Student Preferences and School Choice: A Discrete Choice Experiment

The study adds two randomized information treatments to an online survey given to ninth grade students who are about to choose a secondary school. The main tool in the survey is a discrete choice experiment, and the treatments change the information students see just before it. Each student was randomly placed in one of three groups: a control group (introduction only), a wage premium group, or a career reflection group.
The discrete choice experiment. Each student completed a set of choice tasks. In each task, the student saw two made-up secondary school profiles and a third option that meant leaving school (not enrolling in high school). The student picked the option they liked most. Each school profile was described by six features, and the level of each feature was chosen at random for each profile: schedule (full-day, part-time, or evening), curriculum (vocational training, life project, higher education exam preparation, a focus on social, hard, or life sciences, or no special curriculum), distance from home (15, 30, or 45 minutes), school quality (an IDEB score of 3, 5, or 7), social environment (whether friends, parents, or teachers recommended the school, or none), and a monthly scholarship (from R$0 to R$150). Each task showed only some of the features, chosen at random (a partial profile design). Because the feature levels are set at random and independently, the design shows how each feature changes the chance that a student picks a school. It also lets us measure how much each feature is worth in money, using the scholarship amount as the money scale.
Treatment 1: Wage premium information. Before the choice tasks, students in this group saw information about how much more people earn when they finish high school, compared to finishing only lower secondary school. This information changed in two ways, shown in all combinations: the size of the wage gain (low, shown as a monthly wage of R$1,300 versus R$1,100; or high, shown as R$1,800 versus R$1,100) and the state of the local job market (described as warm, with many job openings, or cold, with few). As a small design detail, the card was shown in one of two colors (blue or red). This was a minor visual framing change that kept the wage and job market information the same. It is not a focus of the study and is not analyzed here. A control group saw no job market information.

Treatment 2: Career reflection. Before the choice tasks, students in this group answered two questions about their future: which job they would like to have, and whether they think that job needs a high school diploma. This treatment brings the student's own career goals to mind right before they choose. It does not give students any outside information. It only asks for their own goal and their own belief about whether that goal needs a high school diploma.
The treatments test whether bringing the returns to schooling, or one's own career goals, to mind changes how much students value staying in school and the school features they care about.

Setting and administration. The study was run as part of an online survey conducted by LEPES/USP. School principals shared the survey link with their ninth-grade students, and the discrete choice experiment was the first task in the survey. The sample includes 14,588 students who completed at least one choice task, giving 85,507 choices and 255,021 choice observations.

The discrete choice experiment. Each student completed six choice tasks. In each task, the student saw two made-up secondary school profiles and a third option that meant leaving school (not enrolling in high school), and picked the option they liked most. The third option was clearly described as leaving the school system, not as "none of the above." Each school profile was described by six features, and the level of each feature was set at random and independently for each profile:

Schedule: full-day, part-time, or evening
Curriculum: vocational training, life project, higher education exam preparation, a focus on social sciences, a focus on hard sciences, a focus on life sciences, or no special curriculum
Distance from home: 15, 30, or 45 minutes
School quality: an IDEB score of 3, 5, or 7 (5 is the national median)
Social environment: friends, parents, or teachers recommended the school, or no one did
Monthly scholarship: from R$0 to R$150
The design uses partial profiles: each task showed only some of the six features, chosen at random. When a feature was not shown, it was set to its reference level in the analysis. Three features were shown in every task (schedule, distance, and curriculum); the other three (quality, social environment, and scholarship) were shown only some of the time. The reference levels are part-time schedule, 15 minutes distance, general curriculum, IDEB 5, student chooses the school, and a R$50 scholarship.

Random assignment to information conditions. Before the choice tasks, each student was randomly placed in a control group (introduction only), a wage premium group, or a career reflection group.

The wage premium group saw information about how much more people earn after finishing high school compared to finishing only lower secondary school. This information varied in three ways, set at random in all combinations: the size of the wage gain (low, shown as R$1,300 versus R$1,100 per month; or high, shown as R$1,800 versus R$1,100), the state of the local job market (warm, with many openings, or cold, with few), and the color of the card (blue or red). The color was a minor visual framing change that kept the wage and job market information the same; it is not a focus of the study and is not analyzed. The analysis groups the cards into four arms by wage gain and job market state, each with about 18,000 to 20,000 observations, against a control group of about 179,000 observations that saw no job market information.

