This was a retrospective cohort study using data from the Vaccine Safety Datalink, a collaborative project involving the Centers for Disease Control and Prevention (CDC) and several integrated health care delivery systems (sites) (13). This study used data from 7 sites, with a combined population of approximately 10.2 million members. We studied influenza vaccinations received from 1 September 2016 through 1 June 2017. We first used administrative health care data to identify the study population and then did manual medical record abstraction to collect additional information on presumptive bursitis cases. The study was approved by institutional review boards at the CDC and each participating site.

Eligibility criteria included receipt of an inactivated influenza vaccine (IIV) in the arm during the 2016–2017 influenza season, age 3 years or older on the date of vaccination, and continuous enrollment in the site’s health care organization from at least 180 days before the IIV dose through 180 days after. For patients with multiple IIV vaccinations during the 2016–2017 influenza season, we included only the vaccination with the earliest date. We excluded children younger than 3 years because they are typically vaccinated in the lateral thigh (14). The enrollment requirement was to allow adequate time before the vaccination to detect preexisting shoulder conditions and after the vaccination for presentation and new bursitis diagnoses to be made. We excluded patients if laterality of vaccination or bursitis diagnosis could not be determined from administrative health care data; if they had International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM), diagnostic codes for bursitis or related shoulder conditions in that arm before vaccination; or if they received any other vaccinations in the same arm in the 180 days after the IIV vaccination. Persons who received other vaccines on the same day were not excluded.

Three ICD-10-CM diagnostic codes are used to indicate shoulder bursitis or a related shoulder bursa diagnosis (M75.5* [bursitis of shoulder], M71.31* [other bursal cyst, shoulder], and M71.81* [other specified bursopathies, shoulder]). These codes are side specific, indicating to which shoulder the diagnosis applies. Patients with any of these codes recorded during the 180 days after receipt of IIV for the arm that was vaccinated were presumptive cases.

We manually abstracted medical records for all presumptive cases at 5 sites, which each had fewer than 100 cases. We selected a random sample of 150 presumptive cases for medical record abstraction at the remaining 2 sites, which had a larger number than was feasible to review manually. We determined a priori that at least 225 charts would have to be abstracted to detect an incidence rate ratio (IRR) of 2.0 or greater with a power of 0.8 and α of 0.05. Our final number of charts abstracted allowed us to detect an IRR of 1.7 or greater (15, 16). We excluded presumptive cases if abstraction identified a miscode for the outcome or a rule-out diagnosis, there was evidence of a preexisting shoulder condition before vaccination in the medical records, or adequate medical records were not available.

On the basis of chart abstraction, we defined cases with 3 levels of diagnostic certainty: definite (bursitis diagnosis confirmed by ultrasonography, magnetic resonance imaging, or surgical findings, or the patient reported a positive response to a corticosteroid injection associated with a bursitis diagnosis), probable (bursitis diagnosis made or confirmed by a musculoskeletal specialty provider, including orthopedic surgeon, rheumatologist, physiatrist, or physical or occupational therapy provider), or possible (bursitis diagnosis made by any other provider).

Chart abstractors collected variables of interest not available in administrative data, including the credentials of the vaccinator; vaccination setting; patient height and weight; and patient statements about date of symptom onset, to what causes the patient attributed their symptoms, and when symptoms resolved. Chart abstractors collected information about symptom onset from up to 4 different medical encounters: the visit on the date of vaccination, the first visit in which the patient mentioned the shoulder symptoms, the first visit in which a bursitis diagnosis was assigned, and the first visit to a specialty provider. We reviewed the symptom onset statements for precision and consistency among visits. We classified symptom onset statements into 3 categories of precision: exact (a specific day was indicated), approximate (symptom duration was stated in definite units of time but did not specify an exact date [for example, “3 weeks”]), or vague (duration was stated using indefinite units of time, such as “a few weeks”). For records with approximate or vague statements, adjudicators assigned an estimated onset date by back calculating from the medical visit date at which the statement was documented.

Medical record abstraction data were collected and managed using REDCap electronic data capture tools (17, 18). The abstraction form is shown in the Supplement (available at Annals.org).

We did a self-controlled risk interval analysis, which compared incidence of subdeltoid bursitis with symptom onset during a postvaccination risk interval versus incidence during a control interval (16, 19). We prespecified the risk interval as the first 3 days after IIV vaccination (days 0 to 2, where day 0 is the day of vaccination), which we selected on the basis of a prior case series of shoulder adverse events related to vaccine administration (8). We defined the control interval as days 30 to 60 after vaccination, which we selected to represent the background rate (that is, a time during which the outcome is not influenced by vaccination). The case ascertainment period of 180 days after vaccination ensured that persons with symptom onset during both the risk and control intervals would have adequate time to present for care and receive a bursitis diagnosis. We calculated the IRR and the Wald 95% CI using Poisson regression (GENMOD procedure) in SAS, version 9.4 (SAS Institute). The model independent variable was the number of cases occurring during the interval. The dependent variable was an indicator of risk or control interval. The offset term was the log of the number of person-days in each interval, which accounts for different interval lengths. Each person contributed person time to both intervals, thereby implicitly acting as their own control, and could only be counted as a case once during the follow-up.

