Results from Two Observational Designs Free of Immortal Time Bias
http://www.100md.com
美国呼吸和危急护理医学 2005年第8期
Worldwide Epidemiology, GlaxoSmithKline, Greenford
National Centre for Health Outcomes Development
Department of Epidemiology and Population Health, London School of Hygiene and Tropical Medicine
Thoracic Medicine, National Heart and Lung Institute, Imperial College, London
North West Lung Centre, South Manchester University Hospital Trust, Manchester, United Kingdom
Worldwide Epidemiology, GlaxoSmithKline, Upper Providence
Center for Clinical Epidemiology and Biostatistics, University of Pennsylvania School of Medicine, Philadelphia, Pennsylvania
Department of Heart and Lung Diseases, Hvidovre Hospital, Hvidovre, Denmark
ABSTRACT
Rationale: Recent cohort studies in chronic obstructive pulmonary disease (COPD) have questioned the validity of previously reported associations between inhaled corticosteroids (ICS) and reductions in mortality and rehospitalization in observational studies. Using time-dependent versions of statistical survival models, these studies have suggested immortal time bias as responsible for the proposed beneficial association. Objectives: We explored the extent of this bias in a study of patients with COPD monitored for a year from COPD discharge with two designs free of any immortal time bias in the General Practice Research Database in the United Kingdom. Methods: In Design 1, we used only patients whose treatment status was defined on the same day of discharge to obtain a matched cohort based on propensity scores, which were derived from the patient-level baseline characteristics. In Design 2, we identified all in the study cohort who experienced death or rehospitalization and then matched each case to up to four noncases by randomly sampling from the cohort risk sets without regard to treatment status. Measurements and Main Results: The propensity scores matched cohort analysis of 786 patients without a wait time found a significant risk reduction associated with use of ICS: hazard ratio, 0.69 (95% confidence interval, 0.52eC0.93). The matched nested case-control analysis of 2,222 patients, designed without regard to exposure status and hence free of immortal time bias, gave a similar association with exposure to ICS in the last 6-month period: hazard ratio, 0.71 (0.56eC0.90). Conclusions: We conclude that immortal time bias cannot account for the risk reduction associated with ICS exposure in observational studies.
Key Words: chronic obstructive pulmonary disease; epidemiologic methods; immortal time bias; nested case-control design; propensity scores matched cohort; time-dependent model
Chronic obstructive pulmonary disease (COPD) is currently the fourth leading cause of death, responsible for more than 2.5 million deaths per year worldwide, and it is estimated to become the third most important cause of death by 2020 (1). COPD is a leading cause of hospitalization in adults in the United States, particularly in older populations. In 1998, almost 662,000 hospitalizations (1.9% of total hospitalizations) were attributed to COPD, and the rate of hospitalizations with COPD as the primary cause of hospitalization was 38.3/10,000 individuals in 1998 (2). An additional 2.5 million hospitalizations (7.0% of total hospitalizations) had COPD listed as a contributing cause.
Inhaled corticosteroids (ICS) reduce exacerbations in patients with moderate to severe COPD (3eC5). The Global Initiative for Obstructive Lung Disease guidelines now recommend long-term use of ICS for patients with COPD with post-bronchodilator FEV1 of less than 50% predicted and repeated exacerbations (6), as do the recent American Thoracic Society/European Respiratory Society COPD treatment guidelines (7).
Several observational studies have suggested that treatment with ICS is associated with risk reduction of rehospitalization or death (8eC12). These pharmacoepidemiologic studies have used automated, linked health care and pharmacy data or electronic medical record databases, and, due to the nonrandom allocation of treatment exposure, they are by nature subject to bias. However, other recent COPD observational studies have failed to find favorable effects of ICS (13eC15). These studies attributed immortal time bias as responsible for the indicated beneficial association. Immortal time bias results from cohort studies with follow-up time during which a subject cannot, by definition, incur the outcome event under study. That is, when the exposure time overlaps follow-up time, patients who die during the exposure time cannot by definition obtain the medication in question and as such will be classified as nonexposed. This is believed to result in underestimation of person-time without ICS treatment leading to overestimation of any treatment effect (13).
Because the applied and suggested analyses vary among the studies published until now, we explored the extent to which immortal time bias can account for the associations found previously. We used the U.K. General Practice Research Database (GPRD) to examine the effect of ICS on the risk of rehospitalization or death among 4,604 patients with COPD within a year of discharge from a first COPD hospitalization, on the basis of two different designs both free from immortal time bias. Some of the results of the study have been previously reported in the form of an abstract (16).
METHODS
Database Setting and Cohort Definition
The GPRD is an automated database of primary care in the United Kingdom that provides a unique source for investigating the implications of therapy on disease from a "real life" perspective. Detailed descriptions of the GPRD data file contents, validity, and research uses are well documented elsewhere (17eC22). We retrospectively identified all patients with newly diagnosed COPD aged 50 years and older from 1990 to 1999. We defined the study cohort as all those hospitalized for a COPD-related condition during this period. Cohort entry was taken as the date of discharge from a first COPD hospitalization, with follow-up restricted to 1 year. The criteria for hospitalization for a COPD-related condition included codes for pneumonia and chest infection, and patients with no record of general practice contact in the 1-year follow-up period were excluded. The Scientific and Ethical Advisory Group of the GPRD approved the study protocol.