The career reflection group answered two questions before the choice tasks: which job they would like to have in the future, and whether they believe that job needs a high school diploma. This brings each student's own career goal to mind right before the choices. It does not give students any outside information; it records the student's own goal and their own belief. Based on their answer, students fall into two groups: those who believe their chosen career needs high school, and those who believe it does not. The second group is small (about 3,555 observations, near 2.4 percent of this sample) because most students name careers that need high school. This part of the analysis uses the students who appear in both the choice data and the student survey (about 12,730 students): control with 73,035 observations, the "needs high school" group with 70,956, and the "does not need high school" group with 3,555.

Estimation. The choices are analyzed with a conditional logit model with choice-set fixed effects, since each student answered several tasks. Two versions are estimated. The first enters schedule and curriculum as separate features. The second combines schedule and curriculum into school types (for example, full-day with vocational training), because in real life schools do not offer a schedule without a curriculum. Each feature's weight is turned into a money value (willingness to pay) by dividing it by the weight on the monthly scholarship. To study group differences and treatment effects, the same model is estimated separately by subgroup and by treatment arm, and the weights and money values are compared.

Note on the career reflection groups. Being in the career reflection treatment was randomly assigned, but whether a student ends up in the "needs high school" or "does not need high school" group is based on the student's own stated belief, not on random assignment. Comparisons across these two groups should be read with that in mind.

2021-04-11

2022-03-30

The main outcome is the choice each student makes in each task: which of the three options they pick, one of the two school profiles or the option to leave school. This is recorded as an indicator for the chosen option. From these choices we estimate how much students value each school feature and how much they value staying in school instead of leaving. We also measure whether the two information treatments change these values, in particular the value of staying in school.

The choices are analyzed with a conditional logit model. For each school feature, the model gives a weight that shows how much that feature changes the chance a school is chosen. We turn these weights into money values (willingness to pay) by dividing each feature's weight by the weight on the monthly scholarship, so the result is in reais per month. The value of staying in school is the weight on the school options relative to the option to leave. To measure treatment effects, we estimate the same model separately for the control group and for each treatment group and compare the weights, focusing on the value of staying in school.

The same choice variable is used to study how preferences differ across groups of students, defined by gender, race, and socioeconomic status, both at baseline and in response to the two treatments.

Subgroups are defined by student gender, student race (White versus Non-White), and a socioeconomic status index built from household assets (such as number of bedrooms, bathrooms, cars, computers, smartphones, and internet access), split into thirds. For each subgroup we estimate the same conditional logit model and compare weights and money values across groups using z-tests for differences between separately estimated models.

This is a survey experiment with ninth grade students who are about to choose a secondary school. It has two layers of random assignment.

The first layer is the discrete choice experiment, run within each student. Each student completed six choice tasks. In each task the student saw two made-up school profiles and a third option that meant leaving school, and picked the option they liked most. Each school profile was described by six features (schedule, curriculum, distance, school quality, social environment, and a monthly scholarship), and the level of each feature was set at random and independently for each profile. Because the levels are random, the design shows how each feature affects the chance a student picks a school, and how much each feature is worth in money.

The second layer is between students. Before the choice tasks, each student was randomly placed in a control group (introduction only), a wage premium group (shown the extra earnings from finishing high school), or a career reflection group (asked about their future job and whether they think it needs high school). This lets us test whether the information changes how much students value staying in school and the school features they care about.

The choices are analyzed with a conditional logit model, and feature weights are turned into money values using the scholarship as the money scale.

Within-student design (choice tasks). Each student answered six tasks, each with two school profiles and a leave-school option. The six features and their levels were: schedule (full-day, part-time, evening); curriculum (vocational training, life project, higher education exam preparation, focus on social sciences, focus on hard sciences, focus on life sciences, or none); distance (15, 30, 45 minutes); school quality (IDEB 3, 5, 7); social environment (recommended by friends, parents, or teachers, or no one); and monthly scholarship (R$0 to R$150). Levels were set at random and independently across profiles. The design uses partial profiles: each task showed only some features, chosen at random, and features not shown were set to their reference level in the analysis. Schedule, distance, and curriculum appeared in every task; quality, social environment, and scholarship appeared only some of the time. Reference levels are part-time, 15 minutes, general curriculum, IDEB 5, student chooses, and R$50.

