The Contribution of Birth Weight and Birth Characteristics to Developmental Outcome

Researchers frequently utilize linked administrative databases to test hypotheses that could also be studied using more time-consuming study designs. While these analyses can provide valid evidence, the results must generally be substantiated with higher-level study designs (cohort studies, randomized controlled clinical or intervention trials). Although many limitations of administrative data for epidemiology and health services research are documented,[1] [2] each database must be carefully scrutinized and evaluated before research is initiated. Thompson et al[3] created a dataset using record linkage methods to link several databases, including birth and infant death certificates, Medicaid eligibility files, and early intervention program records, focusing on an annual statewide birth cohort, followed in death certificate records for a single year and in intervention program records to the 3rd birthday. These data were used to analyze the association between birth weight and developmental disability or delay by age 36 months. This study demonstrates the potential contributions that linked administrative databases can provide. However, while some of the study results are consistent with other findings and bear some relevance for public health, we feel that the article fails to address serious methodological issues that may limit the internal validity of the study and the generalizability of its findings.

First, the datasets utilized in the record linkages that form the basis for this study are insufficiently described, and the quality of these linkages is not discussed. For example, vital statistics data have been repeatedly shown to have poor sensitivity and unacceptable positive predictive values for congenital anomalies overall and for many specific conditions, as well as low sensitivities for many clinical data elements.[4] [5] [6] [7] [8] Although evaluations for early intervention services are not performed for all children, those who survive the neonatal intensive care unit, especially those born very low birth weight or with neurodevelopmental complications, are much more likely to be evaluated, imparting a potential bias to the finding that the lower the birth weight strata, the greater the likelihood of developmental disability or delay. The authors also fail to provide data on the prevalence at age 3 of specific types of developmental disability or developmental delay, making it difficult for those familiar with prevalence rates reported in other studies to interpret these results. There is no discussion of the methods or rationale behind the linkage with Medicaid eligibility files, yet Medicaid eligibility is a moving window that often opens for an infant's delivery retroactively, and also for those infants with special health care needs. Buescher[9] has discussed several issues concerning Medicaid and vital statistics data linkages with special reference to pregnancy and infancy, and it would be interesting to learn how Thompson et al[3] handled these issues.

Turning to the record linkage methods, given the observed prevalence of developmental disability and delay reported by Thompson et al,[3] merge rates in the area of 80% may contribute an ascertainment bias-the authors should have reported how many early intervention records documenting developmental disability or delay could and could not be linked, as well as some data on the characteristics of the linked and not linked records, so that the study results can be read within the context of the reference population.

Second, the paper fails to discuss the potential bias imparted by exclusion of records with missing data. The analysis focuses only on outcomes of singleton live born infants who survived the first year of life (apparently the impact of child deaths during the 2nd or 3rd year of life was deemed of no import), and were born to mothers who had only one completed pregnancy during the study period. An additional 17,479 records were excluded on the basis of missing values on one or more variables. The authors fail to present data on the impact of excluding these cases on their study results. We reanalyzed Table 1 (pp. 324-325), calculating the proportions of studied and excluded cases with developmental disability or delay, no prenatal care, labor/delivery complications, on Medicaid, unmarried, and by race of mother. For all of these comparisons, Chi-square analysis showed significant differences, with higher proportions among the records excluded as missing (p < 0.001). In the case of the dependent variable, among records excluded, 6.4% had developmental disability or delay, compared to 4.2% among records analyzed (p < 0.001), a risk ratio on the order of 1:5. Several researchers have demonstrated that the presence of missing data on key birth certificate variables is an independent risk factor for adverse outcomes;[10] [11] [12] at a minimum, the authors should discuss the implications of excluding these records in interpreting their study results, but ideally they should rerun the analyses with these records included wherever possible, coding missing as a separate category for each relevant variable.