Drug Exposure and Outcome
Drug exposure was defined in terms of prescription of ICS given during a specified treatment window in the follow-up period. Thus, a patient was considered a user in a given window if he or she received at least one such prescription.
The outcome of interest was rehospitalization for a COPD-related medical condition or death within the 1-year follow-up period, whichever occurred first. In the GPRD, a COPD-related hospitalization would have a COPD medical code alongside the hospitalization outcome entry on the patient's medical records on the same day (22). Censoring was defined as end of record on database, hospitalization, death, or 365 days after cohort entry.
Study Designs and Statistical Analysis
Propensity scores matched-cohort design.
In a randomized trial, the randomization of subjects to different treatment groups guarantees that, on average, there should be no systematic differences (i.e., bias) between treatment groups. In observational studies, investigators do not assign the treatment as in a randomized trial. Therefore, differences in observed patient-level characteristics in the two treatment groups may exist, and these differences could lead to biased estimates of treatment effects. The propensity score is the conditional probability of assignment to a particular treatment given a set of observed patient-level characteristics and is increasingly being used to adjust for nonrandom treatment assignment (23eC25). Using only patients whose treatment status was defined on the same day of discharge, we fitted a logistic-regression model that predicted whether a patient would be prescribed ICS on the day of discharge as a function of the 27 variables listed in Table 1, including patient-level baseline characteristics on previous diagnostic label of asthma, smoking, age, sex, comorbidities, and respiratory medications before the first COPD-related hospitalization to obtain a matched cohort. Each ICS user was matched with a nonuser with the closest estimated propensity on the logit scale (i.e., within 0.01 caliper) to reduce differences between treatment groups. We used only patients without waiting times because propensity scores by themselves do not account for immortal time bias (Figure 1).
We fitted the proportional hazards model adjusting for age and sex as well as smoking, use of oral steroids and other respiratory medications at baseline, and the comorbidities specified in Table 1.
Nested case-control design.
Patients in the study cohort who experienced the outcome (death or rehospitalization) were defined as cases. Each case was then matched to up to four noncases by random sampling from the cohort risk sets without regard to treatment status. Cases and their control subjects were matched on four factors: follow-up duration (time between discharge and index dates), sex, age (within 1 year), and discharge date (within 30 days) (26eC28). To reduce exposure misclassification, we excluded patients who were first exposed to ICS beyond the first 90 days after hospital discharge. Index date was the outcome date and comparison was between those with any ICS prescription in the 6 months before the index date (i.e., exposed) versus those without any ICS prescription in the same period (i.e., unexposed). Thus, the design was similar to that of Suissa and coworkers (29) and Huiart and colleagues (30), and there was no bias due to wait time in the assessment of possible association between exposure and outcome. The confounding variables listed in Table 2 were adjusted for using conditional logistic regression.
Confounders.
Other respiratory medications used in the baseline period were treated as eight distinct class (binary) variables. Information on tobacco use was categorized as either nonsmoker (to represent never and ex-smoker), current smoker, or unknown, based on the patient's records in both the medical and prevention databases. Comorbidities were identified from patients' medical records and for the nested case-control design; these were categorized into three groups based on the Charlson comorbidity scores (31) as 0, 1, and 2+, in the direction of increasing burden (Tables 1 and 2).
RESULTS
For the propensity scores matched cohort, we identified 1,629 patients with records of a first COPD-related hospitalization and who either received a prescription for ICS on the day of discharge or were never exposed to ICS in their entire follow-up period, having excluded all those who died within 30 days of discharge, in line with the previous studies (8, 10, 12). Of these, 1,091 patients were given ICS prescriptions. The final matched cohort was made up of 393 exposed and 393 unexposed patients; 210 suffered rehospitalization or death during follow-up and the characteristics of the two groups are given in Table 1. For treatment with ICS, we obtained a hazard ratio estimate of 0.69 (95% confidence interval, 0.52eC0.93) from this propensity scores matched-cohort analysis, with a C-statistic of 0.82 (a measure of the predictive ability of the model in terms of probability).
The nested case-control design was based on the full cohort of 4,604 patients, including those who waited for at least a day and no more than 90 days before receiving an ICS prescription. We identified 4,190 such patients with records of a first COPD-related hospitalization, having also excluded all patients that died within 30 days of discharge. We were able to match 675 of the cases to noncases at a ratio of up to four control subjects per case. Table 2 gives the characteristics of the cases and their 1,547 control subjects. As shown in Table 3, the matched nested case-control analysis gave rate ratio estimates of 0.71 (0.56eC0.90) for those with any prescription of ICS in the last 6 months and decreasing rate ratios with increasing number of prescriptions. This indication of a gradual decrease in the event rate with the number of prescriptions of ICS received could also be expressed using a continuous quantity response analysis estimating that the rate ratio decreased by 6% (rate ratio, 0.94; 0.89eC0.99) for each additional prescription in the period.
DISCUSSION
Our results of the propensity scores matched-cohort and matched nested case-control analyses, both of which by design were free of immortal time bias, indicate a beneficial association of ICS and the risk of death or rehospitalization in COPD, in line with our previous findings using a cohort analysis strategy (11). These findings are strengthened by the observed relationship between increased regularity of ICS prescriptions and reductions in event rates.
