lost in the mail: a field experiment on crime

LOST IN THE MAIL: A FIELD EXPERIMENT ON CRIME

MARCO CASTILLO, RAGAN PETRIE, MAXIMO TORERO and ANGELINO VICEISZA∗

Stealing, shirking, and opportunistic behavior in general can create barriers tothe development of markets. The costs associated with such behavior are shared byboth firms and individuals and can be large enough to even prevent the initiationof trade. Measurement of these costs is difficult because information is not availablefor transactions that fail to occur. We use a field experiment to identify opportunisticcrime in a task that is important and relevant for trade: the delivery of mail. We subtlymanipulate the content and information available in mail sent to households acrossneighborhoods that vary by income, and detected high levels of shirking and stealing.Eighteen percent of the mail never arrived at its destination, and significantly morewas lost if there was even a slight hint of something additional inside the envelope.Our results demonstrate the importance of transaction costs created by crime and thatnot all populations are equally affected. Middle-income neighborhoods suffer the most.(JEL C93, K42, H41, L87, O21)

I. INTRODUCTION

Crime and opportunistic behavior, e.g., shirk-ing on the job or pilfering from the office whenno one is looking, impose high costs on firmsand individuals. Even the time spent in pro-tecting oneself from theft or others trying totake advantage of a situation can be onerous.While the costs associated with each event maynot be large, and may be compensated for bya large number of positive transactions, small

*We thank David Solis for conducting the follow-upsurvey, Cesar Ciudad for coordinating the mail recipients,and Maribel Elias for creating the GIS maps. Seminar par-ticipants at University of Maryland, Iowa State Univer-sity, ICES-George Mason University, Georgia Institute ofTechnology, Virginia Commonwealth University, the WorldBank, the Workshop on Economics Experiments in Develop-ing Countries at CIRANO in Montreal, and the North Amer-ican Economic Science Association Meetings gave helpfulcomments.Castillo: Associate Professor, Interdisciplinary Center for

Economic Science (ICES) and Department of Eco-nomics, George Mason University, Fairfax, VA 22030.Phone 1-703-993-4238, Fax 1-703-993-4831, [email protected]

Petrie: Associate Professor, Interdisciplinary Center forEconomic Science (ICES) and Department of Eco-nomics, George Mason University, Fairfax, VA 22030.Phone 1-703-993-4842, Fax 1-703-993-4831, [email protected]

Torero: Division Director, Markets, Trade and Institutions,International Food Policy Research Institute, Washing-ton, DC 20006. Phone 1-202-862-5600, Fax 1-202-467-4439, E-mail [email protected]

Viceisza: Assistant Professor, Department of Economics,Spelman College, Atlanta, GA 30314. Phone 1-404-270-6055, E-mail [email protected]

losses can add up. Indeed, some markets mightnot even emerge because losses eat up too muchof potential profits. This is especially importantin developing countries where markets are weakor missing. Measuring these types of costs isdifficult because observational data is not avail-able for missed transactions (those that had beena priori assessed to be too risky to enter into).Costs due to corruption, such as at ports, bor-ders, and police stops, have been estimated, butthese measures might reflect local monopolypower (Olken and Barron 2009) rather than thedistribution of costs faced by the population atlarge. In this paper, we address these issues witha field experiment. We examine how an activitythat is important for trade, the delivery of mail,is affected by crime, who in the population suf-fers the most, and the underlying motivations.

We show that the costs of crime are nontriv-ial and that not only shirking but strategic crimeexplain their existence. While sending some-thing in the mail would seem to be an ordinarytask that need not be given much thought, thereliability of delivery has important effects onthe ability of a firm or an individual to safelytransport goods at a low cost. Our experimentsuggests that engaging in exchange in such anenvironment can be costly and that certain cues

ABBREVIATION

OLS: Ordinary Least Squares

285

Economic Inquiry(ISSN 0095-2583)Vol. 52, No. 1, January 2014, 285–303

doi:10.1111/ecin.12046Online Early publication October 17, 2013© 2013 Western Economic Association International

286 ECONOMIC INQUIRY

can make mail more likely to be “lost.” Clearly,when these crimes of opportunity are part ofdaily life, trade is less efficient. Importantly, weshow that these costs are not distributed equally.Middle-class households are disproportionallyaffected.

Our empirical strategy is novel and simple,and it allows us to measure the probability thata piece of mail arrives at its destination, evenin areas where the use of mail services is low.Our design provides counterfactual informationon what would have happened if the servicehad been used. We send identical envelopes todifferent households in Lima, Peru, from twoAmerican cities and record arrivals. Peru is aninteresting case study because it is representativeof middle-level developing countries strugglingto create market-based institutions and inte-grate with the global economy. The experimentincludes a large population of volunteer house-holds across neighborhoods of different socioe-conomic backgrounds. To better understand themotivation behind the commission of crime, wemanipulate the contents and the sender of themail. In particular, every household was sentfour envelopes over the course of a year. Twoenvelopes had a sender with a foreign nameand two had the last name of the sender andrecipient matched (to indicate the letter camefrom a family member). Finally, one of each ofthe two envelopes contained something insidethe enclosed card (a small amount of money)that could not be easily detected without carefulattention.1 The other envelope contained just theenclosed card. All these modifications were assubtle as possible, and the order in which eachdifferent envelope was sent was random.

Our design allows us to develop a behav-ioral measure of crime. We can directly measurecrime and its differential impacts. Because wesend mail across all neighborhoods, we can mea-sure the level of crime even in places wherethe population may not normally use these ser-vices for fear of mail not arriving. Specifically,the design identifies whether crime has occurredor not, which segments of the population aremore likely to suffer from crime, and whetherthis conforms with economic rationality broadlyunderstood.2

1. Sending money through the Peruvian postal serviceand the U.S. Postal Service is not illegal, although it is notadvised.

2. Fried, Lagunes, and Venkataramani (2010) look atbribery by traffic police and how it varies by the incomeclass of drivers.

We concentrate on the delivery of mail forseveral reasons. First, the existence of a reli-able mail sector is considered to be instrumen-tal in the growth of electronic trade (WorldBank 2009). Second, mail services are used byall segments of the population, both rich andpoor. Third, as we will describe, mail deliveryis amenable to field experimentation with littleor no intrusion. This is important as the studyof crime could be constrained by ethical con-siderations and measurement problems. Fourth,mail delivery is a highly decentralized activ-ity and likely suffers from moral hazard prob-lems regardless of firm ownership. For instance,sources of lost noncertified mail are nearlyimpossible to detect. Fifth, in our study environ-ment, mail service, which is normally providedby a public firm, is done by a private entity.Finally, crime in the mail sector is expectedto be highly correlated with the expected gainsand losses of committing a crime and muchless with social pathologies. That is, it is acrime of opportunity that can help us understandeconomic motivations.

By manipulating the information made avail-able to the person handling the mail, we cantest several hypotheses behind the commissionof crime. First, mail can be lost because thecost of delivery is larger than the cost of beingcaught shirking. Lost mail might be a reflectionof apathy rather than crime. Therefore, compar-ing rates of lost mail containing money withthose containing no money permits us to detectif crime, rather than apathy, is taking place.Second, if those handling mail behave strate-gically, one would expect that they will makeuse of information on the socioeconomic char-acteristics of the recipient and the social distancebetween the recipient and the sender. Therefore,comparing similar pieces of mail across sub-groups can potentially reveal the expectationsof those handling the mail. For instance, if aletter from a family member is more likely tocontain something of value (i.e., money), thenletters from family members would be lost at ahigher rate.

The experiments show that the mail servicein Peru is highly inefficient. The overall rateof mail lost is 18%.3 The loss rate, however,

3. Compared to the less than 0.5% of mail reported lostin the United States or the United Kingdom, this is verylarge. Note that loss rates in the United States and the UnitedKingdom are for reported mail lost. This will underestimatethe problem if not all lost mail is reported. Our experimentalmeasure is for all mail lost that should have arrived at adestination.

CASTILLO ET AL.: LOST IN THE MAIL 287

hides the fact that mail containing money islost 21% of the time while mail containing nomoney is lost 15% of the time. That is, we findevidence of shirking as well as crime. Also,crime is targeted primarily at letters comingfrom family members, and the quality of serviceis not independent of socioeconomic status. Mailis lost at the same rate (roughly 18%), whetherit contains money or not, when sent to a poorneighborhood. When sent to a more affluentneighborhood, however, mail without moneyis lost only 10% of the time and mail withmoney is lost 17% of the time. Households inmiddle-income neighborhoods have the highestloss rates. This suggests two things. Crime isstrategic, and not happening randomly, and lossis occurring within Peru rather than the UnitedStates. We ran several robustness checks andconfirmed that our main results still hold.

