Determinants of gang entry

Experiment 1: Treated subjects receive information on expected earning potential among three groups: those who do not complete high school; those who complete high school; and those who complete tertiary education. Half of the treated subjects receive the information in the form of statistics and graphical figures. Half of the treated subjects receive the information in the form of a story about three friends who took different educational and career paths.

Experiment 2: Treated subjects watch a video featuring three actors who provide dialogue directly taken from interviews with current and former gang members. The selected dialogue emphasizes how they had high expectations of enjoyment and status from gang careers but how in the end status and enjoyability were low and friend, family and community relations were difficult.

Note that this experimental design was updated in September 2023 to reflect the second information experiment launched in the same month.

Experiment 3: (On hold). We originally envisioned a third experiment where treated subjects receive information designed to change beliefs about the probability of arrest, incarceration, and death in gang and criminal careers. This is on hold, in the event that the sample size of surveys can be expanded.

2023-05-12

2024-05-17

Our primary survey-based outcome is an indicator variable for high self-reported interest in gang careers.

For Experiment 1, we also pre-specified a primary behavioral outcome--the take-up rate of offers of information on higher education and other educational resources. After concluding Experiment 1, it appeared that the information treatment was not persistent enough to change behavior some days to a week after treatment. Prior to launching Experiment 2, we therefore respecified it as a secondary outcome.

The survey collect data on beliefs about the probability of following four paths: legal careers without high school completion; legal careers requiring high school; legal careers requiring tertiary education; and illegal careers with a gang. The respondents are presented with the four paths and 10 tokens. We ask the respondents to pretend the tokens are 10 clones of themselves and place them according to where they expect their future selves to be at age 25. We repeat this question two additional times that resemble the constraints that some students face: without one career option (legal careers requiring tertiary education); and without two career options (legal careers requiring high school; legal careers requiring tertiary education). We then average the number of clones each student puts in the illegal careers with a gang across the three questions. This average is our survey-based outcome of interest in gang careers. In pilot data, roughly 18% of respondents place 2 or more of their future selves in the gang and roughly 8% place 3 or more. We have created an indicator for this relatively "high" level of interest, with the aim of capturing the (approximately) top decile of risk. We measure this outcome in the survey itself as well as in a subsequent online follow up survey some days after the in-person survey. This follow-up measure (and the effect of time on responses) is a secondary analysis.

Our main secondary analysis is the "first-stage" effect of the information treatment on posterior beliefs. For Experiment 1, this is a measure of their beliefs about the returns to higher education. We expect to look at the mean error, but also the perceived differential between higher education and low-education career earnings. For Experiment 2, this is an index of perceived status and enjoyment (non-pecuniary benefits) of gang entry.

Another secondary outcome is a behavioral outcome --- take-up of a WhatsApp-based message to survey respondents, offering them the opportunity to follow a web link to receive information about educational resources and guidance for higher education. Respondents will have individualized links that will enable us to assess whether they click on the website at all. For the behavioral outcome, we also expect to examine levels of engagement with the information offered, including time spent on the web page. (Note that this was specified as a primary outcome in Experiment 1 but not in Experiment 2, for reasons discussed above.)

We also measure both gang interest and posterior beliefs about income in the follow-up survey some days after the in-person survey. This follow-up measure (and the effect of time on responses) is a secondary analysis (as noted above).

The difference in effectiveness between stories and statistics in Experiment 1 will be a secondary analysis as well, focusing both on posterior beliefs and also gang behavior.

Experiment 1:

We expect treatment effects on our primary outcomes to be concentrated among high-risk pupils, and among pupils who underestimate the returns to high-school completion and higher education. We expect our primary outcome regression to take the following form:

Y = b1 T + b2 T*L + b3 T*E + b4 T*L*E + b5 L + b6 E + b7 X + S + e

where
Y is our primary outcome
T is assignment to the information treatment
L is an indicator for low risk (the number of friends interested in the gang career, using the same threshold as Y)
E is a measure of erroneous prior beliefs, centered at zero, where positive values imply overestimating returns to legal careers
X is a set of pre-information experiment control variables taken from survey and administrative data, selected using the double lasso method on primary outcomes
S is a set of school fixed effects
e is an error term (not clustered)

We expect b1 to be positive when the outcome is legal income, and negative when the outcome is gang interest. We also expect the effects to be larger for youth with priors that overestimate legal incomes, and so we are also interested in b3.

Experiment 2:

Experiment 2 began in September 2023 at the time of this update.

As with experiment 1, we expect treatment effects on our primary outcomes to be concentrated among high-risk pupils. In this case, also among pupils with high expectations of non-pecuniary returns to criminal/gang careers, such as status and enjoyability. We expect our primary outcome regression to take the following form:

Y = a1 T + a2 T*L + a3 T*E + a4 T*L*E + a5 L + a6 E + a7 X + S + e

where
Y is our primary outcome
T is assignment to the information treatment
L is an indicator for low risk (the number of friends interested in the gang career, using the same threshold as Y)
E is a measure of expected non-pecuniary returns to criminal/gang careers
X is a set of pre-information experiment control variables taken from survey and administrative data, selected using the double lasso method on primary outcomes
S is a set of school fixed effects
e is an error term (not clustered)

When Y is a measure of non-pecuniary benefits, we expect a1 to be negative. We also expect the effects to be larger for youth with priors that non-pecuniary benefits are high, and so we are also interested in a3.

Not available

Randomization built into tablet-based survey instrument

Individual

No

Approximately 100 schools (though these are not clusters, given the individual randomization)

We expect an experimental sample size of roughly 3,000-4,000 pupils per information experiment, assuming we are successfully able to recruit roughly 10,000 pupils for the surveys. For Experiment 1, we extended this to nearly 6,000 pupils, because the follow-up survey and behavioral measures were introduced late (due to development time). With a follow-up survey response rate of approximately 40%, this allows us to target roughly 1400 to 1600 completed follow up surveys and behavioral engagements.

Pupils are assigned to the information treatment with a 0.5 probability, and so each experimental sample should include 1,500-2,000 control and treated subjects each. When we also use different means to deliver the information (e.g., statistics vs stories), we randomize within the treatment group with a probability of 0.5 of having one or the other.