Research
ajog.org
GYNECOLOGY
Immortal time bias in drug safety cohort studies: spontaneous abortion following nonsteroidal antiinflammatory drug exposure Sharon Daniel, MD, MPH; Gideon Koren, MD; Eitan Lunenfeld, MD, MHA; Amalia Levy, MPH, PhD OBJECTIVES: Experimental research of drug safety in pregnancy is
generally not feasible because of ethical issues. Therefore, most of the information about drug safety in general and teratogenicity in particular is obtained through observational studies, which require careful methodologic design to obtain unbiased results. Immortal time bias occurs when some cases do not “survive” sufficient time in the study, and as such, they have reduced chances of being defined as “exposed” simply because the durations of their follow-ups were shorter. For example, studies that examine the risk for spontaneous abortions in women exposed to a drug during pregnancy are susceptible to immortal time bias because the chance of drug exposure increases the longer a pregnancy lasts. Therefore, the drug tested may falsely be found protective against the outcome tested. The objective of the current study was to illustrate the extent of immortal time bias using a cohort study of pregnancies assessing the risk for spontaneous abortions following nonsteroidal antiinflammatory drug exposure.
2 definitions of exposure (dichotomous, exposed vs unexposed, regular Cox regression vs Cox regression with time-varying exposure). RESULTS: Significant differences were found in the risk for sponta-
neous abortions between the 2 statistical methods, both for groups and for most specific nonsteroidal antiinflammatory drugs (nonselective Cox inhibitors e hazard ratio, 0.70; 95% confidence interval, 0.61e0.94 vs hazard ratio, 1.10; 95% confidence interval, 0.99e1.22 for dichotomous vs time-varying exposure analyses, respectively). Furthermore, a significant correlation was found between the median misclassified immortal time for each drug and the extent of the bias.
STUDY DESIGN: We assembled 3 databases containing data on
CONCLUSION: Immortal time bias can easily occur in cohort studies assessing the risk for adverse pregnancy outcomes following exposure to drugs. One way to prevent such a bias is by defining exposure only from the time of exposure during follow-up onward using a timevarying exposure analysis.
spontaneous abortions, births and drug dispensions to create the present study’s cohort. The risk for spontaneous abortion was assessed using 2 statistical analysis methods that were compared for
Key words: ibuprofen, immortal time bias, miscarriage, NSAIDs, spontaneous abortions
Cite this article as: Daniel S, Koren G, Lunenfeld E, et al. Immortal time bias in drug safety cohort studies: spontaneous abortion following nonsteroidal antiinflammatory drug exposure. Am J Obstet Gynecol 2014;211:x-ex-x-ex.
E
xperimental research of drug safety is generally not feasible because of ethical issues. Therefore, most of the information about drug safety in general and teratogenicity in particular is obtained through observational studies,1 which require careful methodologic
design to obtain unbiased results on which valid conclusions can be based.2 One source of bias in retrospective cohort studies of drug safety is created by inappropriate definitions of exposure, leading to “immortal time bias.”3-5 An immortal time bias can result when,
From the Departments of Public Health (Drs Daniel and Levy), Pediatrics (Dr Daniel), and Obstetrics and Gynecology (Dr Lunenfeld), Faculty of Health Sciences, Ben-Gurion University of the Negev, Soroka Medical Center, and the Ben-Gurion Motherisk Obstetric Registry of Exposure (BeMORE) collaboration (Drs Daniel, Koren, and Levy), Beer-Sheva, Israel, and the Motherisk Program, Division of Clinical Pharmacology and Toxicology (Dr Koren), Hospital for Sick Children, University of Toronto, Toronto, ON, Canada. Received July 4, 2014; revised Sept. 10, 2014; accepted Sept. 24, 2014. The authors report no conflict of interest. The research was not funded by any company or organization. Corresponding author: Amalia Levy, MPH, PhD. [email protected] 0002-9378/$36.00 ª 2014 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.ajog.2014.09.028
during the follow-up period, patients participating in the study develop the outcome under investigation (eg, death or miscarriage) before they have had the chance to be exposed to the medication tested, simply because the duration of the follow-up period was too short, ie, the patient did not “survive” sufficient time to be able to receive the drug. As a result, patients in the study who survive longer have a higher chance of being defined as “exposed.” Figure 1 shows a hypothetical scenario in which this bias can be created. Suppose there were 2 women in the study who conceived on the same day. After exposure to nonsteroidal antiinflammatory drugs (NSAIDs) in her 15th gestational week, the first woman ended the follow-up period without an “event”
MONTH 2014 American Journal of Obstetrics & Gynecology
1.e1
Research
Gynecology
ajog.org
FIGURE 1
Illustration of immortal time bias in cohort studies of drug safety during pregnancy
Definition of exposure A, not controlling vs B, controlling for immortal time bias. Strand. Immortal time bias in drug safety cohort studies. Am J Obstet Gynecol 2014.