Between-student design (information conditions). Students were randomly assigned to control, wage premium, or career reflection. The wage premium condition used a factorial design that crossed three randomized features in all combinations: wage gain size (low: R$1,300 vs R$1,100; high: R$1,800 vs R$1,100), local job market (warm or cold), and card color (blue or red). Color was a minor visual framing change with no information content and is not analyzed; the analysis groups the cards into four arms by wage gain and job market. The career reflection condition asked two questions (desired job, and whether the student believes that job needs high school). Whether a student lands in the "needs high school" or "does not need high school" group reflects the student's own stated belief and is not randomly assigned.

Samples. The full sample is 14,588 students, 85,507 choices, and 255,021 choice observations. For the wage premium analysis, the control group has about 179,000 observations and each of the four arms about 18,000 to 20,000. The career reflection analysis uses students who appear in both the choice data and the student survey (about 12,730 students): control 73,035 observations, "needs high school" 70,956, and "does not need high school" 3,555.

Estimation. Choices are analyzed with a conditional logit model with choice-set fixed effects. Two specifications are used: one that enters schedule and curriculum separately, and one that combines them into school types (since schools do not offer a schedule without a curriculum). Each feature's weight is divided by the scholarship weight to express willingness to pay in reais per month. Group differences (by gender, race, and socioeconomic status) and treatment effects are estimated by running the same model separately by subgroup and by treatment arm and comparing weights and money values, using z-tests for differences between separately estimated models.

Randomization was done automatically by the online survey software (Qualtrics). For each student, the software randomly assigned the information condition and randomly set the feature levels shown in each choice task. No manual or public lottery was used.

There are two levels of randomization, both handled by the survey software:

Information condition (control, wage premium, or career reflection, including the wage premium sub-arms): randomized at the individual student level. Students in the same school were assigned to different conditions.

School feature levels in the choice tasks: randomized within each student, at the level of the choice task and the school profile, so the two profiles in a task and the six tasks each student saw varied independently.

No

15,170 students. The treatment was not clustered, so the unit of randomization is the individual student. Students came from 332 schools across the state of Paraiba, but assignment to conditions was made at the student level, not the school level.

15,170 students were randomized. The analysis sample is 14,588 students who completed at least one choice task. Because each student completed up to six choice tasks, and each task had three alternatives (two school profiles and a leave-school option), this produces 255,021 school-choice observations, which are the observations used in the conditional logit model.

Control (introduction only): 5,114 students.

Wage premium (shown labor market information): 4,972 students. These were randomly split across eight cells crossing wage gain size (low or high), job market state (warm or cold), and card color (blue or red), with about 520 to 595 students per cell. The analysis groups these into four arms by wage gain and job market state, about 1,240 students per arm. Card color is not analyzed.

Career reflection (asked about their future job): 5,084 students. Based on the student's own answer about whether that job needs a high school diploma, 4,747 believe it does, 218 believe it does not, and 119 did not answer. The "does not need high school" group is small because most students name jobs that require high school.

Why does secondary education fail to engage students despite decades of investment in schools, teachers, and financial incentives? We take a demand-side approach to this question and provide evidence that retaining students requires schools that offer what students value and that support students in clarifying their long-term goals. Using a discrete choice experiment with over 14,000 students at the secondary transition in Brazil, we elicit stated preferences over school attributes, including the option to leave school, and estimate willingness to pay for each. Students value enrollment at a regular school at only 85\% of the minimum wage, suggesting that even entry-level employment could outweigh the perceived value of schooling. Nonetheless, changes in school quality and curriculum might have large effects on the value of schooling. Curriculum changes, for example, double the odds of choosing a school. After taking the average perceived value of schooling, we also test whether long-term goal reminders affect them. Using a randomized experiment, we ask students about their career goals. This nudge halves the perceived cost of opting out for those whose careers do not require a diploma.

Araujo, Daniel, Gil do Vigor, and Leonardo Rosa. 2026. "School Choice Dynamics: Preferences, Heterogeneity, and Information Treatments in Brazilian Secondary Education." Working Paper.

https://drive.google.com/file/d/1uKaGv9bYDdNVb4zGM7El7pXAXpuGxUEk/view?usp=drive_link