INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Salmeterol xinafoate (salmeterol) is the first long-acting -agonist, administered twice daily as maintenance therapy to prevent bronchospasm caused by asthma (1). By contrast, short-acting -agonists are rescue medications to be administered as needed to relieve bronchospasm. Salmeterol was introduced in the United Kingdom (U.K.) in 1990, and a postmarketing study there reported an increased risk of respiratory death associated with salmeterol, but the estimate was based on two deaths in the comparison group and was statistically very imprecise (2). Soon after salmeterol was introduced in the U.S. in 1994, there were a few reports of patients using the drug inappropriately as a rescue medication (3), and instances of respiratory arrests soon after starting salmeterol (4). We undertook this study to evaluate further the risk of severe nonfatal asthma among patients receiving salmeterol.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

We used automated health insurance claims from a New England insurer with about 1.3 million members with comprehensive coverage for medical care and prescription medications from January 1, 1993 to August 31, 1995. The base population includes about 1.5 million people covered by managed care plans or fee-for-service plans. We obtained all claims and enrollment records for insured members who received one or more dispensings of any oral or inhaled -agonist drug between January 1, 1993 and August 31, 1995, and who had insurance coverage for a continuous period of at least 30 d. For all patients, claims indicate date of service, site, provider, type of service, charges, and other details. For drug claims, data include the quantity provided and the National Drug Code indicating drug name, strength, packaging, and route of administration. For physician and hospital visits, claims indicate site of care as well as diagnoses and procedures for medical services. We also obtained records for each patient indicating date of birth, sex, and insurance enrollment and termination dates.

To assess selective prescribing, we compared future salmeterol recipients with the rest of the cohort. In this analysis, each person contributed person-time from a defined start date until the stop date. We defined the start date for each patient as the study start date (January 1, 1993) or start of insurance coverage, whichever came later. We defined the stop date as the study stop date (August 31, 1995), end of insurance coverage, or first salmeterol dispensing, whichever came first. We computed rates of dispensings of asthma medications and rates of acute asthma events. We then computed rate ratio (RR) estimates as the rate in future salmeterol recipients divided by the rate in the rest of the cohort, along with exact 95% confidence intervals (95% CI).

We included in the index group all patients who received salmeterol. To identify patients whose baseline risk was similar to that of the salmeterol patients, we sought for comparison patients who received an asthma medication that has a therapeutic role similar to salmeterol. At the time of this study, there were no other long-acting -agonists available in the U.S.; however, sustained-release (SR) theophylline is used to provide sustained bronchodilation. Therefore, we selected as potential comparison subjects patients who received theophylline. To minimize the effects of possible secular trends in prescribing and in acute asthma, we restricted comparison subjects to patients who received theophylline after the date that salmeterol was first dispensed in this population (April 5, 1994). In addition, owing to the possibility that the decision to add a medication might itself be related to baseline risk, we identified new users of theophylline (no dispensings in the past year) separately from past users.

To assess the effect of salmeterol on acute asthma, we defined a baseline period and a follow-up period, separated by an index date. The index date was defined for each patient as the first dispensing of the drug that defined his or her study group (either salmeterol or theophylline) after April 5, 1994. Each patient contributed person-time during follow-up for up to 1 yr, until the occurrence of an outcome event, termination of insurance coverage, or August 31, 1995, whichever came first. The baseline period was defined as the period 1 yr before the index date, and we used the baseline period to assess and control for differences in baseline risk owing to selective prescribing.

We defined as outcomes emergency care, hospitalization, and intensive care unit (ICU) stays for asthma. Emergency care was defined as an institutional claim with site code indicating emergency room or a professional claim with service type indicating emergency care, and an International Classification of Diseases (ICD) diagnostic code of 493 for asthma (5). We counted as hospitalizations those institutional claims with site code indicating hospital inpatient and primary diagnosis of asthma. ICU stays were the subset of asthma hospitalizations with a site code indicating ICU. To avoid counting a single episode as more than one outcome event, we allowed a maximum of one outcome event in 7 d.

