Immortal time bias in observational studies of drug effects y

Transcription

1 pharmacoepidemiology and drug safety 2007; 16: Published online 25 January 2007 in Wiley InterScience ( ORIGINAL REPORT Immortal time bias in observational studies of drug effects y Samy Suissa PhD 1,2 * 1 From the McGill Pharmacoepidemiology Research Unit, McGill University Health Centre, Montreal, Canada 2 Departments of Epidemiology and Biostatistics and of Medicine, McGill University, Montreal, Canada SUMMARY Purpose Recent observational studies suggest that various drugs are remarkably effective at reducing morbidity and mortality. These cohort studies used a flawed approach to design and data analysis which can lead to immortal time bias. We describe the bias from 20 of these studies and illustrate it by showing that unrelated drugs can be made to appear effective at treating cardiovascular disease (CVD). Methods The illustration used a cohort of 3315 patients, with chronic obstructive pulmonary disease (COPD), identified from the Saskatchewan Health databases, hospitalised for CVD and followed for up to a year. We used the biased approach to assess the effect of two medications, namely gastrointestinal drugs (GID) and inhaled beta-agonists (IBA), both unknown to be effective in CVD, on the risk of all-cause mortality. We also estimated these effects using the proper person-time approach. Results Using the inappropriate approach, the rates ratios of all-cause death were 0.73 (95%CI: ), with IBA and 0.78 (95%CI: ), with GID. These rate ratios became 0.98 (95%CI: ) and 0.94 (95%CI: ), respectively, with the proper person-time analysis. Conclusions Several recent observational studies used a flawed approach to design and data analysis, leading to immortal time bias, which can generate an illusion of treatment effectiveness. Observational studies, with surprising beneficial drug effects should be re-assessed to account for this source of bias. Copyright # 2007 John Wiley & Sons, Ltd. key words biases; cohort studies; drug effectiveness; databases; epidemiology Received 17 October 2006; Revised 15 November 2006; Accepted 22 November 2006 INTRODUCTION Immortal time in epidemiology refers to a period of cohort follow-up or observation time, during which death cannot occur. 1,2 This phenomenon typically arises in cohort studies of drug effects, where exposure to a drug is established from the prescriptions given during the follow-up. Consequently, as the exposure * Correspondence to: Dr S. Suissa, Division of Clinical Epidemiology, Royal Victoria Hospital, 687 Pine avenue west, Ross 4.29, Montreal, Québec, Canada H3A 1A1. samy.suissa@clinepi.mcgill.ca y No conflict of interest was declared. definition is made over time, the period between cohort entry and the first prescription for the drug under study will necessarily be event-free and thus immortal. Indeed, to have been able to receive the prescription, the patient must have survived this period. Immortal time bias can then arise, when the exposure during this immortal time period is either misclassified or simply excluded and not accounted for in the data analysis. 3 5 Over the last few years, several observational database studies have appeared in the literature, suggesting important benefits of certain drugs in reducing morbidity and mortality, with some of the findings highly surprising and unexpected. This perplexity was Copyright # 2007 John Wiley & Sons, Ltd.