Our research makes several important con-tributions. First, it shows that an ordinary task(sending a piece of mail) that is important forinexpensively transporting goods or informationhas hidden costs. This implies that there stillremain barriers to expand commerce that relieson the mail sector to transport goods and ser-vices. Second, it highlights the problems thatdeveloping countries face when trying to solveinefficiencies through privatization of public ser-vices. Private firms suffer the same asymmet-ric information that state-owned enterprises do.Third, our research presents new evidence thatcrime is not shared equally. The middle classseem to be taxed more heavily. Finally, weconfirm that crime is strategic and depends onexpectations and the probability of being caught.

Our results highlight several literatures. Weexamine crime in a traditionally public sec-tor activity: the delivery of mail (however,in our setting, this service is privatized). Inmany developing countries, corrupt behavior inthe provision of public services is not onlywidespread but can also create important ineffi-ciencies and inequities (Hunt and Laszlo 2008;Reinikka and Svensson 2005), such as in obtain-ing a driver’s license (Bertrand et al. 2007),state asset sales and taxes (Fisman and Wang2010; Fisman and Wei 2004), port transactions(Sequeira and Djankov 2010), and traffic tickets(Fried, Lagunes, and Venkataramani 2010). Pub-lic services may be privatized, yet our resultssuggest that a private firm can face the samemoral hazard problems that a public firm would,as well as significant levels of corrupt and strate-gic behavior.

The findings also touch on the large literatureon crime. Theory and casual observation wouldsuggest that people may be affected differen-tially by crime and that crime may be strategic(Becker 1968). Therefore, neither the participa-tion in illegal activities nor the diseconomiescaused by crime are expected to be uniformlydistributed across the population. Indeed, empir-ical studies have shown that crime negativelyaffects economic activity (see Alesina and Per-otti 1996; Abadie and Gardezabal 2003; Barro1991; Gaviria 2002; Pshisva and Suarez 2010,for some examples). And, there is ample evi-dence that people respond to economic incen-tives when committing crimes.4 Deterrence, therisk of being caught, and social norms all seemto be important factors in deciding whether tocommit a crime or not. Our results add to thisliterature by confirming crime to be strategic inthis setting and also providing evidence that notall income groups are equally affected.

The paper is organized as follows. The nextsection presents a model of crime to providea framework for our hypotheses. Section IIIpresents a description of the Peruvian postalsystem. Section IV presents the experimentaldesign and Section V the results. In Section VI,we run robustness checks on our results, and inSection VII, we check that our results are notdue to response bias. Section VIII concludes.

II. A MODEL OF CRIME

Our experiment allows us to look at crimein equilibrium. By varying the content andinformation available on each piece of mail, wecan better understand the strategy of crime andwho is affected the most by it. Below we presenta model of crime (borrowed heavily from Anwarand Fang’s 2006 model of discrimination). Theaim of this section is to derive comparativestatistics from the model to build our hypotheseson how the experimental manipulations willaffect mail loss.

We start from the premise that the incidenceof crime in equilibrium is a function of theprobability of being caught stealing and the

4. For examples of empirical and experimental work,see Erlich (1973), Levitt (1997), Duggan and Levitt (2002),Glaeser, Sacerdote, and Scheinkman (1996), Jacob andLefgren (2003), Di Tella and Schargrodsky (2003, 2004),Olken (2007), Fisman and Miguel (2007), Reinikka andSvensson (2004), Armantier and Boly (2008), Hsieh andMoretti (2006), Olken and Barron (2009), Corman andMocan (2005), among others.


probability that the victim has something ofvalue to steal.5 In the context of the mail sector,this implies that mail loss will be a function ofthe probability that the postal worker will befired or punished if caught stealing mail and theprobability, or the expectation, that the senderof a piece of mail includes something of value.

We assume that individuals choose the mailthey send from a set of possible types ofmail, x ∈ {1, 2, . . . , K}. x represents the phys-ical characteristics of the mail, i.e., the thick-ness of the envelope or whether the mail is aletter, greeting card, manila envelope, or pack-age. Individuals also have to decide whether toplace something of value inside the mail. Thisis represented by a binary variable d that takesthe value of 1 if something of value is in themail and 0 otherwise. To model the expecta-tion that there is something of value (or the riskto the sender of mailing something of value),we assume there is a signal θ associated witheach piece of mail. The mailman observes θ.For instance, mail with valuables might requiremore packaging or might have some irregular-ities that are observable to a careful handler.Signal θ is distributed according to a continuousdensity function f (θ|x, d) over interval [0, 1].

To address the fact that pieces of mail con-taining valuables might be more likely to befound out, we assume that the ratio of f (θ|x, 1)to f (θ|x, 0) is increasing in θ. As enoughinspection might uncover whether a piece ofmail has something of value or not, we assumethat limθ→1 [f (θ|x, 1)]/[f (θ|x, 0)] = ∞ andlimθ→0 [f (θ|x, 1)]/[f (θ|x, 0)] = 0. Finally, wedenote by z the social (nonphysical) character-istics associated with the mail. For instance,z might be the neighborhood of the recipient,where the mail was sent from or the relation-ship between the sender and the recipient. Aperson with social characteristic z sends a pieceof mail with characteristic x and something ofvalue with probability πzx and without valuableswith probability 1 − πzx . Note that the same

5. Becker (1968) and Erlich (1973) present detailedmodels of decision making by individuals considering com-mitting a crime. In the context of their models, the interac-tion between a mailman and a customer can be thought ofas a zero-sum game. If a person sends valuables with prob-ability one and there is moral hazard, mail will be certainlystolen. If the mailman never steals, a customer might feelsafe sending valuables in the mail. In equilibrium, one wouldexpect that those customers that have a larger marginal ben-efit of using the mail to send valuables will face a largeraverage level of crime.

distribution of signals will be interpreted dif-ferently depending on πzx . In other words, thesame evidence might be taken more seriously ina population where πzx is higher.

The mailman decides whether to steal ordeliver the piece of mail. The expected returnof stealing a piece of mail is

P(x, z) Pr(d = 1|x, z, θ) − q(x, z)t.(1)

P(x, z) represents the expected value of thecontent of the mail of type x with characteristicz. q(x, z) represents the likelihood that themailman is caught stealing. This might happenbecause the post office monitors more closelymail of type x with characteristic z or becausethose with characteristic z are more likely tofollow up on the mail sent. The punishment amailman faces when caught stealing or found tobe at fault is t > 0.

We can use Bayes’ rule to determine thatthe probability that a piece of mail containssomething of value when signal θ is observedis:

Pr(d = 1|x, z, θ) = [f (θ|x, 1)πzx]/(2)

f (θ|x, 1)πzx + f (θ|x, 0)(1 − πzx).

The expected return to delivering the mail isnormalized to 0. Given the assumptions, wehave that there is a unique value θ(x, z) suchthat the expected payoff of committing a crime(Equation (1)) equals the payoffs of deliveringthe mail, or P(x, z) Pr(d = 1|x, z,θ(x, z)) −q(x, z)t = 0. As Pr(d = 1|x, z, θ) is increasingin θ, we have a risk that a neutral mailman willsteal the mail if he observes a signal θ such thatθ ≥θ(x, z).

Figure 1 shows the potential effect of achange in x on the behavior of the mailman.6

If a change in x affects only the distribution ofsignals, we would expect that the conditionalprobability that a piece of mail contains some-thing of value will shift for all values of θ. Thiscan be seen in the figure where Pr(d = 1|θ, x, z)shifts up as x changes to x ′. All things constant,this change in expectations will affect the criti-cal value θ(x, z) and, therefore, the probabilityof search (i.e., the likelihood a mailman tamperswith the mail). In the figure, the critical valuechanges from that associated with point A tothat associated with point B, and the probabilityof search increases.

6. The figure graphs the two components of Equation(1): the probability that a piece of mail contains somethingof value and the probability of being caught stealing dividedby the expected value of the piece of mail.


FIGURE 1Incentives to Search

Pr(d = 1|θ, x′, z)

Pr(d = 1|θ, x, z)

q(x′,z)tP(x′,z)

q(x,z)tP(x,z)

B

C

A

SearchNo search

Signal θ

Ben

efit

and

cost

of s

earc

h

0.0

0.5

1.0

0.5 1.0

However, it is possible that a particular pieceof mail of type x is monitored more closelyas well (e.g., certified mail) and so also affectsq(x, z). In this case, it is possible that enhancedincentives to commit a crime are completelyoffset by a change in the probability of beingcaught committing the crime (e.g., moving frompoint A to B to C). For instance, a piece ofmail sent by a family member is more likelyto contain something of value (a shift out ofPr(d = 1|θ, x, z)), but it is also more likelyto be monitored (an increase in q(x, z)). Ourexperimental design includes some changes in xthat are unlikely to affect q(x, z) (e.g., slightlyincreasing the thickness of the mail by placingmoney inside).

Increases in πzx will also shift the curvePr(d = 1|x, z, θ) upwards. Changes in πzx mightreflect the options available to those with char-acteristics z. Those who have safer ways tosend valuables will be less likely to send themthrough the mail. For example, wealthier house-holds might be able to afford to send itemsthrough costlier, yet more secure, services.