We examined as markers of prognosis during the baseline period, history of acute asthma, other chronic respiratory disease, dispensings of asthma medications, and demographic factors. Demographic factors included date of birth, sex, calendar year, and calendar quarter. Asthma medications included number of dispensings during the baseline period of inhaled -agonists, oral -agonists, inhaled corticosteroids, oral corticosteroids, theophyllines, cromolyn sodium, and nedocromil sodium. For inhaled -agonists, we also computed a score to indicate the trend over time in dispensing frequency. The trend score is a generalization of a method described by Suissa and coworkers (6). Each of the acute asthma outcomes also was used to characterize baseline risk. We used diagnostic codes for other illnesses to identify comorbidities, including chronic obstructive pulmonary disease (COPD) (ICD 490-495, except 493). We created variables indicating the frequency of asthma events and dispensings of each asthma medication during the entire baseline period, as well as during the 3 mo immediately preceding the follow-up period. Events occurring during the latter period were ascertained separately to identify any patients who might experience sudden worsening of asthma just before the start of follow-up, and who might have a different risk profile.

In the primary analysis, each patient contributed person-time to the analysis from the index date until the end of the follow-up period, and the person-time was classified according to treatment assignment (salmeterol or theophylline). We also examined dispensing patterns during the follow-up period. We further classified person-time among salmeterol recipients by time since dispensing as current use (0 to 30 d after a dispensing), recent use (31 to 60 d after a dispensing), and past use (> 60 d after a dispensing). Both categories of recent use and past use were interrupted if a new dispensing occurred, which would generate another period of current use. This analysis assumes that patients begin using salmeterol on the dispensing date, and continue using salmeterol according to instructions on the product label, which imply that a metered dose inhaler should contain 30 d of medication. To maintain comparability of effect estimates, we used as a common reference category the entire follow-up period of the theophylline group. In this analysis, we also considered dispensings of other asthma medications during the follow-up period.

To estimate the effect of salmeterol on acute asthma, we computed incidence rates of acute asthma during the follow-up period as the number of people experiencing each type of event divided by the person-years at risk. Rate ratio estimates were computed as the rate in the salmeterol group divided by the rate in the theophylline group. To control for differences between comparison groups in baseline risk, we examined cross-tabulations of treatment assignment by each covariate, and each covariate by each outcome. We used cross-tabulations to identify baseline factors related both to asthma treatments and to asthma outcomes during the follow-up period. We also used Poisson modeling to assess the impact of each covariate on the risk of each outcome while controlling simultaneously for other covariates (7). We considered in the models age (0-14, 15-34, 35-49, 50+), sex, history of COPD, calendar year and quarter, baseline hospitalization and ICU stay for asthma, and baseline use of each asthma drug. Dispensings of asthma medications and acute asthma events were analyzed as both continuous variables and categorical variables. In categorical analyses, indicator variables were used to classify frequency of baseline medical events and drug use as none versus one or more. We also computed standardized effect estimates, with weights taken as the proportional distribution of person-time of the entire study population across covariates.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

We identified 61,712 people who fulfilled the eligibility criteria and who comprised the total cohort. Of these, 2,708 patients received salmeterol, with the earliest dispensing occurring on April 5, 1994. Compared with the rest of the cohort, and adjusting for demographic factors, future salmeterol recipients had two to four times more frequent dispensings of asthma drugs, including inhaled -agonists (RR = 2.80, 95% CI = 2.75, 2.86), inhaled steroids (RR = 3.67, 95% CI = 3.56, 3.78), and oral steroids (RR = 3.38, 95% CI = 3.23, 3.54). In addition, future salmeterol recipients also had more frequent emergency care (RR = 2.85, 95% CI = 2.41, 3.35), hospitalization (RR = 4.34, 95% CI = 3.64, 5.17), and ICU stays (RR = 6.95, 95% CI = 4.31, 11.21) for asthma.

