3.1.2. Models Predicting Total Mortality Incidence
Models of key predictors of COVID-19 deaths by census tract in each of the two phases are presented in Figure 1 (right points, in darker blue). Overall, tract-level characteristics consistently associated with increased risk of COVID-19 mortality included population proportions of Black or Latinx residents, share of population aged over 80 years, and number of LTCF beds per capita. The only variable consistently associated with decreased risk of census-tract mortality rates was proportion of undergraduate students. Crowded housing, proportion of AIAN residents, and rate of uninsured individuals were not statistically significantly associated with tract-level COVID-19 case deaths in either phase. While most variables were statistically stable between Phase 1 and Phase 2 (i.e., consistent directionality and significance), proportion of essential workers went from not statistically significant in Phase 1 (IRR: 0.98 [0.90, 1.06]) to significantly positive in Phase 2 (IRR: 1.08 [1.03, 1.14]), and housing density went from not statistically significant (0.94 [0.85, 1.03]) to significantly negative (0.89 [0.83, 0.97]) . Other variables had substantive attenuations in their point estimates between Phase 1 and Phase 2 (i.e., greater than 10% change toward the null): IRRs decreased for proportion of Black residents (decreasing from 1.20 [1.11, 1.30] in Phase 1 to 1.07 [1.01, 1.13] in Phase 2) and LTCF beds per capita (decreasing from 2.53 [2.33, 2.74] to 1.70 [1.62, 1.78]), while IRRs increased for proportion of residents aged 20 years or younger, albeit to become non-significant (increasing from 0.90 [0.82, 0.98] to 0.98 [0.92, 1.04]).