(a spontaneous abortion) and gave birth. In contrast, the second woman had a spontaneous abortion on the 10th gestational week, before being exposed to the drug. If a nontime-varying exposure statistical analysis (a logistic regression or Cox regression) was used, the first woman, who gave birth, would be assigned to the “exposed” group whereas the second woman, who had a spontaneous abortion at 10 gestational weeks, would be assigned to the “unexposed” group (Figure 1, A). The period of time during which the first woman was followed (20 weeks) was twice that of which the second woman was followed (10 weeks), and therefore, the first woman, who gave birth, was twice as likely to be assigned to the “exposed” group compared with the second women, who had a spontaneous abortion. Hence an association would be found such that NSAIDs had a protective effect against spontaneous abortions. However, this spurious protective effect
was caused not by the drug itself, but rather, by the study’s design. In contrast, with a time-varying exposure analysis, the exposed and unexposed groups are not defined dichotomously (as “exposed” vs “unexposed”) from the beginning of the follow-up period. Rather, exposure is redefined during follow-up, such that a woman is counted for as “exposed” only from the period of time that follows the actual exposure. Woman 1, in that case, would only be counted for as exposed from the 15th gestational week onward (Figure 1, B). Another way of addressing this bias is by presenting the Cox regression model which estimates the hazard for an event (eg, a spontaneous abortion) during follow-up, such that the risk profile (ie, exposure to NSAIDs) is compared at each event time between subjects who developed the event under study (ie, women who experienced a spontaneous abortion) and subjects
1.e2 American Journal of Obstetrics & Gynecology MONTH 2014
who stayed in the study at that time point. By defining exposure dichotomously (“exposed” vs “unexposed”) among women throughout the whole follow-up period, a misclassification of the risk profile occurs—women who were exposed to NSAIDs only in a point of time during follow-up are misclassified as “exposed” for the period of time before the actual exposure. Because women who gave birth “survive” longer in follow-up and hence have a higher chance of using NSAIDs, this misclassified period of time is more likely to occur in women who did not experience a spontaneous abortion, therefore NSAIDs would be found protective against spontaneous abortions. Hence, a varying exposure analysis enables the redefinition of exposure, such that a woman would be counted for as “unexposed” from the beginning of follow-up to the time of actual exposure. Three studies have addressed the potential bias in the association between risk factors and adverse pregnancy outcomes by referring to none time-varying exposure during pregnancy as anytime during-pregnancy exposure.6-9 The aim of the current study was to illustrate the extent of the bias using a recently published retrospective cohort study that assessed the association between exposure to NSAIDs and spontaneous abortions.10
M ATERIALS
AND
M ETHODS
Detailed descriptions of the databases and definitions of the variables used in the present study were published recently.10 We conducted a populationbased retrospective cohort study whose participants included all 15- to 45-yearold women who registered with Clalit Health Services health maintenance organization, who conceived between January 2003 and December 2009, and who were admitted for birth or were diagnosed with spontaneous abortion at Soroka Medical Center. To create the cohort for the current study, we combined 3 computerized databases in which patients are listed according to their personal identification numbers (unique numbers assigned to
Gynecology
ajog.org all citizens by the Israeli Ministry of the Interior and used throughout life). The Soroka Medical Center deliveries and hospitalization databases include information about deliveries and spontaneous abortions that occurred at the medical center. The drug dispension database of Clalit Health Services, the third database, contains electronic records of the drugs dispensed to all registered patients. Gestational age at the time of birth or spontaneous abortion was retrieved from the medical records, and was determined by an obstetrician according to the first day of the last menstrual period as reported by the pregnant woman or according to the estimated gestational age in a first trimester ultrasound imaging, as documented in the medical charts. The exposure period was from the first day of pregnancy until the 20th gestational week for births, or until the day before the diagnosis for spontaneous abortions. For missed abortions, the exposure period continued until the estimated radiologic and morphologic gestational day according to ultrasound imaging performed on admission to the emergency room. Ibuprofen, diclofenac, naproxen, etodolac, lornoxicam and nabumetone were defined as “nonselective Cox inhibitors”; celecoxib, etoricoxib and rofecoxib were defined as “Cox2 selective inhibitors.” We assessed immortal time bias by applying 2 different definitions of exposure, controlling and not controlling for the bias, and comparing the results. In the first instance, exposures were defined dichotomously such that a pregnancy was counted as “exposed” if an NSAID was dispensed from the first day of the last menstrual period to the end of the 20th gestational week or until 2 days before an abortion. Univariate and multivariate Cox regression analyses were performed to assess the risk for spontaneous abortion following exposure to an NSAID. In the second instance, as performed in our published study,10 both univariate and multivariate Cox regression with time-varying exposure analyses were performed to control for immortal time bias. A time-varying
exposure analysis redefined the exposure group at every new exposure that occurred throughout the follow-up period, such that a pregnancy would only be counted as “exposed” from the timeofthefirstreportofexposureforward. All multivariate models were adjusted for diabetes mellitus (International Statistical Classification of Diseases and Related Health Problems-9 Codes 250, 357.2), hypothyroidism (244), obesity (278,649.1), hypercoagulable (286.4, 289.81, 286.53), and inflammatory (710, 714, 720) disorders, recurrent miscarriages (646.33, 629.81), self-report of tobacco use during pregnancy (305.1, 649), the presence of an intrauterine contraceptive device (V4551, Z975), in vitro fertilization of the current pregnancy, maternal age, ethnic group (ie, Jewish vs Bedouin Muslim), and the year of admission. Furthermore, to quantify the extent of the bias according to the length of the misclassified period of immortal time, we correlated the median gestational day at first exposure to the drug and the difference between the adjusted hazard ratio, calculated by Cox regression with time-varying exposure analysis, and the hazard ratio (HR), calculated based on the dichotomous definition of exposure. The Pearson correlation coefficient was also calculated. The study was approved by an ethics committee (institutional review board) according to the declaration of Helsinki.
R ESULTS The cohort included 66,547 pregnancies that began during the study period, of which 6508 (9.9%) ended with spontaneous abortions. Data concerning gestational age at abortion were missing for 1090 (14.3%) pregnancies, and therefore, these were excluded from the analysis. The proportion of pregnancies exposed to NSAIDs during the study period and the mean gestational day at exposure are presented in Table 1. Table 2 presents the differences between both univariate and multivariate HRs using Cox regression analysis with a dichotomous definition of exposure and using Cox regression with time-varying exposure models, following exposure to
Research
nonselective Cox inhibitors, to Cox2 selective inhibitors and to most specific NSAIDs. Significant changes were found between the 2 types of exposure definitions in both univariate and multivariate analyses. As such, NSAIDs were found to be significantly protective against spontaneous abortions by defining exposure dichotomously, but that effect disappeared if the time-varying exposure analysis was performed. The difference between the HR calculated by a time-varying exposure analysis and by a dichotomous definition of exposure using Cox regression was found to be highly correlated with the median gestational age at first reported exposure for all NSAIDs (r ¼ 0.98; P < .001; Figure 2, A) and after omitting indomethacin (r ¼ 0.92; P < .001; Figure 2, B).
