Randomization

The field experiment was conducted in the Philippine provinces of Guimaras, Capiz and Iloilo during the month of April 2015 (see figures 2.1 and 2.2).

Subjects participating in the study are clients of NWTF and were randomly se- lected from the pool of clients in the three provinces, belonging to three branches of NWTF. The randomization procedure had two steps. In the first step, suitable cen- ters were divided into bins according to minimum distance between centers, and the bins where randomly allocated to the training treatment and to a control group. In a second step individuals were randomly selected from the treatment and control pools. All clients selected to be treated as well as clients selected as control, took part in a series of experimental games before the savings training. See Penczynski and Santana (2016) (chapter 1) for more details about the experimental games.

Figure 2.1: Location of experimental sessions

The clients across the three NWTF branches were distributed across 175 centers, located in 155 different villages (called barangays). It was not logistically possible to hold sessions in each barangay, therefore several centers were grouped together in a bin, based on their geographical location. In order to select the barangays (centers) that would form a bin, the minimum geographical distance across barangays was computed, and bins with the minimum distance and with approximately 90 clients were selected.

Bins had on average three barangays, and the sessions were held in one of the barangays of the bin. The selection of the barangay where the session was to be held

depended on the facilities, the proximity to the municipality and the accessibility.48

In order to have the information required for the randomization, recruiters sur- veyed the barangays to collect data about the facilities and resources available. The survey, implemented with tablets, enabled the collection of geographical reference data of the barangays and municipalities as well as pictures of the facilities (see ap- pendix A.2). Recruiters also gathered information about the barangay head and the possibility to acquire permission to hold the sessions in the barangay.

The first step of the randomization process assigned the 60 client bins to one of four treatment arms, a two-by-two design with two institutions and a treatment

48_{Four of the bins were composed of only one barangay. In these cases, the entire village was}

Figure 2.2: Location of experimental sessions in detail

and a control for the savings intervention. The institutions arm was related to the experimental games and implied receiving a later payment from either NWTF or Local Money Lenders (ML) (see Penczynski and Santana (2016) for more details). Given that receiving a promised later payment from the microfinance institution, NWTF, could have a positive effect on the trust level of the client towards NWTF and thus have an effect on savings, I constrain the main analysis to those allocated to the ML treatment.

In order to improve precision of our estimates, a rerandomization procedure based on a set of covariates was implemented, following Morgan and Rubin (2012). The procedure uses available data to check for covariate balance across treatment and control groups. If a lack of balance is present, then rerandomization could help to ensure balance. Figure 2.3, taken from Morgan and Rubin (2012), shows the steps of the procedure.

Figure 2.3: Rerandomization procedure

Source: Morgan and Rubin (2012)

To establish the rerandomization criteria, the Mahalanobis distance M was uti- lized:

M = ( ¯XT − ¯XC)cov( ¯XT − ¯XC)

−1

( ¯XT − ¯XC), (2.1)

where ¯XT− ¯XCis the k-dimensional vector of the difference in covariate means between

the treatment and control groups and cov(x) is the sample covariate matrix of x. A randomization is acceptable whenever M is below a certain threshold.

well as an urban area indicator, the population of the bin, and the average distance from the bin to the municipality were taken as covariates. In the rerandomization procedure the distance across the covariates of the four different treatments arms was compared, allowing for comparison of any two treatment arms against each other.

In order to recruit the participants permission from the municipality mayor was requested and subsequently from the village head, locally known as barangay captain. Barangay captains would additionally authorize the use of the village facilities, usually the barangay hall, where the savings seminar was held. Subjects were invited to participate in the experiment via an invitation letter delivered to their houses.

I define as “injection points” individuals that received the savings training and participated in the experimental games. I denote the treatment status of the injection

points with Ti = 1. Individuals that also participated in the experimental games

but did not receive the savings training are the injection points’ control group, and I identify them as control subjects. I denote the treatment status of control subjects as

Ti = 0.

I am also interested in the effects of the training on the peers of the injection points. I define two groups: group peers denoted by g(i) and center peers denoted by

c(i). I denote the dummy variable T_g(i) as the group treatment indicator. T_g(i)= 1 if

and only if there is a group member that did not take part in the training but that had an injection point in his lending group. I refer to them as “treated group peers”.

Consequently, T_g(i) takes a value of 0 for group members of clients that participated

in the experiment, but did not receive the training, which are the group members of those that form the control group of the injection points. I refer to them as “control group peers”.

Finally, Tc(i)indicates the treatment indicator at the center level, and takes a value

of 1 for members that are in centers with at least one injection point, excluding the lending group to which the injection point belongs. I refer to them as “treated center

peers”. Tc(i)takes a value of 0 for members whose center has at least one member that

attended the experimental games but did not participate in the training, excluding the group member of the control subjects. I refer to them as “control center peers”.

I have 205 injection points and 197 control subjects, a total of 402 individuals.49

They are distributed across 85 centers, out of which 42 have at least one injection

point (Tc(i)= 1), and 43 have at least one control subject (Tc(i)= 0).

I have 352 individuals as peers in treated groups and 360 individuals as peers in control groups, a total of 712 peers at the group level. At the center level, I have 381 individuals as peers in treated centers and 394 individuals as peers in control centers,

49

Initially, we had 256 injection points and 243 control subjects, a total of 499 individuals. I lose 97 individuals due to the fact that their loans end and that I only observe active clients. I focus the analysis on clients that were active at the time of the savings training. Since I have data only on active clients, it is possible that a client that was on a loan cycle break at the time of the training and therefore not marked in April as an active client, appears in subsequent periods in my sample, once she takes up the next loan. For now, I abstract from these clients.

a total of 775 center peers.50