The propensity scores approach is a recommendation of a recent editorial on the subject (32), whereas the nested case-control approach has been described as a method that simplifies the cohort analysis when "exposures vary over time and leads to valid estimates of rate ratios with a negligible loss in precision" (33). Our matched nested case-control design did not involve any patient follow-up and exposure status was only determined in the last 6 months before the index date for each case and his or her corresponding control subject. Consequently, it was free of immortal time bias in its estimation of possible association between treatment and the risk of interest. However, the study base for both sets of analysis was an observational database, and our findings cannot be interpreted as a proxy for an efficacy study. At best, our study tries to measure effectiveness and, even for that, other sources of bias (e.g., group imbalance from unmeasured factors) may be present and affecting our results.
Our findings are distinctly different from those of Suissa (13), and this difference can be due to both the different methods of analysis and the population studied. The study by Suissa (13) did not assess the extent to which immortal time bias was a factor and relied instead on a methodology of restrictive application and allowing subjects to change from "nonexposed" to "exposed" in the observation period. We know from the assumptions underpinning the time-dependent Cox model that regression to the null is a possible indication of inappropriate application. The model used by Suissa is otherwise known as the "treatment switching" method, for which it is a necessary requirement that the reason for switching is unrelated to the subsequent risk of an event (34). Indeed, the fundamental weakness of this approach has recently been illustrated in a study by Rothman (35) that revealed the problem associated with treatment switching, even in a randomized controlled trial setting that allowed for switching at the start of the open-label phase. It is questionable if the condition for appropriate application can be tenable in general practice data. The problems associated with the use of statistical models to estimate time-dependent variables are seldom simple enough to resolve, and most studies are therefore designed to avoid the use of such models (36eC38). That the immortal time bias is a potential source of bias is, however, undoubtedly true and should warrant attention in pharmacoepidemiologic studies. Thus, it is likely that the previous studies by Sin and Tu (8) and Soriano and colleagues (10) were affected by immortal time bias by their inclusion of the 90-day treatment assignment period in patient follow-up, and Suissa was correct to highlight this problem (12).
We are aware that there are likely to be differences in treatment patterns for COPD in the United Kingdom, Canada, and the United States, and that such differences could impact the different risk estimates observed beyond the study design. For instance, the GPRD suggests that 83% of patients with COPD received an ICS within 180 days of discharge. This is considerably higher than in other countries and may explain some of the differences observed; that is, those who do not receive ICS may receive substandard care overall and hence a higher risk of mortality.
As of 2005, there is still an ongoing controversy on whether ICS have significant benefits in COPD. Most likely, a higher level of evidence will be available within a few years (39). Until then, we can only speculate that the epidemiologic signal of the beneficial effects of ICS in COPD in the U.K. GPRD is of such a magnitude that all reanalyses conducted to date with different study designs or criteria, including the propensity scores and nested case-control designs in this study, produce beneficial associations, indicating high internal consistency. It is hoped that research on the effectiveness of ICS in COPD will help those in the emergent discipline of COPD pharmacoepidemiology (40) and ultimately their patients.
In conclusion, our analyses revealed significant influence of study design on the effect of ICS on adverse outcomes after hospitalization. With different study designs reducing potential bias, we consistently found an association between ICS use and reduction of risk of rehospitalization and death.
Acknowledgments
The authors thank the reviewers for their useful comments and suggestions. We also thank Maurille Feudjo-Tapie of the London School of Hygiene and Tropical Medicine and Gilbert MacKenzie of University of Limerick in Ireland for their useful suggestions.
REFERENCES
Murray CJL, Lopez AD, Mathers CD, Stein C. The global burden of disease 2000 project: aims, methods and data sources. Geneva, Switzerland: World Health Organization; 2001. Global Program on Evidence for Health Policy Discussion Paper No. 36.
Mannino DM. COPD: epidemiology, prevalence, morbidity and mortality, and disease heterogeneity. Chest 2002;121:121SeC126S.
Burge PS, Calverley PM, Jones PW, Spencer S, Anderson JA, Maslen TK. Randomised, double blind, placebo controlled study of fluticasone propionate in patients with moderate to severe chronic obstructive pulmonary disease: the ISOLDE trial. BMJ 2000;320:1297eC1303.
Lung Health Study Research Group. Effect of inhaled triamcinolone on the decline in pulmonary function in chronic obstructive pulmonary disease. N Engl J Med 2000;343:1902eC1909.
Calverley P, Pauwels R, Vestbo J, et al. Combined salmeterol and fluticasone in the treatment of chronic obstructive pulmonary disease: a randomised controlled trial. Lancet 2003;361:449eC456.
Pauwels RA, Buist AS, Calverley PM, et al. Global strategy for the diagnosis, management, and prevention of chronic obstructive pulmonary disease. NHLBI/WHO Global Initiative for Chronic Obstructive Lung Disease (GOLD) workshop summary. Am J Respir Crit Care Med 2001;163:1256eC1276.
Celli BR, MacNee W, Agusti A, Anzueto A, Berg B, Buist AS, Calverley PMA, Chavannes N, Dillard T, Fahy B, et al. Standards for the diagnosis and treatment of patients with COPD: a summary of the ATS/ERS position paper. Eur Respir J 2004;23:932eC946.
Sin DD, Tu JV. Inhaled corticosteroids and the risk of mortality and readmission in elderly patients with chronic obstructive pulmonary disease. Am J Respir Crit Care Med 2001;164:580eC584.
Sin DD, Man SFP. Inhaled corticosteroids and survival in chronic obstructive pulmonary disease: does the dose matter Eur Respir J 2003;21:260eC266.