C OMMENT Our study demonstrates that analyzing the results of a retrospective cohort study of the risk for spontaneous abortions following exposure to NSAIDs without using a time-varying exposure statistical model creates a bias in favor of a protective effect of NSAIDs against spontaneous abortions. Similarly, O’Neill et al6 showed that by analyzing urinary tract infection (UTI) during pregnancy as a dichotomous risk factor, UTI would be seen as protective against preterm birth, as the longer a woman stayed pregnant, the higher the chance she would be diagnosed with a UTI. Likewise, Hutcheon et al7 demonstrated that by ignoring immortal time bias, gestational diabetes may be associated with a lower risk of stillbirth, because pregnancy must survive to 28 weeks for a diagnosis of gestational diabetes mellitus to be made, but about half of stillbirths occur before this time. In a recently published study, Matok et al9 have shown that the claimed decreased risk of preterm delivery after gestational use of decongestants was due to immortal time bias, arising from the improper classification of exposure during follow-up. Our findings in the current study also demonstrate a strong correlation between the length of the misclassified
MONTH 2014 American Journal of Obstetrics & Gynecology
1.e3
Research
Gynecology
ajog.org
TABLE 1
Proportion of pregnancies exposed to NSAIDs during the study period and median gestational age at exposure Drug Nonselective Cox inhibitors
Median gestational day at exposure
No. of exposed 4424 (6.75%)
31
Ibuprofen
2732 (4.17%)
40
Diclofenac
919 (1.40%)
23
Naproxen
671 (1%)
27
Etodolac
272 (0.4%)
20.5
Indomethacin
132 (0.2%)
89
71 (0.1%)
25
Cox2 selective inhibitors NSAID, nonsteroidal antiinflammatory drug.
Strand. Immortal time bias in drug safety cohort studies. Am J Obstet Gynecol 2014.
immortal time period and the change in HR following analysis with time-varying exposure model. The longer the period of misclassified time, therefore, the higher the extent of the bias. This type of bias may occur in any observational study in which exposure changes during the follow-up period. Because the follow-up period is shorter for subjects who experience an event, this misclassification of exposure would always underestimate the risk for the event under study. As described, a woman was counted for as “exposed” from the time point of first exposure until the end of follow-up (ie, the end of the 20th gestational week). This definition was based on the biologic
assumption that exposure to NSAIDs might have induced a cascade of events that ended with a spontaneous abortion. Assuming a different biologic scenario, a different definition of exposure could have been made. For example, if NSAIDs were assumed to have a transient effect, a woman could have been defined as exposed only when she was actually using the drug. Although the analysis was performed using Cox regression with time-varying exposure, other methods could have been used to overcome the bias. Assuming follow-up begins from the time of exposure, cases of spontaneous abortions occurring before exposure could have been considered as left
truncated and therefore Cox regression with a correction for left truncation could have been used.11 Moreover, immortal time bias could also be viewed as a selection bias; therefore latent variable analysis could also be performed.12 Previous studies that assessed the risk for spontaneous abortions following the use of NSAIDs measured exposure by the date of prescription rather than the dispension of the drugs.13-15 Filling a prescription, however, does not represent actual use of the drug prescribed. In the present study, exposure was defined by dispension of the drugs. Previous studies found an increasing use of overthe-counter NSAIDs that were not recorded in those studies. Thus, a bias toward the null hypothesis could occur, by misclassification of exposed women as unexposed. Because a vast majority of the pharmacies in the southern district are affiliated to Clalit Health Services, and because over-the-counter dispensions at affiliated pharmacies are recorded in our database, the extent of such a bias in our study is insignificant. Furthermore, in sensitivity analysis that was performed on the same population, the estimated proportion of unrecorded dispensions within the unexposed group was 1.3%.10 Some of the spontaneous abortions occur very early during pregnancy, before the pregnancy is recognized and these escape medical attention. Because
TABLE 2
Corrected and non-corrected risk for spontaneous abortions following exposure to NSAIDs Cox regression with time-varying exposure HR ratio (95% CI)
Cox regression HR (95% CI)
Drug
Unadjusted
Adjusteda
Unadjusted
Adjusteda
Nonselective Cox inhibitors
1.13 (1.01e1.25)
1.10 (0.99e1.22)
0.81 (0.73e0.91)
0.77 (0.69e0.86)
Ibuprofen
1.13 (0.98e1.30)
1.06 (0.93e1.22)
0.75 (0.66e0.86)
0.70 (0.61e0.94)
Diclofenac
1.21 (0.98e1.48)
1.19 (0.97e1.46)
0.98 (0.8e1.21)
0.96 (0.79e1.19)
Naproxen
1.87 (0.66e1.17)
0.97 (0.74e1.28)
0.75 (0.57e0.99)
0.75 (0.57e0.99)
Etodolac
1.26 (0.88e1.80)
1.28 (0.91e1.79)
1.23 (0.88e1.73)
1.12 (0.80e1.57)
Indomethacin
3.54 (2.20e5.71)
2.82 (1.70e4.69)
1.14 (0.68e1.89)
1.05 (0.63e1.75)
Cox2 selective inhibitors
1.97 (1.12e3.47)
1.43 (0.79e2.59)
1.52 (0.84e2.75)
1.24 (0.68e2.25)
CI, confidence interval; HR, hazard ratio; NSAID, nonsteroidal antiinflammatory drug. a
Confidence intervals.
Strand. Immortal time bias in drug safety cohort studies. Am J OBstet Gynecol 2014.