Among patients who did not receive salmeterol, we identified 3,825 patients who received SR theophyllines after April 5, 1994. At baseline, future salmeterol recipients were generally similar to theophylline recipients with regard to age, sex, dispensings of inhaled -agonists, and visits for acute care for asthma (Table 1). However, during the baseline period the salmeterol group was more likely to have received inhaled steroids (44% versus 34%), and oral steroids (30% versus 17%), and less likely to have received theophylline (27% versus 45%).

For the follow-up period, the numbers of cases, person-years, and unadjusted rates of each outcome event for the salmeterol group and the theophylline group are shown in Table 2. In the salmeterol group, 65 patients had emergency care, 69 patients were hospitalized, and 13 patients stayed in the ICU. In the larger theophylline group, 130 patients received emergency care, 100 patients were hospitalized, and 14 patients received intensive care. Overall, the unadjusted rates of emergency care for asthma were similar in the salmeterol group (0.036 case/year) and the theophylline group (0.038 case/year), while the salmeterol group had higher rates of hospitalization (0.038 case/year versus 0.026 case/year), and ICU stays (0.007 case/year versus 0.004 case/year).

The strongest predictor of acute asthma was a recent history of acute asthma. In each treatment group, patients with multiple baseline emergency care visits were about 10 times more likely to be hospitalized for asthma during the risk period compared with patients who had not received emergency care. Similarly, patients with multiple hospitalizations were about 20 times more likely to be hospitalized compared with patients who had no baseline hospitalization, and patients with multiple ICU stays were about 60 times more likely to be hospitalized during the follow-up period than patients who did not have an ICU stay during the baseline period. Baseline use of asthma medications was only weakly related to rate of acute outcomes during the risk period (RR < 2). Trend in use of inhaled -agonists during the baseline period and comorbidities including COPD were not predictive of acute asthma outcomes during the follow-up period. Although previous attacks requiring hospitalization was the most important confounding variable, we also controlled for effects of certain baseline drugs, that also were related to salmeterol use and, though weakly, to acute asthma (e.g., oral steroids).

Standardized analyses and Poisson regression models produced similar adjusted rate ratio estimates, although the standardized estimates produced wider confidence intervals. Because the data were too sparse to provide a secure basis for certain assumptions required by multivariate models (e.g., uniformity of effect estimates across subpopulations) (8), we present the standardized estimates (Table 3). Adjusting for age, sex, baseline asthma outcomes, and baseline asthma drug use decreased the rate ratio estimate for emergency care from 0.95 to 0.69 (95% CI = 0.42, 1.11). Similar adjustment reduced the effect estimate for hospitalization from 1.33 to 1.09 (95% CI = 0.50, 1.98). For ICU stays, the unadjusted rate ratio estimate was 1.77 and decreased with adjustment to 0.81 (95% CI = 0.25, 2.62). Effect estimates were very similar when the analyses were repeated using only new recipients of theophylline, though the confidence intervals were somewhat wider owing to smaller numbers. Therefore, we present the results for the comparison group that included both new and past recipients of theophylline.

When the salmeterol group was restricted to periods within 30 d of a dispensing (current use), the rate ratio estimates for emergency care (RR = 0.27, 95% CI = 0.12, 0.58), hospitalization (RR = 0.94, 95% CI = 0.54, 1.66), and ICU stays (RR = 0.56, 95% CI = 0.12, 2.68) were lower (and less precise) than the corresponding summary rate ratio estimate for the entire follow-up period. In moving from current use to recent use, the effect estimates tend to increase somewhat, and then decline again during past use. Controlling for effects of concurrent asthma medication did not change these results substantially.

