Analysis

Generated by scripts/analyze.py from data/, which is built from Appendix B of TDSB report 11-17-3269 (15 November 2017). No individual-level data is used or available.

1. Who stops answering, and where

The published bases are not constant. Every item has its own denominator, and the spread is wide enough to be worth reading as data rather than noise.

For students the base ranges from 13,699 to 15,444 — a spread of 1,745 respondents, about 11% of the largest base.

The pattern is ordered: demographic items average 14,157 respondents against 15,055 for the SRO items — a drop of about 898. The demographic block sits at the end of the instrument, so ordinary survey fatigue is the obvious reading; the alternative — that students in an anonymous survey about police became less willing to state their race, orientation or disability specifically — is consistent with the same numbers and cannot be separated from fatigue without the microdata.

Within the SRO items the lowest base is 14,853 (uncomfortable to attend school) and the highest 15,444 (did you know your school had an SRO). Items asking students to report harm to themselves draw slightly fewer answers than items asking what the officer does.

2. The size and shape of 'Not Sure'

The largest single feature of this dataset is not approval or disapproval. It is uncertainty, and it is patterned.

Officer-evaluation items, share answering 'Not Sure':

Two things move together here. Uncertainty rises as the item moves from something you can see (visible, 36.0%) to something you must infer about a person's disposition (trustworthy, 53.4%). And the student–staff gap widens along the same axis: on visibility students are 22 points more unsure than staff; on trustworthiness, 26 points. Staff are not merely more positive — they are more willing to answer at all.

3. What contact does and does not explain

78.6% of student respondents report no interaction with the officer at all. That is the ceiling on first-hand knowledge.

Frechet bounds on the joint distribution of 'never interacted' and 'Not Sure', from the two marginals alone:

The lower bounds are the informative half. On trustworthiness at least 60% of the students who answered 'Not Sure' had never spoken to the officer; on visibility at least 41%. But the bounds also close off the simplest story: because non-contact (78.6%) exceeds every uncertainty share, a substantial group of students who never interacted nonetheless returned a definite judgement. On trustworthiness that group is at least 25% of all respondents.

So contact is not a sufficient account of uncertainty, and its absence is not a bar to opinion. Roughly a quarter of the sample rated an officer they never met.

Among the 4,354 students who did interact, 79.7% were comfortable or very comfortable and 20.3% uncomfortable or very uncomfortable. Reported on the full base those same numbers read 22.9% and 5.9%, which is how the same survey supports both a reassuring and an alarming summary.

Conditioning on contact removes the student–staff difference almost exactly. Among staff who interacted, 20.3% were uncomfortable, against 20.3% of students — -0.0pp (95% CI -3.1pp to +3.0pp), an interval that sits on zero. On the unconditional numbers staff look far more comfortable than students; the appearance is produced by who had contact, not by how contact went.

The intensity differs even where the rate does not. Of the 159 uncomfortable staff, 130 chose the extreme category ('very uncomfortable') — 82% — against 539 of 884 students, or 61%. Staff discomfort, when it appears, is more likely to be reported at full strength.

4. Where the three surveys disagree

Matched items only. Two of these comparisons are not like-for-like and are marked: students were asked about themselves, staff about 'some students', parents about 'my child'. A parent reporting on one child cannot produce the same quantity as a student reporting on themselves.

On the four comparable attitude items the ordering is identical every time: parents most favourable, staff next, students least. The gap is widest on value to the school (parents 67.4%, students 29.9%) and on continuation (77.8% against 46.9%). Every one of these differences clears conventional significance, which at n = 15,000 against n = 469 is close to guaranteed and should not be read as evidence of importance; the percentage-point gaps and their intervals carry the substance.

Awareness is the exception to the ordering and sets a floor under everything else. Here staff lead at 82.5%, ahead of parents at 64.8% and students at 46.5% — so on every downstream item a majority of student respondents are answering about a programme they had not registered was there, while the staff they are being compared against mostly had.

The harm items run the other way and are the ones that are not comparable. Students report on themselves and 14.7% agree they felt targeted; staff, asked whether some students felt targeted — a far weaker existential claim, since one student anywhere in the building satisfies it — agree at 18.3%; parents, asked about their own child, agree at 10.4%. The three numbers look like a narrow spread around 15% and are in fact answers to three different questions. Any table that stacks them, including the one on the current website, invites a comparison the instrument does not support.

5. The same students, two referents

Students were asked how valuable the programme is to you (Q7) and to your school (Q8). Same respondents, same scale, adjacent items — so the comparison is within-person in everything but the linkage.

