Understanding the extent to which pregnancy complications such as gestational diabetes mellitus or preeclampsia increase a woman’s risk of stillbirth is important for clinical management. Information on stillbirth risk is used to support patient counseling and to develop clinical guidelines on the appropriate degree of antenatal surveillance if the risk of stillbirth is increased. Estimates of the relative risk of stillbirth associated with various pregnancy complications are typically obtained from large population-based cohort studies in which the risk of stillbirth among deliveries with the pregnancy complication is compared with the risk among deliveries without the complication (after adjusting for relevant confounders).1–6 As mandatory registration for stillbirths begins at or around 20 weeks’ gestation in many jurisdictions (eg, most US States, Canada, Australia, and New Zealand),7–9 the analyses often include all births delivered at or beyond 20 weeks’ gestation.1–4
However, a number of pregnancy-related complications are not diagnosed until mid-to-late pregnancy. For example, gestational diabetes is diagnosed after routine screening at 24–28 weeks,10–12 whereas the majority of preeclampsia cases develop at or near term.13 We hypothesized that these differences between the start of the cohort follow-up period (ie, 20 weeks’ gestation) and the time at which exposure status is established (ie, diagnosis of gestational diabetes or preeclampsia) may create the potential for immortal time bias. Using the example of gestational diabetes, our goals in this report were (1) to describe the theoretical potential for immortal time bias in the specific context of the study of risk factors for stillbirth and (2) to quantify the magnitude of this bias in order to establish its potential impact on substantive conclusions.
We used the conditions for immortal time bias outlined by Lévesque and colleagues14 to assess the theoretical potential for immortal time in the study of gestational diabetes and stillbirth risk. These conditions include (1) treatment status determined after the start of follow-up or defined using follow-up time, (2) different start-up time for the treated and untreated group relative to the start of diagnosis, (3) treatment groups identified hierarchically (one group before the other), (4) subjects excluded on the basis of treatment identified during follow-up, and (5) use of a time-fixed analysis.
We quantified the impact of immortal time bias using 2006 United States live birth-infant death and fetal death files, available from the Centers for Disease Control and Prevention Division of Vital Statistics (www.cdc.gov/nchs/data_access/Vitalstatsonline.htm). We restricted our analyses to births registered using the 2003 revision of the birth certificate in order to differentiate gestational diabetes from type 1 and type 2 diabetes. Multiple births and pregnancies with type 1 or type 2 diabetes were excluded. In accordance with Centers for Disease Control and Prevention Division of Vital Statistics practice,8 stillbirth analyses were restricted to births ≥20 weeks of gestation. We used the clinical estimate of gestational age, as this has been shown to be more accurate than estimates based on last menstrual period.15 Relative risk of stillbirth associated with gestational diabetes was calculated as the risk of stillbirth among all births with this disease divided by the risk of stillbirth among all births without. We also estimated the relative risk of gestational diabetes in a cohort restricted to births >28 weeks, that is, after the recommended screening window for gestational diabetes.10,11
RESULTS AND DISCUSSION
Theoretical Potential for Immortal Time Bias
Immortal time refers to “a period of follow-up during which, by design, death or the study outcome cannot occur.”14 Although the potential for immortal time bias is well-recognized in areas such as transplantation research, nephrology, and pharmacoepidemiology,14,16,17 it has received little attention in the study of stillbirth. The conventional study of stillbirth in gestational diabetes is susceptible to immortal time bias for a number of reasons. First, although most study cohorts include all births from the start of the jurisdiction’s mandatory birth registration period (eg, 20 weeks in most US states, Canada, Australia, and New Zealand),7–9 the diagnosis of gestational diabetes does not usually occur until 24–28 weeks.12 This creates a situation where pregnancies must “survive” until 24–28 weeks in order to be screened for gestational diabetes. As illustrated in Figure 1, the time before the diagnosis becomes “immortal,” as, by definition, stillbirths to women with gestational diabetes cannot be identified before 24–28 weeks.
The difference in the start of follow-up between women with and without gestational diabetes means that for women without this complication, the calculation of stillbirth risk will include all stillbirths from 20 weeks of pregnancy in its numerator, whereas for women with gestational diabetes, the numerator will be primarily restricted to stillbirths ≥24–28 weeks (recognizing that a small fraction of women with risk factors may be screened and diagnosed with gestational diabetes before 24 weeks).12 This would be expected to inflate the risk of stillbirth in the unexposed cohort, leading to an attenuation of the true relative risk of stillbirth associated with gestational diabetes. Using the individual rather than person-time as the unit of analysis (ie, the pregnancy rather than “gestational days at-risk”) prevents this difference in the at-risk periods from being taken into account.