1.e4 American Journal of Obstetrics & Gynecology MONTH 2014
Gynecology
ajog.org
FIGURE 2
Correlation between gestational day at exposure and corrected and non-corrected hazard ratios
Pearson correlation between the median gestational day at first exposure and the difference between the hazard ratios calculated by Cox regression and by Cox regression with time-varying exposure for A, all NSAIDs and B, after omitting indomethacin. Strand. Immortal time bias in drug safety cohort studies. Am J Obstet Gynecol 2014.
our study only included spontaneous abortions that presented to the emergency room at Soroka Medical Center, these cases were not accounted for. Assuming some of the pregnant women have not yet attended a first prenatal visit to an obstetrician, the reported gestational age might be inaccurate. However, the whole Israeli population is medically insured and the
Israeli preventive medicine guidelines indicate a free ultrasound imaging and gestational age determination at the beginning of pregnancy to all pregnant women. Further, the electronic medical records are available for the emergency room gynecologists for all pregnant women registered with Clalit Health Services. Therefore, the extent of such a bias is estimated to be low.
Research
Data on gestational age was missing for 14.3% of pregnancies that ended with spontaneous abortions. There were no significant differences between cases of spontaneous abortions with or without data on gestational age regarding maternal characteristics and risk factors for spontaneous abortions. Moreover, a second analysis was performed after assigning mean gestational age at the time of spontaneous abortion for this group of pregnancies. This analysis yielded similar results. Gestational age at the time of birth or at the time of spontaneous abortions was defined accurately in our study, enabling an accurate definition of exposures throughout the follow-up period. In so doing, we avoided the misclassification bias that could have resulted from defining women who used NSAIDs before pregnancy or in the 2 days before abortion as exposed. Moreover, defining gestational age at the time of first exposure allows for a time-varying exposure analysis. The history of spontaneous abortions, when used as a potential confounder, was previously suspected of biasing the effect between exposure and future perinatal outcomes by influencing exposure on future pregnancies.16 The same question is raised regarding the use of the history of recurrent spontaneous abortions as a potential confounder in our models. Furthermore, findings are inconsistent regarding the association between the use of intrauterine devices and spontaneous abortions, therefore using this variable as a potential confounder might be inappropriate. However, results were not changed after omitting these 2 variables from the analysis. Immortal time bias can easily occur in cohort studies assessing the risk for adverse pregnancy outcomes following exposure to drugs. One way to prevent such a bias is by defining exposure only from the time of exposure during follow-up onward using a time-varying exposure analysis. Thus to produce unbiased results in studies assessing drug safety during pregnancy, the definition of exposure must be carefully considered and the data obtained must
MONTH 2014 American Journal of Obstetrics & Gynecology
1.e5
Research
Gynecology
be analyzed with the appropriate statistical method. REFERENCES 1. Vandenbroucke JP. Benefits and harms of drug treatments. BMJ 2004;329:2-3. 2. Sanderson S, Tatt ID, Higgins JP. Tools for assessing quality and susceptibility to bias in observational studies in epidemiology: a systematic review and annotated bibliography. Int J Epidemiol 2007;36:666-76. 3. Suissa S. Immortal time bias in pharmacoepidemiology. Am J Epidemiol 2008;167:492-9. 4. Levesque LE, Hanley JA, Kezouh A, Suissa S. Problem of immortal time bias in cohort studies: example using statins for preventing progression of diabetes. BMJ 2010;340:b5087. 5. Liu CJ, Hu YW. Immortal time bias in retrospective analysis: is there a survival benefit in patients with glioblastoma who received prolonged treatment of adjuvant valganciclovir? Int J Cancer 2014;135:250-1.
ajog.org 6. O’Neill M, Hertz-Picciotto I, Pastore LM, Weatherley BD. Have studies of urinary tract infection and preterm delivery used the most appropriate methods? Paediatr Perinat Epidemiol 2003;17:226-33. 7. Hutcheon JA, Kuret V, Joseph KS, Sabr Y, Lim K. Immortal time bias in the study of stillbirth risk factors: the example of gestational diabetes. Epidemiology 2013;24:787-90. 8. Hertz-Picciotto I, Pastore LM, Beaumont JJ. Timing and patterns of exposures during pregnancy and their implications for study methods. Am J Epidemiol 1996;143:597-607. 9. Matok I, Azoulay L, Yin H, Suissa S. Immortal time bias in observational studies of drug effects in pregnancy. Birth Defects Res A Clin Mol Teratol 2014;100:658-62. 10. Daniel S, Koren G, Lunenfeld E, Bilenko N, Ratzon R, Levy A. Fetal exposure to nonsteroidal antiinflammatory drugs and spontaneous abortions. CMAJ 2014;186:E177-82. 11. Cain KC, Harlow SD, Little RJ, et al. Bias due to left truncation and left censoring in
1.e6 American Journal of Obstetrics & Gynecology MONTH 2014
longitudinal studies of developmental and disease processes. Am J Epidemiol 2011;173: 1078-84. 12. Burnham AJ, MacGregor JF, Viveros R. Latent variable multivariate regression modeling. Chemomet Intellig Lab Syst 1999;48:167-80. 13. Nakhai-Pour HR, Broy P, Sheehy O, Berard A. Use of nonaspirin nonsteroidal antiinflammatory drugs during pregnancy and the risk of spontaneous abortion. CMAJ 2011;183: 1713-20. 14. Nielsen GL, Skriver MV, Pedersen L, Sorensen HT. Danish group reanalyses miscarriage in NSAID users. BMJ 2004;328:109. 15. Nielsen GL, Sorensen HT, Larsen H, Pedersen L. Risk of adverse birth outcome and miscarriage in pregnant users of nonsteroidal antiinflammatory drugs: population based observational study and case-control study. BMJ 2001;322:266-70. 16. Weinberg CR. Toward a clearer definition of confounding. Am J Epidemiol 1993;137: 1-8.