This model is partial in the sense that weassume that πzx and q(z, x) are determinedexogenously. While the model can be modifiedto make these variables part of an equilibrium,our experimental design does not require forthem to be exogenous. Some of the treatmentsare subtle enough so as to consider the effect tobe minimal. Finally, without explicitly modelingq(x, z), we cannot determine if areas with higherlevels of monitoring are also areas where more

crime is detected. For instance, a monopolistfirm might find it profitable to secure betterservices in some areas by increasing the costof committing a crime.

In our experiment, we manipulate informa-tion that we expect to affect the benefits andcosts of search, as we have outlined in the modelabove. For example, by making the envelopeslightly thicker with something inserted inside(changing x), we expect mail loss to increaseas postal workers think the envelope signalssomething of value. Similarly, by decreasingthe social distance between the sender and therecipient, such as having the sender be a fam-ily member (changing z), we expect mail lossto increase. By sending mail to different neigh-borhoods, we can see if mail loss decreasesin wealthier neighborhoods, relative to poorerneighborhoods, because we would expect thatthe rich are more likely to complain if mail doesnot arrive (probability of being caught rises).The effects, however, may be nonlinear. Whilewe expect service to be better as neighborhoodincome rises, this will interact with the expec-tation that recipients will receive something ofvalue. People who live in wealthier neighbor-hoods may be more likely to receive somethingof value, but they may also be more likely topay a higher price for more secure services(and potentially avoid the post office altogether).Recipients in poor neighborhoods may not beexpected to receive anything of value, so mail-men may not bother to look carefully.7

In sum, our hypotheses are as follows: (1) weexpect mail loss to increase as the probability ofbeing caught stealing declines and the probabil-ity of valuables increases; (2) the relationshipacross neighborhoods may be nonlinear becausethe probability that something of value is beingsent will depend on the recipient’s likelihood ofreceiving something of value, the likelihood thatthe recipient has alternative means for receivingvaluables, and the probability of being caught.

III. MAIL IN PERU

To better understand how and where crimeand opportunistic behavior may be happeningwhen mail is sent from the United States to Peru,

7. A model of crime has enough degrees of freedom asto make it difficult to determine a priori the effect of incomeon public services crime. Our experiments provide data thatmight allow us to determine regularities, but not necessarilyprove or disprove a theory.


in this section we describe the Peruvian postalsystem and how mail is delivered.

The postal system in Peru is a private conces-sion of the Peruvian government and was priva-tized in 1991 (Legislative Decree 685, 1991).The company does not have exclusive rightsto deliver letters, as is the case in the UnitedStates, but practically speaking, they are the onlyprovider of low-cost, nonpackage mail servicein Peru. Indeed, they are bound by law to pro-vide service to isolated areas of the country.There are alternative means for sending mail,including certified services offered by the postoffice, but they are very expensive, costing about100%–200% more, depending on the destina-tion. Mail tracking services are ten times moreexpensive than regular mail and the cost (in dol-lars) is higher than the cost of a similar servicein the United States.

Mail sent from overseas arrives in Peru at acentral processing facility located in the capitalcity of Lima. The facility sorts all mail, domesticand international, and sends it to large district orregional administration offices in Lima or otherregions in Peru for further sorting and delivery.There are also small post office branches wherefurther sorting of mail may occur. There arenine administration offices in Lima and Callaowith an average of 72 employees per office,and there are 39 branches with an average of2.5 employees per branch. The administrationoffices employ mail workers, mail carriers, andmanagement. The post office branches employworkers and carriers.

Mail carriers are paid a fixed salary thatis about two-thirds above the minimum wage,and many have steady employment. However,about one-fourth of the work force is not perma-nently employed. According to the firm’s annualreports, several of the main distribution officesin Lima have installed cameras to monitor thehandling of mail, suggesting that mail is likelynot lost there.8

Table 1 shows the structure of the labor forcein the mail sector in Lima and Callao in 2005according to the company’s publicly availabledata.9 The data is disaggregated by job cat-egory and neighborhood income.10 The table

8. See http://www.serpost.com.pe/transparencia/DocumentacionTransparencia/Docs2009/InformacionAdicional/MemoriaAnual/MemoriaAnual2009.pdf

9. See http://www.serpost.com.pe/transparencia/DocumentacionTransparencia/Docs2005/InformacionAdicional/Memoria Anual/MemoriaAnual2005.pdf

10. We define low-income neighborhoods as ones wherethe percent of the population considered poor is 30% or

shows that the distribution of people acrosstasks is remarkably similar across neighbor-hoods. Around half of the work force dealswith the delivery of mail, a quarter with mailin post offices, and another quarter in admin-istrative tasks. However, we find that wagesand years on the job are different across neigh-borhoods. For instance, postmen and messen-gers earn less in middle-income neighborhoods.This difference is significant when we comparehigh-income neighborhoods to middle-incomeneighborhoods (t-test p value = .0183). Thisdifference amounts to a 5% reduction in earn-ings across these two neighborhoods. Also, theearnings of postal employees is slightly lowerin low-income neighborhoods relative to otherneighborhoods (t-test p value = .0489). Thisdifference is about 1%. Finally, the numberof years on the job of postmen and messen-gers in richer neighborhoods is significantlylarger than those in other neighborhoods (t-testp value = .0129). Postmen and messengers inricher neighborhoods have tenures that are 25%longer than those in other neighborhoods. This isconsistent with the presence of efficiency wagesin richer neighborhoods as a way to enforcehigher quality in delivery in these areas, but it isalso consistent with the firm not paying enoughto retain its work force in more difficult areasor different costs of living or local market con-ditions. In any case, the data suggests that thefirm discriminates across neighborhoods.

In terms of mail loss, there is anecdotal evi-dence that it does occur. Casual observation often post offices in Lima revealed that serviceis slow and cumbersome and that mailing aletter or package requires interaction with unre-sponsive tellers. Customers willing to send mailof intrinsic value went through great lengthsto secure the mail, repeatedly taping the mailclosed after postage had been attached. Con-versations between customers standing in linetended to be about the reliability and risks ofthe mail arriving at its destination. An attendantin a small branch post office revealed to us that,in that particular office, mail from abroad almostnever arrived at its intended final destination or,if it did, it would arrive tampered with.

higher. Middle-income neighborhoods are those where thepercentage is between 10% and 30%, and high-incomeneighborhoods are those where the percentage is less than10%. The proportion of people living in poverty is from theclosest available poverty map to the time of the experiment(2006) calculated by the Peruvian Ministry of Economyand Finance using expenditure surveys from the PeruvianInstitute of Statistics.


TABLE 1Labor Force Characteristics across Neighborhoods

Labor Category

Neighborhood Postmen and Messengers Postal Employees Other

High income Monthly income (soles) 840.9 957.3 1032.6Years on the job 6.5 10.0 6.0Number 128 (51%) 63 (25%) 59 (24%)

Middle income Monthly income (soles) 798.6 959.8 1061.4Years on the job 5.2 10.1 5.9Number 124 (43%) 90 (32%) 72 (25%)

Low income Monthly income (soles) 829.1 947.1 1142.5Years on the job 5.2 9.6 6.5Number 60 (46%) 44 (34%) 26 (20%)

Readers writing to the newspaper El Comer-cio (June 06, 2007, June 23, and July 07, 2007)expressed their frustration for not receivingmagazines and/or letters from family members.The newspaper reported that many additionalletters of similar tenor are received. Indeed,commenting on a letter sent by a reader, thenewspaper noted that many informal magazinevendors offer magazines that show “strangeaddresses.” As there is a market for used mag-azines, there is no way to verify if these mag-azines are stolen or not. Similar comments arereproduced in ForosPeru.net, a Peruvian blog-ging site. Interestingly, comments are not alwaysnegative. Some people state that they have neverexperienced problems while others do. This sug-gests that problems might not be generalized andtherefore unlikely to occur in the central office.

Finally, in terms of discipline or penaltiesfor being caught stealing, there is evidenceto suggest that employees do get fired. Thecompany reports are not complete enough for usto determine if firings are precisely as a resultof stealing, however, there is a lot of evidenceof the company disciplining employees for notdoing their job. For example, according to thecompany’s 2005 annual report, 25 employeeswere fired owing to major offenses.11 Another23 were separated for arbitrary reasons while20 either retired voluntarily or ended their trialperiod at the company. In addition, there havebeen lawsuits brought by former employeesclaiming to have been wrongfully fired becauseof stealing.12

11. Report citation is listed in footnote 9.12. The Peruvian Constitutional Tribunal (the equivalent

of the U.S. Supreme Court) has heard ten cases broughtby ex-workers of the Peruvian mail service asking to bereinstated. Seven out of ten said they were accused of grave

IV. EXPERIMENTAL DESIGN

We send envelopes from the United States toPeru through normal mail services in both coun-tries (U.S. Postal Service and the Peruvian postalservice, respectively). We use a list of residen-tial addresses in metropolitan Lima, Peru that aregeographically representative of low-, middle-,and high-income neighborhoods. A resident ofeach address is the recipient of the envelope andreports to us if the envelope arrives or not.