Quantifying the Magnitude of the Bias
After excluding 67,735 multiple births and 13,883 women with type 1 or type 2 diabetes, 2,001,749 women with known gestational diabetes status were retained for analysis in our cohort. A total of 76,669 women had a documented diagnosis of gestational diabetes, corresponding to a risk of 3.8% in the population.
When analyzing all births ≥20 weeks, we found gestational diabetes to have a protective effect on stillbirth risk (relative risk = 0.88 [95% confidence interval (CI) = 0.79–0.99]) (Table). This result is comparable to results previously reported in the literature based on similar analytic approaches. A study of 120,604 pregnancies in Ontario, Canada, found that gestational diabetes was strongly protective against stillbirth (odds ratio [OR] = 0.33 [95% CI = 0.12–0.71]),18 whereas another hospital-based study from Israel showed similar protection for diet-controlled gestational diabetes (OR = 0.5 [95% CI = 0.4–0.7]).19 Another study also reported an OR of 0.7 (95% CI = 0.6–0.9) associated with gestational diabetes, which the authors attributed to the increased surveillance and active management received by women with the condition.20
However, when analyses were restricted to births at 28 weeks and later, the apparent protective effect of gestational diabetes was eliminated, and the condition became associated with an increased risk of stillbirth (relative risk = 1.25 [95% CI = 1.11–1.41]) (Table). Plotting the gestational age distribution of stillbirths in the unexposed cohort (Figure 2) helps to explain this result. Nearly 30% of stillbirths in the cohort (2658/9165) occurred at 20–23 weeks before the start of routine diabetes screening. These were likely second-trimester losses or pregnancy terminations (following detection of prenatal diagnosis of a congenital anomaly or for other reasons).21 Stillbirths at this gestational age were attributed to women without gestational diabetes, as, by design, the diagnosis of gestational diabetes was not (usually) made until later in gestation.
Restricting analyses to deliveries at 28 weeks and beyond reduces the potential for immortal time bias by eliminating the difference in follow-up period between women with and without gestational diabetes. The reversal of the relative risk reported when analyses were restricted to >28 weeks suggests that the protective effect of gestational diabetes on stillbirth risk obtained in many studies represents immortal time bias, rather than a true protective effect resulting from increased antenatal monitoring and surveillance. This conclusion is supported by the findings of a recent study of stillbirth risk in gestational diabetes at term, in which a relative risk of 1.35 (95% CI = 1.2–1.5) was found in a cohort restricted to pregnancies at or beyond 36 weeks.22 Similarly, a report based on the Swedish Medical Birth Register23 (in which stillbirth registration begins at 28 weeks) found a crude OR of 1.18 (95% CI = 0.87–1.60).
Although these relative risks highlight the role of immortal time bias, details on the specific timing of diabetes diagnosis and adjustments for confounders are needed to generate an unbiased estimate of the association between gestational diabetes and stillbirth. Our study also highlights the importance of carefully considering the outcome definitions used in the study of stillbirth. Definitions of stillbirth other than those used for legal birth registration purposes (eg, definitions that exclude pregnancy terminations, or that contain a gestational age–based criteria such as >28 weeks) may provide more meaningful answers, depending on the specific question under study.
Several approaches have been proposed to account for immortal time.14 In this study, we used a “survivors- only” analysis, restricting the study population to women who remained pregnant beyond 28 weeks. Other potential approaches include the use of a time-matched nested case-control study and the use of a time-dependent analysis. If the amount of immortal time is known, formulas are available that quantify the magnitude of the bias.17 However, these latter approaches are challenging in the study of gestational diabetes. Whereas large population databases are typically required to obtain sufficient statistical precision to study rare outcomes such as stillbirth, they do not usually have information on the age and date of gestational diabetes screening, which is required to match follow-up time from the time of diagnosis. More sophisticated approaches, such as linkage of laboratory records with perinatal outcome data, may be needed to adequately study such questions. In the study of preeclampsia, for example, time of diagnosis could potentially be identified through use of antenatal hospitalization records (as women with preeclampsia are typically admitted to hospital following initial diagnosis) and linked with outcome data contained in delivery records.
We hypothesized that studies of stillbirth risk in gestational diabetes may be susceptible to immortal time bias because the follow-up of women with gestational diabetes begins at a later gestational age than that of women without gestational diabetes. We demonstrated that this bias was large enough to affect the risk of stillbirth associated with gestational diabetes. The implications of our findings extend beyond studies of gestational diabetes. Immortal time bias may be present in any studies of stillbirth in association with pregnancy complications that develop in mid-to-late pregnancy, including preeclampsia, gestational hypertension, intrahepatic cholestasis of pregnancy, and eclampsia, where the bias may attenuate the true relative risk. Analytic approaches that correctly account for the immortal time before the time of diagnosis are critical for ensuring an unbiased understanding of the extent to which pregnancy complications increase a woman’s risk of stillbirth.