Third, following the record linkage, only two pieces of information from files other than the birth certificate records are used (Medicaid status and the outcome variable). The authors treat so many of the birth certificate variables unconventionally that, at best, it is impossible to interpret their results in the context of the related literature, and, at worst, leaves us to speculate that the published results were derived from post hoc model specification rather than formal hypothesis testing. Several examples illustrate this point. Mothers were considered to have received prenatal care on the basis of having one or more prenatal visits, and no care if the birth certificate indicated no visits. While this operational definition may separate those without prenatal care from those with any, few researchers use this approach, preferring to use measures of adequacy of prenatal care.[13] [14] [15] Complications of labor and delivery are measured by the presence of any complication indicated on the birth certificate, ignoring the fact that some complications are more serious than others and some may have no implications whatsoever for infant growth and development. Smoking is measured by number of cigarettes smoked per day, as self-reported by the mother on the birth certificate. Most researchers make this into a categorical variable, or dichotomize into smoker and nonsmoker categories. Perhaps the authors' failure to include results for this variable in Table 3 (p.327) reflects the manner in which this variable was prepared for analysis, as other researchers have identified an independent contribution of smoking to developmental delay.[16] Birth weight, the primary independent variable, was analyzed by 500 g strata to 3000 g, but the reference category is 3000 to 4749 g, with a break point we have never before seen to differentiate normal from macrosomic infants. The authors also use an unconventional approach to operationalizing their previous pregnancy variable, relying far too heavily on the birth certificate fields reporting previous stillbirths, induced and spontaneous abortions and child deaths, and classifying all live births whose mothers had one of these events as having had an adverse outcome in a prior pregnancy. Others have evaluated the quality of these data fields[17] and found them wanting, yet the authors classify fully one-fourth of their subjects into the adverse outcome category on this basis. The authors discuss the modeling of the interpregnancy interval variable in some detail, but present no information on how this variable was calculated, nor is there data on how interpregnancy interval behaves across the range of observed values, leaving readers to accept on faith that the inflection point at 60 months adequately models this variable. Although there is an extensive research literature on measurement and analysis of pregnancy intervals, none of these studies are referenced.[18] [19] [20] And finally, while most perinatal researchers carefully define their race/ethnicity measures, the authors simply present a race variable with categories of white, black, and other, leaving readers to speculate whether the Hispanic ethnicity variable has been properly used to differentiate those mothers who are white Hispanics from those who are white non-Hispanics, and black Hispanics from black non-Hispanics.

An additional issue is the use of an outcome variable that is based on participation in specific public programs. It is highly likely that the statistical findings reflect, to a greater or lesser extent, a bias due to differential opportunities of infants and young children for diagnosis and access to these programs. This bias may be very difficult to evaluate, and may result in under-utilization of these programs by middle- and upper-class families, those living in more isolated areas of the state of Florida, or other issues. Further, the authors apparently made no independent assessment of the reliability and validity of the diagnoses or functional assessments in the early intervention database, yet they make sweeping recommendations for clinical management of high-risk infants by pediatricians and developmental specialists.

The authors suggest that, based on their findings, surviving newborns born at less than 1000 g should be routinely referred for evaluation for early intervention programs or Part C services (p. 329). Apparently, they are unaware that this is already standard of care nationally.[21] The “protective effect” of black race described in the discussion is well known-when demographic and reproductive health variables available on birth certificates are controlled, infants born to black women appear to have lower risks of mortality and adverse developmental outcomes.[22] But we must remember that the birth certificate is and always will be an imperfect instrument, especially with respect to socioeconomic status, living conditions, and other factors for which it currently contains wholly inadequate proxy measures.

In the end, what we learn from the work of Thompson et al[3] is that, in the absence of carefully formulated research hypotheses well grounded in the evidentiary base, the presence of large, linkable datasets may provide a temptation too great to ignore. But, like the forbidden fruit in the Garden of Eden, we must resist the temptation nonetheless, at least until the contents of these databases are well understood and the implications of their imperfections have been carefully assessed. For, unlike Adam and Eve, our fruits in the form of administrative public health records are rarely unblemished, and we must be careful indeed in how we prepare them for the feast.

