Statistical Analysis
In brief, we present encounter characteristics for our overall sample
and a propensity score matched sample (PS-matched sample). We generated
a PS-matched sample to account for potential confounding, specifically
confounding by indication, and selection bias. When modeling the
association between steroid administration and our outcomes, we used
only the PS-matched sample of encounters.
Descriptive statistics were used to summarize characteristics of our
overall sample of encounters and our PS-matched sample; frequencies and
percentages were calculated for categorical variables, while means and
standard deviations were calculated for continuous variables. Encounters
in which patients received a corticosteroid were compared with
encounters in which patients did not receive a corticosteroid by using
t-tests or Somers’ D for continuous variables and chi-square tests for
categorical variables. As a patient may be represented more than once in
our study sample, leading to a potential correlation of encounters
within a patient, we adjusted these tests for clustered errors as
suggested by Donner & Klar 7 and Newsom8.
We estimated average length of stay in our overall sample for a patient
who received a steroid and the average day in which that steroid was
administered using an intercept only regression with cluster-robust
standard errors. We then regressed day of administration on length of
stay to determine the correlation between these clinical course
measures. A cluster-robust variance estimator was used to account for
the possible repeated encounters within a patient.
In order to assess the association between steroid use and our outcomes,
we built a series of regression models, which included an unadjusted
(Model 1) and a fully-adjusted model (Model 2). Restoration of patients
to their baseline FEV1pp after hospitalization and FEV1pp at follow-up
were modeled using linear regression. Time to next APE was modeled using
a Cox’s proportional hazard (PH) regression. In each model, we included
only our PS-matched cohort (see section below for details).
Additionally, we accounted for the correlation of encounters within
patients who were hospitalized more than once using cluster-robust
variance estimators. We utilized a change-in-estimate variable selection
strategy 9 to create our fully-adjusted models (Model
2), in which a covariate or combination of covariates were retained in
the model if they changed the regression coefficient for steroid by
approximately 20% or more. Model assumptions and fit were assessed;
steroid administration violated the proportional hazards assumption,
which we addressed by splitting our follow-up period at the median event
time, creating two Cox PH models 10. We censored
patients still at risk after the median event time in the first Cox PH
model, and included only patients still at risk beyond the median event
time in the second Cox PH model.