Soriano JB, Kiri VA. pride NB, Vestbo J. Inhaled corticosteroids with/without long-acting beta-agonist reduce the risk of rehospitalization and death in COPD patients. Am J Respir Med 2003;2:67eC74.
Soriano JB, Vestbo J, Pride NB, Kiri V, Maden C, Maier WC. Survival of COPD patients after regular use of fluticasone propionate and salmeterol in primary care in the UK. Eur Respir J 2002;20:819eC825.
Kiri VA, Bettoncelli G, Testi R, Viegi G. Inhaled corticosteroids are more effective in COPD patients when used with LABA than with SABA. Respir Med (In press)
Suissa S. Effectiveness of inhaled corticosteroids in chronic obstructive pulmonary disease. Am J Respir Crit Care Med 2003;168:49eC53.
Fan VS, Bryson CL, Curtis RJ, Fihn SD, Bridevaux P, McDonell MB, Au DH. Inhaled corticosteroids in chronic obstructive pulmonary disease and risk of death. Am J Respir Crit Care Med 2003;168:1488eC1494.
Suissa S. Inhaled steroids and mortality in COPD: bias from unaccounted immortal time. Eur Respir J 2004;23:391eC395.
Kiri VA, Soriano JB, Pride NB, Vestbo J. The association between inhaled corticosteroids (ICS) and reduced risk of COPD hospitalization or death cannot be explained by immortal time bias. Am J Respir Crit Care Med 2004;169:A517.
Lawson DH, Sherman V, Hollowell J. The General Practice Research Database: Scientific and Ethical Advisory Group. QJM 1998;91:445eC452.
Jick H, Jick SS, Derby LE. Validation of information recorded on general practitioner based computerised data resources in the UK. BMJ 1991;302:766eC768.
Garcia Rodriguez LA, Perez Gutthann S. Use of the UK General Practice Research Database for pharmacoepidemiology. Br J Clin Pharmacol 1998;45:419eC425.
Nazareth I, King M, Haines A, et al. Accuracy of diagnosis on general practice computer system. BMJ 1993;307:32eC34.
Hansell A, Hollowell J, Nichols T, et al. Use of the General Practice Research Database (GPRD) for respiratory epidemiology: a comparison with the 4th Morbidity Survey in General Practice (MSGP4). Thorax 1999;54:413eC419.
Soriano JB, Maier WC, Visick G, Pride N. Validation of general practitioner-diagnosed COPD in the UK General Practice Research Database. Eur J Epidemiol 2001;17:1075eC1080.
D'Agostino RB. Tutorial in biostatistics: propensity score methods for bias reduction in the comparison of a treatment to a non-randomized control group. Stat Med 1998;17:2265eC2281.
Morant SV, Pettitt D, MacDonald TM, et al. Application of propensity score to adjust for channelling bias with NSAIDs. Pharmacoepidemiol Drug Saf 2004;13:345eC353.
Frolkis JP, Pthier CE, Blackstone EH, et al. Frequent ventricular ectopy after exercise as a predictor of death. N Engl J Med 2003;348:781eC790.
Black C, Kaye JA, Jick H. Relation of childhood gastrointestinal disorders to autism: nested case-control study using data from the UK General Practice Research Database. BMJ 2002;325:419eC421.
Eland IA, Huerta-Alvarez C, Stricker B, et al. The risk of acute pancreatitis associated with acid-suppressing drugs. Br J Clin Pharmacol 2000;49:473eC478.
Nightingale AL, Farmer RD. Ischemic stroke in young women: a nested case-control study using the UK General Practice Research Database. Stroke 2004;35:1574eC1578.
Suissa S, Assimes T, Brassard P, Ernst P. Inhaled corticosteroid use in asthma and the prevention of myocardial infarction. Am J Med 2003;115:377eC381.
Huiart L, Ernst P, Ranouil X, Suissa S. Low-dose inhaled corticosteroids and the risk of acute myocardial infarction in COPD. Eur Respir J 2005;25:634eC639.
Deyo RA, Cherkin DC, Ciol MA. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. J Clin Epidemiol 1992;45:613eC619.
Samet JM. Measuring the effectiveness of inhaled corticosteroids for COPD is not easy. Am J Respir Crit Care Med 2003;168:1eC2.
Suissa S. Novel approaches to pharmacoepidemiology study design and statistical analysis. In: Strom BL, editor. Pharmacoepidemiology, 3rd ed. New York: John Wiley & Sons; 2000. p. 785eC805.
Clayton D, Hills M. Statistical methods in epidemiology. Oxford, UK: Oxford University Press; 1993.
Rothman KJ. A potential bias in safety evaluation during open-label extensions of randomized clinical trials. Pharmacoepidemiol Drug Saf 2004;13:295eC298.
Cox DR, Oakes D. Analysis of survival data. London: Chapman and Hall; 1984.
Kalbfleisch JD, Prentice RL. The statistical analysis of failure time data. New York: John Wiley & Sons; 2002.
Bunday BD, Kiri VA. Analysis of censored recidivism data using a proportional hazards-type model. The Statistician 1992;41:85eC96.
TORCH Study Group. The TORCH (TOwards a Revolution in COPD Health) survival study protocol. Eur Respir J 2004;24:206eC210.