REFERENCES

• 1 Virnig B A, McBean M. Administrative data for public health surveillance and planning.  Annu Rev Public Health. 2001;  22 213-230
• 2 Iezzoni L I. Assessing quality using administrative data.  Ann Intern Med. 1997;  127 666-674
• 3 Thompson J R, Carter R L, Edwards A R et al.. A population-based study of the effects of birth weight on early developmental delay or disability in children.  Am J Perinatol . 2003;  20 321-332
• 4 Watkins M L, Edmonds L, McClearn A, Mullins L, Mulinare J, Khoury M. The surveillance of birth defects: the usefulness of the revised US standard birth certificate.  Am J Public Health. 1996;  86 731-734
• 5 Honein M A, Paulozzi L J. Birth defects surveillance: assessing the “gold standard”.  Am J Public Health. 1999;  89 1238-1240
• 6 Piper J M, Mitchel Jr E F, Snowden M, Hall C, Adams M, Taylor P. Validation of 1989 Tennessee birth certificates using maternal and newborn hospital records.  Am J Epidemiol. 1993;  137 758-768
• 7 Kirby R S. The quality of data reported on birth certificates.  Am J Public Health. 1997;  87 301
• 8 Kirby R S. Invited commentary: using vital statistics databases for perinatal epidemiology: does the quality go in before the name goes on?.  Am J Epidemiol. 2001;  154 889-890
• 9 Buescher P A. Method of linking Medicaid records to birth certificates may affect infant outcome statistics.  Am J Public Health. 1999;  89 564-566
• 10 Gaudino Jr J A, Jenkins B, Rochat R W. No fathers' names: a risk factor for infant mortality in the State of Georgia, USA.  Soc Sci Med. 1999;  48 253-265
• 11 Gould J B, Chavez G, Marks A R, Liu H. Incomplete birth certificates: a risk marker for infant mortality.  Am J Public Health. 2002;  92 79-81
• 12 Wen S W, Chen L M, Li C Y, Kramer M S, Allen A C. The impact of missing birth weight in deceased versus surviving fetuses and infants in the comparison of birth weight-specific feto-infant mortality.  Chronic Dis Can. 2002;  23 146-151
• 13 Kotelchuck M. An evaluation of the Kessner adequacy of prenatal care index and a proposed adequacy of prenatal care utilization index.  Am J Public Health. 1994;  84 1414-1420
• 14 Alexander G R, Cornely D A. Prenatal care utilization: its measurement and relationship to pregnancy outcome.  Am J Prevent Med. 1987;  3 243-253
• 15 Alexander G R, Kotelchuck M. Quantifying the adequacy of prenatal care: a comparison of indices.  Public Health Rep. 1996;  3 408-418
• 16 Faden V B, Graubard B I. Maternal substance use during pregnancy and developmental outcome at age three.  J Subst Abuse. 2000;  12 329-340
• 17 Adams M. Validity of birth certificate data for the outcome of the previous pregnancy, Georgia, 1980-1995.  Am J Epidemiol. 2001;  154 883-888
• 18 Kallan J E. Reexamination of interpregnancy intervals and subsequent birth outcomes: evidence from U.S. linked birth/infant death records.  Soc Biol . 1997;  44 205-212
• 19 Klebanoff M A. The interval between pregnancies and the outcome of subsequent births.  N Engl J Med. 1999;  340 643-644
• 20 Zhu B P, Rolfs R T, Nangle B E, Horan J M. Effect of the interval between pregnancies on perinatal outcomes.  N Engl J Med. 1999;  340 589-594
• 21 American Academy of Pediatrics and American College of Obstetricians and Gynecologists .Guidelines for Perinatal Care. 5th ed. Elk Grove Village, IL; American Academy of Pediatrics 2002: 274-282
• 22 Kirby R S. Perinatal mortality: the role of hospital of birth.  J Perinatol. 1996;  16 43-49