2 242 s. suissa particularly highlighted by a study, using the General Practice Research Database (GPRD) that suggested a 45% reduction (rate ratio 0.55) in fracture incidence with statin use. 6 The same study question was replicated by different authors, using the same GPRD database but a different approach, who found no such reduction (rate ratio 1.0), with no obvious explanation for such a sizable discrepancy. 7,8 Other examples of observational studies finding significant drug effects, that had not been suggested by randomised trials, include the use of low-dose ß-blockers, after acute myocardial infarction associated with a 43% reduction in the rate of heart failure and the use of inhaled corticosteroids (ICS), after hospitalisation for chronic obstructive pulmonary disease (COPD) with a 29% reduction in the rate of all-cause mortality. 9,10 Such findings that can have tremendous implications on clinical practice raise serious questions. Could statins truly reduce the burden of fractures in the elderly by half and how many drugs or medical interventions are currently known to reduce the 1-year rate of all-cause mortality by 30%? Alternatively, these studies may share a common bias. These observational studies, all conducted using large health care databases, used a similar cohort approach that generally attempted to emulate the randomised trial in design and analysis. Such an approach was shown, in recent studies of the effectiveness of drugs used to treat COPD and allergic rhinitis, to be affected by immortal time bias, which can create an impression of remarkable benefit for a drug. 3,5,13 This biased approach may be presently replicated in numerous studies of different medications, commonly prescribed to large numbers of patients. In this paper, we identify several recently published studies that used this approach. We describe immortal time bias and illustrate it, using data from a cohort of patients hospitalised for cardiovascular disease (CVD), by assessing the effectiveness of two drug classes unknown to have any possible beneficial effect in the treatment of CVD. IMMORTAL TIME BIAS IN RECENT STUDIES Table 1 presents 20 recently published observational studies, suggesting that various drugs are either very safe or highly effective at preventing major outcomes of morbidity and mortality. 6,9 12,14 28 These studies, performed with existing computerised databases, found significant rate reductions for various major disease outcomes, ranging generally between 25 and 50%, associated with medication use. We describe below how relevant immortal time periods of follow-up were not appropriately considered in these cohort studies. Instead of discussing each study in length, we classified the studies according to whether the immortal time in these studies were potentially misclassified or excluded. Misclassified immortal time In several cohort studies, the subjects are classified as exposed or unexposed to a drug according to whether or not they receive a prescription for this treatment, within a certain period after cohort entry, the exposure time period. This exposure definition period varies, with some studies using a relatively small period such as 30 days, 18 while others defined exposure as a prescription given at anytime during follow-up, up to 5 years. 9,21,22 The time between cohort entry and the date of the first prescription is unavoidably immortal (Figure 1, top) for the exposed subjects. Indeed, to have received the prescription implies that the subject survived or was event-free, up to the first prescription (had they died or had the outcome event within this period, they would have been classified as unexposed). By defining such immortal subjects as exposed, the analysis will misclassify the immortal time as exposed, when in fact it is unexposed. As a result, the rate ratio that compares exposed subjects to unexposed will be biased downward, because subjects considered as exposed include unexposed immortal person-time. In other words, exposed subjects have an assured advantage over the unexposed subjects: they will be artificially protected until they become exposed. 3 Of course, one must instead consider person-time and not subjects in classifying exposure and outcome events. Several studies listed in Table 1 used this approach, leading to misclassified immortal time. 9,10,12,14 26 For example the study of ICS use in COPD used a cohort design that followed over patients, after discharge for COPD hospitalisation (cohort entry) until death for up to 1 year. 10 Exposure was then defined in a dichotomous way, with the subjects classified as exposed, if they were dispensed a prescription for an ICS during the first 90-day period after cohort entry and unexposed otherwise. Data analysis was based on the time to death. Clearly, however, the time between cohort entry and the time of the first ICS prescription for the exposed subjects is immortal and of course still unexposed. By misclassifying this immortal person-time (corresponding to zero deaths) as exposed, the resulting 29% reduction in the rate of all-cause mortality will be

3 immortal time bias 243 Table 1. Description of 20 observational studies of the effects of various drugs on morbidity and mortality, with rate ratios Reference Year Study subjects Study size Outcome Follow-up Drug exposure Exposure time window Rate ratio (95%CI) Misclassified immortal time Asthma Asthma hospitalisation 3 years Inhaled corticosteroids 3 years 0.5 ( ) 2 years 0.85 ( ) 2000 COPD All-cause death 3 years Ipratropium (over 9 prescriptions) 2000 AMI Heart failure 1 year ß-blockers 1 year 0.57 ( ) 2001 Asthma 6254 Readmission 1 year Inhaled corticosteroids 90 days 0.71 ( ) All-cause death 0.61 ( ) 2001 COPD Readmission 1 year Inhaled corticosteroids 90 days 0.76 ( ) All-cause death 0.71 ( ) 2001 Asthma Asthma hospitalisation 1 year Inhaled corticosteroids 1 year 0.4 ( ) Cromolyn 0.6 ( ) 2002 Asthma/Rhinitis ED visit for asthma 3 years Intranasal CS 3 years 0.70 ( ) 2002 Asthma/COPD Readmission 90 days Physician visit 30 days 0.79 ( ) 2002 AMI 1956 Recurrent coronary event 10 years Statins 180 days 0.59 ( ) Non-statins 0.66 ( ) 2002 Asthma 1293 Readmission 2 years Inhaled corticosteroids 2 years 0.55 ( ) 2002 Heart failure Readmission 5 years ß-blockers 5 years 0.82 ( ) All-cause death 0.72 ( ) Readmission ACE inhibitors 0.93 ( ) All-cause death 0.59 ( ) 2003 COPD 6740 All-cause death 3 years ICS (low dose) 3 years 0.77 ( ) ICS (med dose) 0.48 ( ) ICS (high dose) 0.55 ( ) 2003 COPD 1219 Readmission 1 year N-acetylcysteine 1 year 0.67 ( ) 2003 COPD 4263 All-cause death 1 year ICS LABA 90 days 0.59 ( ) 4 year Anti-inflammatory 100 days 0.49 ( ) 2004 Asthma 7433 Asthma ED visit/hospitalisation Inhaled corticosteroids 0.51 ( ) 2004 Diabetes Insulin initiation >3 years Statins >1 year 0.74 ( ) Excluded immortal time Population Fracture 10 years Statins n/a 0.55 ( ) 2000 Population Dementia 7 years Statins n/a 0.29 ( ) 2002 COPD 4665 All-cause death 3 years ICS n/a 0.62 ( ) LABA 0.79 ( ) ICS þ LABA 0.48 ( ) 2003 Population Depression 9 years Statins n/a 0.4 ( ) Suicidal behaviour 0.5 ( ) n/a, exposure time window not applicable for studies with excluded immortal time.