The 2 × 2 design we employ varies thecontents of the envelope and the sender’s name.The contents of the envelope is a card and eithertwo $1 bills folded in half or no money. Thesender’s name is either a foreign name (i.e.,J. Tucker, M. Scott) or the same family nameas the recipient (i.e., M. Sosa, L. Cordova).13

Varying the sender’s name allows us to test ifnames signal that something of value is in theenvelope (i.e., money). The design is outlined inTable 2 and includes the number of envelopessent in each treatment.14

To get a valid estimate of crime, it isimportant that the envelope look realistic and

misdemeanors and two out of the seven explicitly mentionedstealing.

13. In South America, including Peru, everyone has twolast names. The first is the last name from the father and thesecond is the last name of the mother. We use the first lastname. We only test for family name versus a foreign name,not mismatched Hispanic names. This is another interestingline of research, but not pursued in this paper.

14. There are not an equal number of observations ineach cell for two reasons. First, some households movedbefore the experiment was complete, so they did not receiveall four treatments. Second, because of a clerical error,13 households received four letters, two with money andtwo without, but the sender’s last name was not equallydivided between family and foreign. All main results holdif these households are dropped. The estimates are ofequal magnitude, but some are less precise due to fewerobservations.


TABLE 2Experimental Design

Contents of Envelope

Sender Last Name Money No Money

Foreign n = 136 n = 131Family n = 135 n = 139

like something that would normally be sentin the mail. So, we chose an opaque solid-colored envelope and card (of the same color).The envelope looks like one that would besent for a birthday or other special occa-sion. Keeping with that idea, on the inside ofeach card, we handwrite “Happy Birthday” or“Feliz Cumpleanos”—depending on the returnaddressee’s name—and sign Josh or Mike orMarco or Luis. We do this because if the cardis stolen or opened, we want it to appear, tothe postal worker, like it was actually sent bythe person whose name appears on the front ofthe card. Figure 2 gives examples of two of theenvelopes that were sent to the same addressin the course of the study. To preserve confi-dentiality, we have blackened out the addressesof the sender and recipient and the recipient’sfirst name. The first envelope gives an exampleof mail sent by a foreigner to a recipient andthe second envelope gives an example of mailfrom a family member to a recipient. As can beseen in the Figure, the treatment manipulationof family member is subtle.

Because the envelope is opaque, the greet-ing inside the card cannot be seen. If the cardcontains money, this also cannot be seen, evenif held up to the light. One can feel that thereis something in the envelope because the foldedtwo $1 bills make a very slight bump. However,it is impossible to determine what exactly is inthe envelope without opening it.15 But, there is ahint that the envelope contains something otherthan the card. We chose this subtle manipulationso that anyone looking for something to stealwould need to pay careful attention for signs thatthe envelope contained something that might beworth stealing.

All envelopes have handwritten addresses,stamps for postage, and an airmail stamp on

15. We could very well have placed folded pieces ofpaper in the envelope instead of money, but we want theenvelopes and contents to be realistic, especially in case theenvelope is lost or stolen. Also, it is important to note thatthe envelope is only thicker and still easily slips througha mail slot or a sorting machine, as would the envelopewithout money.

FIGURE 2Examples of Envelopes Sent

Note: To preserve confidentiality, addresses and firstnames are blocked.

the front of the envelope. There are two returnaddresses in Atlanta and two in Washington,DC. All addresses are real so that we couldmonitor if the card was returned to the UnitedStates for any reason. The envelopes are gluedshut, making it difficult to steam open, resealand deliver. Envelopes are always mailed fromone of two locations. Envelopes with a returnaddress from Atlanta were mailed from the mainpost office in downtown Atlanta, and envelopeswith a return address of Washington, DC weremailed from a post office mailbox in Washing-ton, DC. The color of the envelope, the returnaddress, and the handwriting on the envelopeare randomized across the four treatments.16

16. We did this to insure that each envelope sent to ahousehold by a different person was indeed handwritten bya different person.


Envelopes were sent during the period Novem-ber 2006–November 2007.17

Mailboxes in Peru are secure and not exposedto theft from people passing by on the street.Typically, mail is placed in a locked mail boxinside a locked gate or entryway. Or, it is placedunder the door of the locked residence. Mail isnot left in post boxes on the streets, as is thecase in the United States.

To find recipient addresses, we tapped intotwo networks of people who engage in researchto recruit volunteers willing to receive the cardsand report to us. These people are trained in dataanalysis and data collection and are aware ofthe importance of honest reporting. In SectionVII, we describe steps taken to insure dataquality. The two networks include people froma variety of demographic and income groups.The important design element for us was thatthe addresses where the mail was sent weregeographically diverse. So, even though themail recipients might know one another, theaddresses are dispersed across locations. Tominimize the number of addresses in the studyfor any given post office, no more than fourhouseholds were within a 1-km radius of eachother (i.e., 0.62 miles or the equivalent of tenblocks). We mapped all the recipient addresseson a GIS updated street map of MetropolitanLima to minimize agglomeration and also toverify that the addresses were correct and active.This ensures that nonarrival of mail is not dueto an incorrect address.

Recipients of the mail reported the arrivalor nonarrival of each envelope and kept anymoney if an envelope with money arrived. Thisremoved any incentive to misreport the contents.They were instructed to not ask the mailmanabout the card or go to the post office toinquire. After the envelope was put in the mailin the United States, we sent an e-mail to therecipients telling them that an envelope wassent. They were instructed to inform us whenthe mail arrived, who it was from, the color,and the contents. They were not told ahead oftime the characteristics of the envelope or ifthe envelope contained money. This was doneto ensure no a priori bias in reporting and toallow us to check that reporting was accurate.In addition to collecting information on whetherenvelopes were received or not using severalmethods, we also asked for replies by e-mail.

17. Mailings were sent at various dates in November,June, July, and August.

This provided a simple way for us to check theresponsiveness of the recipients. We use thesedata later in Section VII to evaluate potentialnonresponse biases. We also had ten supervisorswho contacted various recipients to check if theenvelopes were received or not.

To compensate recipients for their time andhelp, at the end of the experiment, we con-ducted a lottery with cash prizes for recipientswho reported. Recipients knew of the lotterybefore we began sending envelopes. Also, toverify mail receipt responses, in December 2007we conducted a follow-up survey and collectedmore individual data on mail recipients. Thisalso allowed us to verify for a second time thataddresses were correct. All addresses were veri-fied, and all previous responses were confirmed.This gives us confidence that our data are accu-rate.18

V. RESULTS

In this section, we explore the patterns ofmail loss geographically, across various demo-graphic characteristics and across our treat-ments. First, we turn to a description of thesample, then main findings, and finally evidenceof strategic behavior.

A. Sample Characteristics

Table 3 shows descriptive statistics on theindividual and geographical characteristics ofthe mail recipients. The sample is split roughlyhalf and half between male and female recipi-ents. The distribution of residents across low-,middle-, and high-income neighborhoods is noteven, with more people living in middle-incomeneighborhoods.19 Most recipients have a univer-sity education, are married or living with theirpartner and have a family member that lives inthe United States.20 This latter result is impor-tant as it makes receipt of a card from the UnitedStates not seem strange and also attests to the

18. We asked the recipient if they received an envelopeduring a certain period of time and asked them to report thereturn address and color of the envelope. Recipients wereable to correctly confirm reports from 4 to 5 months earlier.

19. See footnote 10 for definitions of neighborhoodincome. We show later that our results are robust to otherdefinitions of economic status.

20. This is a reflection of our sample procedure; whilerecipients resided across many neighborhoods, they allwere contacted through research institutions. Also, recipientsregularly received mail, so the receipt of four envelopes overthe course of a year was not unusual.


TABLE 3Descriptive Statistics of Sample Recipients

Percent SD # Obs

Male recipients 47.5 141Low income 34.8 49Middle income 39.7 56High income 25.5 36Age (mean, years) 37.2 10.1 124University education 57.4 136Married or cohabitating 44.1 136Family size (mean, number) 4.1 1.5 124Family in United States 47.1 136Time in residence (mean, years) 16.5 12.8 124Minutes to post office (mean) 3.0 8.4 140

Note: Some variables have missing values because ofsurvey nonresponse.

degree of mail that could potentially come fromthe United States. Recipients have lived in theircurrent residence for an average of 16.5 years,and the nearest post office is 3 minutes away.

