Risk factors for pneumonia and influenza hospitalizations in long-term care facility residents: a retrospective cohort study

This was a retrospective cohort study using Medicare enrollment, and Parts A and D claims linked to Minimum Data Set (MDS) for 100% of LTCF residents enrolled in fee-for-service Medicare during 2013–2015. Medicare Part A data were used to identify hospitalizations involving P&I, and Part D claims enabled the ascertainment of prescribed medications. The MDS is a federally required clinical assessment completed at admission and at least quarterly thereafter among all residents in Medicare or Medicaid certified nursing homes. MDS data provide a comprehensive and standardized assessment of the functional capabilities and health needs of LTCF residents [21, 22]. Specifically, MDS data include demographics, clinical conditions, treatments, behaviors, physical function, and cognitive status. We applied the residential history file algorithm to track the timing and location of health services utilization [23]. Facility-level variables were obtained from Online Survey and Certification And Survey Provider Enhanced Reports (OSCAR/CASPER) and LTCFocus data collected for all Medicare- and Medicaid-certified LTCFs. This study was approved by the Brown University Institutional Review Board.

The study cohort was derived from a national source population of Medicare beneficiaries residing in LTCFs between January 1, 2013 and December 31, 2015. Eligible residents were categorized as short-stay (total stay < 100 days in the same LTCF), or long-stay (total stay ≥100 consecutive days with ≤10 days outside the facility). Index dates were defined as the LTCF admission date for short-stay residents and day 100 of a stay for long-stay residents. We sampled the first LTCF stay, and followed residents from their respective index dates until hospitalization, discharge from the LTCF, disenrollment from Medicare, death, or end of the study period, whichever occurred first. The cohort inclusion criteria were 1) continuous enrollment in Medicare Parts A and D 6 months prior to index; 2) age at index ≥65 years; and 3) ≥1 MDS assessment within 100 days before the index date for long-stay residents and upon entry to the facility for short-stay residents. We excluded residents with Medicare Advantage enrollment, who received hospice services, or had missing data on any covariate used in analyses.

Risk factors were selected based on prior literature and our clinical experience related to what factors could influence P&I risk [6, 19, 24‐26]. Resident characteristics were measured during the 6-month period prior to, or at, the index date to ensure they were not influenced by the outcome. We evaluated demographic, tobacco use, body mass index, clinical (diagnoses and geriatric syndromes), medication use and health service use variables as potential risk factors for P&I hospitalizations. Demographic factors included age, sex, race and ethnicity. Clinical diagnoses from MDS included, e.g., cancer, atrial fibrillation, history of pneumonia, diabetes mellitus, arthritis, Alzheimer’s disease, asthma/chronic obstructive pulmonary disease (COPD)/chronic lung disease. Among others [27], geriatric syndromes included, e.g., cognitive function scale score [28], Changes in Health, End-stage disease and Symptoms and Signs (CHESS) scale score [29], and activities of daily living (ADL) 28-point scale score [30]. The validated CHESS score is primarily used as a risk adjustment tool to identify LTCF residents with high health instability who are likely to have adverse health outcomes, including death [31]. Medication use was defined as receiving ≥1 qualifying prescription for antipsychotics, opioid analgesics [32], antibiotics, corticosteroids, or proton pump inhibitors as well as for Beers criteria medications [33]. The Beers criteria identify specific medications and prescribing practices (e.g., excessive dose, prolonged treatment duration, harmful drug combinations, and coexisting health conditions) with evidence to suggest they should be avoided or used with caution by older adults due to unfavorable risk/benefit profiles or questionable efficacy [34]. Examples of drug classes in the Beers criteria are first generation antihistamines, barbiturates, benzodiazepines, proton-pump inhibitors, and estrogens. We measured the status of influenza vaccination for the season of cohort entry based on index date and up to date pneumococcal vaccination counting vaccinations received within or outside the LTCF. We assessed health service use as hospitalization and intensive care unit (ICU) use.

As with resident-level factors, we considered facility features based on prior literature and clinical experience, including: 1) structural characteristics (urbanicity of facility location, total bed size, for-profit status); 2) staffing hours (total nursing hours/resident/day); 3) staffing type; and, 4) quality of care measures. Staffing type included proportion of registered nurses (RNs), on-site presence of a licensed independent practitioner (LIP) - either a physician assistant (PA) or an advanced practice RN (APRN), and speech language pathologist (SLP) on-staff hours per 100 beds. Quality of care measures included the percent of residents receiving antipsychotics, percent of residents restrained, and percent of residents with a pressure ulcer [19].

We identified P&I hospitalizations by the presence of ICD-9 or ICD-10 diagnostic codes for pneumonia or influenza-like-illness (480–488.XX, J09-J18) [35, 36]. The main analysis focused on P&I diagnoses in the principal position on the claim. Secondarily, we analyzed P&I identified from any diagnosis position.

We report the distribution of baseline characteristics of the study cohort with means and percentages for the entire cohort and among short- and long-stay residents.

The process of identifying the risk factors for P&I hospitalization proceeded in three steps. First, we grouped the variables into domains as follows: demographics, admission characteristics (location resident was admitted from, LTCF admission is new), cardiovascular conditions (atrial fibrillation, coronary artery disease, heart failure, hypertension, cerebrovascular accident), respiratory conditions (asthma/COPD/chronic lung disease, respiratory failure, pneumonia), other medical conditions (cancer, Parkinson’s disease, depression, diabetes mellitus, arthritis), cognition (Alzheimer’s and non-Alzheimer’s dementia, cognitive function scale), physical function (ADLs, urinary/bowel continence), overall health stability (CHESS scale score, Charlson comorbidity score, prognosis, prior hospitalization and/or ICU stay), breathing (shortness of breath, ventilator/respirator use), eating (tube feeding, swallowing disorders), medication use, vaccinations, and facility characteristics.

Second, we examined intercorrelations of variables within domains using a Pearson’s correlation coefficient matrix. None of the bivariate correlations reached a level (r > 0.8) indicating severe multicollinearity. We included state fixed effects to help account for potential state-level differences in LTCFs' propensity to hospitalize residents and code for P&I on hospital claims.

Finally, all variables from the domains identified in the first step were entered into a Cox proportional hazards model specified to account for clustering of residents within facilities using the Huber-White sandwich estimator. A stability analysis assessed an alternative Fine and Gray competing risk regression modeling approach with death as a competing outcome. Considering the large sample size, an alpha = 0.01 significance level was used to guide identification of potential P&I risk factors in the final model.

Data preparation and analyses were conducted using SAS version 9.4 (SAS Institute, Inc., Cary, NC) and Stata version 15 (StataCorp, College Station, TX). We secured administrative permission to access Medicare data through a Data Use Agreement with the Centers for Medicare and Medicaid Services (CMS). Informed consent was neither relevant nor feasible in this secondary data analysis.