3Sample Design—Consideration of Multiple Cohorts

Publication Details

This chapter begins with information on sample design that was provided to workshop participants in advance via Kwan et al. (2013, pp. 4-7), followed by the questions panelists were asked to address on this topic.1 The third section of the chapter provides the individual viewpoints of the four panelists, followed by open discussion.


Geographical Area Sampling

The original National Children's Study (NCS) plan called for 105 primary sampling units (PSUs) consisting of whole counties or groups of counties, with each PSU expected to generate about 2,000 births over the recruitment period based on 1999–2002 birth statistics, and with stratification by county size, percent black, Hispanic, Asian, and low-weight births. After sampling of segments (groups of census blocks), and door-to-door household screening, the PSUs were expected to generate about 1,000 births for inclusion in the NCS over a 4-year recruitment period.2

The NCS has abandoned the use of house-to-house screening methods due to projections based on the unexpectedly high and unsustainable resources that were expended in the initial phase of the Vanguard Study. The NCS is still planning to base the Main Study on a probability-based sample with current plans to start with a probability-based geographic area sample, though other probability-based options are under consideration. The optimal balance between number of PSUs, number of births per PSU, and environmental variety is not yet known. Different contract teams are working on scenarios, but this will not be resolved prior to the workshop and therefore will not be presented or discussed. The important point for the remaining discussion is that the NCS is currently planning an area probability sample of PSUs that is expected to generate 100,000 live births for participation in the NCS.

The Birth Cohort consists of births collected from a sample of hospitals and birthing centers and a subsample of women giving birth at those selected centers. This is tentatively planned to be about 45,000 participants of the overall sample. A 2-year initial recruitment period is proposed.

Within the current proposal, for each selected geographic PSU, a list of all hospitals and birthing centers will be prepared as a sampling frame for the birth cohort. Based on data from 2006, roughly 98 percent of all births in the United States take place at hospitals or birthing centers. A random sample of hospitals and birthing centers will be selected, with probability proportional to the number of births, and recruited to participate in the study. All women who give birth at the selected hospitals and birthing centers during specific times within the planned 2-year initial recruitment period will be eligible to be sampled while at the hospital, regardless of whether they live within the selected PSU or not. A systematic sample of women giving birth will be selected.

The NCS has documented multiple studies that recruit new mothers (and fathers) in the hospital and some that collect specimens (the Fragile Families Study is one of these). The acceptance rate is high and in some cases over 90 percent. The NCS has several strategies for collecting the relatively few specimens of interest (maternal blood and urine, cord blood, placenta, and perhaps an infant second dried blood spot following newborn screening), including collection from all sampled women during the recruitment windows and then discarding specimens from women who do not consent. The NCS is also piloting a few methods in the Provider-Based Sampling Vanguard sites to give some empirical data on acceptance, logistics, and costs. It is also possible in the birth cohort to attempt to collect medical records not only from the hospital or birthing center, but also from the sample member's prenatal care provider (if any).

The birth cohort will be a nationally representative sample of births in the United States. It can include stillbirths as well as live births.

While the recruitment of a relatively unbiased sample at acceptable cost is attractive, a knowledge gap that needs to be addressed is prenatal exposure data. An essential question for the NCS is what is the scope and integrity of data that can be captured indirectly that would inform the prenatal history for a child recruited in the birth cohort.

The birth cohort mothers can be followed over time and subsequent children added to the sample. This provides an opportunity to include a sibling cohort and to collect both preconception and prenatal measures for some births. Information from the Fragile Families Study indicates that about 4.5 percent of women who have a child have another within 18 months, and 25 percent have another within 3 years.

Applying a similar analysis to the NCS would project the following scenario. For a sample of 45,000 births recruited over a 2-year period, allowing for about 350 stillbirths (1 in 150 of pregnancies) and, say, 4,150 attriters after the hospital interview, would leave 40,500 in the sample (assuming that women at the hospital would be oversampled to allow for refusals to participate). Of these 40,500, about 10,000 would be expected to have another child within a 3-year follow-up recruitment period. The subsequently born children could have prospective documentation of preconception and prenatal exposures.

The Prenatal Cohort is a sample of the prenatal care providers that are linked to the sampled hospitals or birthing centers from the birth cohort and a subsample of women who visit a prenatal care provider and expect to deliver at one of the selected hospitals or birthing centers. This is tentatively planned as about 45,000 births.

The primary purpose of a prenatal cohort is to obtain prospectively collected exposure data. There is some evidence that exposures within the first 8 weeks of pregnancy are the most critical. However, a prenatal sample enrolled from community care providers is unlikely to recruit very many women this early in their pregnancy. It has been estimated that at 8 weeks, only about 10 percent of pregnant women may have sought prenatal care, and these are likely to be those seeking fertility assistance, or those who are trying to get pregnant or have preexisting medical conditions and are monitoring. The prenatal cohort could, however, provide a reasonable sample of women in their third trimester of pregnancy.

The NCS will work with hospitals and birthing centers selected into the sample to identify the prenatal care providers including clinics, family practitioners, midwives, etc. that refer women to the hospitals and birthing centers. A sample of these prenatal care providers will be selected, using probability proportional to number of births. All women who are expected to give birth in one of the selected hospitals or birthing centers are eligible to be selected into the sample.

The NCS is currently exploring several options in the field with the Provider-Based Sample Vanguard sites, including working with county medical societies and other professional societies and licensing bureaus, as well as using birth records (where available) to construct a list of prenatal care providers in the PSU. Birth records are available in some sites but not all sites. The logistics and resources required to prepare sampling frames of prenatal care providers as documented in the NCS Vanguard Study experience combined with information from other studies and the desire to work with selected hospitals and birthing centers for collecting birth information led to the approach described above.