Burney P, Suissa S, Soriano JB, et al. The pharmacoepidemiology of COPD: recent advances and methodological discussion. Eur Respir J 2003;21:1eC43.(Victor A. Kiri, Neil B. P)
National Centre for Health Outcomes Development
Department of Epidemiology and Population Health, London School of Hygiene and Tropical Medicine
Thoracic Medicine, National Heart and Lung Institute, Imperial College, London
North West Lung Centre, South Manchester University Hospital Trust, Manchester, United Kingdom
Worldwide Epidemiology, GlaxoSmithKline, Upper Providence
Center for Clinical Epidemiology and Biostatistics, University of Pennsylvania School of Medicine, Philadelphia, Pennsylvania
Department of Heart and Lung Diseases, Hvidovre Hospital, Hvidovre, Denmark
ABSTRACT
Rationale: Recent cohort studies in chronic obstructive pulmonary disease (COPD) have questioned the validity of previously reported associations between inhaled corticosteroids (ICS) and reductions in mortality and rehospitalization in observational studies. Using time-dependent versions of statistical survival models, these studies have suggested immortal time bias as responsible for the proposed beneficial association. Objectives: We explored the extent of this bias in a study of patients with COPD monitored for a year from COPD discharge with two designs free of any immortal time bias in the General Practice Research Database in the United Kingdom. Methods: In Design 1, we used only patients whose treatment status was defined on the same day of discharge to obtain a matched cohort based on propensity scores, which were derived from the patient-level baseline characteristics. In Design 2, we identified all in the study cohort who experienced death or rehospitalization and then matched each case to up to four noncases by randomly sampling from the cohort risk sets without regard to treatment status. Measurements and Main Results: The propensity scores matched cohort analysis of 786 patients without a wait time found a significant risk reduction associated with use of ICS: hazard ratio, 0.69 (95% confidence interval, 0.52eC0.93). The matched nested case-control analysis of 2,222 patients, designed without regard to exposure status and hence free of immortal time bias, gave a similar association with exposure to ICS in the last 6-month period: hazard ratio, 0.71 (0.56eC0.90). Conclusions: We conclude that immortal time bias cannot account for the risk reduction associated with ICS exposure in observational studies.
Key Words: chronic obstructive pulmonary disease; epidemiologic methods; immortal time bias; nested case-control design; propensity scores matched cohort; time-dependent model
Chronic obstructive pulmonary disease (COPD) is currently the fourth leading cause of death, responsible for more than 2.5 million deaths per year worldwide, and it is estimated to become the third most important cause of death by 2020 (1). COPD is a leading cause of hospitalization in adults in the United States, particularly in older populations. In 1998, almost 662,000 hospitalizations (1.9% of total hospitalizations) were attributed to COPD, and the rate of hospitalizations with COPD as the primary cause of hospitalization was 38.3/10,000 individuals in 1998 (2). An additional 2.5 million hospitalizations (7.0% of total hospitalizations) had COPD listed as a contributing cause.
Inhaled corticosteroids (ICS) reduce exacerbations in patients with moderate to severe COPD (3eC5). The Global Initiative for Obstructive Lung Disease guidelines now recommend long-term use of ICS for patients with COPD with post-bronchodilator FEV1 of less than 50% predicted and repeated exacerbations (6), as do the recent American Thoracic Society/European Respiratory Society COPD treatment guidelines (7).
Several observational studies have suggested that treatment with ICS is associated with risk reduction of rehospitalization or death (8eC12). These pharmacoepidemiologic studies have used automated, linked health care and pharmacy data or electronic medical record databases, and, due to the nonrandom allocation of treatment exposure, they are by nature subject to bias. However, other recent COPD observational studies have failed to find favorable effects of ICS (13eC15). These studies attributed immortal time bias as responsible for the indicated beneficial association. Immortal time bias results from cohort studies with follow-up time during which a subject cannot, by definition, incur the outcome event under study. That is, when the exposure time overlaps follow-up time, patients who die during the exposure time cannot by definition obtain the medication in question and as such will be classified as nonexposed. This is believed to result in underestimation of person-time without ICS treatment leading to overestimation of any treatment effect (13).
Because the applied and suggested analyses vary among the studies published until now, we explored the extent to which immortal time bias can account for the associations found previously. We used the U.K. General Practice Research Database (GPRD) to examine the effect of ICS on the risk of rehospitalization or death among 4,604 patients with COPD within a year of discharge from a first COPD hospitalization, on the basis of two different designs both free from immortal time bias. Some of the results of the study have been previously reported in the form of an abstract (16).
METHODS
Database Setting and Cohort Definition
The GPRD is an automated database of primary care in the United Kingdom that provides a unique source for investigating the implications of therapy on disease from a "real life" perspective. Detailed descriptions of the GPRD data file contents, validity, and research uses are well documented elsewhere (17eC22). We retrospectively identified all patients with newly diagnosed COPD aged 50 years and older from 1990 to 1999. We defined the study cohort as all those hospitalized for a COPD-related condition during this period. Cohort entry was taken as the date of discharge from a first COPD hospitalization, with follow-up restricted to 1 year. The criteria for hospitalization for a COPD-related condition included codes for pneumonia and chest infection, and patients with no record of general practice contact in the 1-year follow-up period were excluded. The Scientific and Ethical Advisory Group of the GPRD approved the study protocol.
Drug Exposure and Outcome
Drug exposure was defined in terms of prescription of ICS given during a specified treatment window in the follow-up period. Thus, a patient was considered a user in a given window if he or she received at least one such prescription.