1. Ahmad AS, Samuelsen SO. Hypertensive disorders in pregnancy and fetal death at different gestational lengths: a population study of 2 121 371 pregnancies. BJOG. 2012;119:1521–1528
2. Ananth CV, Savitz DA, Bowes WA Jr. Hypertensive disorders of pregnancy and stillbirth in North Carolina, 1988 to 1991. Acta Obstet Gynecol Scand. 1995;74:788–793
3. Liu S, Joseph KS, Liston RM, et al.Maternal Health Study Group of Canadian Perinatal Surveillance System (Public Health Agency of Canada). Incidence, risk factors, and associated complications of eclampsia. Obstet Gynecol. 2011;118:987–994
4. Simpson LL. Maternal medical disease: risk of antepartum fetal death. Semin Perinatol. 2002;26:42–50
5. Smulian JC, Ananth CV, Vintzileos AM, Scorza WE, Knuppel RA. Fetal deaths in the United States. Influence of high-risk conditions and implications for management. Obstet Gynecol. 2002;100:1183–1189
6. Basso O, Rasmussen S, Weinberg CR, Wilcox AJ, Irgens LM, Skjaerven R. Trends in fetal and infant survival following preeclampsia. JAMA. 2006;296:1357–1362
7. Public. Health Agency of Canada. Canadian Perinatal Health Report, 2008 Edition. 2008 Ottawa: Minister of Public Works and Government Services Canada
8. Barfield W, Martin J, Hoyert D. Racial/Ethnic trends in fetal mortality- United States, 1990–2000. MMWR Morb Mortal Wkly Rep. 2004;53:529–532
9. Perinatal and Maternal Mortality Review Committee (PMMRC). Sixth Annual Report of the Perinatal and Maternal Mortality Review Committee. Reporting Mortality 2010. 2012 Wellington, New Zealand: PMMRC
10. Berger H, Crane J, Farine D, et al.Maternal-Fetal Medicine Committee; Executive and Coundil fo the Society of Obstetricians and Gynaecologists of Canada. Screening for gestational diabetes mellitus. J Obstet Gynaecol Can. 2002;24:894–912
11. ACOG Practice Bulletin. . Clinical management guidelines for obstetrician-gynecologists. Number 30, September 2001 (replaces Technical Bulletin Number 200, December 1994). Gestational diabetes. Obstet Gynecol. 2001;98:525–538
12. Meltzer SJ, Snyder J, Penrod JR, Nudi M, Morin L. Gestational diabetes mellitus screening and diagnosis: a prospective randomised controlled trial comparing costs of one-step and two-step methods. BJOG. 2010;117:407–415
13. Sibai BM. Diagnosis and management of gestational hypertension and preeclampsia. Obstet Gynecol. 2003;102:181–192
14. Lévesque 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
15. Joseph KS, Kramer MS, Marcoux S, et al. Determinants of preterm birth rates in Canada from 1981 through 1983 and from 1992 through 1994. N Engl J Med. 1998;339:1434–1439
16. Shariff SZ, Cuerden MS, Jain AK, Garg AX. The secret of immortal time bias in epidemiologic studies. J Am Soc Nephrol. 2008;19:841–843
17. Suissa S. Immortal time bias in pharmaco-epidemiology. Am J Epidemiol. 2008;167:492–499
18. Peticca P, Keely EJ, Walker MC, Yang Q, Bottomley J. Pregnancy outcomes in diabetes subtypes: how do they compare? A province-based study of Ontario, 2005-2006. J Obstet Gynaecol Can. 2009;31:487–496
19. Karmon A, Levy A, Holcberg G, Wiznitzer A, Mazor M, Sheiner E. Decreased perinatal mortality among women with diet-controlled gestational diabetes mellitus. Int J Gynaecol Obstet. 2009;104:199–202
20. Ohana O, Holcberg G, Sergienko R, Sheiner E. Risk factors for intrauterine fetal death (1988-2009). J Matern Fetal Neonatal Med. 2011;24:1079–1083
21. Liu S, Joseph KS, Kramer MS, et al.Fetal and Infant Health Study Group of the Canadian Perinatal Surveillance System. Relationship of prenatal diagnosis and pregnancy termination to overall infant mortality in Canada. JAMA. 2002;287:1561–1567
22. Rosenstein MG, Cheng YW, Snowden JM, Nicholson JM, Doss AE, Caughey AB. The risk of stillbirth and infant death stratified by gestational age in women with gestational diabetes. Am J Obstet Gynecol. 2012;206:309.e1–309.e7
Copyright © 2013 Wolters Kluwer Health, Inc. All rights reserved.
23. Fadl HE, Ostlund IK, Magnuson AF, Hanson US. Maternal and neonatal outcomes and time trends of gestational diabetes mellitus in Sweden from 1991 to 2003. Diabet Med. 2010;27:436–441