Rating the programme very or extremely valuable rises from 18.1% for themselves to 29.9% for their school, a gap of +11.8pp (95% CI +10.8pp to +12.7pp). Dismissal moves the other way: 19.9% call it not valuable to them, 11.2% not valuable to the school.

Students consistently locate the benefit somewhere other than themselves. Whether that is displaced civic endorsement or an accurate report that the officer's work touched other people cannot be settled from marginals — but the gap is large, it is internal to one survey, and it does not depend on any cross-survey comparison.

6. Counts and rates point in different directions

The board discontinued the programme. Read as rates the survey does not obviously support that; read as counts it does. Both readings are in the report, and the executive summary uses counts for harm and percentages for approval.

A rate of 14.7% is a minority finding. A count of 2,207 students is a policy problem with a constituency. Nothing separates them except the denominator, and the denominator is a survey base, not the student body — the 45 schools enrolled well over 15,217 students, so every rate here is a rate among volunteers.

The report's own framing tracks this. Approval appears as percentages ('57% said that having an SRO made them feel safer'); harm appears as frequencies ('2207 students noted...'). Both are accurate. The choice of form is doing argumentative work, and it is visible in the text.

7. The one disaggregated number, and what it implies

The report publishes exactly one figure broken down by any student characteristic: 'including 50% of Black students who completed the Student Survey' felt the programme helped them feel safe (p.4). Everything else is reported for all respondents together.

Black respondents are 15.9% of the demographic sample (2,267 of 14,272). Overall agreement that the SRO made school safer is 57.0%. If the published 50% for Black students is right, then agreement among all other students works out at about 58.4% — a gap of roughly 8 percentage points.

Three cautions, and they are not small. The race item has a different base (14,272) from the safety item (15,119), so the arithmetic assumes the non-overlap is ignorable. The published 50% is rounded, and a true value anywhere in 49.5–50.4 moves the implied other-student figure by about a fifth of a point — small. And 'all other students' is not a group anyone belongs to; it pools eight categories that almost certainly differ from each other.

Still, this is the only place in the published record where the differential impact the board acted on can be quantified at all, and it survives the arithmetic.

What the microdata would and would not add

Frechet bounds show which unobserved crosstabs are already pinned down by the marginals and which are wide open. For a demographic group of size p and an outcome of size q, the within-group rate is bounded below by max(0, p+q−1)/p and above by min(p, q)/p.

Every row runs 0–100%. A group is constrained from below only when its share exceeds 43.0% — one minus the outcome share — and the largest racial category here is South Asian at 23.7%, well under that. No group clears it. The marginals therefore place no restriction whatever on the race-by-outcome pattern: the published tables are consistent with Black students agreeing at any rate from 0% to 100%.

This is the precise sense in which the microdata are necessary, and it is worth stating exactly, because the intuition that 'the aggregates must constrain the subgroups somewhat' is wrong here. The 50% figure in the report body is not recoverable from Appendix B — it comes from a tabulation the board ran and did not publish. Whoever has that tabulation has the file we want.

8. Where the qualitative consultation happened

The surveys are board-wide; the focus groups and community meetings are not. Their geography is itself a finding, because the report's qualitative conclusions are markedly more negative than its survey results.

The 306 focus-group participants are about 2.0% of the student survey base, and they are not drawn evenly: Northwest accounts for the largest share of sessions, and three of the sixteen sessions are with Emery CI students alone. The report is explicit that it set out to centre students 'traditionally excluded, marginalized and/or discounted', so the skew is by design and not a defect — but it does mean the qualitative and quantitative strands describe different populations, and the divergence between them is partly a sampling artefact.

9. What cannot be done without the microdata

• Any crosstab. Race by outcome, gender by outcome, contact by evaluation, grade by anything. The bounds in section 7 show the marginals do not constrain these enough to be useful.
• Any regression or adjustment. No covariates can be held constant.
• School-level analysis. The 45 schools are never identified in the survey tables, so nothing can be linked to school composition, the Learning Opportunities Index, or officer assignment.
• Anything about non-respondents. The base is volunteers within schools that had an officer; there is no sampling frame and no weights.
• Change over time. One cross-section, collected while the programme was suspended — students were reporting on an officer who was not currently there.

That last point is easy to miss and it conditions everything above. The board suspended the SRO programme on 30 August 2017 pending the review; fieldwork ran 18 September to 25 October. Every response is retrospective.