4 244 s. suissa Figure 1. Description of typical exposed and unexposed subjects from cohort studies. Top: misclassified immortal time: drug exposure is defined after cohort entry, so that the time between cohort entry and the first drug exposure, for those who become exposed, is immortal, since the subject must survive to receive the drug and misclassified as exposed, when in fact is unexposed. Bottom: excluded immortal time: cohort entry for the exposed is defined by the first drug exposure and by an event for the unexposed. The time between the event and first drug use is immortal, since the subject must survive to receive the drug and not accounted for in the unexposed group biased. 3 The solution is to use a time-dependent definition of exposure that properly classifies this immortal person-time as unexposed, until the start of drug use and exposed thereafter. 3,4,13 Excluded immortal time Another approach to cohort studies is to first form the exposed cohort from all subjects who received the drug under study, thus entering the cohort at the time of their first prescription (Figure 1, bottom). All other subjects are then considered unexposed and their cohort entry is defined arbitrarily by some entry event. Patients from both cohorts are then followed up from cohort entry to the occurrence of the outcome event. It is important to recognise that the exposed subjects may also have had the same entry event prior to their exposure-defined cohort entry date. Therefore, the time between this prior entry event and cohort entry for the exposed group is immortal (Figure 1, bottom) and must thus be accounted for. Indeed, the exposed subjects who previously had the event could not have incurred the outcome event before their cohort entry, since, had they had this event during this immortal period, they would have been classified in the unexposed group. Thus, by simply excluding this pre-cohort entry immortal person-time generated by the exposed subjects will introduce a bias, since it should have been accounted for as unexposed. This period would normally add persontime to the unexposed denominator but not add any new events to its numerator since immortal. By excluding this immortal time, the outcome rate in the unexposed group will be artificially increased, because the rate will be missing part of the denominator so that the rate ratio will be biased downwards. 5 Such bias from excluded immortal time is identified in some studies of Table 1. 6,11,27,28 For example the GPRD study of the use of statins, associated with the risk of fractures used a cohort approach, with a nested case-control analysis, where exposure cohorts

5 were identified in a hierarchical manner. 6 First, all subjects exposed to a statin or to another lipid-lowering drug (LLD) at anytime up to September 1998 were identified and formed the LLD-exposed cohort, with follow-up starting at the time of their first prescription. The subjects who were not treated for their hypercholesterolemia formed an unexposed cohort, along with a sample of randomly selected subjects from the remainder. The follow-up of these two latter cohorts started a year after their start in the GP practice. Thus, the cohort follow-up starts at different times for the exposed and unexposed cohorts. Clearly, the time between a year after their start in the GP practice and the date of the first LLD for the subjects exposed to a LLD is immortal and unexposed. This immortal person-time is not accounted for, either as exposed or unexposed, when in fact it should be counted as unexposed and added to the reference group. Here again, the solution is to use a person-time definition of exposure that does not exclude, but rather accounts for this immortal person-time as unexposed until the start of LLD use. 5 ILLUSTRATION OF THE BIAS We illustrate the bias engendered by both misclassified and excluded immortal time. To illustrate the bias, we verify whether two drug classes bearing no plausible beneficial effect in the treatment of CVD can appear effective at reducing the 1-year rate of death from all causes. Cohort definition and outcome The base cohort was formed from an existing cohort of COPD patients, identified using health insurance databases from Saskatchewan, Canada, described previously. 