An important component of our experimentaldesign, in addition to a diverse and represen-tative distribution of individual mail recipientcharacteristics, is that the distribution of recipi-ent addresses is geographically dispersed acrossneighborhoods and post offices and is repre-sentative of metropolitan Lima. Figure 3 showsthe geographical distribution of residents in ourstudy. The residents cover the majority of thecity. There are fewer residents in some of theperi-urban areas of the city, but the addresses arenicely distributed across neighborhoods. Thisgives us observations across most areas of Limaand confidence that our results apply to thelarger, city-wide mail sector.

B. Main Findings

Turning to loss rates, we see that mail ser-vice in Lima is inefficient and subject to crime.Table 4 shows loss rates overall and by experi-mental treatments. Overall, 18% of all envelopessent through the mail never arrived at their desti-nation.21 Envelopes with money were less likelyto arrive than envelopes without money, so itdoes not appear that mail loss is solely due

21. For the mail that arrived, the average arrival timeof a piece of mail was 7.2 days (4.2 s.d.). In high-incomeneighborhoods, the arrival time is 7.2 days (4.3 s.d.). Inmiddle-income neighborhoods, it is 7.3 days (4.6 s.d.) and6.9 days (3.7 s.d.) in poorer-income neighborhoods. Thedifferences across neighborhoods in arrival times are notsignificantly different.

FIGURE 3Distribution of Addresses across Lima, Peru

to bad service. This confirms our first hypoth-esis and hints more of criminal activity. Over21% of envelopes with money did not arrive,whereas 14.8% of envelopes without money didnot arrive. This 50% increase in loss is statisti-cally significant (t-test p value = .047).

How was mail lost across our four treat-ments? The bottom panel of Table 4 showsloss rates by the contents of the envelope andthe sender’s last name. Again, envelopes withmoney were more likely to be lost than thosewithout money, whether the sender’s last namewas foreign or a family name. The differencebetween money and no money envelopes forenvelopes with a foreign sender is not signif-icantly larger, but the almost 10 percentagepoint difference for envelopes with a familylast name is (t-test for difference in mean p-value = .047). Indeed, most loss happened forenvelopes with money sent by a family mem-ber. Almost one in four of those envelopes neverarrived.


TABLE 4Average Loss Rates Overall and by

Experimental Treatment (%)

Number

Overall 18.1 98(1.7)

Money 21.4 58(2.5)

No money 14.8 40(2.2)

Experimental TreatmentsContents of

EnvelopeDifferencein Mean

Money No Money p value

Foreign 19.8 16.0 .4181(3.4) (3.2)

Family 23.0 13.7 .0467(3.6) (2.9)

Difference in meanp value .5343 .5868

Notes: Standard error of the mean in parentheses. Chi-square distribution test validates the null hypothesis that theloss rates across treatments are the same.

TABLE 5Average Loss Rates by Income Groups (%)

LowIncome

MiddleIncome

HighIncome

Overall 18.9 20.4 13.5Money 19.8 25.7 16.9No money 18.0 15.3 10.0Foreign sender name 18.9 18.3 16.4Family sender name 18.9 22.4 10.3

Across low-, middle-, and high-incomeneighborhoods, mail is lost at different rates.Table 5 shows that residents in middle-incomeneighborhoods lose mail at the highest rate,20.4%, and those in high-income neighborhoodslose mail at the lowest rate, 13.5%. The lossrate in middle-income income neighborhoods issignificantly larger than in high-income neigh-borhoods.22

One might wonder if mail loss can beattributed to the Peruvian mail service or tothe U.S. Postal Service. The results in Table 5suggest that loss is happening on the Peruvianside. While it may be reasonable to think that

22. t-tests yield p values of .696 comparing low- tomiddle-income loss rates, .080 comparing middle- to high-income, and .189 comparing low- to high-income loss rates.

envelopes with money might be lost on the U.S.side, it is highly unlikely that the significantlydifferent loss rates we see across middle- andhigh-income neighborhoods is due to the U.S.Postal Service. Such loss rates cannot exist with-out knowledge of neighborhoods in Lima. Thenext section provides further evidence of this.

Looking at the content of the envelopes, mailwith money is significantly more likely to be lostthan that without money in middle-income neigh-borhoods. In middle-income neighborhoods, theloss rate of envelopes with money is over 10 per-centage points larger than for envelopes withoutmoney (t-test p value = .057). The loss rate inpoor neighborhoods is around 18% and is similarfor envelopes with and without money. High-income neighborhoods have an almost 7 percent-age point increase for envelopes with money, butthis is not significantly different (t-test p value =.232).

This pattern of loss confirms our secondhypothesis and is consistent with an expectationthat the poor are not receiving valuables by mail,so loss rates are no different with and withoutmoney. It seems to be more a reflection of badservice than crime. Loss rates in middle-incomeand high-income neighborhoods, however, areconsistent with the expectation that these pop-ulations have valuable items to receive throughthe mail. Search is relatively larger in middle-income neighborhoods, and this may reflect anexpectation that people in middle-income neigh-borhoods have few alternatives for receivingand sending mail. The positive but insignificanteffect of money in envelopes in rich neighbor-hoods may reflect the expectation that wealthierhouseholds have more alternatives.

The last two rows in Table 5 show that thepattern of loss across neighborhoods is primarilydriven by envelopes where the sender and recipi-ent share the same last name. Mail from a familymember appears to be given special attentionas loss rates vary across neighborhoods. Thisresult is important, and consistent with our firsthypothesis, because it suggests that those han-dling the mail attribute a similar probability ofbeing caught across neighborhoods when dis-posing of mail sent by nonfamily members. Itis also consistent with the expectation that mailfrom nonfamily members is unlikely to containanything of value.

C. Evidence of Strategic Behavior

The main findings in the previous sectionshow that mail with money is lost more


TABLE 6Average Loss Rates by Money, Income

Groups, and Sender Name (%)

Envelope with Money

LowIncome

MiddleIncome

HighIncome

Foreign sender name 21.3 20.4 17.1Family sender name 18.2 30.9 16.7

Envelopes with No Money

LowIncome

MiddleIncome

HighIncome

Foreign sender name 16.3 16.0 15.8Family sender name 19.6 14.8 3.1

frequently and that crime is not distributedequally across the population. The mechanismfor loss seems to be that envelopes coming fromfamily members are scrutinized more closelythan those from a foreigner. In this section, welook more closely at the patterns of loss and whythey might exist.

Table 6 shows the joint effect of incomeand money on mail loss. The numbers in thetable also allow us to calculate a difference-in-difference estimate of the effect of incomeon crime using our treatment variables. Thepresence of money in envelopes sent by afamily member increases the rate of mail lostby 16.1 percentage points (30.9%–14.8%) inmiddle-income neighborhoods and decreases itby 1.4 percentage points (18.2%–19.6%) inpoor neighborhoods. In other words, mail is 17.5percentage points less likely to arrive when sentto middle-income neighborhoods when there issuspicion of valuable content. A comparisonof the richer neighborhoods and poorer neigh-borhoods gives a similar estimate (15.0). Thisincrease in the likelihood of loss from poor tomiddle and poor to rich neighborhoods couldbe due to expectations that something of valuemight be sent in the mail or the perceivedsmaller risk of being caught.

In order to test whether the differential impactof crime across income groups is explainedby differences in the quality of service andnot expectations, in Table 7, we compare lossrates across types of envelopes using regres-sion analysis. This analysis assumes that qualityof service by neighborhood affects loss ratesuniformly independent of who sends the mail.The results show that both expectations that theenvelope contains something of value (money,

no money) and the probability of being caught(which varies by the socioeconomic level ofthe neighborhood) explain loss rates. The tablepresents results from fixed-effects linear proba-bility (OLS) regressions of loss on whether theenvelope contained money, if it was sent by afamily member, and interactions with neighbor-hood income (percent classified as poor). Thedependent variable equals 1 if the mail did notarrive at its destination and 0 otherwise.23

It is important to note that the analysisincludes recipient fixed effects, to control forany idiosyncrasies of our small sample, and ourmain results still hold.24 The first column inTable 7 illustrates this. Envelopes with moneyare more likely to be lost. The second and thirdcolumns show results for the subsamples of mailsent by a family member and by a foreigner. Bylooking only at envelopes from family membersor from foreigners, we attempt to keep con-stant the expected cost of committing a crime,so that we can focus on expectations that theenvelope contains something of value.25 We seethat the effect of money for envelopes comingfrom family members, controlling for recipi-ent fixed effects, is stronger as neighborhoodsbecome wealthier. There is no significant effectof envelopes coming from foreigners. This con-firms the results in Table 6 and suggests thatexpectations that an envelope contains some-thing of value matter. This is so because weexpect not only that both the rich and the poorcare about receiving mail but that the cost ofbeing caught stealing should not decrease withthe wealth of the neighborhood. The fact thatenvelopes with money are lost at a higher rateas neighborhood income goes up says there isan expectation that mail from family members

23. The same results hold with a fixed-effects Logitmodel. We report OLS results because the regressions donot drop observations and give us more robust results.

