Figure 2. Propensity score matching plot.
Continuous variables were summarized using descriptive statistics such
as mean and standard error. All categorical variables were summarized
using frequencies and percentages. To further investigate the balance of
prognostic variables between the case and control groups, univariate
analyses were performed using the t-test or chi-square test upon
specific requirements. All the variables associated with prognostic
factors and disease history demonstrated no significant impact on the
outcome, except one variable, LVEF (see Table I). The conditional
logistic regressions (an extension of
logistic
regression which considers stratification and matching) were used to
examine the potential association between the exposure(s) and the
outcome while adjusting for the covariate LVEF. A unique matching ID for
each pair was generated to indicate the stratum. Thus, the matched
variables were controlled simultaneously because their combined
information was already incorporated into the matched pairs by this ID.
To identify the best-fitting model, we conducted a model selection
process based on 6 candidate models whose exposure variables varied (see
Table II). The goodness-of-fit measurement Akaike information criterion
(AIC) was used to compare these candidates. The preferred model is the
one with the minimum AIC value. Thus, Model 1 was the final model for
our statistical results and inference. Odds ratios and 95% confidence
intervals were reported in Table III and displayed in the forest plot
(Figure 3). All analyses were performed at a two-sided 5% Type I error
rate by SAS 9.4 (SAS Institute, Cary, NC).