ajog.org
GYNECOLOGY
Immortal time bias in drug safety cohort studies: spontaneous abortion following nonsteroidal antiinflammatory drug exposure Sharon Daniel, MD, MPH; Gideon Koren, MD; Eitan Lunenfeld, MD, MHA; Amalia Levy, MPH, PhD OBJECTIVES: Experimental research of drug safety in pregnancy is
generally not feasible because of ethical issues. Therefore, most of the information about drug safety in general and teratogenicity in particular is obtained through observational studies, which require careful methodologic design to obtain unbiased results. Immortal time bias occurs when some cases do not “survive” sufficient time in the study, and as such, they have reduced chances of being defined as “exposed” simply because the durations of their follow-ups were shorter. For example, studies that examine the risk for spontaneous abortions in women exposed to a drug during pregnancy are susceptible to immortal time bias because the chance of drug exposure increases the longer a pregnancy lasts. Therefore, the drug tested may falsely be found protective against the outcome tested. The objective of the current study was to illustrate the extent of immortal time bias using a cohort study of pregnancies assessing the risk for spontaneous abortions following nonsteroidal antiinflammatory drug exposure.
2 definitions of exposure (dichotomous, exposed vs unexposed, regular Cox regression vs Cox regression with time-varying exposure). RESULTS: Significant differences were found in the risk for sponta-
neous abortions between the 2 statistical methods, both for groups and for most specific nonsteroidal antiinflammatory drugs (nonselective Cox inhibitors e hazard ratio, 0.70; 95% confidence interval, 0.61e0.94 vs hazard ratio, 1.10; 95% confidence interval, 0.99e1.22 for dichotomous vs time-varying exposure analyses, respectively). Furthermore, a significant correlation was found between the median misclassified immortal time for each drug and the extent of the bias.
STUDY DESIGN: We assembled 3 databases containing data on
CONCLUSION: Immortal time bias can easily occur in cohort studies assessing the risk for adverse pregnancy outcomes following exposure to drugs. One way to prevent such a bias is by defining exposure only from the time of exposure during follow-up onward using a timevarying exposure analysis.
spontaneous abortions, births and drug dispensions to create the present study’s cohort. The risk for spontaneous abortion was assessed using 2 statistical analysis methods that were compared for
Key words: ibuprofen, immortal time bias, miscarriage, NSAIDs, spontaneous abortions
Cite this article as: Daniel S, Koren G, Lunenfeld E, et al. Immortal time bias in drug safety cohort studies: spontaneous abortion following nonsteroidal antiinflammatory drug exposure. Am J Obstet Gynecol 2014;211:x-ex-x-ex.
E
xperimental research of drug safety is generally not feasible because of ethical issues. Therefore, most of the information about drug safety in general and teratogenicity in particular is obtained through observational studies,1 which require careful methodologic
design to obtain unbiased results on which valid conclusions can be based.2 One source of bias in retrospective cohort studies of drug safety is created by inappropriate definitions of exposure, leading to “immortal time bias.”3-5 An immortal time bias can result when,
From the Departments of Public Health (Drs Daniel and Levy), Pediatrics (Dr Daniel), and Obstetrics and Gynecology (Dr Lunenfeld), Faculty of Health Sciences, Ben-Gurion University of the Negev, Soroka Medical Center, and the Ben-Gurion Motherisk Obstetric Registry of Exposure (BeMORE) collaboration (Drs Daniel, Koren, and Levy), Beer-Sheva, Israel, and the Motherisk Program, Division of Clinical Pharmacology and Toxicology (Dr Koren), Hospital for Sick Children, University of Toronto, Toronto, ON, Canada. Received July 4, 2014; revised Sept. 10, 2014; accepted Sept. 24, 2014. The authors report no conflict of interest. The research was not funded by any company or organization. Corresponding author: Amalia Levy, MPH, PhD. [email protected] 0002-9378/$36.00 ª 2014 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.ajog.2014.09.028
during the follow-up period, patients participating in the study develop the outcome under investigation (eg, death or miscarriage) before they have had the chance to be exposed to the medication tested, simply because the duration of the follow-up period was too short, ie, the patient did not “survive” sufficient time to be able to receive the drug. As a result, patients in the study who survive longer have a higher chance of being defined as “exposed.” Figure 1 shows a hypothetical scenario in which this bias can be created. Suppose there were 2 women in the study who conceived on the same day. After exposure to nonsteroidal antiinflammatory drugs (NSAIDs) in her 15th gestational week, the first woman ended the follow-up period without an “event”
MONTH 2014 American Journal of Obstetrics & Gynecology
1.e1
Research
Gynecology
ajog.org
FIGURE 1
Illustration of immortal time bias in cohort studies of drug safety during pregnancy
Definition of exposure A, not controlling vs B, controlling for immortal time bias. Strand. Immortal time bias in drug safety cohort studies. Am J Obstet Gynecol 2014.