The outcome of interest was rehospitalization for a COPD-related medical condition or death within the 1-year follow-up period, whichever occurred first. In the GPRD, a COPD-related hospitalization would have a COPD medical code alongside the hospitalization outcome entry on the patient's medical records on the same day (22). Censoring was defined as end of record on database, hospitalization, death, or 365 days after cohort entry.
Study Designs and Statistical Analysis
Propensity scores matched-cohort design.
In a randomized trial, the randomization of subjects to different treatment groups guarantees that, on average, there should be no systematic differences (i.e., bias) between treatment groups. In observational studies, investigators do not assign the treatment as in a randomized trial. Therefore, differences in observed patient-level characteristics in the two treatment groups may exist, and these differences could lead to biased estimates of treatment effects. The propensity score is the conditional probability of assignment to a particular treatment given a set of observed patient-level characteristics and is increasingly being used to adjust for nonrandom treatment assignment (23eC25). Using only patients whose treatment status was defined on the same day of discharge, we fitted a logistic-regression model that predicted whether a patient would be prescribed ICS on the day of discharge as a function of the 27 variables listed in Table 1, including patient-level baseline characteristics on previous diagnostic label of asthma, smoking, age, sex, comorbidities, and respiratory medications before the first COPD-related hospitalization to obtain a matched cohort. Each ICS user was matched with a nonuser with the closest estimated propensity on the logit scale (i.e., within 0.01 caliper) to reduce differences between treatment groups. We used only patients without waiting times because propensity scores by themselves do not account for immortal time bias (Figure 1).
We fitted the proportional hazards model adjusting for age and sex as well as smoking, use of oral steroids and other respiratory medications at baseline, and the comorbidities specified in Table 1.
Nested case-control design.
Patients in the study cohort who experienced the outcome (death or rehospitalization) were defined as cases. Each case was then matched to up to four noncases by random sampling from the cohort risk sets without regard to treatment status. Cases and their control subjects were matched on four factors: follow-up duration (time between discharge and index dates), sex, age (within 1 year), and discharge date (within 30 days) (26eC28). To reduce exposure misclassification, we excluded patients who were first exposed to ICS beyond the first 90 days after hospital discharge. Index date was the outcome date and comparison was between those with any ICS prescription in the 6 months before the index date (i.e., exposed) versus those without any ICS prescription in the same period (i.e., unexposed). Thus, the design was similar to that of Suissa and coworkers (29) and Huiart and colleagues (30), and there was no bias due to wait time in the assessment of possible association between exposure and outcome. The confounding variables listed in Table 2 were adjusted for using conditional logistic regression.
Confounders.
Other respiratory medications used in the baseline period were treated as eight distinct class (binary) variables. Information on tobacco use was categorized as either nonsmoker (to represent never and ex-smoker), current smoker, or unknown, based on the patient's records in both the medical and prevention databases. Comorbidities were identified from patients' medical records and for the nested case-control design; these were categorized into three groups based on the Charlson comorbidity scores (31) as 0, 1, and 2+, in the direction of increasing burden (Tables 1 and 2).
RESULTS
For the propensity scores matched cohort, we identified 1,629 patients with records of a first COPD-related hospitalization and who either received a prescription for ICS on the day of discharge or were never exposed to ICS in their entire follow-up period, having excluded all those who died within 30 days of discharge, in line with the previous studies (8, 10, 12). Of these, 1,091 patients were given ICS prescriptions. The final matched cohort was made up of 393 exposed and 393 unexposed patients; 210 suffered rehospitalization or death during follow-up and the characteristics of the two groups are given in Table 1. For treatment with ICS, we obtained a hazard ratio estimate of 0.69 (95% confidence interval, 0.52eC0.93) from this propensity scores matched-cohort analysis, with a C-statistic of 0.82 (a measure of the predictive ability of the model in terms of probability).
The nested case-control design was based on the full cohort of 4,604 patients, including those who waited for at least a day and no more than 90 days before receiving an ICS prescription. We identified 4,190 such patients with records of a first COPD-related hospitalization, having also excluded all patients that died within 30 days of discharge. We were able to match 675 of the cases to noncases at a ratio of up to four control subjects per case. Table 2 gives the characteristics of the cases and their 1,547 control subjects. As shown in Table 3, the matched nested case-control analysis gave rate ratio estimates of 0.71 (0.56eC0.90) for those with any prescription of ICS in the last 6 months and decreasing rate ratios with increasing number of prescriptions. This indication of a gradual decrease in the event rate with the number of prescriptions of ICS received could also be expressed using a continuous quantity response analysis estimating that the rate ratio decreased by 6% (rate ratio, 0.94; 0.89eC0.99) for each additional prescription in the period.
DISCUSSION
Our results of the propensity scores matched-cohort and matched nested case-control analyses, both of which by design were free of immortal time bias, indicate a beneficial association of ICS and the risk of death or rehospitalization in COPD, in line with our previous findings using a cohort analysis strategy (11). These findings are strengthened by the observed relationship between increased regularity of ICS prescriptions and reductions in event rates.