The NCS's Vanguard Pilot Study data indicate that the proportion of women that providers inform about the study and that actually enroll is between 35 and 50 percent. In other words, for the most efficient providers, about 1 in 2 women enroll, and for others it is about 1 in 3. Prenatal cohort mothers can also be followed over time and subsequent children added to the sample. This provides an opportunity to include preconception measures and additional prenatal measures for some births as described in the birth cohort example. See above for further discussion.

It is not clear what population a prenatal care cohort would represent on a probability basis. If the prenatal cohort were limited to women visiting the sample of prenatal care providers within a specified time period who were in their third trimester, then the enrolled population would likely cover close to the entire population of pregnant women who receive prenatal care. However, the NCS would not be able to obtain measures of exposures earlier in the pregnancy except to the extent medical records contained relevant information. If the cohort were extended to include all women visiting the sample of prenatal care providers within a specified time period, then its representation of women in their first or second trimesters would be incomplete and could be biased given that women vary in the stage of pregnancy at which they seek prenatal care. In either case, the prenatal cohort is likely to have the measures most uniform for women in the third trimester. Although the proportion of women who receive prenatal care is relatively high, there are women who for multiple reasons do not receive prenatal care. Women who do not receive prenatal care could only be enrolled into the Study at a hospital or birthing center.

Preconception and Special-Purpose Cohorts are currently undefined, but may include a preconception group or a special-purpose group. About 10,000 sample cases are reserved for these, as yet unspecified, cohorts.

Neither the birth nor the prenatal care provider cohort can obtain information on preconception exposures for first-order births (except for what may be available in medical records for some sample members), nor will they necessarily include geographic areas of special interest (e.g., environmental “hot spots,” such as areas where natural gas fracking is under way). The originally proposed design of household screening was intended to generate a preconception sample but proved infeasible on grounds of excessive costs and time for recruitment. Following women in one or both cohorts will generate samples of subsequent births that occur within a window (of 2-3 years) that will provide both preconception and prenatal exposures. In addition, it might be possible to consider some special cohorts that could be sampled purposively or on a probability basis. For example, it might be useful to identify a small number of known environmental “hot spots” and seek to enroll all or a large sample of women of child bearing age at these locations. This cohort will be a convenience sample in addition to the other enrolled participants and is not intended as a topic of discussion for the workshop.



What should be the criteria for the cohort allocation decision and what evidence is available to support an assessment of each criterion? Examples include the following:


Recruitment costs, which include the costs of constructing the frame and the relative costs and efficiency of enrolling a participant.


Generalizability. What population is being represented?


Extent of exposures and other information that can be gathered. By definition, a birth cohort will have more limited data on prenatal exposures than a prenatal cohort, while a prenatal cohort will have less information on prenatal exposures (and much less information on preconception exposures) than the cohort of subsequent births to already enrolled mothers or a separate preconception cohort.


What should be the allocation of sample cases among the various cohorts? Assume that 10 percent of the sample is reserved for preconception and special studies; then, the allocation involves the remaining 90,000.


One option is the current proposal, which is about a 50-50 split or 45,000 participants in each.


Another option is something like an 80-20 split allocated between birth and pregnancy, with the pregnancy sample used to form the basis for imputing prenatal exposures (after using medical records for the mothers to get as much prenatal information as possible).


Yet another option is like an 80-20 split allocated between pregnancy and birth, with the birth sample used to form the basis for providing generalizability to the data analysis.


One extreme could be the entire initial enrollment allocated to the birth cohort, with studies of prenatal and preconception exposures using primarily the 25 percent cohort of subsequent births to originally enrolled mothers.


At the other extreme, most of the sample could be allocated to the prenatal cohort with a small birth sample consisting of women who did not receive any prenatal care and are enrolled at the hospital.


The moderator for this panel was Barbara Carlson (Mathematica Policy Research) and panelists were Michael Bracken (Center for Perinatal, Pediatric, and Environmental Epidemiology, School of Public Health, Yale University), Naihua Duan (Division of Biostatistics, Department of Psychiatry, Columbia University Medical Center, Columbia University), Irwin Garfinkel (School of Social Work, Columbia University), and Nancy Reichman (Robert Wood Johnson Medical School, University of Medicine and Dentistry of New Jersey and Department of Economics, Princeton University). Carlson introduced the session saying that it was set up as a debate with each panelist stating his or her own views about the questions. The individual viewpoints of the four panelists, discussion among the panel, and open discussion are summarized below.

First Viewpoint

The first viewpoint was presented by Irwin Garfinkel (Columbia University). Garfinkel reminded the audience that the NCS is likely to be the most important birth cohort study in the United States for several decades, and, in retrospect, the protracted struggle and very expensive pretesting over how to conduct it is not surprising. Reconciliation of different objectives and different disciplinary traditions will be important. In particular, it is not obvious how to collect prenatal and preconception exposure data from a population-based probability sample of births at reasonable cost. He characterized the recent evolution of the study design as very positive and on the brink of reconciling conflicting objectives.

Garfinkel observed that he and fellow panelist Bracken were on opposite ends of the spectrum about the appropriate balance between the birth and prenatal cohorts, but they have a fundamental area of agreement: Probability sampling is essential to the quality of the NCS.

