The school climate survey, 2018-19 to 2024-25

1. Coverage

Schools with a safety composite by respondent group and year, against every active school in the sector (the “district” column therefore includes alternative, ESE-center, technical and adult schools, most of which are not surveyed). Charter participation is partial; district-run participation is near-complete.

Like for like, in 2024-25: among active regular schools with students in grades 4 and up, 310 of 314 district-run schools and 76 of 130 charter schools have a student safety composite.

2. How the composites are built

• Each workbook lists, per school and item, the share of respondents agreeing. Item wording and numbering change between years, so items are matched on wording into canonical items (the crosswalk); a composite uses only the core items asked in every year for that group, so it is comparable over time. Negatively worded items (violence, gangs, class size, lack of support) use the disagreement share; positive items use the agreement share. Undecided or unknown responses remain in the denominator and receive no favorable credit in either direction.
• Domains follow the district’s own sections: safety and order, leadership, teaching and learning, resources and support. The staff workload item is matched across its older and newer wordings and retains the older publisher’s teaching-and-learning category. A school needs at least half of a domain’s core items unsuppressed, and the district withholds schools with fewer than 10 responses.
• The 2021-22 workbooks publish only pivot tables; their per-school records were read from the workbooks’ pivot caches. Percent agreeing is agree ÷ (agree + don’t know + disagree) where the three are published and sum further from 100 than rounding allows (the 2024-25 workbooks carry rows whose published shares exceed 100%; the response counts on the same rows confirm the corrected values). Rows whose three answer shares all equal zero have no usable distribution and are treated as missing. Response rates for parents and staff use the workbook’s eligible count where it exists; the students’ rate is computed against the same year’s October membership in grades 4–12 because the workbooks’ own eligible column alternates between all-grade and grades-4-and-up counts.

Normaliser: pipeline_miami/parse/climate_survey.py · loader: climate_survey.ts · metric pages: students, parents, staff.

3. Does sentiment predict enrollment loss?

Pre-registered before the regression was run: for district-run schools open in both years, the log change in October membership from the survey year to 2025-26 is regressed on the survey measure, controlling for that year’s utilization (headcount ÷ permanent capacity), the school-grade letter, school level and log enrollment. The expectation was a positive coefficient that is small next to the utilization and grade terms. The state's first capacity report is 2022-23, so the 2021-22 control is that cohort's own October 2021 headcount over the 2022-23 permanent capacity.

Survey associations with enrollment change, using overlapping school samples and different follow-up periods. In the 2021-22 survey, a one-standard-deviation higher parent recommend share (12.13 points) goes with +6.9% enrollment four years later (p <0.001), adding 10% of the remaining variance after the controls, which themselves explain 6% on the same 256 schools. In the 2022-23 survey, one year shorter, the same parent measure gives +2.0% per standard deviation (n = 265, p 0.113). The two cohorts as published are different school sets (the 2022-23 workbook withholds fewer schools under 10 responses); on the 242 schools in both, the parent effect is +7.1% per standard deviation in 2021-22 and +3.5% in 2022-23 (t 2.9). A paired school bootstrap puts the difference at p 0.003. The parent share has year-to-year correlation r 0.60 (the table gives each measure). These correlations describe stability; they do not distinguish sampling noise from changes in sentiment or who responds, and do not explain the difference between coefficients. The student recommend estimates are: +5.9% per standard deviation from 2021-22 (p <0.001) and +5.3% from 2022-23 (n = 289, p <0.001). 12 coefficients are reported; after a Holm correction over that family, 7 remain significant at 5%: Parents: would recommend 2021-22, Students: would recommend 2021-22, Parents: teaching and learning 2021-22, Students: safety and order 2021-22, Staff: leadership 2021-22, Students: would recommend 2022-23, Staff: leadership 2022-23. Both windows end in the same enrollment year and reuse many of the same schools. One window spans four years and the other three; the total-change coefficient comparison is not an independent replication of an identical follow-up period. (Until the 2026-09-16 review the 2022-23 parent measure was missing from the load, so the parent result had been reported on 2021-22 alone; the same-school comparison, the bootstrap and the Holm correction were added after the 2026-09-16b review and were not pre-registered.)

This is an association in observational data. Parents who are about to leave may already rate the school lower, and a school losing students may be losing the families most likely to answer. The test asks whether the survey adds information about later enrollment beyond utilization and the school grade; it does not estimate the effect of changing sentiment.

Engine: pipeline_miami/analysis/survey_enrollment.py · packet dd43e33b8f8c generated 2026-09-16 22:43:08 UTC · packet JSON.

4. What this cannot show

• Response rates are low and uneven, and only some years publish the eligible count; the survey’s main bias is who answers.
• Composites are this project’s unweighted means of favorable item shares (agreement for positive items, disagreement for negative items); no respondent-level data exist to weight or to split by student group.
• Charter schools are under-represented, so sector comparisons on the survey are not like for like.

Corrections log

2026-09-16: zero-total response distributions are missing, unknown answers receive no favorable credit on negative items, and the staff workload item is matched across all seven years. Earlier analytical results remain hidden until recomputed with those corrections. 2026-09-16 (later): the 2018-19 staff workbook was retired. It publishes agreement and disagreement only, with no don’t-know column, and the two sum to a median of 83% (10th percentile 50%); on the negative-worded core items its disagreement is 25% against 65–70% in every later year while agreement is stable, so under favorable-response scoring the 2018-19 staff composites sat 13–20 points below 2019-20 and under the earlier 100 − agreement rule the missing category counted as favorable. The staff series now starts in 2019-20 and its core item set is recomputed over six years (32 items, one more than before).