We used the Merative MarketScan Commercial Database (2006–2019), which includes longitudinal, patient-level data on enrollment, adjudicated inpatient and outpatient medical insurance claims, and outpatient pharmacy-dispensed medications. The database includes individuals with employer-sponsored commercial insurance and their spouses and dependents [14]. The Washington University School of Medicine Institutional Review Board deemed this study exempt from human subjects review.

We identified a cohort of adults aged 18–64 years with an International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM), or an International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM), diagnosis code for pneumonia (Supplementary Table 1) (ignoring diagnostic/rule-out claims) diagnosed in an outpatient setting from 1 July 2007 to 30 November 2019. We required same-day treatment (index date) with a guideline-recommended oral antibiotic (see “Antibiotic Exposure” section) per the 2007 and 2019 Infectious Diseases Society of America / American Thoracic Society (IDSA/ATS) CAP guidelines [2, 3]. We also required continuous medical and prescription drug insurance enrollment during the 180-day baseline period before the index date. We implemented several exclusion criteria to create a final cohort of otherwise healthy adults with CAP that was unlikely to be caused by a multidrug-resistant pathogen (Supplementary Tables 1-6) [2, 3, 15-20]. The Supplementary Methods provide additional details on these exclusions.

We categorized antibiotic regimens in accordance with the 2007 and 2019 IDSA/ATS guidelines on the clinical management of adults with CAP [2, 3]. We included antibiotic regimens recommended in both guidelines (macrolide, doxycycline, respiratory fluoroquinolones, β-lactam plus macrolide, β-lactam plus doxycycline) and added β-lactam monotherapy, given common real-world prescribing for CAP. We did not include amoxicillin monotherapy in the primary analysis, as this new recommendation from the 2019 guidelines [3] occurred very late in the study period. Rather, amoxicillin initiators were included with β-lactam monotherapy in the main analysis, and further studied in subgroup analyses. We additionally categorized antibiotic regimens by coverage type, defined as narrow-spectrum (macrolide, doxycycline) or broad-spectrum (respiratory fluoroquinolones, β-lactam monotherapy, β-lactam plus macrolide, β-lactam plus doxycycline) [2]. Supplementary Table 5 presents antibiotic generic names and coverage by antibiotic regimen category.

We used ICD-9-CM and ICD-10-CM diagnosis codes to identify individual ADE outcomes, including nausea/vomiting/abdominal pain, rash/urticaria, non–C difficile diarrhea, vulvovaginal candidiasis/vaginitis, unspecified allergy, renal failure, tendinopathy (including tendon rupture), CDI, and anaphylaxis/angioedema/laryngeal edema [21]. This approach captures ADEs that are serious enough for the patient to seek medical care. Outcome-specific follow-up periods ranged from 2 to 90 days, based on expert knowledge of anticipated onset after antibiotic initiation (Supplementary Table 7) [9, 10, 22]. We also used negative-control outcomes—ankle/knee sprain and influenza vaccination—which have been used previously to assess the comparability of treatment groups (Supplementary Table 7) [9, 23]. Each negative-control outcome met the criteria of being (1) causally unrelated to the treatment under study and (2) sharing common confounding structures with the treatments and outcomes of interest (ie, the same confounders affect the negative control as the outcome of interest) [24, 25]. To ensure identification of new-onset outcomes, we excluded index events coded for the outcome of interest within 30 days before the index event date, except for influenza vaccination.

Potential confounders of the association between antibiotic exposure and ADE outcomes were identified a priori based on clinical knowledge. The following patient- and provider-level characteristics were assessed at the index event date: age, sex, geographic region, health insurance plan type, urban residence (ie, residence in a metropolitan statistical area), calendar month, calendar year, provider specialty, provider outpatient location (ie, office, emergency department, urgent care center, other), and ceftriaxone administration on the date of CAP diagnosis (since patients with CAP frequently receive a one-time dose of intravenous [IV] ceftriaxone). The presence of baseline emergency department encounters, number of physician office visits (0, 1, 2, 3+), number of unique medication therapeutic groups (0, 1, 2, 3+), and frailty markers (Supplementary Table 8) were assessed during the 180-day baseline period [26]. Comorbidities (other than those used to exclude adults with comorbidities per CAP treatment guidelines) were captured using an all-available look-back period (Supplementary Table 6) [18, 19].