29 All subjects, newly treated for COPD and at least 55 years of age during were included. Data were obtained until 31 December 1999, death or emigration from the Province. The study cohort included those patients hospitalised for a primary diagnosis of CVD, including ischemic heart disease, heart failure, cardiomyopathy, dysrythmia and cerebrovascular disease, on or after 1 January The cohort for the misclassified immortal time situation had cohort entry defined as the date of discharge from their CVD hospitalisation. All subjects were followed up for 1 year. The outcome was death from any cause during the 1-year follow-up. The cohort for the excluded immortal time situation had cohort entry taken as the date of the first dispensed immortal time bias 245 prescription for the study drug, within a year after discharge from their hospitalisation for CVD. For subjects not receiving the study drug, cohort entry was the discharge date and all subjects were followed for up to 1 year. The outcome was death from any cause, during the 1-year follow-up from the respective cohort entry dates. Drug exposure and data analysis As the exposure, we selected two classes of frequently used drugs in this population, but unknown to be effective in treating CVD, namely inhaled betaagonists (IBA) and gastrointestinal drugs (GID). IBA included albuterol, fenoterol, terbutaline. The gastrointestinal agents included H-2 antagonists, proton pump inhibitors, as well as misoprostol and sucralfate. For the cohort with misclassified immortal time, in accordance with the approach used in most studies, a patient was considered exposed to IBA if they filled a prescription during the first 180 days of follow-up, and were considered unexposed otherwise. To simplify the illustration and emphasise the bias, we selected an equal number of exposed and unexposed subjects from the base cohort. Thus, since there were more unexposed than exposed cohort members, we used all exposed subjects and randomly selected an equal number of unexposed subjects from the base cohort. The resulting data were first analysed by the biased approach used in all studies, namely by classifying exposure, according to the subject and allocating this exposure to the entire follow-up time. Second, a person-time approach was used, where the unexposed follow-up prior to the start of exposure (the immortal time) was classified as unexposed and the subsequent follow-up time was classified as exposed. For the cohort with excluded immortal time, a subject was hierarchically classified in the exposed cohort, if they were dispensed the study drug during the 1-year period after cohort entry. Their cohort entry date was the date of their first prescription for this drug and they were followed up for a year after that. All other subjects were included in the unexposed cohort and their cohort entry date was the date of discharge, with a 1-year follow-up as well. Here again, we selected an equal number of exposed and unexposed subjects to simplify the illustration. The data were first analysed by the biased approach, namely by starting follow-up at the hierarchically defined entry date. The second approach used the immortal unexposed person-time prior to the start of exposure that was not accounted for in the first analysis.

6 246 s. suissa Table 2. Rate ratios of death for the misclassified immortal time approach, estimated by the misclassified and correctly classified approaches, with exposure defined by use of inhaled beta-agonist (IBA) drugs in the 180-day period after cohort entry Number of Subjects Number of deaths Person-years Rate (per 100 per year) Rate ratio (95% confidence interval) Misclassified approach No IBA during 180-day period Reference IBA during 180-day period ( ) Correctly classified approach Unexposed to IBA Reference No IBA during 180-day period From cohort entry to IBA Exposed after IBA prescription ( ) In all instances, rate ratios of death associated with exposure to gastrointestinal medications or IBA were estimated by employing the corresponding number of deaths and person-time. The Poisson distribution was used to estimate all confidence intervals. RESULTS The base cohort included the same 3315 subjects who were hospitalised for CVD after 1 January Table 2 shows that for the cohort with misclassified immortal time, there were 771 subjects who received IBA during the first 180 days of follow-up and were classified as exposed, of which 114 died within the 1-year follow-up. The unexposed group included 771 subjects who did not receive an IBA during the first 180 days of follow-up, of whom 148 died within the 1-year follow-up. The resulting rate ratio is 0.73 (95%CI: ), suggesting a reduced mortality rate with IBA use. Table 2 also shows, however, that the 713 person-years of follow-up, generated by the IBA-exposed group includes person-years, during which the subjects were yet to be exposed. This immortal time (no deaths) should, in fact, have been classified as unexposed with the result that the corrected rate ratio becomes 0.98 (95%CI: ) and no longer significant. Figure 2 displays the corresponding Kaplan Meier survival