He noted that in Kwan et al. (2013), the proposal was to enroll 45,000 mothers and children at birth from separate probability samples of hospitals and prenatal providers. He characterized the 50-50 split between hospitals and prenatal clinics as a huge step forward from sampling only from prenatal clinics in terms of cost and scientific value. He argued that neither the prenatal nor a birth cohort would produce excellent prenatal data, and that only sibling data could produce them at reasonable cost. He said if NCS used an almost 100 percent birth cohort that enrolled subsequent sibling births, this would save even more costs than the current 50-50 split and would immeasurably increase the scientific value of the study when it is completed 21 years from now.

Garfinkel said he and Bracken also agree that collecting prenatal data is a critical component of the NCS and that his own understanding was reinforced by the discussion of the first panel (see Chapter 2) that first-trimester data are the most valuable part of prenatal data for many questions of interest and, for some questions, preconception data may be equally important. He stated that data produced by the prenatal sample would fail on these grounds because first-trimester data would not be collected from a sufficient percentage of sampled women.

Garfinkel offered a potential design and the data needed to implement it. He suggested an area probability design with a sample of hospitals in each selected primary sampling unit (PSU), and a sample of births within each selected hospital. The ultimate probability of selection would be known. He assumed that 60,000 mothers and their children would be enrolled in hospitals at birth, and that placentas and cord blood would be collected, as would breast milk. For these selected women, all subsequent sibling births over the course of the 21 years of the study would be enrolled in the study. Women would be appropriately monitored to determine when they become pregnant. He stated that this design would provide nearly as large a sample of children with prenatal data as the proposed 50-50 design. Further, if first births were oversampled in the birth cohort, the sample of siblings with prenatal data would be as large as a pure prenatal sample. He said these sibling prenatal data would be superior to the prenatal data provided by the prenatal sample because the sibling prenatal data can be collected earlier during the first trimester and may also include data on preconception conditions as well as data on a previous birth.

Garfinkel said if the fundamental biology of harm from environmental exposures is the same for first and subsequent births (observing the first panel provided no evidence to the contrary) and early prenatal data and preconception periods are critical, his suggested design is nearly optimal. He noted even if this assumption were not true, his proposed design enormously simplifies what is otherwise an extremely complex sampling problem. A third virtue is that the design points to the importance of finding out what is known about this assumption. Since it is possible that the biology of exposures is different among first-born and siblings, a small prenatal sample of first births may be justified.

Enrolling sibling births from a birth cohort has enormous virtues, he said, because it is the most cost-efficient method of sampling births during preconception and very early pregnancy. Within 3 years of all births, nearly 30 percent of mothers have a subsequent birth. Within 5 years, the figure is about 44 percent. Within 21 years, the overwhelming majority of mothers would have completed their childbearing. Assuming that completed fertility is about 2 children, a birth cohort of first births would have sibling births with preconception and prenatal data on about the same number of births as a 100 percent prenatal cohort.

Further, according to Garfinkel, each observation generated by the sibling sample would be superior to the prenatal sampled observation because it would contain data not only on preconception, but also earlier prenatal data and data on a mother's previous births, including placentas and cord blood. This information would be invaluable for imputing missing exposure for the first-birth prenatal period. He argued that as long as the biology of exposure is the same, the best data, not just on preconception but also the prenatal period as a whole, would come from siblings and not from births sampled prenatally.

He described two other advantages of the sibling sample. First, although sibling-based estimates would be less precise than corresponding non-sibling estimates, the sibling sample would allow researchers to control for or rule out confounding from genetic and environmental circumstances shared by siblings. Second, collecting sibling data would be cheaper from start to finish than collecting data from two children from different mothers and different household circumstances. Each birth enrolled in a prenatal sample cohort would be de nova. Each sibling enrolled from a birth cohort would be enrolled from a mother who was previously recruited and is a loyal member of the study.

He further stated that a birth cohort would be superior to a prenatal cohort in terms of cost and sample size for two reasons. First, enrollment costs of a birth cohort would be smaller than enrollment costs of a prenatal cohort because of economies of scale. Second, the prenatal data collected from the first births enrolled in a prenatal cohort would be very expensive. NICHD estimates that the cost per child of prenatal enrollment and collection of prenatal data is at least three to four times, and may be as much as 10 times, the cost of enrolling a child at birth. He illustrated how these ratios could be so large, assuming a cost of $1,000 to enroll a mother in either a prenatal sample or birth hospital sample and another $5,000 to collect prenatal data from the mother. The ratio of total cost would be 6 to 1. If enrollment costs were $2,000 and prenatal data collection costs $18,000, the ratio would be 10 to 1. In other words, he said, for every child enrolled in a prenatal cohort, 3 to 10 children could be enrolled in a birth cohort for the same cost.

Although costs would be incurred for every sibling birth enrolled, he said total costs are lower for four reasons. First, enrollment costs of already loyal members of a longitudinal study would be lower than enrollment costs of de nova prenatal mothers. Second, the costs of collecting data on family circumstances would be lower for siblings. Third, the siblings would be followed for a shorter period of time. Fourth, the enrollment and data costs of siblings come later than the enrollment costs for a prenatal sample, which means they are lower because the later-incurred costs would be discounted.

He noted the only parts missing from a birth cohort with siblings are the prenatal and preconception data on first births. The prenatal missing part could be efficiently filled in by a relatively small sample of first-time pregnant mothers drawn from prenatal providers. He estimated that roughly 10,000 would suffice and might indeed be too high. Every additional birth to a first-time pregnant mother drawn from the prenatal providers would reduce the number of sibling births that could be enrolled in the study by between 4 and 10 children.