Because the diagnosis of asthma is reportedly most accurate among nonelderly adults (9), we computed age-specific effect estimates. Among people 15 to 49 yr of age, in whom the diagnosis of asthma should be most accurate, the adjusted rate ratio estimate for emergency care was 0.69 (95% CI = 0.46, 1.03), for hospitalization the estimate was 0.97 (95% CI = 0.62, 1.51), and for ICU stay the estimate was 0.63 (95% CI = 0.18, 2.13). Rate ratio estimates also were not elevated for these outcomes among children or elderly adults.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

We used automated claims records from a large New England insurer to evaluate the relation between salmeterol and severe asthma requiring emergency care, hospitalization, or ICU stay. Virtually all of the salmeterol patients had received a short-acting -agonist before they received salmeterol. Compared with -agonist recipients as a whole, salmeterol was prescribed preferentially to those patients with a recent history of emergency care and hospitalization and who had received asthma medications more frequently. That is, patients who received salmeterol appeared to have more severe asthma than patients who received short-acting -agonists.

In consideration of selective prescribing of salmeterol to high-risk patients, we selected as a comparison group patients who, in addition to receiving a short-acting -agonist, also received theophylline. This comparison group provided greater baseline comparability with the salmeterol group than did recipients of short-acting -agonists, but did not eliminate baseline differences. During the follow-up period, unadjusted rates of emergency care were similar in both treatment groups, but rates of hospitalization and ICU stays were higher in the salmeterol group. After adjusting analytically for baseline history of an asthma hospitalization, ICU stay, and dispensings of inhaled -agonists, inhaled steroids, oral steroids, and theophyllines, rate ratio estimates for each outcome declined, and salmeterol was not associated with an increased risk of any of these outcomes. Indeed, rates tended to be similar or lower in salmeterol recipients than in comparison subjects. Restricting the analysis to current use of salmeterol produced lower effect estimates for each outcome. These analyses indicate that salmeterol is prescribed to patients at high risk of acute asthma and that, after adjusting for baseline risk and compared with patients who receive theophylline, salmeterol is not associated with an increased risk of emergency care, hospitalization, or ICU stay for asthma.

One limitation of this analysis is that treatments were not randomly assigned, and salmeterol was prescribed preferentially to patients at high baseline risk of acute asthma. Inaccuracy in measuring baseline risk would reduce the ability to control confounding. We measured baseline risk using history of emergency care, hospitalization, and ICU stay during the past year, as well as dispensings of asthma medications and demographic factors. We verified that these markers were associated with both salmeterol and with acute asthma, and a history of baseline hospitalizations was the only factor strongly related both to salmeterol and to subsequent risk of acute asthma. To the extent that these markers do not fully capture baseline risk of acute asthma, controlling for these markers would not eliminate the effects of differences in baseline risk. Because salmeterol was prescribed preferentially to patients at increased baseline risk, controlling for these markers reduced the effect estimates. It is reasonable to conclude, therefore, that controlling for more accurate markers of baseline risk should reduce the effect estimates even further. If so, improved assessment of baseline risk would likely result in lower effect estimates for salmeterol. The difficulty of measuring baseline risk may have reduced to some degree our ability to identify reduced risks of acute asthma among recipients of salmeterol.

One would expect that insurance claim records give rise to certain kinds of measurement error. For instance, a dispensing of salmeterol does not necessarily indicate use of salmeterol, and a diagnosis code for asthma may not indicate presence of the disease. These concerns are important because a general (nondifferential) misclassification of exposure or outcome would be expected to introduce bias that would tend to diminish the strength of any real association (7). We addressed this concern by restricting the analysis to groups of patients among whom measurement error would be reduced. For instance, to reduce exposure misclassification, we examined the period 30 d after a dispensing for salmeterol. We reason that regular salmeterol use would be more likely immediately after a dispensing than during subsequent periods. To address possible diagnostic inaccuracy, we conducted an alternative analysis, restricted to nonelderly adults, among whom an asthma diagnosis is more accurate (9). The impact of both of these restrictions was to decrease the magnitude of the effect estimates for salmeterol and acute asthma outcomes. Thus, although some bias from measurement error is likely, analyses in which we attempted to control for errors consistently produced lower effect estimates. These results argue against the hypothesis that bias from misclassification of exposure or disease masked an increased risk of salmeterol.

