Note: Shared synthetic release 2.0. Real hospital names
are reference labels only; all patient records and outcomes are
simulated.
Overview
Research
questions
Primary question: After adjustment for patient risk,
does the probability of admission for adult asthma ED patients differ
between hospitals in the Calgary region, the Edmonton region, and other
regions? Secondary question: Does the admission
difference between asthma and non-asthma visits vary by hospital
region?
The analysis uses 24,926 adult ED visits from
unified data release 2.0-2026-10-09, including
8,170 asthma visits, across 30 hospitals. A09 does not
reuse A07’s matched subset, so that the regional comparison is not
affected by match selection made for a different question.
Simulation only. Hospital names and the official ED
service directory are real public reference information; hospital IDs,
patient distributions, admission rates, and regional effects are all
simulated. A higher or lower admission rate does not directly indicate
quality of care, and this memo does not evaluate or rank any real
hospital.
Hospital reference and
region definitions
As specified for this project, Calgary and Edmonton use
regional boundaries that include surrounding towns. The
grouping follows public AHS geographic materials as a fixed study
definition and does not claim to represent the current administrative
structure. hospital_city keeps the specific municipality.
Other consists of the selected Central, North, and South hospitals. The
10 hospitals per group are a deliberately chosen teaching sample, not
all hospitals, and do not reflect real hospital weights.
All selected facilities are listed in the AHS
Emergency Services directory. The source and verification date for
each hospital are stored in the hospital_reference sheet of
both Excel workbooks and in config/hospital_reference.csv.
The directory was checked on 2026-10-09; being “listed
with an ED” does not guarantee there were no temporary service changes
at any given time.
This study does not automatically classify Urgent Care as an ED and
does not include dedicated children’s hospitals. Strathcona Community
Hospital does have an ED, but it was not selected for this round’s
simplified same-hospital admission sample; its absence from the list
does not mean the hospital has no ED. Strathcona
ED listing
Primary model: asthma
patients
The primary analysis includes only patients aged ≥18 with a main
diagnosis of J45; the outcome is admission following this ED visit.
admitted_flag ~ hospital_region + age10 + sex + smoking + obesity
+ cardiometabolic_history + asthma_history + prior_ed_visits
+ patient_residence + season + ed_arrival_severity + (1 | hospital_id)
A binomial / logit GLMM is used: region is a fixed
effect with Other as the reference, and hospital enters as a normally
distributed random intercept. Patients at the same hospital share one
baseline admission propensity. With only three regions, region is not
treated as a random effect. An ordinary linear mixed model is not used
for the binary admission outcome. lme4 model
documentation
Because this question compares hospital regions, it adjusts for the
simulated arrival severity, which is measured before treatment and is
comparable across hospitals. This severity is downstream of the current
asthma episode, so the A07/A08 total-effect analyses of asthma do not
adjust for it; a variable’s role depends on the research question. A
real study would first need to confirm when severity is measured and
that it is consistent across hospitals, and could not simply substitute
a post-treatment hospital measure.
Standardized risks
and contrasts
Risks are standardized to the shared empirical patient mix of all
adult asthma visits, with each patient weighted equally. In each
region’s prediction, the same estimated hospital random-effect
distribution is integrated over to obtain the population-average risk
across hospitals. This is not setting the random effect to 0 and
calling it a marginal risk, nor is it weighted by real hospital
case volumes. The estimand is the regional difference, for the hospital
population described by the model, under a common patient mix.
The risk difference for Calgary versus Other is -5.72
percentage points (95% CI -10.53 to 0.25); for Edmonton versus
Other it is 7.23 percentage points (95% CI 2.16 to
12.99). These results come from fitting a simulation and cannot be used
to infer which real region performs better.
Based on these exploratory intervals, the Calgary versus Other
comparison does not yet establish a direction (the interval includes 0);
the Edmonton versus Other comparison supports a higher probability of
admission. These conclusions apply only to the current simulation and
the defined standardization target.
Secondary model: asthma
by region interaction
All adults in the same main NACRS file are used, with the hospital
random intercept kept and asthma_visit * hospital_region
plus pre-visit background factors added. Arrival severity and
CTAS are not added, so that the severity pathway of asthma is
preserved. This question differs from the primary model’s “regional
difference among asthma patients of comparable severity”, and the two
sets of risk differences should not be conflated.
Regions must be compared directly through the difference in RDs and
its interval; “significant in one region but not in another” does not
show that regions differ. The joint asthma-by-region interaction test (2
df, log-odds scale) is below; interaction on the log-odds scale and
interaction in absolute risk differences are not the same claim.
Because the secondary model implicitly averages over unobserved
individual severity, its marginal logit form is a working-model
approximation. It describes risk-adjusted heterogeneity and does not
guarantee recovery of causal effects in real data.
Uncertainty and model
checks
Each model is refitted to 200 parametric bootstrap replicates that
regenerate the hospital random effects and binary outcomes while holding
the target patient mix fixed. Intervals are the bootstrap 2.5% and 97.5%
quantiles and include the uncertainty in the estimated hospital
variance; this is not a row-wise independent bootstrap. Seeds are
20261014 for the primary model and 20261015 for the secondary model. bootMer
documentation
The primary model returned finite estimates in 200/200 replicates and
the secondary model in 200/200. Tail quantiles from 200 replicates carry
simulation error, and the intervals are not family-wise simultaneous
confidence intervals; the multiple regional contrasts are reported as
prespecified exploratory results.
Neither bootstrap recorded any fitting messages or warnings.
The models use 9-point adaptive Gauss–Hermite quadrature. Switching
the primary model to 17 points changes the fixed effects by at most
2.18e-06. The estimated SD of the hospital random intercept is 0.340,
against a generating value of 0.35. The singular-fit check returns FALSE
for the primary model and FALSE for the secondary model.
Repeated
known-scenario checks
Separately, 100 simulations were run under each of a positive-effect
scenario and a “no systematic regional effect” scenario, each with 6,000
adults and 30 hospitals. The latter sets the region main effects and the
asthma-by-region interaction to 0 while keeping random differences
between hospitals. The checks cover the conditional regional
coefficients of the primary asthma model, Wald interval coverage, and
the joint test. This check is not the same as validating the coverage of
this memo’s bootstrap intervals 100 times.
The simulation is only a finite-sample methods check. Complete
failures, convergence warnings, and boundary fits are kept in the result
bundle, and seeds were not changed in response to results.
Interpretation and
limitations
The probability of admission is affected by illness, referral, bed
availability, and case mix; neither a higher nor a lower rate means a
hospital is “better”. Admission is currently simplified to admission at
the same hospital; a real study should specify transfers, observation
beds, and the final admitting hospital. Real hospital weights, service
scope, and case volumes did not enter the generating mechanism.
The GLMM handles hospital clustering and some observed differences,
but by itself it cannot turn the regional comparison into causal
inference. Common support, unmeasured confounding, the timing of
diagnosis recording, and the hospital selection mechanism still need
independent justification. The 30 deliberately selected facilities
cannot directly represent all of Alberta.
Reproducibility
The only observed data source remains the unified NACRS/DAD Excel
workbooks. Case data, the public hospital reference table,
counterfactual truth, and methods simulations each have a defined
purpose; the truth and the hospital random effects used at generation
never enter fitting. A09’s input checks match A07/A08, and the full
models, bootstrap, and validation results are saved in
outputs/asthma_a09/.
Back to project home · Data definitions A00 · Matching A07 · Outcome regression A08
