Propensity Score Methods
As corticosteroid use was not randomly assigned to each encounter, we
created a propensity score model 11 for steroid
prescription in order to account for potential confounding and selection
bias; we used the propensity score estimated from this model to create
our PS-matched sample. In this study, the propensity score was the
conditional probability that a patient would receive a steroid during
their hospitalization, given a set of covariates. For each of our
encounters, we estimated the propensity for steroid prescription using a
non-parsimonious multivariable logistic regression model (C statistic=
0.8779). We used the following variables in our propensity score model:
sex, genotype, inhaled steroids, IgE value, asthma, reactive airway
disease or impaired glucose tolerance diagnosis, best baseline
spirometry, admission FEV1, change from baseline to
admission FEV1, change from admission to midpoint
FEV1, change in antibiotics treatment during
hospitalization, positive fungal sputum culture, history of
nontuberculous mycobacteria, and bacteria present in sputum cultures.
As some patients were represented more than once in our study sample, we
initially modeled the propensity score using a random-effects logistic
regression model. Likelihood ratio tests indicated that the
random-effects model did not outperform the traditional logistic
regression model. As well, the estimated ICC from the random-effect
model indicated that the odds of steroid administration was only
slightly correlated within the individual patient. As few propensity
score methodologies and applied works exist using clustered data12, and greater than half of our patient pool was
represented by one encounter (54%; only 26% of patients had
>=3 encounters), we chose to use a traditional logistic
regression when estimating our propensity score and a simple matching
algorithm when matching encounters.
We matched encounters 1:1 using a greedy algorithm on the logit of the
propensity score and a caliper width of 0.2 the standard deviation of
the logit of the propensity score. This resulted in a PS-matched sample
of 25 non-steroid encounters and 25 steroid encounters, representing 19
and 17 patients in each group, respectively. We evaluated the balance in
the distribution of encounter characteristics between the two groups
using t-tests or Somers’ D for continuous variables and chi-square tests
for categorical variables; we adjusted these tests for clustered errors
as described in the previous section. When modeling the association
between our outcomes and steroid administration, we used this PS-matched
sample to compare our outcomes among encounters with equivalent
likelihood of corticosteroid prescription. All analyses were conducted
in Stata/SE, version 15 (StataCorp, College Station, TX).