(a spontaneous abortion) and gave birth. In contrast, the second woman had a spontaneous abortion on the 10th gestational week, before being exposed to the drug. If a nontime-varying exposure statistical analysis (a logistic regression or Cox regression) was used, the first woman, who gave birth, would be assigned to the “exposed” group whereas the second woman, who had a spontaneous abortion at 10 gestational weeks, would be assigned to the “unexposed” group (Figure 1, A). The period of time during which the first woman was followed (20 weeks) was twice that of which the second woman was followed (10 weeks), and therefore, the first woman, who gave birth, was twice as likely to be assigned to the “exposed” group compared with the second women, who had a spontaneous abortion. Hence an association would be found such that NSAIDs had a protective effect against spontaneous abortions. However, this spurious protective effect
was caused not by the drug itself, but rather, by the study’s design. In contrast, with a time-varying exposure analysis, the exposed and unexposed groups are not defined dichotomously (as “exposed” vs “unexposed”) from the beginning of the follow-up period. Rather, exposure is redefined during follow-up, such that a woman is counted for as “exposed” only from the period of time that follows the actual exposure. Woman 1, in that case, would only be counted for as exposed from the 15th gestational week onward (Figure 1, B). Another way of addressing this bias is by presenting the Cox regression model which estimates the hazard for an event (eg, a spontaneous abortion) during follow-up, such that the risk profile (ie, exposure to NSAIDs) is compared at each event time between subjects who developed the event under study (ie, women who experienced a spontaneous abortion) and subjects
1.e2 American Journal of Obstetrics & Gynecology MONTH 2014
who stayed in the study at that time point. By defining exposure dichotomously (“exposed” vs “unexposed”) among women throughout the whole follow-up period, a misclassification of the risk profile occurs—women who were exposed to NSAIDs only in a point of time during follow-up are misclassified as “exposed” for the period of time before the actual exposure. Because women who gave birth “survive” longer in follow-up and hence have a higher chance of using NSAIDs, this misclassified period of time is more likely to occur in women who did not experience a spontaneous abortion, therefore NSAIDs would be found protective against spontaneous abortions. Hence, a varying exposure analysis enables the redefinition of exposure, such that a woman would be counted for as “unexposed” from the beginning of follow-up to the time of actual exposure. Three studies have addressed the potential bias in the association between risk factors and adverse pregnancy outcomes by referring to none time-varying exposure during pregnancy as anytime during-pregnancy exposure.6-9 The aim of the current study was to illustrate the extent of the bias using a recently published retrospective cohort study that assessed the association between exposure to NSAIDs and spontaneous abortions.10
M ATERIALS
AND
M ETHODS
Detailed descriptions of the databases and definitions of the variables used in the present study were published recently.10 We conducted a populationbased retrospective cohort study whose participants included all 15- to 45-yearold women who registered with Clalit Health Services health maintenance organization, who conceived between January 2003 and December 2009, and who were admitted for birth or were diagnosed with spontaneous abortion at Soroka Medical Center. To create the cohort for the current study, we combined 3 computerized databases in which patients are listed according to their personal identification numbers (unique numbers assigned to
Gynecology
ajog.org all citizens by the Israeli Ministry of the Interior and used throughout life). The Soroka Medical Center deliveries and hospitalization databases include information about deliveries and spontaneous abortions that occurred at the medical center. The drug dispension database of Clalit Health Services, the third database, contains electronic records of the drugs dispensed to all registered patients. Gestational age at the time of birth or spontaneous abortion was retrieved from the medical records, and was determined by an obstetrician according to the first day of the last menstrual period as reported by the pregnant woman or according to the estimated gestational age in a first trimester ultrasound imaging, as documented in the medical charts. The exposure period was from the first day of pregnancy until the 20th gestational week for births, or until the day before the diagnosis for spontaneous abortions. For missed abortions, the exposure period continued until the estimated radiologic and morphologic gestational day according to ultrasound imaging performed on admission to the emergency room. Ibuprofen, diclofenac, naproxen, etodolac, lornoxicam and nabumetone were defined as “nonselective Cox inhibitors”; celecoxib, etoricoxib and rofecoxib were defined as “Cox2 selective inhibitors.” We assessed immortal time bias by applying 2 different definitions of exposure, controlling and not controlling for the bias, and comparing the results. In the first instance, exposures were defined dichotomously such that a pregnancy was counted as “exposed” if an NSAID was dispensed from the first day of the last menstrual period to the end of the 20th gestational week or until 2 days before an abortion. Univariate and multivariate Cox regression analyses were performed to assess the risk for spontaneous abortion following exposure to an NSAID. In the second instance, as performed in our published study,10 both univariate and multivariate Cox regression with time-varying exposure analyses were performed to control for immortal time bias. A time-varying
exposure analysis redefined the exposure group at every new exposure that occurred throughout the follow-up period, such that a pregnancy would only be counted as “exposed” from the timeofthefirstreportofexposureforward. All multivariate models were adjusted for diabetes mellitus (International Statistical Classification of Diseases and Related Health Problems-9 Codes 250, 357.2), hypothyroidism (244), obesity (278,649.1), hypercoagulable (286.4, 289.81, 286.53), and inflammatory (710, 714, 720) disorders, recurrent miscarriages (646.33, 629.81), self-report of tobacco use during pregnancy (305.1, 649), the presence of an intrauterine contraceptive device (V4551, Z975), in vitro fertilization of the current pregnancy, maternal age, ethnic group (ie, Jewish vs Bedouin Muslim), and the year of admission. Furthermore, to quantify the extent of the bias according to the length of the misclassified period of immortal time, we correlated the median gestational day at first exposure to the drug and the difference between the adjusted hazard ratio, calculated by Cox regression with time-varying exposure analysis, and the hazard ratio (HR), calculated based on the dichotomous definition of exposure. The Pearson correlation coefficient was also calculated. The study was approved by an ethics committee (institutional review board) according to the declaration of Helsinki.
R ESULTS The cohort included 66,547 pregnancies that began during the study period, of which 6508 (9.9%) ended with spontaneous abortions. Data concerning gestational age at abortion were missing for 1090 (14.3%) pregnancies, and therefore, these were excluded from the analysis. The proportion of pregnancies exposed to NSAIDs during the study period and the mean gestational day at exposure are presented in Table 1. Table 2 presents the differences between both univariate and multivariate HRs using Cox regression analysis with a dichotomous definition of exposure and using Cox regression with time-varying exposure models, following exposure to
Research
nonselective Cox inhibitors, to Cox2 selective inhibitors and to most specific NSAIDs. Significant changes were found between the 2 types of exposure definitions in both univariate and multivariate analyses. As such, NSAIDs were found to be significantly protective against spontaneous abortions by defining exposure dichotomously, but that effect disappeared if the time-varying exposure analysis was performed. The difference between the HR calculated by a time-varying exposure analysis and by a dichotomous definition of exposure using Cox regression was found to be highly correlated with the median gestational age at first reported exposure for all NSAIDs (r ¼ 0.98; P < .001; Figure 2, A) and after omitting indomethacin (r ¼ 0.92; P < .001; Figure 2, B).
