Comparación en los estudios de caso.

To test the main hypotheses discussed in the previous section, we implement a (2x2)-design consisting of two control treatments. The …rst is a "bait" treatment that randomly assigns money and no money to pieces of mail. The second is a "fa- miliar" treatment that randomly assigns same and di¤erent sender and recipient

last names to pieces of mail. We refer to the case in which the sender and the re- cipient have the same last name as "familiar" and the case in which they do not as "foreign."

The experimental design is within-households in the sense that each recipient household is sent four pieces of mail: (1) a familiar money envelope, (2) a familiar no-money envelope, (3) a foreign money envelope and (4) a foreign no-money en- velope. This is opposed to a between-households design in which households would strictly be assigned to one of the following treatments: "money," "no money," "fa- miliar" and "foreign."59

While a potential bene…t of a between-households design is that it mitigates any type of order e¤ects ("red ‡ags"), we choose a within-households design because of the following reasons. First, a between-households design requires a substantially larger sample of recipient households; in fact, four times as much as what we cur- rently have if households are to be randomly assigned to only one of the four treat- ments and if we seek balanced comparisons. Furthermore, the larger the group of recipient households, the more we need to be concerned about "tipping o¤" mail handlers. As discussed previously, this is a legitimate concern when mitigating bias in the data. So, we choose a within-households design.

To further mitigate any data bias resulting from order and/or subject-pool ef- fects as well as any sample selectivity bias, we implement the following controls. First, we send four sets of envelopes to each household. Any set of envelopes to all households is sent as one batch and the four batches are sent within a range of three to …ve weeks apart.60 _{The envelopes are identical in terms of size and appear-}

ance, and di¤er only by the characteristics of interest, which are (1) the treatment

59_{Note that the terms within- and between-subjects are traditionally used in the experimen-} tal literature to indicate designs in which subjects’identities are known. It should be noted that in this case, subjects’identities are completely unknown; in particular, the subjects of interest are handlers of mail whose identities are unobserved. What we observe is a reduced-form of their actions–i.e., whether or not mail is delivered.

characteristics (i.e., content and sender last name) and (2) additional observable characteristics, which include the sender’s address, the color of the envelope and the sender’s handwriting. We discuss the treatment characteristics below. The addi- tional observable characteristics are discussed in the next subsection.

The main treatment characteristics are implemented as follows. First, all pieces of mail comprise a birthday-like envelope containing a card with enscription "Fe- liz Día" or "Happy Birthday" depending on whether the sender is familiar or for- eign. Secondly, if the mail is money mail, it contains two U.S. dollar bills ($2.00) folded in half. This serves as bait. If the mail is no-money mail, it obviously does not contain money. However, it does contain a small lottery number in the lower right-hand corner. The relevance of the lottery number is explained later.

Secondly, since the mail is sent in four batches, we need four names (two famil- iar and two foreign) if we are to randomize (i.e., sample without replacement from) sender last names across batches. The familiar names are "L. Last Name" and "M. Last Name." The foreign names are "P. Thomson" and "J. Scott."61 _{So, for exam-}

ple, if for a particular batch an envelope is randomly chosen to go to a household with last name Perez coming from familiar1, then the sender is L. Perez. If famil- iar2, then M. Perez. The enscription in the envelope is adjusted according to the sender last name.

For each household, we randomize the main treatment observables across batches by sampling without replacement and keeping in mind that each household must be sent a money and a no-money envelope from both a familiar and a foreign sender. So, ultimately, each household is sent two money and two no-money envelopes at random points in time and each envelope randomly comes from one of the four senders.

61_{The experiment was conducted in two stages–a pilot stage and a large stage. In particular,} the pilot stage only comprised money mail. In the pilot stage, we did not randomize across all observables. Furthermore, the foreign name was Mike Tucker. We control for these discrepancies in the empirical analysis.

According to the theoretical model, if loss of mail is random, then there should be no di¤erence between the probability that money is opened and the probability that no-money mail is opened; i.e., Pr(open_jno_money) = Pr(open_jmoney). So, by comparing the likelihood of delivery across money and no-money envelopes ceteris paribus, we are able to test H1. Furthermore, if variants of a particular observable characteristic x have no e¤ect on the likelihood of mail delivery, then Pr(open_jx) = Pr(open_jx0_{). In particular, if xrepresents the sender’s last name and a handler}

infers nothing from the fact that mail has the same sender and recipient last name (familiar) as opposed to a distinct last name (foreign), there should be no discrep- ancy between the above conditional probabilities controlling for other factors. So, by comparing the likelihood of delivery across familiar and foreign ceteris paribus, we are able to test a special case of H2.

Two …nal comments are necessary. First, note that as the experimenter we con- trol two main aspects given what we want to learn from conducting the experiment. In particular, the theoretical model tells us that two aspects determine corruptibility– bait and an observable characteristic such as matching of sender/recipient last name. Thus, we control those and vary them systematically to tease apart any treatment e¤ects. Under usual circumstances a sender of mail does not observe the counter- factuals, which can be a confounding e¤ect.

Secondly, note that the theoretical model is a one-shot simultaneous-move game of complete information. So, strictly speaking, it does not allow for repeated inter- actions between the sender and a handler. However, with an average of one month between mail batches, we …nd that the model represents a reasonable approxima- tion. Furthermore, since di¤erent senders randomly send di¤erent pieces of mail at di¤erent points in time, it is not unreasonable to expect that distinct handlers handle distinct pieces of mail. This is consistent with the mail game and with the interpretation of mixed-strategy equilibrium discussed earlier.