24. That is, our results are not due to idiosyncrasiesin recipient reporting, postal carriers, or household location(e.g., at the end of the postal carrier’s route).

25. This is a difference-in-difference estimate on thenet benefit (expectation of something of value less theexpected cost of being caught) of an envelope with andwithout money across neighborhoods. Or, re-written, thisis the net expectation of encountering something of valueless the net cost of being caught taking something ofvalue (as in Equation (1) in Section II). By splitting thesample into envelopes from family members and fromforeigners, we can control for the net cost of being caughtacross neighborhoods. We expect this to be constant, sinceotherwise this would say that the rich care less aboutreceiving mail with money. So, any significant effect onmoney or money interacted with neighborhood will be dueto expectations.


TABLE 7Evidence of Strategic Behavior Probability of Mail Loss by Sender Name and Envelope Contents

OLS Fixed-Effects Regressions

(1) (2) (3) (4) (5)Variables All Family Foreign Money No Money

Money 0.062∗∗ 0.197∗∗∗ 0.035(0.028) (0.070) (0.063)

Family 0.000 0.050 −0.081(0.028) (0.060) (0.066)

Money × percent poor −0.005∗ −0.001(0.003) (0.002)

Family × percent poor −0.001 0.003(0.002) (0.003)

Constant 0.150∗∗∗ 0.139∗∗∗ 0.176∗∗∗ 0.199∗∗∗ 0.156∗∗∗

(0.024) (0.029) (0.027) (0.025) (0.027)Recipient fixed effects Yes Yes Yes Yes YesObservations 541 274 267 271 270R2 .012 .059 .003 .007 .013

Notes: Standard errors in parentheses. Dependent variable is Mail Loss (=1 if mail never arrived and =0 if mail arrived).Independent variables: Money = 1 if envelope contained money, Family = 1 if sender’s last name was the same as recipient’s,Percent Poor = percent of population in neighborhood living in poverty.

∗p < .10, ∗∗p < .05, ∗∗∗p < .01.

contains something of value when sent towealthier neighborhoods.

The results in the fourth and fifth columnsof Table 7 of the regressions on money andno-money envelopes are also indicative ofincentives. The previous regression on familyenvelopes suggests that family envelopes sentto richer neighborhoods are perceived to containvaluables. So, we would expect that the effectof family on money envelopes to be strongerin richer neighborhoods, not weaker. The factthat we do not find this suggests that there isa countervailing force limiting the incentive tocommit a crime. Since larger expected costsof being caught reduce the incentives to steal,this result is consistent with the belief that theprobability of being caught stealing is larger inricher neighborhoods.26 This result is also con-sistent with our model predictions that changesin social characteristics, z, can affect both theprobability of something of value being therein the mail and the probability of being caughtstealing.

26. As described in the section on the Mail Sector inPeru, postal workers in richer neighborhoods earn higherwages and have longer job tenures. This efficiency wagestory would be consistent with a higher probability of beingcaught stealing. However, with the evidence we have fromnaturally occurring data as in Table 1, we only observeequilibrium outcomes, so we cannot say with certainty thatthis is the case. Postal workers in richer neighborhoodsmight also be more honest.

All together, these results are consistent witha story that, as neighborhood income rises,expectations that valuables are sent through themail increase and interact with an increasedprobability of being caught.

VI. ROBUSTNESS CHECKS ON MAIN FINDINGS

This section presents regression analysis ofmail loss rates to test the robustness of ourmain results to omitted variables and specifi-cation assumptions. We check that our resultsare not due to recipient-level fixed effects, ourdefinition of neighborhood grouping by income,correlation between neighborhood income andthe way mail is processed, misreporting, or othersocioeconomic variables. Table 8 presents OLSregressions of mail loss on covariates, includingtreatment variables, nonlinear effects of neigh-borhood income, distance to the closest postoffice branch, whether there is an administra-tive center located in the neighborhood, numberof post office branches in the neighborhood, andthe time effect of when the mail was delivered.27

The first column in Table 8 (and the firstcolumn in Table 7) confirms our main find-ings. Envelopes containing money are more

27. The results hold if specified as Logit regressions andare robust to autocorrelation.


TABLE 8Robustness Check on Main Findings Probability of Mail Loss by Envelope Contents and Sender

Name OLS Regressions

(1) (2) (3) (4) (5)Variables All Money No Money Family Foreign

Money 0.065∗ 0.100∗∗ 0.029(0.034) (0.050) (0.048)

Family 0.006 0.031 −0.020(0.033) (0.049) (0.043)

Percent poor 0.011∗∗ 0.013 0.009 0.013∗ 0.010(0.005) (0.008) (0.007) (0.008) (0.007)

Percent poor squared −0.000∗ −0.000 −0.000 −0.000 −0.000(0.000) (0.000) (0.000) (0.000) (0.000)

Minutes to post office 0.001 0.004 −0.001 −0.000 0.003(0.002) (0.003) (0.003) (0.003) (0.003)

Administrative center 0.115∗∗∗ 0.133∗∗ 0.096∗∗ 0.075 0.157∗∗∗

(0.035) (0.053) (0.047) (0.050) (0.050)Number of post office branches −0.029∗∗ −0.023 −0.034∗∗ −0.033∗ −0.023

(0.012) (0.018) (0.016) (0.017) (0.017)Mailing number (1–4) 0.002 −0.003 0.005 0.011 −0.011

(0.015) (0.023) (0.021) (0.022) (0.022)Constant 0.041 0.086 0.062 0.012 0.078

(0.068) (0.086) (0.094) (0.099) (0.090)Observations 537 269 268 272 265R2 .039 .040 .037 .041 .050

Notes: Standard errors in parentheses. Dependent variable is Mail Loss (=1 if mail never arrived and =0 if mail arrived).Independent variables: Money = 1 if envelope contained money, Family = 1 if sender’s last name was the same as recipient’s,Percent Poor = percent of population in neighborhood living in poverty, Minutes to post office = number of minutes fromresidence to closest post office, Administrative Center = 1 if an administrative center is located in neighborhood, Number ofPost Office Branches = number of branches in neighborhood, Mailing Number = 1 if first mailing, =2 if second mailing,etc.

∗p < .10, ∗∗p < .05, ∗∗∗p < .01.

likely to get lost. The results in Table 7 con-firm that this holds with recipient-level fixedeffects, and the results in Table 8 confirm thatit holds at the same time as does the non-linear relationship with neighborhood income(percent classified as poor). This latter resultstill remains even, controlling for the manner inwhich mail is processed across neighborhoods.The results are intuitive. Mail sent to neighbor-hoods with an administrative center is lost morefrequently because, presumably, there are manymore employees handling the mail and this helpsto dissipate responsibility. Neighborhoods withmore post office branches lose less mail becauseit is easier to identify responsibility.

In Table 8, columns 2 and 3 show thatthe effect of neighborhood income is slightlystronger for the envelopes with money, butthis is not significant. The regressions dividingthe population receiving envelopes from familyand nonfamily members (columns 4 and 5)confirm that it is the envelopes with moneycoming from family members that are more

likely to get lost. Finally, the patterns of lostmail across all regressions do not seem torespond to the proximity of post offices. Thenumber of minutes it takes to get to the closestpost office is not significant.28

The results reported above also eliminatetwo alternative explanations for higher loss ofenvelopes with money: systematic misreport-ing by recipients and misreporting envelopeswith money. First, since the results hold whencontrolling for recipient fixed effects (Table 7),individual misreporting is not causing our mainresult that envelopes with money are lost more.Second, since the recipients did not know aheadof time what kind of envelope was sent, thereis no reason to believe that the money effect isdue to people not paying attention to the moneyenvelopes more than no-money envelopes. Also,

28. The number of minutes to the closest post office isbased on an accessibility model which calculates the leastcost path surface (based on time) from any place, using GIS.The accessibility measure uses three different levels of roadswith different speeds of movement.


and this is crucial, they could keep the moneyand therefore did not need to say it was lost.

Finally, the nonlinear effect of neighborhoodincome on overall loss rates is not what onewould expect if it were only due to differen-tial shirking by neighborhoods. Table 1 showedthat years on the job for postmen and mes-sengers were lower in low- and middle-incomeneighborhoods. If years on the job reflects theconsequences of shirking (i.e., getting fired), wewould expect loss rates to be similar in low- andmiddle-income neighborhoods, not different.

The results in Table 8 are also robust to learn-ing by the mail carrier and the inclusion of othersocioeconomic information, such as family size,time in residence, and marital status.29 Overall,envelopes with money are more likely to be lostand neighborhood income is nonlinearly relatedto loss rates. Also, the fact that money envelopesare more likely to be lost, even when controllingfor recipient fixed effects, is strong evidence thatthis result is robust.