Finally, although the effect estimates reported here reveal no increased risk of severe acute asthma, the events studied are rare and the effect estimates are somewhat imprecise, especially for ICU stays. While we note that increasingly accurate analyses tended to produce progressively lower effect estimates, confidence intervals are still compatible with a fairly broad range of values. In addition, because most asthma patients receive treatment, epidemiologic studies compare the risk among salmeterol patients with the risk among patients receiving other medications (e.g., recipients of theophylline) which, in turn, could be related to acute asthma.

Postmarketing surveillance of a possible relation between an asthma medication and severe asthma presents a difficult challenge. Case reports have shown that salmeterol patients sometimes suffer respiratory failure, and that deaths have occurred within a few months after starting salmeterol (3, 4). There have been two epidemiologic studies in the U.K. that assessed the relation between salmeterol and acute asthma (2, 10). A randomized study found no relation between salmeterol and hospitalization for asthma (RR = 0.95, 95% CI = 0.75, 1.20), and an imprecise association with asthma-related death (RR = 3.0, 95% CI = 0.70, 20) (2). These results indicate the absence of an elevated risk of hospitalization among salmeterol users; such reassurance cannot be derived regarding asthma-related death, although the effect estimate is very imprecise.

A nonrandomized study reported higher rates of respiratory death among patients prescribed ipratropium bromide (RR = 1.8, 95% CI = 0.4, 9.6) and theophylline (RR = 3.0, 95% CI = 0.4, 22.4) than among patients prescribed salmeterol (10). These results do not indicate any increased risk of respiratory death associated with salmeterol but, again, the effect estimates are very imprecise.

A prescription event monitoring study of 15,407 salmeterol patients in the U.K. whose physicians completed questionnaires reported elevated rates of a variety of outcomes, including respiratory, cardiovascular, alimentary, and central nervous system outcomes, during the first month after starting therapy compared with the subsequent 5 mo (11). The absence of a comparison group makes it difficult to interpret these results, but the temporal associations were even stronger for psychiatric outcomes (e.g., malaise) than respiratory outcomes. Although it is possible that salmeterol is associated with psychiatric outcomes, a methodologic explanation, such as more complete reporting of all outcomes soon after a dispensing, seems more likely.

A recent case-control study in the U.K. compared the risk of ICU admissions for asthma among users of salmeterol compared with nonusers (12). The overall relative risk estimate was 2.32 (95% CI = 1.05, 5.16), but there was evidence of selective prescribing, as among control subjects salmeterol had been prescribed three times more frequently to patients with a recent history of asthma hospitalization. When the analysis was stratified by a recent history of hospitalization, the relative risk declined; among patients with a recent hospitalization, the relative risk estimate was 1.42 (95% CI = 0.49, 4.10). These results indicate that any association between salmeterol and ICU admission has been exaggerated by selective prescribing of salmeterol to high-risk patients. Although effect estimates tended to indicate a slightly higher risk for salmeterol users, and the confidence intervals do not exclude a strong association, in consideration of the effect of selective prescribing, the investigators concluded that there was little evidence that salmeterol increases the risk of ICU admission.

The overall picture with regard to a possible adverse effect of salmeterol on severe acute asthma is gaining coherence. Case reports suggest a possible association between salmeterol and respiratory arrest. Asthma death is difficult to study because it is rare, and effect estimates from formal studies are too imprecise to either rule-in or rule-out a relation between salmeterol and asthma death. The current study strengthens the hypothesis that an association between salmeterol and severe asthma has been created by selective prescribing, and that salmeterol is not related to an increased risk of severe nonfatal asthma requiring emergency care, hospitalization, or ICU care for asthma. This study does not address mechanisms by which salmeterol would be related to asthma death but not to nonfatal asthma. Typically wide variation in biologic responses suggests that many mechanisms of acute asthma that resulted in death would, in at least some instances, be survivable. If so, they would also produce an association with severe nonfatal asthma. For mechanisms that imply a relation with both fatal and nonfatal asthma, mounting evidence against an association with hospitalization also would argue against an association with asthma death.