What predicts a school's ties to outside organizations

Choose which schools to look at

Loading the survey…

Every percentage is weighted with the School Survey on Crime and Safety's own weights, because unweighted percentages of it do not describe US schools. The unweighted number of schools behind each figure is shown beside it, and a selection under 50 schools is suppressed rather than drawn.

This page recomputes no regression. It shows weighted shares and averages for whichever schools you select. Every coefficient elsewhere on the site comes from the analysis code, with standard errors clustered on the sampling stratum within each survey.

⬇ Download all 13,167 schools (CSV)

other comparisons

Low-crime against high-crime areasthe paper's main finding: schools in areas principals call high-crime work with more organizations, 3.62 against 3.34, and fewer of them work with the police, 72% against 86%. Raw shares with nothing controlled; the estimates are on Full results
2005-06 against 2021-22which organizations schools worked with changed while how many did not: juvenile justice agencies down from 42% of schools to 26%, mental-health providers up from 55% to 78%, and 3.44 organizations per school then against 3.24 at the end
City against rural schoolswhere a school is shows up in which organizations far more than in how many: 3.46 per city school against 3.40 per rural school, while 75% of city schools work with the police against 87% of rural ones
Primary against high schoolswhat a school runs for itself predicts the count at both grade levels, +0.42 of an organization per standard deviation in primary schools and +0.26 in high schools, which is one of the paper's checks. High schools work with more to begin with, 3.90 against 3.21

$ data & code run real R in this tab no install · R, Python, curl, JSON, citation

static files · CORS open · no key · no rate limit

This is a real R session, running in your browser through webR. Nothing is sent anywhere. The survey is already loaded as schools, a data frame of 13,167 rows, with the survey weight in w. The first run downloads R itself, about 30 MB.

R not started

Press Run. The first run takes a moment while R downloads.

# the analytic file: 12,691 school responses, six surveys, 39 columns
base <- "https://school-connectivity.pages.dev/data/"
schools <- read.csv(paste0(base, "school_connectivity.csv"))

# every statistic is weighted with the survey weight; raw percentages are wrong
weighted.mean(schools$police, schools$w)      # 0.841
weighted.mean(schools$deg, schools$w)         # 3.40

# the paper's largest coefficient, rebuilt: the three standing arrangements as one index
z <- function(x) as.vector(scale(x))
schools$capacity <- z(z(schools$security) + z(schools$parent_process) + z(schools$equipment))
coef(lm(deg ~ capacity + z(crime_num) + factor(wave) + factor(level) + factor(urban),
        data = schools, weights = schools$w))["capacity"]
# +0.386. The paper reports +0.377 for the same index, holding school size as well.

# standard errors cluster on sw, the stratum crossed with the survey: 365 clusters
length(unique(schools$sw))                    # 365

# pandas
import pandas as pd
base = "https://school-connectivity.pages.dev/data/"
schools = pd.read_csv(base + "school_connectivity.csv")
edges   = pd.read_csv(base + "edges.csv")      # one row per school and organization

# weighted, always
w = schools["w"]
(schools["police"] * w).sum() / w.sum()        # 0.841
(schools["deg"] * w).sum() / w.sum()           # 3.40

# the published coefficients, with their clustered standard errors and caveats
import json, urllib.request
json.load(urllib.request.urlopen(base + "coefficients.json"))

curl -s https://school-connectivity.pages.dev/data/school_connectivity.csv -o schools.csv
curl -s https://school-connectivity.pages.dev/data/tie_rates.json | jq .overall
curl -s https://school-connectivity.pages.dev/data/coefficients.json | jq .capacity_index

# everything the site serves, static, CORS open, no key
data/school_connectivity.csv    # the analytic file: 12,691 school responses, 39 columns
data/school_connectivity.json   # the same, as records
data/explorer.csv               # the lean subset this page parses
data/edges.csv                  # the two-mode structure: one row per school and organization
data/codebook.json              # every column, its SSOCS variable, and the six coding traps
data/tables.json                # all 22 generated tables, rows and notes
data/coefficients.json          # the headline coefficients, each with its caveat
data/tie_rates.json             # weighted shares of each organization, overall and by survey
data/hero.json                  # the count distribution, for the sketch on the front page

# READ codebook.json FIRST. Yes/no items are 1 and 2, the crime items are reversed,
# security staffing has a seam at 2015-16, and the cluster variable is sw, not strata.

# the source files are free from NCES; this site redistributes a subset
https://nces.ed.gov/surveys/ssocs/data_products.asp

# The working paper is under review and its authors are not listed here yet.
# Write to ron.levi@utoronto.ca for the current draft and how to cite it.
# To cite this site and its data:
@misc{school_connectivity_site,
  author = {{Community Safety and Policing Research Program, Global Justice Lab}},
  title  = {What predicts a school's ties to outside organizations},
  year   = {2026},
  note   = {University of Toronto},
  url    = {https://school-connectivity.pages.dev}
}

# and the source data
National Center for Education Statistics. School Survey on Crime and Safety (SSOCS),
public-use data files, 2005-06 through 2021-22. U.S. Department of Education.