functions that demonstrate the greater incidence of deaths, during the early follow-up period of the unexposed group, particularly during the first 180-day period used to define exposure. On the other hand, the exposed group has a lower incidence that is the result of the advantage they received by having an immortal period, during the first 180 days of follow-up. For the cohort with excluded immortal time, we used exposure to gastrointestinal drugs (GID) for Figure 2. Kaplan Meier survival functions for the misclassified immortal time approach according to exposure or non-exposure to inhaled beta-agonist drugs dispensed, during the first 180 days after discharge for a hospitalisation for cardiovascular disease

7 Table 3. Rate ratios of death for the excluded immortal time approach, using the gastrointestinal drug (GID) defined cohort, estimated by the incorrectly excluded and correctly classified approaches Number of subjects immortal time bias 247 Number of deaths Person-years Rate (per 100 per year) Rate ratio (95% confidence interval) Excluded immortal time approach No GID during follow-up Reference GID during follow-up ( ) Correctly classified approach Unexposed to GID Reference No GID during follow-up From cohort entry to GID Exposed after GID prescription ( ) illustration. Table 3 shows that the exposed cohort includes 640 subjects who received GID in the year after discharge, of whom 99 died during the 1-year follow-up that started on the day of the first GID exposure. The unexposed group also included 640 subjects who did not receive a GID prescription, of whom 125 died within the 1-year follow-up that started on the day of discharge. The resulting rate ratio, based on the comparison of these two cohorts is 0.78 (95%CI: ), suggesting a reduced mortality rate with GID use. Table 3, however, shows that the person-years of follow-up from the GID-exposed group were preceded by personyears, during which the subjects were yet to be exposed, but were eligible to be counted in the unexposed group. This immortal time (no deaths) should, in fact, have been classified as unexposed, with the result that the corrected rate ratio becomes 0.94 (95%CI: ). DISCUSSION We found that several recently published observational studies, suggesting that various medications are very effective at preventing major outcomes or safe, use a design that is seriously biased. These studies were all based on health care databases and employed one of two types of cohort definition that resulted in an important amount of immortal time that was either misclassified or excluded. This biased approach resulted in an artificial overestimation of the outcome rate in the reference group. Consequently, the group exposed to the drug under study appeared to have a lower rate of the outcome, when compared with the reference group, producing the illusion that the drug is effective. The illustration we used to depict this bias exemplifies the phenomenon. We showed that this biased approach could be used to suggest that several dummy drugs such as GID and IBA, both unknown to be in any way effective in treating CVD, reduce the 1-year rate of all-cause death by up to 27%. This rate reduction is of similar magnitude as that found with beta-blockers or LLDs, that have been proposed by studies using this design as truly beneficial, after hospitalisation for acute myocardial infarction. 9,19 Our study shows that this appearance of effectiveness is simply the result of immortal time bias. Two of the 20 studies singularly emphasise this bias. The study suggesting a rate ratio of 0.55 for fracture incidence, associated with statins was in fact independently replicated. 6 Other investigators also used the same GPRD, but found a radically different result, namely a rate ratio of An accompanying editorial proposed several explanations for this wide discrepancy between the studies, including the different number of practices and sizes of the exposure periods, but did not address the role of immortal time. 8 While common biases with claims data, such as unmeasured indication, could also have played a role in these differences, the major difference between the two designs is in the definition of cohort entry: The first study was based on the hierarchical approach that defined cohort entry by the exposure, thus not accounting for unexposed immortal time, 6 while the second study defined cohort entry in the same way for statin users and non-users, and was thus able to properly account for the entire immortal time span prior to the start of statin exposure as unexposed. 7 The study that assessed whether a visit to a physician during the 90 days after an emergency department (ED) visit for asthma or COPD could reduce the rate of subsequent readmissions to the ED, also underscores the bias. 18 Indeed, had one alternatively inquired whether a prescription for any drug at all (including respiratory and non-respiratory) during the same 90-day period could reduce the rate of subsequent readmissions to the ED, the analysis would have produced exactly the same results, since a prescription necessarily requires a visit to a physician. Thus, any action or exposure, concomitant with the