Garfinkel concluded that his analysis identifies the key scientific questions underlying the choice between the size of the prenatal and birth cohorts: How important are early prenatal data? How important are preconception data? Is the fundamental biology of harm different for environmental exposures for first and subsequent births? The key operational questions all relate to cost. Is the ratio of the cost of enrolling the prenatal sample as opposed to the birth sample 3 to 1 or 10 to 1? How costly will it be to collect placenta and cords on the first births? Finally, time is important. The birth-cohort sibling design would collect prenatal data in later years than would a prenatal cohort. Once these issues are clarified, a formal sample design would provide a precise optimal allocation.

Second Viewpoint

Naihua Duan (Columbia University) supported Garfinkel's suggestion about incorporating the sibling cohort into the study and thanked the workshop's first panel for laying the groundwork on samples to be collected and important time periods. He said he agreed with the suggestion by some Vanguard investigators that objectives are a good basis for the design of a study, noting the NCS has the potential to go beyond being a descriptive study. Specific hypotheses will help everyone understand how design decisions are made. He concurred with Roderick Little's comment during the open discussion in the previous panel (see Chapter 2) that to the extent possible, maximizing the dispersion of potential exposures in the sample to get both high- and low-exposure measurements is a good idea.

Duan observed that a large, complex, and multifaceted study like the NCS needs to balance across multiple study objectives, and the theory of optimal design may shed some light on this issue. The optimal design literature mainly started with Kiefer (1959) and has evolved into a major literature in statistical methodology, mainly in experimental design. The main idea is to use the methodology to balance across multiple study objectives. He described how he has used this approach in several studies. For example, the Human Immunodeficiency Virus (HIV) Cost and Services Utilization Study (HCSUS), initially sponsored by the Agency for Healthcare Research and Quality, recruited HIV-positive patients through care providers, somewhat similar to the way the prenatal sample is proposed for the NCS. Another example is the National Latino and Asian American Study, one of the surveys sponsored by the National Institute of Mental Health Collaborative Psychiatric Epidemiology Surveys (CPES). Duan noted the potential for wider applications of this methodology in sampling applications.

He also provided several examples of using the theory of optimal design. In his first example, he used what he termed is a naïve simplistic model for what the study might want to accomplish. Here, Y represents the outcome, perhaps the cognitive function for children at age 5, and E1, E2, E3 represent measures of exposure at different time periods.3 E1 might represent exposure in the first trimester, E2 might stand for exposure in the third trimester, and E3 might stand for postnatal exposure. The relationship among these variables is given in the regression model

Image p32

The magnitudes of the estimated regression coefficients b0, b1, b2, b3 reflect the relative importance of each study objective in explaining Y. Duan noted that Kwan et al. (2013) listed five candidate designs under consideration, which he labeled D1, D2, up to D5. An alternative, D6, might be to take 40,000 from a birth cohort, 40,000 from a prenatal cohort, 10,000 from a sibling cohort, and maybe 10,000 from hot spots. Other designs might allocate the cohorts differently. An exercise in optimal design would specify performance criteria for each design and then, somewhat similar to doing power calculations, calculate those criteria for each design to determine which has the best performance according to the specified criteria. The key is to specify common criteria across multiple study objectives.

One simple performance measure or criterion in the optimal design literature is called the “A-Optimality Criterion.” It is the sum of the variances of the parameters associated with each design:

Image p32a

He noted this might not be a very good criterion for the NCS because it does not take into account the relative importance of the different study objectives.

His next example was a weighted version of the performance metric. Each W stands for the relative importance or weight the investigators want to attach to each study objective:

Image p33

Duan stated this example does not really answer the question, instead transforming it to a question that might be more tangible for the investigators to think about, namely asking how important it is to reduce uncertainty in the estimated parameters.

In his final example, he stated that instead of considering the variances that are usually used in the optimal design literature, it is conceivable to use a performance measure like the mean square error—the variance plus the square of the bias—to incorporate both the sampling error and also the non-sampling error:

Image p33a

Duan said non-probability sample strategies might be able to be incorporated into this framework to assess how bias and variance tradeoff, and suggested the study recruit a statistician who is familiar with the issues involved to work on the design.

Duan argued lifetime costs are a more useful basis for decision-making than only recruitment (or up-front) costs. The ultimate product in 21 years or so is what the study overall has accomplished and what it cost. He suggested taking follow-up costs into consideration in choosing sampling strategies, with future costs discounted in today's dollars. He also observed the multicohort study uses a multiframe sampling strategy. The multiframe sampling strategy does not require each cohort to be representative of the entire population. Instead, the combination represents the entire population. It would be possible, for example, to include the sibling cohort that does not cover the entire population but gives good coverage for an important part of a population as long as the rest of the population is covered otherwise.

He suggested integrating special cohorts into the overall design. The 10,000 special cohort may be from hot spots; however, it will likely be analyzed together with a main cohort comparing hot spot exposure to exposures among the general population. It would be advantageous, he argued, to have a single probability sample that covers all special populations, with perhaps other special populations useful to include. For example, the first panel noted the potential higher exposure to pesticides in multifamily housing. Integration of the sample might allow NCS to be more flexible in thinking about sampling strategies, Duan concluded.

Third Viewpoint

The next viewpoint was presented by Nancy Reichman (Princeton University). She said that as a research associate at the then-new Center for Research on Child Wellbeing at Princeton University in 1997, she was involved in the new birth cohort study called Fragile Families. She described how Fragile Families involved interviews with new parents and medical record data collection at 75 hospitals in 20 cities across the United States with a success rate of over 90 percent of sampled hospitals. Reichman was responsible for gaining hospital access for the Fragile Families Study.

