Methods

Study design and data sources

The design of the study is already described elsewhere for community-dwellers [22, 23]. Briefly, this observational cohort study, based on the National Health Data System (Système National Des Données de Santé; SNDS) from the French National Health Insurance system, was designed to describe the characteristics of individuals who received a seasonal influenza vaccine in nursing homes between 1 September 2022, and 31 March 2023, and to assess the rVE of HD vs. SD. The SNDS encompasses anonymous, individual-level data for all healthcare claims for more than 99% of the population residing in France, close to 65 million people [24–28]. Nursing homes admission and discharge date can be identified through the Resid’Ehpad database integrated into the SNDS. Medication consumption at individuals’ level is only available for nursing homes without an in-house Pharmacy (around 77.5% of total nursing homes in France, i.e. 5690 facilities and a theoretical number of 469 369 residents in 2023) [29]. The nursing home unit was unavailable, making impossible to conduct a cluster-level analysis. Furthermore, vaccine choice was likely made at the facility level rather than at the individual resident level, representing a potential source of confounding that could not be fully addressed within the SNDS framework and is discussed as a limitation.

Study population and study period

All individuals aged ≥65 years, living in nursing homes and with an influenza vaccine dispensed between 1 September 2022, and 31 March 2023 (official end of the French influenza vaccination campaign) were included. Individuals who experienced a hospitalisation for the outcomes of interest between the start of the 2022–2023 influenza season (1 September 2022) and 14 days after vaccine dispensing were excluded. Individuals were followed-up from vaccine dispensing (index date) until 30 June 2023, nursing home discharge, medico-social housing admission or death, whichever came first. Hospitalisations for the outcomes of interest were not treated as censoring events; participants remained in follow-up after a hospitalisation event.

Variables of interest

Exposure

We used pharmacy dispensing records as a proxy for influenza vaccination (Supplementary Methods 1). Vaccine type was classified using medication codes (Supplementary Table S1).

Outcomes

The outcomes of interest were the total count of hospitalisations of at least one night due to influenza, pneumonia and influenza and/or pneumonia (Supplementary Table S2). A new event was defined as any admission after discharge from a preceding stay. Outcomes were identified based on the International Classification of Diseases, 10th Revision (ICD-10) discharge diagnosis code and were recorded from 14 days after the index date (start of vaccine protection [30]) until the end of the follow-up period.

Covariates

A fixed 5-year pre-index date period was used to capture baseline demographics, comorbidities/medical history and previous treatments and vaccinations (Supplementary Table S3) [31].

Statistical methods

Patient matching

To reduce the bias due to confounding factors, individuals with HD were matched to individuals with SD using a 1:1 ratio propensity score (PS) with an exact constraint on sex, age groups (65–74/75–84/85 years and over), geographical region of residence and week of dispensing of the vaccine. The PS was estimated by logistic regression, which included socio-demographic variables, proxies for health behaviours and comorbidities. The balance between covariates was evaluated using standardised mean differences, with a maximum accepted difference of 10% (Supplementary Methods 2).

Main analysis

In the main analysis, we reported endpoints based on stays identified with a primary ICD-10 discharge code, as it should be the main reason for hospitalisation [32]. Hospitalisations with a discharge diagnosis code associated with COVID-19 were excluded.

To ascertain the association between vaccination with HD or SD and hospitalisation for the outcomes of interest, Poisson models, negative binomial models and their zero-inflated counterparts were employed to calculate incidence rate ratios (IRR) with corresponding 95% confidence intervals (95%CI) [33]. The model with the lowest Akaike information criterion was selected [34] (Supplementary Methods 3 & Tables S8–S9). The models included an offset for the log of the follow-up time (starting from day 15 post vaccination), which permitted the calculation of a rate model. The rVE was calculated using the formula ((1–IRR) × 100), with the corresponding 95% CI obtained through the Taylor series variance approximation [35, 36]. P-values <.05 were considered as statistically significant and no adjustment for multiplicity was done.

Sensitivity analyses

Sensitivity analyses were conducted to assess the robustness of the findings in the main analysis. Given that the PMSI (Programme de médicalisation des systèmes d’information) hospital administrative database is maintained for reimbursement purposes, the selection of coding in PMSI could be influenced by the level of severity of the outcome and associated with it by the level of reimbursement that can be claimed by the hospital. We thus used both primary ICD-10 discharge diagnosis codes and non-primary diagnosis codes to identify the study outcomes as a first sensitivity analysis.

Given the high prevalence of SARS-CoV-2 during the 2022–2023 influenza season and the robust surveillance measures in place, we examined individuals with outcomes of interest who also exhibited a diagnosis of SARS-CoV-2 infection (ICD-10 code) in either primary or non-primary diagnosis codes.

To increase the statistical power and maximise the number of HD that would be matched, we also calculated rVE results with a matching ratio varying from 1:1 to 1:4 as a third sensitivity analysis. rVE models were weighted by the proportion of HD recipients matched in each ratio (1:1 to 1:4), scaled to the mean of the matching weights distribution.

Finally, to assess the potential extent of residual bias due to unmeasured confounding, negative control outcomes (NCO) were also analysed; these outcomes shared the same potential sources of bias as the primary outcome but were not plausibly related to the exposure of interest. The effect of HD vs. SD on hospitalisations unrelated to influenza, its complications or influenza vaccination was examined (i.e. urinary tract infection [UTI], cataract surgery or erysipelas) (Supplementary Table S2) [37].

The statistical analyses were performed with SAS Software, Version 9.4 (SAS Institute Inc., Cary, NC, USA).