C OMMENT Our study demonstrates that analyzing the results of a retrospective cohort study of the risk for spontaneous abortions following exposure to NSAIDs without using a time-varying exposure statistical model creates a bias in favor of a protective effect of NSAIDs against spontaneous abortions. Similarly, O’Neill et al6 showed that by analyzing urinary tract infection (UTI) during pregnancy as a dichotomous risk factor, UTI would be seen as protective against preterm birth, as the longer a woman stayed pregnant, the higher the chance she would be diagnosed with a UTI. Likewise, Hutcheon et al7 demonstrated that by ignoring immortal time bias, gestational diabetes may be associated with a lower risk of stillbirth, because pregnancy must survive to 28 weeks for a diagnosis of gestational diabetes mellitus to be made, but about half of stillbirths occur before this time. In a recently published study, Matok et al9 have shown that the claimed decreased risk of preterm delivery after gestational use of decongestants was due to immortal time bias, arising from the improper classification of exposure during follow-up. Our findings in the current study also demonstrate a strong correlation between the length of the misclassified
MONTH 2014 American Journal of Obstetrics & Gynecology
1.e3
Research
Gynecology
ajog.org
TABLE 1
Proportion of pregnancies exposed to NSAIDs during the study period and median gestational age at exposure Drug Nonselective Cox inhibitors
Median gestational day at exposure
No. of exposed 4424 (6.75%)
31
Ibuprofen
2732 (4.17%)
40
Diclofenac
919 (1.40%)
23
Naproxen
671 (1%)
27
Etodolac
272 (0.4%)
20.5
Indomethacin
132 (0.2%)
89
71 (0.1%)
25
Cox2 selective inhibitors NSAID, nonsteroidal antiinflammatory drug.
Strand. Immortal time bias in drug safety cohort studies. Am J Obstet Gynecol 2014.
immortal time period and the change in HR following analysis with time-varying exposure model. The longer the period of misclassified time, therefore, the higher the extent of the bias. This type of bias may occur in any observational study in which exposure changes during the follow-up period. Because the follow-up period is shorter for subjects who experience an event, this misclassification of exposure would always underestimate the risk for the event under study. As described, a woman was counted for as “exposed” from the time point of first exposure until the end of follow-up (ie, the end of the 20th gestational week). This definition was based on the biologic
assumption that exposure to NSAIDs might have induced a cascade of events that ended with a spontaneous abortion. Assuming a different biologic scenario, a different definition of exposure could have been made. For example, if NSAIDs were assumed to have a transient effect, a woman could have been defined as exposed only when she was actually using the drug. Although the analysis was performed using Cox regression with time-varying exposure, other methods could have been used to overcome the bias. Assuming follow-up begins from the time of exposure, cases of spontaneous abortions occurring before exposure could have been considered as left
truncated and therefore Cox regression with a correction for left truncation could have been used.11 Moreover, immortal time bias could also be viewed as a selection bias; therefore latent variable analysis could also be performed.12 Previous studies that assessed the risk for spontaneous abortions following the use of NSAIDs measured exposure by the date of prescription rather than the dispension of the drugs.13-15 Filling a prescription, however, does not represent actual use of the drug prescribed. In the present study, exposure was defined by dispension of the drugs. Previous studies found an increasing use of overthe-counter NSAIDs that were not recorded in those studies. Thus, a bias toward the null hypothesis could occur, by misclassification of exposed women as unexposed. Because a vast majority of the pharmacies in the southern district are affiliated to Clalit Health Services, and because over-the-counter dispensions at affiliated pharmacies are recorded in our database, the extent of such a bias in our study is insignificant. Furthermore, in sensitivity analysis that was performed on the same population, the estimated proportion of unrecorded dispensions within the unexposed group was 1.3%.10 Some of the spontaneous abortions occur very early during pregnancy, before the pregnancy is recognized and these escape medical attention. Because
TABLE 2
Corrected and non-corrected risk for spontaneous abortions following exposure to NSAIDs Cox regression with time-varying exposure HR ratio (95% CI)
Cox regression HR (95% CI)
Drug
Unadjusted
Adjusteda
Unadjusted
Adjusteda
Nonselective Cox inhibitors
1.13 (1.01e1.25)
1.10 (0.99e1.22)
0.81 (0.73e0.91)
0.77 (0.69e0.86)
Ibuprofen
1.13 (0.98e1.30)
1.06 (0.93e1.22)
0.75 (0.66e0.86)
0.70 (0.61e0.94)
Diclofenac
1.21 (0.98e1.48)
1.19 (0.97e1.46)
0.98 (0.8e1.21)
0.96 (0.79e1.19)
Naproxen
1.87 (0.66e1.17)
0.97 (0.74e1.28)
0.75 (0.57e0.99)
0.75 (0.57e0.99)
Etodolac
1.26 (0.88e1.80)
1.28 (0.91e1.79)
1.23 (0.88e1.73)
1.12 (0.80e1.57)
Indomethacin
3.54 (2.20e5.71)
2.82 (1.70e4.69)
1.14 (0.68e1.89)
1.05 (0.63e1.75)
Cox2 selective inhibitors
1.97 (1.12e3.47)
1.43 (0.79e2.59)
1.52 (0.84e2.75)
1.24 (0.68e2.25)
CI, confidence interval; HR, hazard ratio; NSAID, nonsteroidal antiinflammatory drug. a
Confidence intervals.
Strand. Immortal time bias in drug safety cohort studies. Am J OBstet Gynecol 2014.