In her opinion, it would not be harder to get hospital access today than for Fragile Families. In the late 1990s, institutional review board (IRB) policies were in flux. Some of the biggest problems Fragile Families had were in hospitals that initially had the easiest application procedures but then retroactively decided that the approved study was not acceptable. The issue of changing IRB procedures is likely to be much less of a problem today because procedures have universally become much more formalized. However, substantial resources and a well-chosen team would be needed to get through the necessary processes. She noted that for Fragile Families, whether it would be logistically possible to collect placental material and cord blood when mothers give consent after they give birth was not an issue, although it may be now.

She said she and Garfinkel, who consulted with doctors and hospital administrators, think it would be possible to collect placental material and cord blood when sampling is done in hospitals for several reasons. First, there are no risks to the mother from collecting the needed materials. Second, it is apparently not unusual for mothers to take the initiative to have their placentas preserved and banked. She said they have been told by hospital administrators and research deans that if fairly compensated, and the burden to the hospital minimized, the hospitals would likely agree to a system in which placentas that might be needed for the study are preserved and stored, at the study's expense, with those of non-consenting mothers later destroyed. She suggested storage of the placental material and cord blood could be an incentive to consenting mothers, made available to them if needed in the future.

Reichman agreed that the question about optimal allocation cannot be answered without a clear accounting of projected costs and benefits of each type of sampling. She asked about administrative approval or outside institutional approval at prenatal care providers. She suggested that if the provider is in the hospital, such as at a hospital clinic, the hospital IRB approval is needed. She asked about the different types of prenatal sites, the average number of mothers expected to be recruited per site, and the costs to maintain quality control and standardization of protocols across possibly a large number of small sites, noting that keeping track of case dispositions and response rates, particularly the denominators, could be a logistical challenge. She observed that it is difficult to compare the cost of securing institutional access and running the study without having better information on these aspects.

Sampling done at hospitals, she noted, would require a hospital encounter. If, on the other hand, sampling is done at prenatal care providers, mothers could request placental material, cord blood, and medical records from the delivery hospital for purposes of the NCS, hence eliminating the need for a hospital encounter. If a hospital encounter would still be needed for some reason, sampling from prenatal care providers would be enormously expensive compared to sampling from hospitals because of the added cost of the prenatal data collection encounters, access to both prenatal care providers and hospitals, and the logistics of coordinating the study across so many sites. She said key pieces of information are missing, including the cost of access and recruitment for prenatal care providers and hospitals, cost of obtaining placental material under both options, and detail about whether women in a prenatal cohort require a hospital evaluation. Once the relevant information is available, the two approaches could be compared via full cost benefit calculation that includes the sibling cohort.

Reichman added that while the NCS will be truly pioneering by collecting prenatal and preconception data in addition to birth data and beyond, she urged obtaining additional information, such as information on the health or death of the mother's and father's parents. Examples of information that might be collected from death certificates include parent's name, cause of death, education, date of birth, and year of death. Brief information about the parent's lifetime smoking and drinking could also be collected from mothers, since, as animal and human studies increasingly demonstrate, determinates of health can originate well before the parent's generation.

Fourth Viewpoint

Michael B. Bracken (Yale University), the final panelist, said his remarks went in a different direction based on his experience and biology background. He proposed recruiting the majority of the NCS participants (85–90 percent) in pregnancy through the prenatal sample. He explained that fetal origins of disease are dominant issues in studying both childhood and adult disease and only very large pregnancy cohorts could provide the information to study them. As examples, he cited Herbst et al. (1971) as a crucial paper showing how female fetuses exposed to diethylstilbestrol tended to develop vaginal adenocarcinoma when they grew up. Antibiotic use in pregnancy is known to increase risk for asthma. Five percent of all pregnancies result in children with mental birth defects and physical defects. Pregnancy cohorts are needed to study all of these.

Pregnancy cohorts could also provide data to answer such questions as why people born at low birth weight have higher risks of adult cardiovascular mortality or seem to have lower risks of cancer mortality, and the influence of many drugs used by millions of women in pregnancy on their children's physical and mental disabilities remains uncertain. Other public health concerns can only be understood by studying pregnancy cohorts. As examples, what are the effects of exposure in pregnancy to antidepressants, antiepileptics, antiemetics, or pesticides on the developing fetus? Causes of autism, cerebral palsy, attention deficit hyperactivity disorder (ADHD), and many other so-called perinatal conditions actually have origins earlier in pregnancy, but they are not understood. He emphasized that all of this research could be supported by pregnancy cohorts.

He said exposure data are poorly recalled in questionnaires, even when women are asked at birth about exposure during pregnancy. Further, infant mortality in the United States ranks 34th in the world and is becoming worse (compared to 12th in 1960 and 23rd in 1990), but the causes will not be found in birth cohorts. Rather, he said, they are due to associations in pregnancy, including disparities in prenatal care, and only the prenatal cohort would provide the data to study these issues.

Referring to an earlier publication he wrote, Bracken noted, “We know that the vicissitudes in our own uterine existence may profoundly influence the rest of our lives, both physically and behaviorally” (Bracken, 1984). He said nothing has changed since then; moreover, pregnancy itself merits study. Miscarriages occur in about 15 percent of clinically recognized pregnancies, and fetal death and stillbirth are outcomes of great concern. Again, he stressed, only the pregnancy cohort would provide the information to study their causes.

He suggested the NCS could make a real contribution by looking at pregnancy cohorts, as many birth cohorts are being completed around the world. In his view, the proposed mixed cohort, the layered sample, is too cumbersome, unnecessary, and misses the real scientific goals.

Bracken said there is no evidence that a pregnancy cohort is more expensive or more costly to recruit in provider practices. The table in Annett et al. (2013) showed evidence from 16 cohorts, three of which were Dr. Bracken's. Most collected biospecimens for an average cost, including indirect costs, of $2,000. Even with inflation, the costs could not possibly exceed more than $5,000, still two orders of magnitude less than the NCS Vanguard costs.