VII. ROBUSTNESS CHECKS ON REPORTING BIAS

This section presents additional evidence thatthe results in the paper are not due to misre-porting by the mail recipients. As discussed inthe data section, recipients in our experimentwere recruited among people involved in fieldresearch. They are trained in survey methodsand are aware of the problems associated withmisreporting. In addition we had ten monitorscoordinating the collection of data as a way tokeep close vigilance on the process. Despite allthis, it is always possible that mail recipientsmistakenly reported lost mail, or worse, pur-posely reported losses when they did not exist.While our post-experiment survey gives us con-fidence that the envelopes received were indeedreceived, it is harder to check if reported lossesdid indeed occur. This section presents a seriesof tests that suggest that the patterns of lost dataare not biased and therefore the main results ofthe paper are not due to misreporting.

29. Learning is tested by the inclusion of a lagged termfor loss or a dummy variable that equals one if the firstenvelope did not arrive (whether it contained cash or not).We also tested for whether the likelihood of loss varied bywhether the envelope was the first, second, third, or fourthmailing and for the correlation between lost mailings. Thereis a slightly higher likelihood a mailing is lost if the first onewas stolen. The most important outcome of these checks isthat our main results still hold. These results are not includedin the paper but are available from the authors.

Table 9 checks the significance of the effectof money on loss rates. These are OLS regres-sions of mail loss on the treatment variable andresponse time.30 The estimations exploit the factthat we had a measure of recipients’ responsive-ness with the time they took to respond to andcheck their e-mail. One hypothesis is that lessresponsive recipients or those unable to respondquickly might be more likely to report missingmail owing to distraction. Since we sent an e-mail each time the envelopes were sent out, italso allows us to see if our results are associ-ated in some way to periods in which subjectswere busier or more distracted. We see that, evenwhen controlling for response time, our resultsstill hold. Columns 1–3 show the rate of mailloss as a function of whether it contained moneyand whether the subject responded to the e-mailor not. As expected, recipients who do not sendan e-mail response are less likely to receivethe envelope. However, we see that envelopeswith money are still more likely to be lost andthat this result is explained mainly by the lossof envelopes sent by family members (coeffi-cients are similar to those reported in Table 8).Columns 4–6 show the rate of mail loss as afunction of money and the time it took subjectsto respond to e-mails. Again, we find that moneyenvelopes are lost more frequently and that thisis significant among the envelopes sent by fam-ily members.

Table 10 checks the significance of the effectof money on loss rates. The estimations exploitthe variation in time to confirm receipt ofthe mail to see if those confirming relativelylater were driving the results. Confirmationtimes are measured from the day the envelopeswere mailed from the United States and takeadvantage of the fact that those receiving themail were not equally easy to reach. This isso because some recipients had to travel forwork or because they had no Internet connectionat home or work (and therefore checking e-mail regularly was less convenient).31 Table 10makes the assumption that the longer the timethe mail is not reported the higher the likelihoodthat reasons other than crime or bad serviceexplain the results. Columns 1–3 estimate theeffect of money after all observations which

30. Note that all results of the effect of money onloss rates in Tables 9 and 10 hold with Logit fixed-effectsregressions.

31. Part of the research was done concurrently withthe 2007 National Census which required several of ourrecipients to be away from home.


TABLE 9Robustness Check on Reporting Probability of Mail Loss Controlling for Response Time OLS

Fixed-Effects Regressions

(1) (2) (3) (4) (5) (6)Variables All Family Foreign All Family Foreign

Money 0.073∗∗∗ 0.102∗∗ 0.017 0.062∗∗ 0.079∗ 0.011(0.028) (0.042) (0.038) (0.028) (0.042) (0.039)

Responded by e-mail in a week −0.165∗∗∗ −0.198∗∗ −0.097(0.047) (0.082) (0.076)

Responded by e-mail in 2 weeks −0.104∗∗ −0.156∗ −0.031(0.050) (0.093) (0.080)

Responded by e-mail in 3 or more weeks −0.022 0.043 −0.032(0.050) (0.087) (0.082)

Responded to e-mail −0.121∗∗∗ −0.130∗ −0.089(0.037) (0.069) (0.056)

Constant 0.210∗∗∗ 0.203∗∗∗ 0.219∗∗∗ 0.203∗∗∗ 0.199∗∗∗ 0.203∗∗∗

(0.026) (0.045) (0.038) (0.026) (0.043) (0.037)Recipient fixed effects Yes Yes Yes Yes Yes YesObservations 541 274 267 541 274 267R2 .039 .059 .020 .047 .106 .013Number of recipients 141 139 139 141 139 139

Notes: Standard errors in parentheses. Dependent variable is Mail Loss (=1 if mail never arrived and =0 if mail arrived).Independent variables: Money = 1 if envelope contained money, Responded by e-mail in a week = 1 if recipient respondedto initial e-mail informing the mailing of a card within a week of us sending the e-mail, Responded by e-mail in 2 weeks =1 if it took 1–2 weeks, Responded by e-mail in 3 or more weeks = 1 if it took 3 or more weeks to respond.

∗p < .10, ∗∗p < .05, ∗∗∗p < .01.

TABLE 10Robustness Check on Reporting Probability of Mail Loss with Restricted Sample OLS

Fixed-Effects Regressions

Observations Confirmed Losses Confirmed AfterAfter 4 Weeks Dropped 4 Weeks Switched to No Loss

(1) (2) (3) (4) (5) (6)All Family Foreign All Family Foreign

Money 0.070∗∗ 0.062∗ 0.052 0.048∗∗ 0.065∗∗ 0.027(0.029) (0.035) (0.045) (0.023) (0.033) (0.035)

Constant 0.093∗∗∗ 0.091∗∗∗ 0.108∗∗∗ 0.063∗∗∗ 0.052∗∗ 0.076∗∗∗

(0.020) (0.023) (0.023) (0.016) (0.023) (0.024)Recipient fixed effects Yes Yes Yes Yes Yes YesR2 within .025 .042 .022 .011 .028 .005N 366 189 177 541 274 267

Notes: Standard errors in parentheses. Dependent variable is Mail Loss (=1 if mail never arrived and =0 if mail arrived).Independent variable: Money = 1 if envelope contained money.

∗p < .10, ∗∗p < .05, ∗∗∗p < .01.

confirmed receipt of the mail after 4 weeks aredropped. The regressions show that envelopeswith money are still more likely to be lost in thisrestricted sample. The regressions also confirmthat the effect is only significant in the envelopesfrom family members. Columns 4–6 repeat theanalysis under the strong assumption that laterreports were false positives (i.e., made-up data).The regressions show that the effect of money

on loss rates persists. In all, these regressionsshow that if misreporting is correlated with timeto confirm, misreporting is not serious enoughto eliminate our main results.

Finally, we test for random misreporting.These estimates assume that some of thereported losses were mistakes and that thesemistakes were random. To do this, we randomlyselect some reported losses and change them to


TABLE 11Robustness Check on Reporting Probability of Mail Loss with Reported Losses Changed to No

Loss Average and 90% Confidence Interval of Bootstrapped OLS Recipient Fixed-EffectsRegressions

20% of Reported Losses 30% of Reported LossesChanged to No Loss at Random Changed to No Loss at Random

(1) (2) (3) (4) (5) (6)All Family Foreign All Family Foreign

Money 0.042 0.066 0.006 0.036 0.056 0.004[0.011,0.070] [0.026,0.103] [−0.033,0.044] [0.004,0.065] [0.017,0.101] [−0.039,0.045]

Constant 0.107 0.098 0.123 0.091 0.084 0.107[0.087,0.125] [0.075,0.122] [0.096,0.150] [0.070,0.110] [0.056,0.112] [0.072,0.137]

Notes: 90% CI in brackets. 10,000 bootstraps. Dependent variable is Mail Loss (=1 if mail never arrived and =0 if mailarrived). Independent variable: Money = 1 if envelope contained money.

no loss. This amounts to assuming that reportedlosses are measured with error. Table 11 presentsthe estimates of 10,000 repetitions of lin-ear regressions with recipients’ fixed effects.Columns 1–3 report results when 20% of thereported losses are assumed to be mistakes andcolumns 4–6 report results when 30% of thereported losses are assumed to be mistakes. Asexpected, the estimated parameters are smaller.However, neither the significance nor directionof the results are affected. Money envelopes arelost more frequently, and the losses are sig-nificant in the envelopes coming from familymembers.

While none of our robustness checks is proofthat no misreporting occurred, the checks shoulddispel concerns that any misreporting is largeenough as to invalidate our results.

VIII. CONCLUSIONS

Using a simple and novel field experimentthat opens the door for opportunistic behavior,we examine strategy and crime in the mail sec-tor in Lima, Peru. We hypothesize that the verynature of mail delivery gives an opportunity tothose who handle the mail to “lose” mail if itis beneficial to do so. Our design allows us todifferentiate poor service from targeted crimeand to investigate what information is pertinentto committing crimes and who suffers the mostfrom it.