8 248 s. suissa physician visit would have appeared to be beneficial by this design. Of course, immortal time bias shows this is only so, because the exposed patient had to have a period of immortality (no readmission) to permit a visit their physician, while there was no such advantage for the unexposed patients who were readmitted before they could visit their physician. The appropriate approach to data analysis for the studies employing these designs requires that all immortal time be accounted for fully, including that before the start of exposure, and that it be correctly classified in terms of exposure. When this approach was used in our illustration, the rate ratios correctly became no different than unity. Thus, the apparent effectiveness suggested by the biased approach disappeared with proper consideration of immortal time. Of course, the extent of the bias will depend directly on the amount of total person-time misclassified or excluded. Thus, the longer the exposure window, the more immortal time is misclassified, the larger the bias. Alternatively, if the exposure window is short, the impact of the bias will be negligible. 3 While our illustration used a simple Poisson rate approach to data analysis based on classifying person-time, more sophisticated techniques, such the Cox proportional hazards model, can be used. Such models, however, must include time-dependent exposures that properly classify the patients timeto-event as unexposed, until they become exposed and exposed thereafter, to provide the proper estimates. Indeed, when such a Cox model is used, the risk sets must include a patient during their immortal period as alive and unexposed, and this status must be included until the exposure for that patient starts. Otherwise, their exclusion or misclassification will cause an underascertainment of the unexposed non-cases in the risk set, leading to an underestimation of the hazard ratio. Several alternative approaches to the analysis of cohort studies to avoid misclassified immortal time were recently compared. 30 That paper was to some extent, confusing by its inaccurate designation of immortal time bias as survival bias, which has another meaning altogether in epidemiology. 31,32 Observational database research is challenging for researchers, peer-reviewers, editors and readers of medical journals. The 20 databases studies we identified, published in respectable journals, had all the attractive properties for acceptance and credibility. However, since their potential impact can be sizable, particularly as treatment and practice guidelines are becoming more popular, an in-depth re-evaluation and re-analysis of these and similar studies should be performed, before their potentially inaccurate conclusions are used to influence clinical practice. KEY POINTS Immortal time bias can affect observational cohort studies when improper design or analysis methods are used. Immortal time bias has been recognised in 20 recently published observational pharmacoepidemiology database studies. Improper data analysis methods can be used to show absurd findings such as a reduction in mortality with the use of gastrointestinal drugs and inhaled beta-agonists in cardiovascular disease. A proper person-time approach to data analysis eliminates this bias. Observational cohort studies with surprising beneficial effects should be re-assessed for immortal time bias. ACKNOWLEDGEMENTS This research was funded by grants from the Canadian Institutes of Health Research (CIHR) and Fonds de la recherche en santé du Québec (FRSQ). The author is the recipient of a Distinguished Investigator award from the CIHR. The database used for the illustration was acquired thanks to prior grants from AstraZeneca, Boehringer-Ingelheim and GlaxoSmithKline. DISCLAIMER This study uses data provided by the Saskatchewan Department of Health. The interpretation and conclusions contained herein do not necessarily represent those of the government of Saskatchewan or the saskatchewan Department of Health. REFERENCES 1. Walker AM. Observation and Inference: an Introduction to the Methods of Epidemiology. Epidemiology Resources, Inc.: Newton Lower Falls, MA, Rothman KJ, Greenland S. Modern Epidemiology. (2nd edn), Lippincott-Raven: Hagerstown, MD, Suissa S. Effectiveness of inhaled corticosteroids in chronic obstructive pulmonary disease: immortal time bias in observational studies. Am J Respir Crit Care Med 2003; 168(1): Samet JM. Measuring the effectiveness of inhaled corticosteroids for COPD is not easy. Am J Respir Crit Care Med 2003; 168(1): Suissa S. Inhaled steroids and mortality in COPD: bias from unaccounted immortal time. Eur Resp J 2004; 23(3): Meier CR, Schlienger RG, Kraenzlin ME, Schlegel B, Jick H. HMG-CoA reductase inhibitors and the risk of fractures. JAMA 2000; 283(24):