1.e4 American Journal of Obstetrics & Gynecology MONTH 2014
Gynecology
ajog.org
FIGURE 2
Correlation between gestational day at exposure and corrected and non-corrected hazard ratios
Pearson correlation between the median gestational day at first exposure and the difference between the hazard ratios calculated by Cox regression and by Cox regression with time-varying exposure for A, all NSAIDs and B, after omitting indomethacin. Strand. Immortal time bias in drug safety cohort studies. Am J Obstet Gynecol 2014.
our study only included spontaneous abortions that presented to the emergency room at Soroka Medical Center, these cases were not accounted for. Assuming some of the pregnant women have not yet attended a first prenatal visit to an obstetrician, the reported gestational age might be inaccurate. However, the whole Israeli population is medically insured and the
Israeli preventive medicine guidelines indicate a free ultrasound imaging and gestational age determination at the beginning of pregnancy to all pregnant women. Further, the electronic medical records are available for the emergency room gynecologists for all pregnant women registered with Clalit Health Services. Therefore, the extent of such a bias is estimated to be low.
Research
Data on gestational age was missing for 14.3% of pregnancies that ended with spontaneous abortions. There were no significant differences between cases of spontaneous abortions with or without data on gestational age regarding maternal characteristics and risk factors for spontaneous abortions. Moreover, a second analysis was performed after assigning mean gestational age at the time of spontaneous abortion for this group of pregnancies. This analysis yielded similar results. Gestational age at the time of birth or at the time of spontaneous abortions was defined accurately in our study, enabling an accurate definition of exposures throughout the follow-up period. In so doing, we avoided the misclassification bias that could have resulted from defining women who used NSAIDs before pregnancy or in the 2 days before abortion as exposed. Moreover, defining gestational age at the time of first exposure allows for a time-varying exposure analysis. The history of spontaneous abortions, when used as a potential confounder, was previously suspected of biasing the effect between exposure and future perinatal outcomes by influencing exposure on future pregnancies.16 The same question is raised regarding the use of the history of recurrent spontaneous abortions as a potential confounder in our models. Furthermore, findings are inconsistent regarding the association between the use of intrauterine devices and spontaneous abortions, therefore using this variable as a potential confounder might be inappropriate. However, results were not changed after omitting these 2 variables from the analysis. Immortal time bias can easily occur in cohort studies assessing the risk for adverse pregnancy outcomes following exposure to drugs. One way to prevent such a bias is by defining exposure only from the time of exposure during follow-up onward using a time-varying exposure analysis. Thus to produce unbiased results in studies assessing drug safety during pregnancy, the definition of exposure must be carefully considered and the data obtained must
MONTH 2014 American Journal of Obstetrics & Gynecology
1.e5
Research
Gynecology
be analyzed with the appropriate statistical method. REFERENCES 1. Vandenbroucke JP. Benefits and harms of drug treatments. BMJ 2004;329:2-3. 2. Sanderson S, Tatt ID, Higgins JP. Tools for assessing quality and susceptibility to bias in observational studies in epidemiology: a systematic review and annotated bibliography. Int J Epidemiol 2007;36:666-76. 3. Suissa S. Immortal time bias in pharmacoepidemiology. Am J Epidemiol 2008;167:492-9. 4. Levesque LE, Hanley JA, Kezouh A, Suissa S. Problem of immortal time bias in cohort studies: example using statins for preventing progression of diabetes. BMJ 2010;340:b5087. 5. Liu CJ, Hu YW. Immortal time bias in retrospective analysis: is there a survival benefit in patients with glioblastoma who received prolonged treatment of adjuvant valganciclovir? Int J Cancer 2014;135:250-1.
ajog.org 6. O’Neill M, Hertz-Picciotto I, Pastore LM, Weatherley BD. Have studies of urinary tract infection and preterm delivery used the most appropriate methods? Paediatr Perinat Epidemiol 2003;17:226-33. 7. Hutcheon JA, Kuret V, Joseph KS, Sabr Y, Lim K. Immortal time bias in the study of stillbirth risk factors: the example of gestational diabetes. Epidemiology 2013;24:787-90. 8. Hertz-Picciotto I, Pastore LM, Beaumont JJ. Timing and patterns of exposures during pregnancy and their implications for study methods. Am J Epidemiol 1996;143:597-607. 9. Matok I, Azoulay L, Yin H, Suissa S. Immortal time bias in observational studies of drug effects in pregnancy. Birth Defects Res A Clin Mol Teratol 2014;100:658-62. 10. Daniel S, Koren G, Lunenfeld E, Bilenko N, Ratzon R, Levy A. Fetal exposure to nonsteroidal antiinflammatory drugs and spontaneous abortions. CMAJ 2014;186:E177-82. 11. Cain KC, Harlow SD, Little RJ, et al. Bias due to left truncation and left censoring in
1.e6 American Journal of Obstetrics & Gynecology MONTH 2014
longitudinal studies of developmental and disease processes. Am J Epidemiol 2011;173: 1078-84. 12. Burnham AJ, MacGregor JF, Viveros R. Latent variable multivariate regression modeling. Chemomet Intellig Lab Syst 1999;48:167-80. 13. Nakhai-Pour HR, Broy P, Sheehy O, Berard A. Use of nonaspirin nonsteroidal antiinflammatory drugs during pregnancy and the risk of spontaneous abortion. CMAJ 2011;183: 1713-20. 14. Nielsen GL, Skriver MV, Pedersen L, Sorensen HT. Danish group reanalyses miscarriage in NSAID users. BMJ 2004;328:109. 15. Nielsen GL, Sorensen HT, Larsen H, Pedersen L. Risk of adverse birth outcome and miscarriage in pregnant users of nonsteroidal antiinflammatory drugs: population based observational study and case-control study. BMJ 2001;322:266-70. 16. Weinberg CR. Toward a clearer definition of confounding. Am J Epidemiol 1993;137: 1-8.