Baseline covariates were summarized by antibiotic regimen. We used an as-treated approach to evaluate the association between antibiotic regimens and ADE outcomes. We estimated the weighted cumulative incidence of individual ADE outcomes using Kaplan-Meier methods. Adults were followed until the first occurrence of the ADE outcome or a censoring event (health insurance disenrollment, diagnosis of a non-CAP bacterial or viral infection, diagnosis of viral pneumonia or pneumonia due to a rare organism, influenza diagnosis or oseltamivir prescription, hospitalization, end-of-outcome-specific follow-up period, or end-of-study period on 31 December 2019). In as-treated analyses, adults are considered to be exposed to the initial treatment until changes in treatment. We censored for diagnosis of non-CAP infections as a proxy for treatment changes, given the possibility of ADE outcomes related to other infections (eg, nausea/vomiting/abdominal pain, rash/urticaria) or medication use not recorded in the database (eg, over-the-counter medication, prescription paid out-of-pocket). To account for potential confounding, we used standardized mortality ratio (SMR) weights applied to the Kaplan-Meier estimator. SMR weights standardize the covariate distribution of a comparator cohort (eg, doxycycline initiators) to a reference cohort (ie, macrolide initiators), resulting in unconfounded treatment comparisons, assuming no unmeasured confounding [27]. For each macrolide/comparator contrast, SMR weights were calculated using propensity scores (p) estimated from multivariable logistic regression, where p represents the probability of initiating macrolides versus each other antibiotic of interest, accounting for baseline covariates. Age was modeled with restricted cubic splines. We used absolute standardized mean differences to assess the balance of measured covariates between treatment groups, with values less than 0.10 considered adequate [28]. If balance was not initially achieved for all covariates, we applied asymmetric trimming of the propensity score as needed to account for areas of non-overlap. We estimated both absolute and relative effect measures to assess the association between different antibiotic regimens and ADE outcomes. Risk differences (RDs) per 1000 treatment episodes and risk ratios (RRs) of the cumulative incidence functions were calculated over follow-up for each ADE outcome, with corresponding 95% confidence intervals (CIs) calculated using nonparametric bootstrap sampling with replacement (N = 1000). The number needed to harm (NNH) was calculated by inverting the weighted RD.

We conducted prespecified subgroup analyses for broad- and narrow-spectrum β-lactam monotherapy, with amoxicillin and ampicillin considered narrow-spectrum and other β-lactam agents considered broad-spectrum (Supplementary Table 5). We used analogous methods to evaluate the association between macrolides (vs comparator regimen) and the risk of negative-control outcomes, wherein each outcome is causally unrelated to the antibiotic exposure. In the absence of a biologically plausible mechanism for antibiotics to cause each negative-control outcome, effect estimates should be null in the absence of confounding [29]. Analyses were conducted using SAS software version 9.4 (SAS Institute, Inc.) and R Statistical Software version 4.3 (R Core Team) with R packages dplyr, forestplot, ggplot2, and patchwork.

After applying eligibility criteria, the outpatient CAP cohort included initiators of macrolides (n = 63 474), doxycycline (n = 11 301), fluoroquinolones (n = 56 388), β-lactam monotherapy (n = 9652), β-lactam + macrolide (n = 4322), and β-lactam + doxycycline (n = 612) (see the flow diagram in Figure 1). Due to rare use, the β-lactam + doxycycline treatment group was excluded from further analyses. Thus, the final study sample included 145 137 otherwise healthy adults with outpatient CAP. The distribution of antibiotic regimens was as follows: macrolides (44%), doxycycline (8%), fluoroquinolones (39%), β-lactam monotherapy (7%), and β-lactam + macrolide (3%). Overall, adults with CAP had a mean (SD) age of 44 (13) years and 47% were female. The majority were seen in an office setting (69%), as well as the emergency department (16%) or urgent care (11%). Baseline characteristics are reported by antibiotic regimen in Table 1. The distribution of antibiotic agents within each antibiotic treatment group is presented in Supplementary Table 9.

After exclusions for ADE outcomes occurring in the 30 days before antibiotic initiation (Supplementary Table 10) and propensity score weighting, all measured characteristics were well balanced between antibiotic treatment groups (Supplementary Figure 1). The crude risk of ADEs observed during follow-up varied widely (Figure 2, Supplementary Table 11). Clostridioides difficile infection and anaphylaxis/angioedema/laryngeal edema were rare (Supplementary Table 12); thus, no further analyses were conducted for these outcomes. Figure 3 and Supplementary Figure 2 present the weighted cumulative incidence curves for individual ADEs by antibiotic regimen and type of coverage.

Of the narrow-spectrum antibiotic regimens (doxycycline vs macrolide), we observed similar risks for most individual ADEs, including rash/urticaria, non–C difficile diarrhea, vulvovaginal candidiasis/vaginitis, unspecified allergy, and tendinopathy; the exception was nausea/vomiting, which was associated with increased risk in doxycycline users (weighted RD per 1000 treatment episodes, 2.18; 95% CI, .17–4.54) (Figure 2). The NNH suggests that 1 additional case of nausea/vomiting/abdominal pain would occur during a 14-day follow-up period for every 455 doxycycline treatment episodes.

