Infections, Hospitalizations, and Deaths Among US Nursing Home Residents With vs Without a SARS-CoV-2 Vaccine Booster

Key Points Question What is the estimated vaccine effectiveness up to 12 weeks for the first COVID-19 mRNA booster vaccines administered in US nursing homes? Findings In this cohort study of 10 949 residents of 202 community nursing homes and 4321 residents of 128 Veterans Health Administration community living centers, booster vaccination was associated with significant reductions in SARS-CoV-2 infections, hospitalizations, and the combined end point of hospitalizations or deaths. Meaning These findings suggest that administration of a SARS-CoV-2 mRNA vaccine booster among nursing home residents may have played an important role in preventing COVID-19–associated morbidity and mortality.

When a resident receives a booster vaccination on a target trial date, they are "assigned" to the booster arm. Once a resident is boosted, they are ineligible for future trial dates because one criteria is not already being boosted (see Table 1 and consort diagram in Figure 1). Therefore, a person only appears once in the booster arm. Alternatively, residents might be eligible as controls on multiple emulated trial dates up until they received a booster. If these repeated observations by person are pooled together it is more statistically efficient, but also requires greater computational effort because of inflated sample size and the bootstrapping procedure. Because we had a reasonably large sample, and wished to avoid very long computation times (days to weeks), we randomly selected one eligible target trial date per control resident. This doesn't impact the total number of unique persons in the study, or which events are counted, but randomly selects a "time zero" for each control person. Controls are censored if they subsequently receive a booster dose, but contribute follow-up time until the point of censoring on the date of booster administration. If that control person is otherwise eligible and the date they receive a booster is a target trial date, they will be included in the treatment arm as well. Persons are unique within assignment arm (boosted or unboosted), but can appear in both arms of the study. This is accounted for in the statistical analysis by bootstrapping, where the resampling occurs at the person-level cluster.
Because actual assignment of vaccine on a given day is non-random, inverse probability of treatment weights are used to weight the sample and create a pseudopopulation of observations where probability of assignment to vaccine is similar between groups. Additionally, because control residents censor for treatment, this (by design) informative censoring between groups must be accounted for. We use inverse probability weighting for the probability of remaining uncensored to account for this mechanism.
Why employ this method versus traditional retrospective analytical methods?
The reason the authors elected for this method was for two reasons: 1) Target trial emulation is a helpful tool for refining your study question and identifying a causal contrast of interest, with clearly laid out inclusion/exclusion criteria for your cohort, 2) There is an inherent challenge of comparing persons receiving a vaccine versus non-receipt. If not carefully addressed immortal time bias can be introduced, where there exists a misalignment of eligibility assessment, start of follow-up and assignment of treatment periods between groups. Target trial emulation is useful because it helps pinpoint eligibility assessment and treatment assignment on at the same timepoint and avoids this bias. There are many problems in observational data, unmeasured confounding, model misspecification etc. that target trial emulation doesn't solve but it is a useful and practical tool for causal inference.

Statistical analysis
We estimated two weights for our statistical analysis, the first were stabilized inverse probability of treatment weights (IPTW). This was the probability of receiving the vaccine on each target trial date, adjusting for measured baseline confounders. A stabilized propensity score was estimated with a pooled logistic regression model (eTable 4) for each system separately. The second weight was an inverse probability of remaining uncensored weight (IPCW, eTable 5). This is using the probability of remaining uncensored at each timepoint, conditional on remaining uncensored up to that point (cumulative probability). The mechanism of censoring was different by treatment group, for example unboosted residents can censor when they subsequently receive the vaccine but boosted residents do not. Pooled logistic regression models were used to estimate the IPCW for boosted and unboosted arms separately. The product of the weights was taken (IPW = IPTW * IPCW) and used in an outcome regression model. Weights were truncated at the 1% and 99% intervals. We fit a pooled logistic regression model of a dataset which was expanded for each discrete time interval (days) with dummy variables for each timepoint, f(t), and an interaction with treatment arm, weighted by IPW.
[Pr( +1 | = 0, )] = 0 + 1 ( ) + 2 + 3 [ * ( )] An event-free survival probability at each timepoint and outcome was calculated from the model. Both cohorts used the same analytical approach, but statistical models differ because of underlying differences in the data and separate analysts performing model selection for the two nursing home systems (e.g. the Veteran population is mostly male, different age distribution). eFigure 1. Flowchart of Eligibility for Target Trials of mRNA Boosters in 2 US Nursing Home Systems Description. Each system include residents present in the home, and meeting long-stay definitions (100 days in the home, with a gap of no more than 10 days). Further exclusion criteria are outlined.

System 2 -Veteran Affairs, Community-Living Centers
Description. The plot depicts the standardized mean differences (SMD) in boosted and non-boosted residents. Unadjusted refers to differences before applying probability weights. Adjusted is after applying inverse probability weighting. Description. Each panel represents the vaccine effectiveness as 1the cumulative incidence rate ratios for each outcome. Shaded regions represent 95% confidence intervals. The confidence interval is omitted for certain timepoints and panels due to instability and difficulty in estimation. The point estimates at week 12 are reported in Table 3.  Description. Models where the dependent variable was receipt of booster dose on index date. Fixed effects for states were not included in System 2, because most states only have one facility. The logistic regression models were estimated separately in the two datasets by different analysts but followed similar principles for model selection. *,**,*** refer to spline basis terms, a '-' means that term was not included in the model. I=Interaction term. Description. Models where the dependent variable was not censoring at a follow-up timepoint in a pooled regression model of person-time observations. The data for System 1 was expanded in time units of days, while in System 2, units of weeks (this makes the scale of coefficients different in either model, so values are not directly comparable). The logistic regression models were estimated separately in the two datasets by different analysts who followed similar principles in model selection. *,**,*** refer to spline basis terms, a '-' means that term was not included in the model. I=Interaction term. The time coefficients are excluded from the output for brevity. a In System 1, a restricted cubic spline with 4 degrees of freedom was used for time, b System 2 used fixed effects for each timepoint.