Hospitalizations and Mortality From Non–SARS-CoV-2 Causes Among Medicare Beneficiaries at US Hospitals During the SARS-CoV-2 Pandemic

This cohort study investigates how hospitalizations and mortality rates for non–SARS-CoV-2 conditions have changed during the COVID-19 pandemic.

Both methods have similar estimation and fit statistics. However, RSPL takes about 1 hour per model, while LAPLACE requires approximately 16 hours per model. We chose RSPL for the analysis.
We performed the same main three-level logistic regression analyses for different cohorts, including the cohort in 2019 and April 2020 to March 2021 in Table 1, and the 6 quarterly cohorts in Table 2. All models included the same fixed effects and the same random effects.
We further included interaction terms between time period and patient/hospital characteristics into the models. For Table 3, we added an interaction term for each variable into the model for Table 1. We checked the F statistics for the type 3 test of fixed effects for the interaction term.
If the p-value was <0.05, we considered that interaction term significant. For significant interaction terms, we further performed stratified analyses by stratifying the cohort into subcohorts by that stratification variable. We then re-performed the same main model and reported the odd ratios (95% CI) for each sub-cohort in Table 3. For Table 4, we stratified the 6 quarterly cohorts into sub-cohorts by hospital SARS-CoV-2 prevalence during that quarter.
Hospital SARS-CoV-2 prevalence was categorized by quartiles as a hospital-level variable.
However, if >25% of hospitals in a time period had prevalence, we categorized the prevalence into 3 groups. We performed the three-level logistic regression model for the sub-cohorts (high prevalence vs. low prevalence) hospitals for each quarterly cohort separately and reported odd ratios (95% CI) in Table 4.
The SAS code for the main analysis is below (Table 1). We initially tried the quad method (QPOINTS=X). We abandoned this method because of insufficient memory capacity. We then compared the LAPLACE method and the default RSPL method; both introduced similar estimations. However, the LAPLACE approximation takes about 16 hours computing time and the RSPL takes about 1 hour, so we choose the RSPL technique for our analysis.
For the interaction term analysis in Table 3, we added one interaction term at a time into the main model. For example, for the interaction term of Medicaid and time in Table 3, the SAS command is below.
F statistic was significant, we performed stratified analysis for the sub-cohorts by each characteristics.
The models for Table 2 are the same as the main model in Table 1. The only difference was the cohort. Table 1 used the cohort of year 2019 and April 2020 to March 2021. Table 2 Table 4 used the same 6 quarterly cohorts as in Table 2. We further stratified each quarterly cohort by the hospital SARS-CoV-2 prevalence in that pandemic quarter. We then performed the same model as that used in Table 1 for the sub-cohorts with high prevalence and low prevalence (12 models total). The odds ratios between 2020/2021 and 2019 were obtained for each sub-cohort. Since the model was the same, the SAS code is not included here again. 15%) a For analyses exploring interactions between admission characteristics and hospital prevalence of SARS-CoV-2, (Tables e5-e10), we substituted a "disease severity" measure for the 20 individual admission diagnoses. In order to generate a measure of severity for the 20 admission diagnoses, we calculated the unadjusted and adjusted mortality in the 30 days post admission using 2018 fee for service Medicare data. In the adjusted analyses we controlled for all the admission characteristics used in the analyses in Table 1 including the comorbidities listed in Table e3. Thus, the mortality rate associated with each diagnosis is independent of other admission characteristics. CI: confidence interval; UTI: urinary tract infection; COPD: Chronic obstructive pulmonary disease; CHF: congestive heart failure; AMI: acute myocardial Infarction. Given the very large sample size, all differences in the characteristics in 4/1/20-3/3121 were statistically significant compared to those in 2019 at P < 0.05 level. Also, all differences in mortality rates between categories of characteristics were statistically significant at P<0.05. Table 1 presented in Tables e5-e10, we first tested for interactions between admission or hospital characteristics and the percentages of SARS-CoV-2 admissions at each hospital (prevalence of SARS-CoV-2). Tables 5-10 present the stratified analyses based on those interactions. Each table presents analyses from a three-month period, because SARS-CoV-2 prevalence in hospitals changed over time. If an admission or hospital characteristic showed a significant interaction with hospital SARS-CoV-2 prevalence in any time period, then it was included in the stratified analyses for all the time periods. . OR, odds ratio; CI, confidence interval.