2.5 | Statistical analyses
Propensity score matching was adopted to optimize comparability between insulin users and nonusers [17]. The propensity score was estimated for every person through nonparsimonious multivariable logistic regression, with insulin treatment as the dependent variable. We used 26 clinically related variables in the analysis as controlling variables (Table 1). The nearest-neighbor algorithm was adopted to construct matching pairs under the assumption that the proportion of 0.995–1.0 was perfect [18].
Crude and multivariate-adjusted Cox proportional hazard models with robust sandwich standard error estimates were used to compare the risk of outcomes between insulin users and nonusers. The results are presented as hazard ratios (HRs) with 95% confidence intervals (CIs) for insulin users versus nonusers. To assess the risk of all-cause mortality, we checked the persons’ time of death or the end of the study, whichever occurred first. For other outcomes, we checked the persons’ date of respective outcomes or end of follow-up on December 31, 2013, whichever occurred first. We compared the cumulative incidence of all-cause mortality, MACE, decompensated cirrhosis, and hepatic failure over time between insulin users and nonusers using the Kaplan–Meier method and log-rank tests.
We conducted a sensitivity test by excluding persons with hypoglycemia before or after the index date; matching insulin users and nonusers; and calculating the incidence and hazard ratio of death, MACE, and liver-related outcomes to avoid the interference from hypoglycemia on other main outcomes.
Two-tailed P < .05 was considered as significant. SAS v9.4 (SAS Institute, Inc., Cary, NC, USA) was used for the analyses.