¹ Center for Clinical Epidemiology and Biostatistics, University of Pennsylvania School of Medicine, Philadelphia, PA.
² Developing Evidence to Inform Decisions about Effectiveness (DEcIDE) Center, University of Pennsylvania School of Medicine, Philadelphia, PA.
³ Center for Education and Research on Therapeutics, University of Pennsylvania School of Medicine, Philadelphia, PA.
⁴ Department of Medicine, University of Pennsylvania School of Medicine, Philadelphia, PA.
⁵ Center for Outcomes and Evidence, Agency for Healthcare Research and Quality, U.S. Department of Health and Human Services, Rockville, MD.
⁶ Department of Veterans Affairs, Philadelphia, PA.

Identifiable information on which this report, presentation, or other form of disclosure is based is confidential and protected by Federal law, section 903(c) of the Public Health Service Act, 42 U.S.C. 299a-1(c). Any identifiable information that is knowingly disclosed is disclosed solely for the purpose for which it has been supplied. No identifiable information about any individual supplying the information or described in it will be knowingly disclosed except with the prior consent of that individual.

We performed an observational case-control study nested within the General Practice Research Database (GPRD) from the UK. GPRD data constitute the primary outpatient electronic medical records from approximately two thousand general practitioners (GPs) in the UK. Within the UK, approximately 98% of the population is registered with a GP, and virtually all patient care is coordinated by the GP through the National Health Service. The database is broadly representative of the UK population in terms of age, sex, and geography. More than 500 published epidemiologic studies have been performed using this database. The GPRD records outpatient prescriptions, outpatient diagnoses, and hospital diagnoses that are entered in response to receipt of a consultant's letter or discharge summary [8]. We used GPRD data entered from its establishment in 1987 through April 2002, and included only data classified as "up-to-standard" for identifying exposures and outcomes [8].

2.2. Study Base

The person-time eligible for inclusion in this nested case-control study (i.e., the study base within which the case-control study was nested) was all up-to-standard person-time occurring in individuals age 65 years and above with at least six months of prior up-to-standard follow-up in whom no instance of the study outcome had yet occurred. We required at least six months of data prior to study eligibility to ensure an adequate period in which to measure baseline variables. Cases and controls were sampled from this study base using incidence density sampling, as described below, which results in odds ratios (ORs) that are interpretable as incidence rate ratios [9].

2.3. Identification of Cases

Cases of the primary outcome consisted of all individuals who were hospitalized with pneumonia at any time following the six month anniversary date of that person's up-to-standard entry date. Cases of the secondary outcome consisted of persons who were hospitalized with aspiration pneumonia or aspiration pneumonitis at any time following the six month anniversary date of that person's up-to-standard entry date. The index date for cases was the date of the case event. For persons with more than one case event, we included only the first event. Diagnostic codes for the study outcomes are listed in the Appendix.

2.4. Identification of Controls

For each case, we randomly selected four controls from the same GP practice using incidence density sampling [9] based on amount of time in the study base. In particular, potential controls were persons who contributed person-time to the study base (defined above), who had not yet experienced the relevant study outcome before the corresponding case's event, assessed with respect to entry into the study base. The index date for controls was defined as the date of entry into the study base plus the number of days of event-free follow-up contributed by the corresponding case. This is analogous to time-to-event analyses of cohort studies, and ensures the interpretability of the resulting odds ratios as incidence rate ratios [9].

2.5. Ascertainment of Exposure

The principal exposure definition was a recorded prescription for any antidepressant medication in the thirty days preceding the index date. Prescriptions for chronically administered medications in the UK are issued in one-month increments. In order to obtain repeat prescriptions (i.e., prescription refills), patients must telephone or visit the GP office, which then issues and records a repeat prescription. To confirm the tendency to provide a one-month supply per prescription, we examined the frequency distribution of the number of days between subsequent prescriptions for commonly used antidepressants for the same patient. There was a very strong central tendency (i.e., mode and median) for repeat prescriptions to be issued approximately every thirty days (data not shown). In addition to considering exposure to any antidepressant, we also performed sub-analyses of the two major subclasses: cyclic agents (amitriptyline, amoxapine, butriptyline, clomipramine, desipramine, dibenzepin, dosulepin, doxepin, imipramine, iprindole, lofepramine, maprotiline, mianserin, mirtazapine, nortriptyline, opipramol, protriptyline, trimipramine) and selective serotonin uptake inhibitors and selective serotonin and norepinephrine inhibitors (SSRI/SSNRIs; citalopram, duloxetine, escitalopram, fluoxetine, fluvoxamine, paroxetine, sertraline, venlafaxine, zimeldine).