The propensity scores approach is a recommendation of a recent editorial on the subject (32), whereas the nested case-control approach has been described as a method that simplifies the cohort analysis when "exposures vary over time and leads to valid estimates of rate ratios with a negligible loss in precision" (33). Our matched nested case-control design did not involve any patient follow-up and exposure status was only determined in the last 6 months before the index date for each case and his or her corresponding control subject. Consequently, it was free of immortal time bias in its estimation of possible association between treatment and the risk of interest. However, the study base for both sets of analysis was an observational database, and our findings cannot be interpreted as a proxy for an efficacy study. At best, our study tries to measure effectiveness and, even for that, other sources of bias (e.g., group imbalance from unmeasured factors) may be present and affecting our results.
Our findings are distinctly different from those of Suissa (13), and this difference can be due to both the different methods of analysis and the population studied. The study by Suissa (13) did not assess the extent to which immortal time bias was a factor and relied instead on a methodology of restrictive application and allowing subjects to change from "nonexposed" to "exposed" in the observation period. We know from the assumptions underpinning the time-dependent Cox model that regression to the null is a possible indication of inappropriate application. The model used by Suissa is otherwise known as the "treatment switching" method, for which it is a necessary requirement that the reason for switching is unrelated to the subsequent risk of an event (34). Indeed, the fundamental weakness of this approach has recently been illustrated in a study by Rothman (35) that revealed the problem associated with treatment switching, even in a randomized controlled trial setting that allowed for switching at the start of the open-label phase. It is questionable if the condition for appropriate application can be tenable in general practice data. The problems associated with the use of statistical models to estimate time-dependent variables are seldom simple enough to resolve, and most studies are therefore designed to avoid the use of such models (36eC38). That the immortal time bias is a potential source of bias is, however, undoubtedly true and should warrant attention in pharmacoepidemiologic studies. Thus, it is likely that the previous studies by Sin and Tu (8) and Soriano and colleagues (10) were affected by immortal time bias by their inclusion of the 90-day treatment assignment period in patient follow-up, and Suissa was correct to highlight this problem (12).
We are aware that there are likely to be differences in treatment patterns for COPD in the United Kingdom, Canada, and the United States, and that such differences could impact the different risk estimates observed beyond the study design. For instance, the GPRD suggests that 83% of patients with COPD received an ICS within 180 days of discharge. This is considerably higher than in other countries and may explain some of the differences observed; that is, those who do not receive ICS may receive substandard care overall and hence a higher risk of mortality.
As of 2005, there is still an ongoing controversy on whether ICS have significant benefits in COPD. Most likely, a higher level of evidence will be available within a few years (39). Until then, we can only speculate that the epidemiologic signal of the beneficial effects of ICS in COPD in the U.K. GPRD is of such a magnitude that all reanalyses conducted to date with different study designs or criteria, including the propensity scores and nested case-control designs in this study, produce beneficial associations, indicating high internal consistency. It is hoped that research on the effectiveness of ICS in COPD will help those in the emergent discipline of COPD pharmacoepidemiology (40) and ultimately their patients.
In conclusion, our analyses revealed significant influence of study design on the effect of ICS on adverse outcomes after hospitalization. With different study designs reducing potential bias, we consistently found an association between ICS use and reduction of risk of rehospitalization and death.
Acknowledgments
The authors thank the reviewers for their useful comments and suggestions. We also thank Maurille Feudjo-Tapie of the London School of Hygiene and Tropical Medicine and Gilbert MacKenzie of University of Limerick in Ireland for their useful suggestions.
REFERENCES
Murray CJL, Lopez AD, Mathers CD, Stein C. The global burden of disease 2000 project: aims, methods and data sources. Geneva, Switzerland: World Health Organization; 2001. Global Program on Evidence for Health Policy Discussion Paper No. 36.
Mannino DM. COPD: epidemiology, prevalence, morbidity and mortality, and disease heterogeneity. Chest 2002;121:121SeC126S.
Burge PS, Calverley PM, Jones PW, Spencer S, Anderson JA, Maslen TK. Randomised, double blind, placebo controlled study of fluticasone propionate in patients with moderate to severe chronic obstructive pulmonary disease: the ISOLDE trial. BMJ 2000;320:1297eC1303.
Lung Health Study Research Group. Effect of inhaled triamcinolone on the decline in pulmonary function in chronic obstructive pulmonary disease. N Engl J Med 2000;343:1902eC1909.
Calverley P, Pauwels R, Vestbo J, et al. Combined salmeterol and fluticasone in the treatment of chronic obstructive pulmonary disease: a randomised controlled trial. Lancet 2003;361:449eC456.
Pauwels RA, Buist AS, Calverley PM, et al. Global strategy for the diagnosis, management, and prevention of chronic obstructive pulmonary disease. NHLBI/WHO Global Initiative for Chronic Obstructive Lung Disease (GOLD) workshop summary. Am J Respir Crit Care Med 2001;163:1256eC1276.
Celli BR, MacNee W, Agusti A, Anzueto A, Berg B, Buist AS, Calverley PMA, Chavannes N, Dillard T, Fahy B, et al. Standards for the diagnosis and treatment of patients with COPD: a summary of the ATS/ERS position paper. Eur Respir J 2004;23:932eC946.
Sin DD, Tu JV. Inhaled corticosteroids and the risk of mortality and readmission in elderly patients with chronic obstructive pulmonary disease. Am J Respir Crit Care Med 2001;164:580eC584.
Sin DD, Man SFP. Inhaled corticosteroids and survival in chronic obstructive pulmonary disease: does the dose matter Eur Respir J 2003;21:260eC266.
