Primary Outcomes (explanation)
1) Exercise habits: We will use scanner data from the student gym to test whether there is a causal effect on gym habits in subsequent semesters. We will consider two dimensions: i) do people buy a gym card/visit the gym at least one time (capturing their intention); ii) total number of gym visits per semester (capturing actual habit). We expect the largest effect in the fall 2016, and will use that period as the main test, but we will also analyze the accumulated effects until the spring 2018. As a point of reference, in the intervention period (spring 2016) 11.6 percent of control group students visited the gym at least one time, whereas the same number in the treated group was 68.6 percent (a 57 pp. difference), while the treatment effect on the number of gym visits was 0.63 S.D. We expect smaller effects in subsequent semesters when treated students also need to pay for their gym card.
2) Academic performance: In the short-term analysis we pre-specified that we would investigate the effects on both completed study points in graded courses (CSP) and the grade average conditional on passing a course (GA). We also pre-specified that we expected a larger treatment effect among students that scored low on life satisfaction, health, study hours and self-control. The analysis revealed that students passed more courses as shown by a significant treatment effect on CSP (≈0.15SD), while maintaining the same grade average on completed courses (no significant effect on GA, but positive point estimate). We also found that the effect on CSP was larger (≈0.40SD) in the subsample that struggled along the four dimensions mentioned above. In the long-run analysis on academic performance we will continue to focus on both of these measures, to understand whether the increase in completed study points in the spring 2016 helped students to complete more courses / receive higher grades in the next semester, or whether it has a crowding-out effect implying that treated students complete fewer courses in the next semester. We will also report treatment effects for the full sample and separately for the subsample that struggled initially. As for exercise, we expect the largest long-term effect (in absolute terms) during the fall 2016, and we will use that period for the main test. In addition, we will analyze accumulated effects from the spring 2016 until the spring 2018. It should be noted that we used registered study points (RSP) as a control variable in the short-term analysis. That was uncontroversial since students registered for courses prior to the intervention, and the variable balanced across treatment. Because of the treatment effect on completed study points in the spring 2016, there could be a difference in how many courses students register for at the beginning of the academic semester in the fall 2016 and onwards. Hence, RSP in the fall 2016 is a potential outcome variable in the long-term analysis, and to use it as a control variable (to increase power) we need to confirm that it balance across treatment. It should also be mentioned that we will record CSP and RSP as zero if there is no data in the academic records for a person in a given semester. The rationale is that the person has neither registered nor completed any study points at one of the two academic institutions. We will discuss in detail how to interpret the estimated effect on academic success. Given our definition of how to record CSP and RSP, people who have successfully finished their Bachelor or Master degree and entered the labor market will get the same value (i.e. zero) as university dropouts. We will therefore compare the average treatment effect on the full sample with the average treatment effect on the subsample of students that was in their first or second year of study in the spring 2016. These students should be unlikely to have completed their degree before the fall semester 2016, and we can thus assume that any missing observations in this subsample can be ascribed to the person dropping out of university rather than anything else. Importantly, we intentionally over-sampled bachelor students in the first two years of their studies, to facilitate the study of long-term effects on academic performance, and they turned out to make up about 75% of the total sample. Hence, although power is slightly reduced in this analysis we have a meaningful sample size to make qualitative judgements that will inform us about the size and sign of the bias in the overall sample. In explorative analysis on academic performance we will also study the effects on the likelihood and time to complete a Bachelor/Master degree.
3) Overall life outcomes: In May 2018, two years after the intervention, participants answered an endline survey (92% response rate ). The survey covered many aspects including: health (BMI, and a lifestyle index variable consisting of self-reported satisfaction with health, sleep habits and alcohol usage); a self-control index consisting of the answer to four questions (resist temptations, follow plans, tendency to procrastinate, think before acting); life satisfaction (overall, social sphere, financial, academic, work); and occupational status (full time employed, part time employed, in higher education, other). In the short-term analysis we found a 0.15 SD treatment effect on both the lifestyle index and the self-control index. The question is whether those effects remain, and, in addition, if there are any effects on BMI, the likelihood to be (un)employed, and life satisfaction.