In contrast, each broad-spectrum antibiotic regimen was associated with increased risk of several individual ADEs, compared with a macrolide. Specifically, fluoroquinolones were associated with increased risk of nausea/vomiting/abdominal pain and unspecified allergy; β-lactam monotherapy was associated with increased risk of nausea/vomiting/abdominal pain, rash/urticaria, non–C difficile diarrhea, and vulvovaginal candidiasis/vaginitis; and β-lactam + macrolide use was associated with increased risk of non–C difficile diarrhea and vulvovaginal candidiasis/vaginitis (Figure 2). On the absolute scale, the highest increased risk of individual ADEs was observed for vulvovaginal candidiasis/vaginitis, among users of β-lactam monotherapy (NNH = 278) or combination therapy with a macrolide (NNH = 204); and for non–C difficile diarrhea, among users of β-lactam monotherapy (NNH = 217) or combination therapy with a macrolide (NNH = 175). The interpretation of crude and weighted results was consistent for analyses on the absolute and relative scales, with few exceptions (Supplementary Table 11).

Analyses restricted to β-lactam broad-spectrum monotherapy revealed similar results as the primary analysis that included all β-lactam monotherapy (Supplementary Figure 3, Supplementary Table 13). The number of ADE outcomes among adults treated with narrow-spectrum β-lactam monotherapy was small; thus, the effects of this antibiotic regimen on most ADEs were not estimable.

In analyses of negative-control outcomes, the effects were not estimable for 3 of 15 comparisons due to small sample size (Figure 4). Otherwise, weighted RD estimates were null for all treatment comparisons for ankle/knee sprain outcomes and all treatment comparisons for influenza vaccination, except for β-lactam + macrolide versus macrolide (weighted RD per 1000 treatment episodes, 7.71; 95% CI, 1.40–14.90).

Our results are subject to several possible limitations. First, the effect estimates may be subject to residual confounding given the nonrandomized study design. The unadjusted and weighted estimates were generally similar, indicating little confounding by measured factors. However, our analysis was not able to account for certain risk factors for ADE outcomes that are unmeasured or poorly measured in health insurance claims data (eg, common noninfectious causes of rash/urticaria). Bias may be present if patients with these risk factors experienced differential exposure to the antibiotic regimens under study, although we expect this to be unlikely. To mitigate measured and unmeasured confounding, we implemented several established epidemiologic methods, including an active-comparator new-user design by restricting the cohort to a homogeneous population with indication for treatment [34, 35]. We further restricted the study population to otherwise healthy adults [36] and applied propensity score methodology to account for potential confounders [37]. Last, we used negative-control outcomes to assess the comparability of initiators of different CAP-related antibiotics under study; the results were generally null, indicating that residual confounding was minimal [29].

Second, study inclusion was based on diagnosis codes for pneumonia that have not been validated in the outpatient setting, and there is general diagnostic uncertainty related to pneumonia [38]. To further refine the study population, we required a same-day chest X-ray and CAP-related antibiotic prescription, as well as the absence of codes for other bacterial and viral infections within 30 days before the CAP diagnosis. Despite possible inclusion of patients without true CAP, our results are useful because they generalize to adults treated for CAP in the outpatient setting, all of whom are at risk of antibiotic-related ADEs. Third, the antibiotic exposure was based on pharmacy-dispensing billing claims. While this is the gold standard of prescription drug ascertainment, information on adherence was not available and failure to consume the medication could result in bias toward the null. Fourth, our claims-based algorithms to identify ADEs have been used in previous studies [8-10] but have not been validated. While claims databases typically have excellent ascertainment of serious events, milder safety events are less likely to be reported to providers, as demonstrated by a study that identified 10 times more ADEs via telephone calls to families versus manual review of electronic health record data [39]. It is unlikely that outcome misclassification would be differential by exposure. Fifth, we did not have the statistical power to evaluate the effect of narrow-spectrum β-lactam monotherapy (ie, amoxicillin or ampicillin monotherapy) on the risk of individual ADEs. Subgroup analyses restricted to β-lactam broad-spectrum monotherapy revealed similar results as the primary analysis that included all β-lactam monotherapy. Sixth, we did not have the statistical power to evaluate the effect of antibiotic duration on ADEs. Last, the study was restricted to otherwise healthy, commercially insured adults aged 18–64 years; thus, results may not be generalizable to other populations such as adults with comorbidities who may be at increased risk of ADEs due to underlying health conditions or drug interactions.