He explained that within a PSU, a list of providers and the hospitals in which their patients deliver is developed into a cluster. These clusters are then sampled to form a probability sample. There is no cost to the sampling process itself, and recruiting sampled providers in a hospital would be no more costly than recruiting a convenience sample of providers in a hospital. Sampling fractions and denominators could be obtained from birth certificate data.

He said a blood sample collected prenatally carries the same cost as a blood sample at birth but contains more valid pregnancy information. There may be additional costs in prenatal exposure assessment versus estimating prenatal exposures at birth, but these costs are related to sample collection, rather than subject recruitment. To him, they are costs worth bearing because they relate to collecting more valid data.

Bracken posed the question about how early gestations could be studied in a pregnancy cohort. He said he has had four Yale cohorts, a total of almost 17,000 pregnancies. In one, where the researchers restricted gestational age to week 16, they recruited 30 percent at 8 weeks and 91 percent by 12 weeks. In another cohort restricted to 22 weeks gestation, it was almost the same at 8 weeks—29 percent—and at 12 weeks, 76 percent. Extrapolating to a cohort of 100,000 pregnancies, as many as 30,000 women could be assigned for interview by 8 weeks, and 75,000 to 90,000 by 12 weeks. He stated collecting first-trimester exposures in pregnancy cohorts is well documented and is not particularly complicated.

Bracken stated collecting prenatal information on first births, not just the subsequent births in women already enrolled, is crucial because first pregnancies are biologically different from subsequent pregnancies. For example, preeclampsia is a first-pregnancy disease and fetal growth restriction is more severe in first pregnancies; in addition, as children are followed up through childhood, birth order becomes important. He stated that collecting prenatal data only for children who already have a sibling would be a detriment to the NCS.

He noted that biological exposures may not differ between first and subsequent pregnancies, but the scientific interest is in the interaction between these exposures and the fetus, and the fetus changes from one pregnancy to another. There are important biological effects that are already being studied in gene environment studies and using epigenetics. He said these are areas with more hypothesis than fact but warned against never being able to study these questions because of assumptions made at the sample design phase. An assumption-free strategy for sampling and recruitment places fewer constraints on the way pregnant women are sampled so they are representative of all pregnancies in the United States.

He labeled the preconception cohort as a particularly interesting group because many hypotheses concern exposures at the time of conception or before. It is also a difficult cohort to recruit. Women in fertility clinics who may know and plan the exact date of conception are highly selected exactly by virtue of their infertility. Preconception probability samples are almost impossible to obtain and likely not worth the effort. He viewed the sibling cohort as a natural way to obtain data to support preconception studies. It uses women who are already recruited to the NCS and is based on an original probability sample. Although it has the significant disadvantage of including only preconceptions after a prior pregnancy, it may be the NCS' only feasible alternative to a preconception cohort before first pregnancies.

Bracken emphasized that he sees no advantage to the birth cohort because, to him, it misses the unique opportunity offered by the NCS to study the most important scientific questions. He stated that recruitment at hospitals is only worthwhile for women who receive no prenatal care, an important group of women who are often at high risk for poor health and have problems in child rearing. He supported development of special recruitment strategies for these women.

He said, in his experience, recruitment is easier in a prenatal clinic than at a hospital. One has to consider the provider, the hospital, and the research subjects. Providers are easy to recruit and do not have IRBs (although, in answer to a question from Reichman, it was noted that the researcher's own IRB would have jurisdiction). Providers are also more homogeneous than hospitals in the way they deliver care. Bracken said he has had more hospitals refuse to join research than providers. He noted refusal by a hospital to participate eliminates many more women from a sample than does refusal of a private practice, because all of the associated practices are eliminated.

Bracken said consent is more readily obtained from subjects prenatally, and it is unethical to try to obtain consent when a woman is in labor. In hospitals, after a mother has delivered, either she or the child may be indisposed. Twelve-hour discharges are very common in hospitals, which would mean missing a large number of women. Regarding obtaining consent after labor to get cord blood and placentas, he said it may be possible to get consent, but the cord blood and placenta would probably have disappeared. In addition, the presence of families and the excitement of postbirth are other barriers to obtaining consent after labor. He noted in contrast, when recruiting in a prenatal practice, the medical records of study subjects are flagged when they go into the hospital so the delivery room staff know to keep the placenta and cord blood.

In conclusion, he stated the most sophisticated sampling design will fail utterly unless the practical details of how obstetrical care is delivered in this country are taken into account, both in provider offices and in hospitals.

Discussion Among Panelists About Sample Design

Garfinkel noted he and the other three panelists in the session agreed that NCS needs to get prenatal data, and first-trimester prenatal data are important. He suggested one approach might be to compare which approach would result in the most first-trimester prenatal data. He expressed his opinion that when all up-front costs are considered, NCS could get as many or more women with early prenatal data from the birth cohort with sibling follow up, because collecting prenatal data on the first birth is so expensive.

Bracken replied that it is a matter of the scientific questions, not just cost. Prenatal information on first births is important; if the study resulted in no prenatal data on pregnancies to women delivering for the first time, important scientific questions would remain unanswered.

Garfinkel referred to the previous panelists (see Chapter 2), who stated they are not aware of evidence that the biology of exposures differs by first and subsequent births. Bracken replied that the biology of exposures is only half of the question: The question remains how these exposures interact with a developing fetus and said many examples may indicate that the developing first-pregnancy fetus is not identical to subsequent developing fetuses.

