri anes et ., in the year pre eding the survey.
2 More drasti ally interpreted, our example ould suggest, as a referee kindly pointed out,
that reating intrinsi aggregate un ertainty ould be welfare improving.
3 Furthermore, a naive and areless insuran e s heme in some ir umstan es might bring
about a redu tion in welfare by simply over{insuring the agents in the e onomy.
4 There are simple onditions that we assume and that guarantee that the stationary
distribution exists, is unique and is the limit for any initial ondition.
5 See Casta~neda, Daz-Gimenez, and Ros-Rull (1998) for details about the modelization
of joint aggregate and idiosyn rati sho ks.
6 We make suÆ ient assumptions on the 0 s to ensure that there is a unique stationary
distribution and no y li ally moving subsets.
7 It just follows from stri t on avity and the possibility of transferring resour es a ross
dates and states.
8 Alternatively, we ould hara terize the degree of enfor eability of ontra ts between the
two agents by a fun tion P (y ). This fun tion denotes the ost for an individual of breaking
the agreed arrangement. When P (y ) = 1 we are in the standard perfe t enfor ement ase.
When P (y ) = 0 we have the ase studied here. This fun tion ould be used to study spe ial
institutions su h as the family where ertain so ial a tivities an be used to in rease the
osts of breaking the agreement. This spe i ation is parti ularly well suited for empiri al
work.
9 Note that the allo ations attained through the enfor ement indu ed by the fear of the
reversion to autarky (trigger strategies) are not renegotiation proof.
10 The hara terization of the optimal ontra ts in a model with imperfe t enfor eability is
stated in dierent terms in Ligon et al. (1997), Ligon et al. (1998) and Alvarez and Jermann
(1998). We nd the approa h that keeps tra k of the urrent ratio of utility weights is both
more transparent and omputationally easier.
11 There will typi ally be only one solution given the monotoni ity of all the fun tions
involved.
12 The requirement of smallness is required so that all agents have to be able to pay it
regardless of their unobservable idiosyn rati sho k.
13 Negative so that it is in reasing, log so that it is symmetri around zero. There is no
loss of generality in this.
14 As mentioned above, Ligon et al. (1997) show that the presen e of storage might de-
rease welfare through the same me hanism. This possibility arises in our model with the
introdu tion of aggregate insuran e.

34
Figure 1:

35
Figure 2:

36
Figure 3:

37
Figure 4:

38
Figure 5:

39
Figure 6:

40