Comparative effectiveness of macrolides and quinolones for patients hospitalized with acute exacerbations of chronic obstructive pulmonary disease (AECOPD)

Data Elements

For each patient, we assessed age, gender, race, marital and insurance status, principal diagnosis, comorbidities, and specialty of the attending physician. Comorbidities were identified from ICD‐9 secondary diagnosis codes and Diagnosis Related Groups using Healthcare Cost and Utilization Project Comorbidity Software (version 3.1), based on the work of Elixhauser et al.17 In addition, to assess disease severity we recorded the presence of chronic pulmonary heart disease, the number of admissions for COPD during the 12 months prior to the index admission, and arterial blood gas testing.18^,19 We also identified pharmacy or diagnostic charges for interventions that were recommended in current guidelines (beta‐adrenergic and anticholinergic bronchodilators, steroids, and noninvasive positive‐pressure ventilation); those that were not recommended or were of uncertain benefit (methylxanthine bronchodilators, spirometry/pulmonary function testing, mucolytic medications, chest physiotherapy, and sputum testing); and drugs that might be associated with severe exacerbations or end‐stage COPD (loop diuretics, morphine, and nutritional supplements).1^,69 Hospitals were categorized by region (Northeast, South, Midwest, or West), bed size, setting (urban vs. rural), ownership, and teaching status.

Antibiotic Class and Outcome Variables

Our primary predictor variable was the antibiotic initiated during the first 2 hospital days and continued for at least 2 days, regardless of other antibiotics the patient may have received during the course of hospitalization. Because we anticipated low in‐hospital mortality, our primary outcome was a composite measure of treatment failure, defined as initiation of mechanical ventilation after hospital day 2, in‐hospital mortality, or readmission for COPD within 30 days of discharge.20 Secondary outcomes included hospital costs and length of stay, as well as allergic reactions identified by ICD‐9 code, and antibiotic‐associated diarrhea, defined as treatment with either metronidazole or oral vancomycin begun after hospital day 3.

Statistical Analysis

Summary statistics were computed using frequencies and percents for categorical variables; and means, medians, standard deviations, and interquartile ranges for continuous variables. Associations between antibiotic selection and patient and hospital characteristics were assessed using chi‐square tests for categorical variables and z‐tests for continuous variables.

We developed a series of multivariable models to evaluate the impact of initial antibiotic selection on the risk of treatment failure, length of stay, and total cost. In order to account for the effects of within‐hospital correlation, generalized estimating equation (GEE) models with a logit link were used to assess the effect of antibiotic selection on the risk of treatment failure, and identity link models were used for analyses of length of stay and cost. Unadjusted and covariate‐adjusted models for treatment failure were evaluated with and without adjustments for propensity score. A propensity score is the probability that a given patient would receive treatment with a macrolide, derived from a nonparsimonious model in which treatment with a macrolide was considered the outcome. The propensity model included all patient characteristics, other early treatments and tests, comorbidities, hospital and physician characteristics, and selected interaction terms.21 Length of stay and cost were trimmed at 3 standard deviations above the mean, and natural log‐transformed values were modeled due to extreme positive skew. In addition, we carried out matched analyses in which we compared the outcomes of patients who were treated with a macrolide to those with similar propensity scores (ie, with similar likelihood of receiving a macrolide) who received a quinolone.22

Finally, to reduce the threat of residual confounding by indication, which can occur if sicker patients are more likely to receive a particular antibiotic, we developed a grouped treatment model, in which all patients treated at the same hospital were assigned a probability of treatment with a macrolide equal to the overall treatment rate at that hospital.23 This is an adaptation of instrumental variable analysis, a well‐accepted technique in econometrics with growing use in health care.24^,25 It attempts to assess whether patients treated at a hospital at which quinolones are used more frequently have better outcomes than patients treated at hospitals at which macrolides are used more frequently, while adjusting for other patient, physician, and hospital variables. It ignores the actual treatment the patient received, and instead substitutes the hospital's rate of macrolide use. By grouping treatment at the hospital level, this method greatly reduces the possibility of residual selection bias, unless hospitals that use a lot of macrolides have patients who differ in a consistent way from hospitals which use mostly quinolones.

All analyses were performed using SAS version 9.1 (SAS Institute, Inc., Cary, NC).