9 immortal time bias van Staa TP, Wegman S, de Vries F, Leufkens B, Cooper C. Use of statins and risk of fractures. JAMA 2001; 285(14): Hennessy S, Strom BL. Statins and fracture risk. JAMA 2001; 285(14): Rochon PA, Tu JV, Anderson GM, et al. Rate of heart failure and 1-year survival for older people receiving low-dose betablocker therapy after myocardial infarction. Lancet 2000; 356(9230): 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(4): Soriano JB, Vestbo J, Pride NB, Kiri V, Maden C, Maier WC. Survival in COPD patients after regular use of fluticasone propionate and salmeterol in general practice. Eur Respir J 2002; 20(4): Adams RJ, Fuhlbrigge AL, Finkelstein JA, Weiss ST. Intranasal steroids and the risk of emergency department visits for asthma. J Allergy Clin Immunol 2002; 109(4): Suissa S, Ernst P. Bias in observational study of the effectiveness of nasal corticosteroids in asthma. J Allergy Clin Immunol 2005; 115(4): Donahue JG, Weiss ST, Livingston JM, Goetsch MA, Greineder DK, Platt R. Inhaled steroids and the risk of hospitalization for asthma. JAMA 1997; 277(11): Sin DD, Tu JV. Lack of association between ipratropium bromide and mortality in elderly patients with chronic obstructive airway disease. Thorax 2000; 55(3): Sin DD, Tu JV. Inhaled corticosteroid therapy reduces the risk of rehospitalization and all-cause mortality in elderly asthmatics. Eur Respir J 2001; 17(3): Adams RJ, Fuhlbrigge A, Finkelstein JA, et al. Impact of inhaled antiinflammatory therapy on hospitalization and emergency department visits for children with asthma. Pediatrics 2001; 107(4): Sin DD, Bell NR, Svenson LW, Man SF. The impact of followup physician visits on emergency readmissions for patients with asthma and chronic obstructive pulmonary disease: a population-based study. Am J Med 2002; 112(2): Klungel OH, Heckbert SR, de Boer A, et al. Lipid-lowering drug use and cardiovascular events after myocardial infarction. Ann Pharmacother 2002; 36(5): Sin DD, Man SF. Low-dose inhaled corticosteroid therapy and risk of emergency department visits for asthma. Arch Intern Med 2002; 162(14): Sin DD, McAlister FA. The effects of beta-blockers on morbidity and mortality in a population-based cohort of 11, 942 elderly patients with heart failure. Am J Med 2002; 113(8): Sin DD, Man SF. Inhaled corticosteroids and survival in chronic obstructive pulmonary disease: does the dose matter? Eur Respir J 2003; 21(2): Gerrits CM, Herings RM, Leufkens HG, Lammers JW. N-acetylcysteine reduces the risk of re-hospitalisation among patients with chronic obstructive pulmonary disease. Eur Respir J 2003; 21(5): Soriano JB, Kiri VA, Pride NB, Vestbo J. Inhaled corticosteroids with/without long-acting ß-agonists reduce the risk of rehospitalization and death in COPD patients. Amer J Respir Med 2003; 2(1): Smith MJ, Rascati KL, McWilliams BC. Inhaled antiinflammatory pharmacotherapy and subsequent hospitalizations and emergency department visits among patients with asthma in the Texas Medicaid program. Ann Allergy Asthma Immunol 2004; 92(1): Yee A, Majumdar SR, Simpson SH, McAlister FA, Tsuyuki RT, Johnson JA. Statins use in Type 2 diabetes mellitus is associated with a delay in starting insulin. Diabet Med 2004; 21(9): Jick H, Zornberg GL, Jick SS, Seshadri S, Drachman DA. Statins and the risk of dementia [In Process Citation]. Lancet 2000; 356(9242): Yang CC, Jick SS, Jick H. Lipid-lowering drugs and the risk of depression and suicidal behavior. Arch Intern Med 2003; 163(16): Bourbeau J, Ernst P, Cockcoft D, Suissa S. Inhaled corticosteroids and hospitalisation due to exacerbation of COPD. Eur Respir J 2003; 22(2): Zhou Z, Rahme E, Abrahamowicz M, Pilote L. Survival bias associated with time-to-treatment initiation in drug effectiveness evaluation: a comparison of methods. Am J Epidemiol 2005; 162(10): Encyclopedia of Epidemiologic Methods. In Encyclopedia of Epidemiologic Methods, Gail MH, Benichou J (eds). John Wiley & Sons, Ltd.: New York, Epidemiology: Beyond the Basics. In Epidemiology: Beyond the Basics, Szklo M, Nieto FJ (eds). Aspen Publishers, Inc.: Gaithersburg, Maryland, 2004.