Soriano JB, Kiri VA. pride NB, Vestbo J. Inhaled corticosteroids with/without long-acting beta-agonist reduce the risk of rehospitalization and death in COPD patients. Am J Respir Med 2003;2:67eC74.
Soriano JB, Vestbo J, Pride NB, Kiri V, Maden C, Maier WC. Survival of COPD patients after regular use of fluticasone propionate and salmeterol in primary care in the UK. Eur Respir J 2002;20:819eC825.
Kiri VA, Bettoncelli G, Testi R, Viegi G. Inhaled corticosteroids are more effective in COPD patients when used with LABA than with SABA. Respir Med (In press)
Suissa S. Effectiveness of inhaled corticosteroids in chronic obstructive pulmonary disease. Am J Respir Crit Care Med 2003;168:49eC53.
Fan VS, Bryson CL, Curtis RJ, Fihn SD, Bridevaux P, McDonell MB, Au DH. Inhaled corticosteroids in chronic obstructive pulmonary disease and risk of death. Am J Respir Crit Care Med 2003;168:1488eC1494.
Suissa S. Inhaled steroids and mortality in COPD: bias from unaccounted immortal time. Eur Respir J 2004;23:391eC395.
Kiri VA, Soriano JB, Pride NB, Vestbo J. The association between inhaled corticosteroids (ICS) and reduced risk of COPD hospitalization or death cannot be explained by immortal time bias. Am J Respir Crit Care Med 2004;169:A517.
Lawson DH, Sherman V, Hollowell J. The General Practice Research Database: Scientific and Ethical Advisory Group. QJM 1998;91:445eC452.
Jick H, Jick SS, Derby LE. Validation of information recorded on general practitioner based computerised data resources in the UK. BMJ 1991;302:766eC768.
Garcia Rodriguez LA, Perez Gutthann S. Use of the UK General Practice Research Database for pharmacoepidemiology. Br J Clin Pharmacol 1998;45:419eC425.
Nazareth I, King M, Haines A, et al. Accuracy of diagnosis on general practice computer system. BMJ 1993;307:32eC34.
Hansell A, Hollowell J, Nichols T, et al. Use of the General Practice Research Database (GPRD) for respiratory epidemiology: a comparison with the 4th Morbidity Survey in General Practice (MSGP4). Thorax 1999;54:413eC419.
Soriano JB, Maier WC, Visick G, Pride N. Validation of general practitioner-diagnosed COPD in the UK General Practice Research Database. Eur J Epidemiol 2001;17:1075eC1080.
D'Agostino RB. Tutorial in biostatistics: propensity score methods for bias reduction in the comparison of a treatment to a non-randomized control group. Stat Med 1998;17:2265eC2281.
Morant SV, Pettitt D, MacDonald TM, et al. Application of propensity score to adjust for channelling bias with NSAIDs. Pharmacoepidemiol Drug Saf 2004;13:345eC353.
Frolkis JP, Pthier CE, Blackstone EH, et al. Frequent ventricular ectopy after exercise as a predictor of death. N Engl J Med 2003;348:781eC790.
Black C, Kaye JA, Jick H. Relation of childhood gastrointestinal disorders to autism: nested case-control study using data from the UK General Practice Research Database. BMJ 2002;325:419eC421.
Eland IA, Huerta-Alvarez C, Stricker B, et al. The risk of acute pancreatitis associated with acid-suppressing drugs. Br J Clin Pharmacol 2000;49:473eC478.
Nightingale AL, Farmer RD. Ischemic stroke in young women: a nested case-control study using the UK General Practice Research Database. Stroke 2004;35:1574eC1578.
Suissa S, Assimes T, Brassard P, Ernst P. Inhaled corticosteroid use in asthma and the prevention of myocardial infarction. Am J Med 2003;115:377eC381.
Huiart L, Ernst P, Ranouil X, Suissa S. Low-dose inhaled corticosteroids and the risk of acute myocardial infarction in COPD. Eur Respir J 2005;25:634eC639.
Deyo RA, Cherkin DC, Ciol MA. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. J Clin Epidemiol 1992;45:613eC619.
Samet JM. Measuring the effectiveness of inhaled corticosteroids for COPD is not easy. Am J Respir Crit Care Med 2003;168:1eC2.
Suissa S. Novel approaches to pharmacoepidemiology study design and statistical analysis. In: Strom BL, editor. Pharmacoepidemiology, 3rd ed. New York: John Wiley & Sons; 2000. p. 785eC805.
Clayton D, Hills M. Statistical methods in epidemiology. Oxford, UK: Oxford University Press; 1993.
Rothman KJ. A potential bias in safety evaluation during open-label extensions of randomized clinical trials. Pharmacoepidemiol Drug Saf 2004;13:295eC298.
Cox DR, Oakes D. Analysis of survival data. London: Chapman and Hall; 1984.
Kalbfleisch JD, Prentice RL. The statistical analysis of failure time data. New York: John Wiley & Sons; 2002.
Bunday BD, Kiri VA. Analysis of censored recidivism data using a proportional hazards-type model. The Statistician 1992;41:85eC96.
TORCH Study Group. The TORCH (TOwards a Revolution in COPD Health) survival study protocol. Eur Respir J 2004;24:206eC210.
Burney P, Suissa S, Soriano JB, et al. The pharmacoepidemiology of COPD: recent advances and methodological discussion. Eur Respir J 2003;21:1eC43.(Victor A. Kiri, Neil B. P)