2.6. Ascertainment of Potential Confounding Factors

Potential confounding factors were ascertained based on the index date. These are listed in Tables 1 and 2, and included demographic factors, diagnoses ever recorded in the past (focusing on conditions previously implicated as risk factors for pneumonia, including chronic neurologic and pulmonary conditions [4]), drug prescribed ever in the past (as markers of disease), currently prescribed drugs (i.e., within the past thirty days), and measures of morbidity and health care utilization (e.g., number of non-antidepressant prescriptions in the prior thirty days, number of GP visits for non-depression diagnoses in the prior thirty days, number of hospitalizations in the prior six months).

2.7. Statistical Analysis

We first compared cases to controls with respect to antidepressant exposure and potential confounding factors. We next used conditional logistic regression to fit a "minimally adjusted" model that estimated the odd ratio (OR) for antidepressant use adjusting for age (expressed as a continuous variable), sex, and calendar year (to account for potential secular trends in pneumonia hospitalization and antidepressant use). We then examined each potential confounding factor individually using a change-in-estimate criterion [10]. In particular, we individually added each potential confounding factor to the minimally adjusted model. If any factor changed the OR of interest by ten percent or more, we considered it to be a confounding factor and retained it in the "change-in-estimate" adjusted model. We used this same strategy to examine the association with each antidepressant subclass. However, because of the small number of events of the secondary outcome (aspiration pneumonia), we were unable to examine associations for this outcome with specific antidepressant subclasses. We used SAS version 9.1 for all analyses.

The incidence of pneumonia is expected to peak during the winter months and decline during the summer, whereas antidepressant therapy might be prescribed year-round. We compared the seasonal distribution of hospitalization for pneumonia in patients with and without exposure to antidepressants using the method of Edwards [11]. This sums the vectors for the number of cases in each month extending from a common origin to the midpoints of twelve equal sectors of a circle from January (0 to 30°) through December (330 to 360°).

Based on pilot data, our a priori power calculations suggested that using a two-tailed α of 5%, we would have 85% statistical power to detect an OR for exposure to any antidepressant of 1.07 or higher for pneumonia hospitalization, and an OR of 1.5 or higher for hospitalization for aspiration pneumonia or pneumonitis [12].

We identified 12,044 cases of hospitalization for pneumonia. For each case, we identified four controls, for a total of 48,176 controls. Table 1 presents a comparison of cases and controls for the primary outcome. Cases were older than controls (median age 81 vs. 75), and had a substantially higher number of non-depression GP visits in the prior six months (average 2.8 vs. 0.8), hospital visits over the prior six months (mean 0.72 vs. 0.13), and number of non-antidepressant prescriptions in the prior month (mean 3.6 vs. 2.0). Most diagnoses, drugs prescribed ever in the past, and currently prescribed drugs examined were substantially more common among cases than controls (Table 1).

The OR (95% CI) for the association between pneumonia hospitalization and any antidepressant use, adjusted for age, sex, and calendar year (i.e., the minimally adjusted model) was 1.61 (1.46 to 1.78). However, when further adjusting for factors in Table 1 that changed the OR by ten percent or more (number of non-depression GP visits in the prior six months, ever past benzodiazepine prescription, and ever past antipsychotic prescription) the OR (95% CI) was 0.89 (0.79 to 1.00) (Table 1).

In the analysis of the major antidepressant subclasses, the ORs for cyclic antidepressants, SSRIs/SSNRIs, and other agents were similar to the corresponding OR for any antidepressant in baseline model (Table 1). The change-in-estimate adjusted ORs for cyclic agents and SSRI/SSNRIs were also similar to the corresponding OR for antidepressants as a whole (Table 1).

The seasonal vector of hospitalization for pneumonia in patients without prescriptions for antidepressants showed the expected pronounced wintertime peak (0.088 units at 27°, i.e., late January). This vector was closely matched in both relative amplitude and direction (0.103 units at 23°) by that of hospitalization for pneumonia in patients who were receiving antidepressants (data not shown).

3.2. Secondary Outcome: Hospitalization for Aspiration Pneumonia or Aspiration Pneumonitis

We identified 159 cases of hospitalization for diagnosed aspiration pneumonia or pneumonitis, and 636 controls. Table 2 presents a comparison of cases and controls for the secondary outcome. The relationships of the factors presented in Table 2 with hospitalization for aspiration pneumonia or pneumonitis were similar to those seen for the primary outcome, although the confidence intervals are much wider in Table 2 because of the small number of events. The OR (95% CI) for any antidepressant use, adjusted for age, sex, and calendar year was 1.45 (0.65 to 3.24). When further adjusting for factors in Table 2 that changed the OR by ten percent or more (number of non-depression GP visits in the prior six months, ever past benzodiazepine prescription, and ever past antipsychotic prescription) the OR (95% CI) was 0.63 (0.23 to 1.71) (Table 2). There were not enough users of specific antidepressant subclasses with this outcome to analyze these subclasses separately.