Data collection

A retrospective cohort study was conducted from July 1, 2016, through June 30, 2017, at 32 primary care practices in the University of Pennsylvania Health System, which included 227 primary care providers. These practices serve a diverse population in urban and suburban settings in southeastern Pennsylvania and southern New Jersey with over 90,000 RTD office visits annually. Office visits were eligible for inclusion if: (1) there was an ICD-10 code for at least one RTD; (2) the patient was ≥18 years of age; and (3) the clinician was either an attending physician or advanced practice provider (APP). Only adults were included because only a minority of prescribers saw children. Although teaching practices were included in this study, only office visits in which a patient was seen primarily by an attending physician or APP were included, given that residents were not present in all practices. Office visits were excluded if there was an infection other than an RTD coded in the same encounter. These resulting office visits are subsequently referred to as the entire cohort.

Antibiotic(s) prescribed and ICD-10 code(s) for each office visit were collected. Demographics and relevant comorbidities using ICD-10 codes were collected. Charlson comorbidity index for each patient was calculated using methods previously described. ^{Reference Degnan, Cluzet and David11} Each visit occurring during a fall or winter month (October through March) was characterized as winter and during a spring or summer month (April through September) was characterized as summer. Prescriber demographics were collected, and ambulatory care practices were characterized as either teaching or nonteaching depending on whether residents saw patients at the clinic and as either urban or nonurban.

Each RTD was further characterized by diagnostic grouping (sinusitis, pneumonia, acute bronchitis, pertussis, otitis media, pharyngitis, and “other RTDs,” which included nonspecific respiratory symptoms such as cough and rhinorrhea and nonspecific diagnoses such as viral upper respiratory tract infection) and by prescribing tier based on the likelihood of that diagnosis to require an antibiotic (tier 1: antibiotic almost always indicated; tier 2: antibiotic sometimes indicated; tier 3: antibiotic rarely indicated) (Supplemental Table 1). ^{Reference Suda, Hicks, Roberts, Hunkler and Danziger1}

Determination of Appropriateness of Antibiotic Prescribing

To determine a reference standard of antibiotic appropriateness, encounters were randomly selected from the entire cohort in which there was an oral antibacterial (excluding rifamycins, ethambutol, isoniazid, pyrazinamide, vancomycin, rifaximin, fidaxomicin, fosfomycin, and nitrofurantoin) or intramuscular penicillin prescribed. Sixty prescribers were randomly selected from prescribers with at least 20 RTD visits in the summer months and at least 20 RTD visits in the winter months. Then, 20 office visits for each of these prescribers were randomly selected for a total of 1,200 office visits. Visits were intentionally sampled so that each prescriber had 10 visits in the summer months and 10 visits in the winter months. These office visits are subsequently referred to as the random cohort.

To determine the appropriateness of antibiotic prescribing, clinical documentation and diagnostic testing from each office visit in the random subset were reviewed in the electronic health record (EHR) by at least one member of the research team. Diagnostic criteria for determination of appropriateness were adapted from relevant professional society and public health guidelines (Supplemental Table 2). ^{Reference Mandell, Wunderink and Anzueto13–18}

Antibiotics were classified as appropriate if the patient fit the diagnostic criteria for a bacterial respiratory tract infection. Otherwise, they were classified as inappropriate. Only the decision to prescribe an antibiotic was considered when determining appropriateness. Antibiotic choice, dose, and duration were not considered. Chart abstraction was performed using REDCap electronic data capture tools hosted at the University of Pennsylvania. ^{Reference Harris, Taylor, Thielke, Payne, Gonzalez and Conde19} Fifteen percent of office visits were double-coded by two members of the research team. If there was a discrepancy in the determination of appropriateness, two members of the research team discussed the case and came to a consensus on appropriateness.

Statistical analysis

For the entire cohort, categorical variables were summarized using frequencies, and continuous variables were summarized using mean and median when appropriate. Proportions of antibiotic prescribing for diagnostic groupings and for tier groupings were compared for the summer and winter months. Similarly, the proportion of prescriptions by antibiotic class was compared for the summer and winter months. The associations between season and antibiotic used and between season and RTD diagnoses were determined using a χ² test.

Each prescriber from the entire cohort with at least 20 RTD visits in summer months and at least 20 RTD visits in winter months was included in an analysis to determine the association of prescriber demographic characteristics with seasonal variation in antibiotic prescribing. To assess whether prescribers with more RTD office visits in winter months or greater changes in RTD visits from summer to winter months were more likely to prescribe antibiotics for these indications in winter months, linear regression was performed on the entire cohort. This analysis evaluated whether the dependent variable prescriber’s seasonal change in percent prescribing from summer to winter is correlated with the prescriber’s total RTD visits in winter months, the absolute seasonal difference in number of RTD visits from summer to winter, and seasonal percent change in RTD visits from summer to winter. The association of prescriber demographic characteristics with the difference in the proportion of RTD visits in which an antibiotic was prescribed between winter and summer months was determined for the entire cohort using t test or Wilcoxon rank sum test when appropriate.

The association between season and appropriateness of antibiotic prescribing was determined for the random cohort using a χ² test. To determine if various potential administrative metrics were correlated with inappropriate prescribing, univariate and multivariate fractional logistic regression was used for the random cohort with the proportion of inappropriate prescriptions as the dependent variable and various possible administrative metrics and prescriber characteristic covariates as independent variables. Analysis was performed at the level of the prescriber, and possible administrative metrics included difference in the proportion of prescribing for RTDs between winter and summer months, the proportion of prescribing for all RTDs, and the proportion of prescribing for tier 1, 2, and 3 diagnoses.

Given the large sample size, significance for summary statistics performed on the entire cohort was reported using standardized difference (SD) with SD≥0.1 considered a significant difference. Significance for other analyses was reported using P-values with P ≤ .05 considered a significant difference except in univariate analyses where more liberal P ≤ .20 criteria were used for inclusion in multivariate models.

Data were analyzed using STATA v15.0 (StataCorp, College Station, TX). This study was approved by the Institutional Review Board at the University of Pennsylvania.