To estimate the attributable risk, we first calculated the incidence rate in each interval as the number of cases divided by the number of vaccinated person-days in the interval. We used the Byar method to calculate the CIs of the difference between risk and control interval incidence rates and multiplied the results by the length of the risk interval. To account for random sampling of chart reviews done at 2 sites, we adjusted the vaccinated person-days by the chart-sampling fraction at each site.

We did sensitivity analyses to see how the strength of association would change with changes in risk or control intervals, levels of diagnostic certainty, or precision of symptom onset date information. We also performed an exploratory, non–self-controlled analysis of potential risk factors for vaccine-associated bursitis by comparing characteristics of cases with symptom onset during the risk interval with cases with symptom onset during the control interval using the Fisher exact and Kruskal–Wallis tests.

This study was funded by internal funds of the CDC. The authors had complete control over design, analysis, and the decision to submit the manuscript for publication.

Our cohort included 2 943 493 persons who received IIV during the 2016–2017 influenza season and met study eligibility criteria (Table 1). Among these, we identified 1098 presumptive cases of bursitis. We used random sampling to select 526 presumptive cases for medical record abstraction, of which 421 met the case definition for an incident diagnosis of subdeltoid bursitis (Figure). Bursitis symptom onset began after vaccination for 257 cases: 16 (6.2%) had symptom onset in the risk interval, 51 (19.8%) had symptom onset in the control interval, and 190 (74.0%) had symptom onset in neither prespecified interval (Appendix Figure, available at Annals.org). The distribution of cases by study site is shown in Table 2.

The 421 cases of subdeltoid bursitis were classified as definite (n = 135 [32%]), probable (n = 44 [10.5%]), or possible (242 [57.5%]). All cases occurred in adults except for 6, which occurred in adolescents aged 14 to 17 years. The median number of days between reported symptom onset and the first medical visit for bursitis symptoms was 20 (range, 0 to 3653 days).

Characteristics of cases with symptom onset during the prespecified risk or control intervals are shown in Table 3. The median age of patients in the risk interval was 57.5 years (range, 24 to 98 years), and 68.8% were women. We did not observe any differences between cases in the risk interval and those in the control interval by sex, age, body mass index, race/ethnicity, vaccination setting, or laterality of the vaccinated arm. Almost half of the vaccinators for cases in the risk interval were medical assistants. Compared with cases in the control interval, those in the risk interval were more likely to be vaccinated by medical assistants than nurses (including registered nurses, licensed practical nurses, and licensed vocational nurses) (P = 0.052).

The self-controlled risk interval analysis found an increased risk for subdeltoid bursitis with symptom onset in the 3 days after vaccination compared with the 30 to 60 days after vaccination (IRR, 3.24 [95% CI, 1.85 to 5.68]). This corresponds to an attributable risk of 7.78 (CI, 2.19 to 13.38) excess cases of subdeltoid bursitis during the 3 days after vaccination per 1 million persons vaccinated. The association between influenza vaccination and subdeltoid bursitis remained elevated in all sensitivity analyses (Table 4).

The median follow-up after vaccination for cases was 576 days (range, 189 to 781 days). Among definite cases in the risk interval, 4 had surgical procedures for bursitis or mentioned bursitis on surgical reports. The intervals from vaccination to surgery were 90, 115, 210, and 266 days. Four definite cases had findings of bursitis on magnetic resonance imaging at 11, 101, 167, and 217 days after vaccination. One definite case reported a positive effect of a corticosteroid injection in the affected shoulder. The 1 probable case was diagnosed with bursitis by an orthopedic surgeon. Resolution of symptoms was documented in the medical record for 2 of the 16 (12.5%) risk interval cases (by 62 and 207 days after symptom onset). A similar proportion of control interval cases (5 of 51 [9.8%]) had symptom resolution documented (by days 81, 106, 136, 190, and 322 after symptom onset). The 14 risk interval cases without documented resolution of symptoms had a median time to the last visit with persistent shoulder symptoms of 386 days (range, 21 to 781 days), whereas the median for control interval cases was 95 days (range, 4 to 690 days).

Documentation in medical records showed that 12 case patients specifically mentioned the influenza vaccine as being the cause of their symptoms (Appendix Table, available at Annals.org). Nine of the 16 case patients (56%) in the risk interval attributed their injuries to vaccination; 2 of the 9 voiced specific concerns about the technique by which the vaccine was administered. One patient expressed that the vaccination was given too high, and the other that it was too deep. Three case patients outside of the risk interval also attributed their shoulder symptoms to vaccination: 2 identified their symptom onset outside of the risk interval (at days 3 and 5), and 1 did not have specific statements about symptom onset in the medical record. The median time from reported symptom onset to the first medical visit among patients attributing their symptoms to vaccination was 21 days (range, 3 to 144 days), which was similar to that for control interval cases (median, 22 days; range, 0 to 123 days) (P = 0.90).