Table 4 Potential sources of bias in vital event data in population-based cohort studies
Source of bias | Potential bias on mortality estimates | Indicators to assess presence of bias | Approaches to reduce bias in cohort studies |
|---|---|---|---|
Pregnancy and birth outcomes | |||
Missed pregnancies: Certain pregnant women, potentially eligible for a study, are missed by pregnancy surveillance | - Underestimate mortality Reason: Women missed by the survey may be more likely to experience stillbirth or neonatal/infant mortality | - Comparison of estimates to DHS, HDSS, or other data - Study protocols for identifying pregnancies through multiple avenues | - Conduct census or utilize vital registration and health care information systems - Enroll participants through multiple sources (e.g., ANC sites, health facilities, at home) - Conduct follow-up surveys |
Birth under-reporting: Differential enrollment, loss to follow-up, or identification of the birth outcome for pregnant women who deliver at home vs. in a health facility | - Underestimate or overestimate mortality - Underestimate very early/early neonatal mortality Reason: Infants born at home may be less likely to be identified and included. They may be more likely to experience mortality due to reduced access to health care or lower socioeconomic status. In resource-limited settings with low facility delivery rates, women with a complicated pregnancy or delivery may be more likely to seek care, especially last minute, creating a spurious association between care seeking or facility delivery and adverse outcomes, such as stillbirth or early neonatal mortality [36] | - Facility delivery rate - Study protocols for identifying births at multiple sites and home - Comparison of characteristics of women delivering at home vs. facility | - Use community-based enrollment in settings without high facility delivery rate - Enroll participants through multiple sources - Ask participants about birth plan - Conduct early/frequent visits on day of birth if possible |
Selection bias associated with loss to follow-up/out-migration, in-migration, or censoring: Certain pregnant women move from the home where they were originally enrolled and cannot be reached to obtain vital status and date of death if the infant died Conversely, certain pregnant women move into a study area and may be included in birth and mortality estimates Women still pregnant at the time of the study end date have an unknown birth outcome | - Underestimate mortality - Underestimate very early/early neonatal mortality Reason: Loss to follow-up due to out-migration in pregnancy is a common cause of selection bias. Women that are lost may differ in ways that impact mortality estimates. E.g., in some communities, women return to their parental homes late in pregnancy for delivery and the postnatal period. The characteristics of these women may differ from those who do not practice this custom in ways that affect birth and mortality outcomes Reason: Similarly, women who enter the study area due to in-migration may differ in ways that impact mortality estimates Reason: Some studies follow infants to a fixed age until the last infant has been followed, but others set an end date for the study. Deaths might occur among infants of censored women. However, this may not be associated with risk for mortality and therefore is not likely a significant source of bias | - Comparison of characteristics of women who were successfully followed vs. those who out-migrated or in-migrated - Comparison of characteristics of women censored vs. not censored | - Use comprehensive participant tracking (e.g., via digital technology or cell phone) - Conduct follow-up surveys for basic mortality information - Include in-migration to offset out-migration for women moving to parental homes to deliver in the study area - Carefully consider the potential biases associated with an open vs. closed cohort in the specific study context |
Inclusion/exclusion criteria: Certain participants or special populations are excluded for the purpose of achieving the study’s primary aim | - Underestimate or overestimate mortality - Generate unbiased estimates of mortality for a special, non-representative population Reason: Exclusion of certain participants could introduce selection bias if their mortality risk differs from those who are included. This includes exclusion of certain participants not reached within some time frame after delivery (e.g., < 72 h). If a certain special population is the focus of a study, mortality estimates may be unaffected by selection bias, but still non-representative of the underlying population | - Review evidence on mortality risk in included vs. excluded populations - Compare characteristics in those included vs. excluded | - Limit exclusion criteria if possible - Collect key characteristics and vital event data on all participates, to allow for comparisons, even if a subset of participants are excluded from the primary study and/or analysis - Assess and report the potential generalizability of mortality estimates from special populations and how they may differ from the underlying population |
Mortality outcomes | |||
Misclassifying very early neonatal deaths: Differential misclassification of neonatal deaths as stillbirths or vice versa | - Underestimate or overestimate very early neonatal mortality Reason: Very early deaths occurring at home may be less likely to be correctly classified and reported (may be misclassified as stillbirths or vice versa). There are many reason this happens, including absence of skilled birth attendance at delivery and stigma or other reluctance related to reporting neonatal deaths | - Compare estimates and ratio of stillbirths to neonatal deaths to DHS, HDSS, or other data | - Conduct early/frequent visits on day of birth if possible - Use verbal autopsy surveys to classify antepartum stillbirths, intrapartum stillbirths, and early neonatal deaths |
Under-reporting of neonatal or infant deaths: Neonatal or infant deaths are not reported or hidden from data collectors and/or health local authorities | - Underestimate mortality - Underestimate very early neonatal mortality Reason: In many settings there is stigma associated with reporting neonatal or infant deaths that could result in substantial under-reporting of birth outcomes. Under-reporting may also be associated with specific participant characteristics, particularly those indicative of disadvantaged populations | - Examine mortality pattern and rates (e.g., across first day, week, month, and year of life) - Compare estimates to DHS, HDSS, or other data | - Conduct comprehensive follow-up and participant tracking to reduce number of missed mortality outcomes - Describe cultural factors associated with potential bias and design customized strategies to reduce their impact |
Selection bias associated with loss to follow-up/out-migration, in-migration, or censoring: Certain mothers and infants move from the home where they were originally enrolled and cannot be reached to obtain vital status and date of death if the infant died Conversely, certain mothers and infants move into a study area and may be included in mortality estimates Infants not having reached 28 days or 1 year of life (or other benchmark) at the time of the study end date have an unknown vital status at that time point | - Underestimate mortality - Underestimate very early neonatal mortality Reason: Loss to follow-up due to out-migration is often the most common cause of selection bias in infancy. Mothers and infants that are lost may differ in ways that impact mortality estimates. Live birth cohorts that allow enrollment of newborns beyond the day of delivery may introduce substantial bias, leading to underestimated very early and early neonatal mortality Reason: Similarly, mothers and infants who enter the study area due to in-migration may differ in ways that impact mortality estimates Reason: Some studies follow infants to a fixed age until the last infant has been followed but others set an end date for the study. Deaths might occur among censored infants. However, this may not be associated with risk for mortality and therefore is not likely a significant source of bias | - Compare maternal or infant characteristics of participants successfully followed vs. those who out-migrated - Compare numbers of deaths and LTF by age category - Comparison of characteristics of infants censored vs. not censored | - Use comprehensive participant tracking (e.g., via cell phone) - Conduct follow-up surveys for basic mortality information - Exclude mothers and infants who are censored from mortality estimates - Utilize survival analysis to include and appropriately apportion time contributed by censored infants - Carefully consider the risks of missed birth outcomes and out-migration, and, especially, in-migration in the neonatal period, among live birth cohorts, which do not have the benefits of pregnancy enrollment |
Date heaping and other recall error: Dates of death reported by mother, parents, or data collectors are sometimes rounded up to the 15th or 30th of the month when being recorded during data collection due to recall bias | - Underestimate of early or infant mortality Reason: Rounding up of dates of death to 15th or 30th of the month could shift deaths above age specific cut-offs for mortality estimates, such as 28 days or 1 year, reducing rates for early mortality categories | - Create histographs to explore presence of date heaping | - Use locally appropriate methods to improve date recall (e.g., event calendars) - Increase frequency of follow-up visits - Consider other methods to improve recall (e.g., diary for important dates) - Apply analytical techniques to adjust date heaping |