Reichman provided another argument for the prenatal sample, stating collection is structured as part of the prenatal care of the mother during her regularly scheduled visits to the provider. With subsequent siblings, she noted, there is no connection to the provider, which might complicate collections.

Duan reiterated that he and the other three session panelists agree that exposure data as early as possible are important. He suggested a prenatal sample and sibling sample are not mutually exclusive, with the question how to combine and make the best use of them. He also noted that missing first births is an important question to address. The prenatal cohort might offer the best solution, unless there is a practical way to get a preconception cohort. Strategies may be combined, instead of trying to use one or the other.

In response, Bracken stated complex designs are more difficult to manage operationally. Managing the schedules of women in different sub-samples is complicated, increasing the chance of mistakes. He suggested a straightforward sampling strategy where women are only recruited during pregnancy. This would simplify the study and remove errors that might occur in the field in trying to implement sophisticated subgroup study designs.

Duan stated that he appreciates the argument for simplicity, but noted advances in information technology may make some approaches more feasible. He noted in a provider sample, the appropriate design needs to take into account both the response rate at the provider level and at the individual patient level; when a provider refuses to participate; all its patients are automatically non-respondent to the study. He pointed to the American Association of Public Opinion Research (AAPOR) definitions of response rates and cooperation rates, saying that measuring and monitoring them is likely to be an important component of the NCS.

Carlson noted she had an opportunity to listen to one of the weekly calls among the Vanguard principal investigators during the week before the workshop. She related that they are giving prenatal care practices four choices for the prenatal cohort, and each practice selects the one that works best for it as a way to sample and recruit women. Most practices are choosing a temporal type of sampling, although there are cases where they feel that they may not be completely keeping track of the denominator. The denominator is probably harder to track in prenatal providers than in a hospital.

Bracken noted that birth certificates will eventually provide the data. He went on to say that the document provided by the Children's Study on power calculations (National Children's Study, 2012) uses very broad categories of defects and shows how well they could be estimated. Included are nervous system defects, major birth defects, neurocognitive development, neurodevelopment disability groups, and developmental disabilities. However, this is not the way people study malformations. For birth defects, the important issue is congenital heart malformations; even then, there are numerous subgroups. When these are studied in the proposed pregnancy cohort, by going from a sample size of 100,000 children to 45,000, even the bit of (inadequate) power presented in National Children's Study (2012) has been reduced.

Open Discussion About Sample Design

Dorr Dearborn (Case Western University) noted that Vanguard recruitment was not limited to first births and asked if the dataset would support a comparison of first births and subsequent births. Garfinkel agreed the question is worth testing. To compare first births to subsequent births, good early data from a small prenatal cohort would be important. Garfinkel asked how big that cohort has to be, saying that he doubts it has to be more than 10,000. He added that determining the size of the cohort is a scientific question.

Bracken noted that with the birth cohort plus siblings, no real-time pregnancy data on first births are obtained, which he termed a dangerous position for the study going forward. In contrast, he noted that the pregnancy cohort would advantageously include first, second, third, and all other births. Further, the ability of analysts to examine the effects of covariates on child and adult health would be severely limited if the study data are confounded by the lack of a representative sample of first births. This is one more reason, in his opinion, not to rely on the birth cohort.

Nigel Paneth (Michigan State University) reminded the audience about Bracken's experience in enrolling 17,000 pregnancies in four cohorts, 30 percent of them as early as eight weeks, for under $2,000 a person. He related his own experience in seven such studies, six of which were funded by the National Institutes of Health. His research has concentrated on enrolling either births or pregnancies, with four birth cohorts and three pregnancy cohorts. He said it is much easier to recruit in pregnancy than birth, and it is administratively simpler to deal with prenatal care providers because they do not have IRBs. In his experience in Wayne County, he had one refusal of a prenatal care provider in some 70 different prenatal care settings. In contrast, working for two years in Wayne County, he could not get 25 percent of 28 hospitals to agree to even a protocol where the woman consented in advance to placenta collection. He said a random sample of hospitals would be unlikely to agree to alter their protocol in the delivery room to do something different with the placenta and cord blood. Some academic hospitals may participate, but he said he doubted many others would.

Duan noted the sibling cohort does not necessarily have to come from a hospital birth cohort but could very well come from a prenatal cohort. NCS could recruit a prenatal cohort and then go on to recruit the siblings. The advantage of the sibling cohort is that the mother has already agreed to participate in the study so there may be some economy of scale in recruiting her for the next child. In some sense, he said, it is not a question about hospital versus prenatal care, but, rather, once the first child is in the sample, what can be done to recruit additional children.

Graham Kalton (Westat) described the provider-based sampling (PBS) methodology now being implemented in three NCS Vanguard sites. An argument that has been made for the birth cohort is that a prenatal provider sample alone does not have complete coverage. In fact, the PBS as it is currently being implemented with a hospital component gives marginally better coverage than a birth cohort. The current approach lists as many prenatal providers as possible within an area, and it recruits women from a sample of those providers. The women are sampled at their first prenatal care visits.

Kalton said existing data indicate that approximately 70 percent of women report that they have their first prenatal visit during the first trimester. The question for the NCS is how quickly it can enroll and interview the women once they have been sampled. Very few women have no prenatal care, but the design is such that they are covered by treating their birthing hospitals as their “first prenatal care” visits. Thus, women who have had no prenatal care are picked up at the hospital. The current approach also provides coverage for any deficiencies in the prenatal provider frame, because women who use only prenatal care providers that are not on the frame are sampled at the hospital. Thus, this approach provides virtually complete coverage. It also enrolls women as early as possible when using a provider-based frame.