We have several key findings. First, lossrates are very high. Over 18% of all mail sentnever arrived at its destination. These rates arehuge and would imply large barriers for thedevelopment of trade that relies on mail services.

Second, these loss rates are partially explainedby poor service but not completely. Envelopescontaining money were 50% more likely to belost than those without money. So, mail lossis not random and hints at strategic behavior.Third, when the sender’s last name matchedthe recipient’s last name, the mail was almosttwice as likely to be lost if it contained money.Clearly, those who handle the mail are look-ing for clues that might suggest that an enve-lope holds something of value. Fourth, middle-income neighborhoods suffer the highest lossrates and high-income neighborhoods suffer thelowest. This result (and the previous) lends sup-port for the crime occurring in Peru rather thanthe United States since it would require the U.S.Postal Service to know which neighborhoodswere rich or poor.

Finally, the patterns of behavior we observeare consistent with expectations that the recip-ient could receive something of value and theperceived probability of being caught stealing.This results in a nonlinear effect of neighbor-hood income on loss. Looking only at mail fromfamily members, we see that loss increases asneighborhood income rises if the mail containsmoney, suggesting that there is an expectationthat residents in middle-income and wealthyneighborhoods may have something valuable tosteal. The magnitude of loss is higher in middle-income neighborhoods though, supporting thenotion that the probability of being caught ismore likely in wealthy neighborhoods. That is,while it is more likely to get something of valuewhen stealing from middle- and high-incomeresidents, stealing in the highest-income neigh-borhoods may be tempered by a higher likeli-hood of being caught. Middle-income residents


then suffer a compounding effect that results inthe highest loss rates.

Put together our results suggest a model ofbehavior where those who handle the mail arelooking for items of value to steal, but theytake into account the likelihood that valuablesare being sent in the mail and the probabilityof being caught stealing. Moreover, and impor-tantly, crime is not independent of the neighbor-hood’s characteristics.

While our study cannot speak for the pres-ence of large inefficiencies in all the sectorsdealing with the transaction of goods and ser-vices, it highlights the large barriers to marketdevelopment that developing economies face.Certainly, our results suggest barriers are highfor e-commerce to emerge since it relies onthe mail sector to transport goods. Other, moresecure, transportation services are available, butthey can be two to five times more expensive.Also, while the cost of losing a piece of mailmay be low (depending on what was lost), rela-tive to other crimes, it is the unreliability of theservice that has a larger cost because it hampersmarket development.

The sophistication in criminal activity foundin our research suggests that there is a needto design monitoring mechanisms (e.g., securitycameras) and appropriate incentives to minimizestrategic behavior. Indeed, in a field experimenton auditing, Nagin et al. (2002) find that anincreased perception of monitoring can reduceshirking. Cameras would work in mail sortingfacilities (where they are currently installed atsome locations), but it is more difficult andcostly to monitor mail delivery on foot. Randomaudits of the nature of our experimental designcould be an effective mechanism for identifyingproblem neighborhoods and facilities.

Our study further shows that private firmsproviding public services also face incentiveproblems due to moral hazard in the same waystate-owned enterprises do. The nature of thegood seems to be as important as the natureof ownership. Incentive problems may well pre-vent markets from developing. How severe theseproblems are relative to other factors, such asinefficiencies in governance or lack of competi-tion or infrastructure, remains to be studied.

REFERENCES

Abadie, Alberto, and Javier Gardeazabal. “The EconomicCosts of Conflict: A Case Study of the Basque Coun-try.” American Economic Review, 93, 2003, 113–32.

Alesina, Alberto, and Roberto Perotti. “Income Distribution,Political Instability and Investment.” European Eco-nomic Review, 40, 1996, 1203–28.

Anwar, S., and H. Fang. “An Alternative Test of RacialProfiling in Motor Vehicle Searches: Theory andEvidence.” American Economic Review, 96, 2006,127–51.

Armantier, Olivier, and Amadou Boly. “Can Corruption BeStudied in the Lab? Comparing a Field and a LabExperiment.” Working Paper, Federal Reserve Bankof New York, 2008.

Barro, Robert J. “Economic Growth in a Cross Sectionof Countries.” Quarterly Journal of Economics, 106,1991, 407–43.

Becker, Gary. “Crime and Punishment: An EconomicApproach.” Journal of Political Economy, 76, 1968,169–217.

Bertrand, Marianne, Simeon Djankov, Rema Hanna, Send-hil Mullainathan. “Obtaining a Driving License inIndia: An Experimental Approach to Studying Corrup-tion.” Quarterly Journal of Economics, 122(4), 2007,1639–76.

Corman, Hope, and Naci Mocan. “Carrots, Sticks, andBroken Windows.” Journal of Law and Economics,2005, 235–66.

Di Tella, Rafael, and Ernesto Schargrodsky. “The Role ofWages and Auditing during a Crackdown on Corrup-tion in the City of Buenos Aires.” Journal of Law andEconomics, 46(1), 2003, 269–92.

. “Do Police Reduce Crime? Estimates Using theAllocation of Police Forces after a Terrorist Attack.”American Economic Review, 94(1), 2004, 115–33.

Duggan, Mark, and Steven Levitt. “Winning Isn’t Every-thing: Corruption in Sumo Wrestling.” American Eco-nomic Review, 92(5), 2002, 1594–605.

Erlich, I. “Participation in Illegitimate Activities: A Theoret-ical and Empirical Investigation.” Journal of PoliticalEconomy, 81, 1973, 521–65.

Fisman, Raymond, and Edward Miguel. “Corruption,Norms, and Legal Enforcement: Evidence from Diplo-matic Parking Tickets.” Journal of Political Economy,115, 2007, 1020–48.

Fisman, Raymond, and Yongxiang Wang. “Corruption inState Asset Sales-Evidence from China.” WorkingPaper, Columbia University, 2010.

Fisman, Raymond, and Shang-jin Wei. “Tax Rates and TaxEvasion: Evidence from ’Missing Imports’ in China.”Journal of Political Economy, 112, 2004, 471–97.

Fried, Brian, Paul Lagunes, and Atheendar Venkataramani.“Corruption and Inequality at the Crossroads: A Mul-timethod Study of Bribery and Discrimination in LatinAmerica.” Latin American Research Review, 45(1),2010, 76–97.

Gaviria, Alejandro. “Assessing the Effects of Corruptionand Crime on Firm Performance: Evidence fromLatin America.” Emerging Markets Review, 3, 2002,245–68.

Glaeser, Edward L., Bruce Sacerdote, and Jose A.Scheinkman. “Crime and Social Interactions.” Quar-terly Journal of Economics, 111, 1996, 507–48.

Hsieh, Chang-Tai, and Enrico Moretti. “Did Iraq Cheat theUnited Nations? Underpricing, Bribes, and the Oil forFood Program.” Quarterly Journal of Economics, 121,2006, 1283–309.

Hunt, Jennifer, and Sonia Laszlo. “Is Bribery Really Regres-sive? Bribery’s Costs, Benefits and Mechanisms,”Working Paper, McGill University, 2008.

Jacob, Brian, and Lars Lefgren. “Are Idle Hands theDevil’s Workshop? Incapacitation, Concentration, andJuvenile Crime.” American Economic Review, 93,2003, 1560–77.


Levitt, Steven. “Using Electoral Cycles in Police Hiring toEstimate the Effect of Police on Crime.” AmericanEconomic Review, 87(3), 1997, 270–90.

Nagin, Daniel, James Rebitzer, Seth Sanders, and LowellTaylor. “Monitoring, Motivation, and Management:The Determinants of Opportunistic Behavior in aField Experiment.” American Economic Review, 92(4),2002, 850–73.

Olken, Benjamin. “Monitoring Corruption: Evidence froma Field Experiment in Indonesia.” Journal of PoliticalEconomy, 115(2), 2007, 200–49.

Olken, Benjamin, and Patrick Barron. “The Simple Eco-nomics of Extortion: Evidence from Trucking inAceh.” Journal of Political Economy, 117(3), 2009,417–52.

Pshisva, Rony, and Gustavo Suarez. “Capital Crimes: Kid-nappings and Corporate Investment in Colombia,” in

The Economics of Crime: Lessons for and from LatinAmerica, edited by Rafael Di Tella, Sebastian Edwards,and Ernesto Schargrodsky. Chicago: University ofChicago Press, 2010, 63–100.

Reinikka, Ritva, and Jakob Svensson. “Local Capture: Evi-dence from a Central Government Transfer Program inUganda.” Quarterly Journal of Economics, 119, 2004,679–705.

. “Fighting Corruption to Improve Schooling: Evi-dence from a Newspaper Campaign in Uganda.” Jour-nal of the European Economic Association, 3, 2005,259–67.

Sequeira, Sandra, and Simeon Djankov. “An EmpiricalStudy of Corruption in Ports,” Working Paper, LondonSchool of Economics, 2010.

lost in the mail: a field experiment on crime

Documents