Kalton noted that in comparing prenatal and birth cohort approaches, an important question to be addressed is which methodology is more acceptable in practice: Is it better to recruit through prenatal practices or is it better done through the hospitals? He noted Fragile Families obtained a very high response rate in the sampled hospitals, but the study did not collect biospecimens and did not cover situations where the woman or the baby was ill. Recruiting in the hospital may not work in these situations. Enrolling sampled prenatal care practices is also challenging because they are generally very busy. Enrolling the pregnant women presents additional challenges. Kalton endorsed the idea of a sibling cohort, but identified some missing operational details, particularly concerning how to efficiently collect preconception data. The plan for the NCS is to collect data on the child fairly frequently after the birth, every three months in the first year, then every six months, and finally less frequently. For the sibling cohort to be effective in collecting data on the women very early in their pregnancies, the women will need to be identified shortly after conception. The logistics associated with accomplishing this would need to be worked out. Carlson said she thought the original NCS had a pre-pregnancy data collection plan and suggested examining the previous strategy for collecting this type of data under the house-to-house recruitment plan.

Duan noted that PBS, a combination of the prenatal cohort and what has been called the birth cohort, sounds like a very good approach. For the operation of the sibling cohort in the detection of pregnancy, he said some of his colleagues make use of information technology, such as mobile devices, to encourage or invite the participants to send feedback to the study when an important event occurs. He said with careful planning and technology, it is possible to get close to desired event timing.

Reichman asked about the incentive for prenatal care providers to participate in the study. Nina Markovic (University of Pittsburgh) said her institution has an NCS site and, in her experience, providers like to have a plaque on the wall as recognition. She said her study featured providers and their children in their brochures, providing public recognition that they were supportive of the study. Providers felt that they were affiliated and contributing to good science. She also noted she has participated in studies in which recruitment occurred at hospitals and found buy-in at the hospital was top down. They did not get good cooperation in labor and delivery until they placed a 24/7 research staff team in the hospital to collect the samples. With her current cohort, they pay the woman and/ or her significant other $25 to call them when the woman is headed to the hospital, so her staff can collect the samples, much less expensive than 24/7 staffing.

Markovic commented on the first-born issue by noting that from a woman's perspective, many significant changes occur during the first pregnancy. For example, she may continue to work or may be smoking or drinking or have other exposures during the periconception time that do not occur with a second or third pregnancy because there is a toddler in the house. Duan speculated that the issue is not just the biology of the exposure health outcome but potentially also the sociology. The first-born's parents are getting on-the-job training, and later, children are exposed to more experienced parents who might be better able to cope with child-rearing issues. He added, based on his experience with various provider-based studies, one approach is to compensate providers for the time and resources they had to devote. Some studies pay for a staff member to help with the recruitment or offer providers the opportunity to be collaborators as part of research teams. In a sense, he said, this is an extension of recognition such as a plaque, and many collaborators were genuinely interested in the topic of the study. He noted that to acknowledge local collaborators, the author list for the study's papers included “HCSUS Research Team” (a long list of all the study's local collaborators).

Bracken agreed that providers do contribute to research, and how a study manages providers varies. This is why local knowledge is useful in working with providers. He said it is going to be difficult for NCS contractors to come in from the outside to manage this process, because of the role of personal relationships. Providers are more likely to be receptive to a colleague talking to them about research than they would be to an outside group. He considers this is an area, one of many, where losing the local academic centers will be a real detriment to NCS recruitment. Duan agreed, saying that in his experience with the HIV study, prominent HIV providers in each sample geographical area were recruited to serve as site captains. This was a collaborator model, and the site captain helped identify and recruit the other providers.

Sara McLanahan (Princeton University) asked how well the two sampling cohorts generate good representative samples based on actual cooperation and response rates. To her, the most important things are the overall response rates and representativeness of the sample. She shared her sense that many stories about success with providers and people's own hospitals are based on convenience samples and asked whether there is a difference between the provider-based cohort and the hospital-based cohort in terms of representativeness and overall participation rates. Bracken said most of the provider examples that he has worked with are from convenience samples, but, since he had 100 percent acceptance, he finds it hard to believe that a random model would result in large numbers of defections.

McLanahan asked about the importance of eight-week first-trimester measures. If the provider sample can do as well on response rates, this provider cohort might be preferable because it would result in more data on prenatal care. But if it turns out that the most important data are in the first six weeks of pregnancy, then there is a question about whether it is worth the extra cost to get very early data. Information from the scientific community would help to make the decision about how to allocate the sample. Bracken said that the importance of getting data at eight weeks entirely depends on the hypothesis. Some exposures, such as cigarette smoking and the outcome of low birth weight, exert a lot of their effects in the third trimester, so third-trimester exposures are very important, but it is crucial to be able to measure early exposures as well. He stated restricting this massive study to look only at late trimester exposure is unnecessary. According to his data and estimates, he said, if the study is conducted efficiently, they can expect about 30,000 women to be recruited by eight weeks of pregnancy in the all-provider cohort.

McLanahan said Fragile Families found that most mothers received prenatal care but not always in the first trimester. She observed, however, that there is a big difference in access to early prenatal care by race and ethnic minorities. She urged NCS to consider the ability of the design to address disparities. She asked whether starting with a provider sample would produce consistency across race and ethnic groups, income groups, and other subpopulations, in terms of representativeness and response rates.



Subsequent to the workshop, the sampling plan for the NCS is undergoing revision.


The duration of the recruitment period is still under consideration.


Exposure measures are standardized to have zero mean and unit standard deviation. Hence, the effects are standardized, and b0 has the interpretation of the population mean.