Commentary on Disparities in Infant Mortality Due to Congenital Anomalies on Guam

Dear Editors,

Recently, we read the article on congenital causes of death in Guam.¹ As public health professionals (ie, a physician, biostatistician, a biomedical laboratory officer and three public health officers) we support highlighting the importance of congenital causes of death. Explained or unexplained, the death of a child is a sad tragedy and provides motivation for a comprehensive public health approach to mitigate potential risk factors on the island and elsewhere. Although the rate of congenital causes of death in Guam in the studied population was comparable to similar populations, there are still likely to be opportunities for preventive efforts. Well established, evidence based recommendations for lowering infant mortality should be considered a public health priority.²

Although it is commendable to evaluate the geographic pattern of congenital causes of death, the authors could have communicated many of their points using an editorial format with supporting anecdotes and data points. The authors acknowledge several of the scientific weaknesses in their paper, but still chose to publish the data as though they were creating generalizable knowledge in an original research paper. Unfortunately, this format creates unnecessary confusion because of methodologic weakness, scientifically inappropriate language, and poorly documented ethical considerations.

The central methodologic weaknesses in the study include the ecologic design, the small sample size, the inappropriate statistical analysis (and interpretation) and the high risk of measurement bias and confounding. Ecologic study designs are among the weakest sources of evidence because they do not identify an association among individuals (only among populations), greatly limiting causal inferences. Villages differ for many reasons, making the exposure of interest only one of many differences. Herbicide is typically used for a purpose and so village characteristics that impact congenital death might also be related to the reasons for differential spraying (ie, confounding by indication). The sample size is low (n=19) and the villages are unequally balanced between “exposed” (n=12) and “unexposed” (n=7), further reducing statistical efficiency. Within villages, some had too few births to find any cases and some had too few births to identify an accurate rate per 1000 births.

The authors' use of linear regression further weakens the scientific veracity of the study. Linear regression is based on an assumption of independence of each village. Analysis of autocorrelation among neighboring villages could have identified spatial clusters (ie, hotspots) so that spatial units were appropriately representing the underlying spatial variance in the outcome of interest. Otherwise, the spatial categories chosen may have created a modifiable area unit problem (MAUP). Looking at Figure 1, it can be seen that 18 villages border at least one other village with their same risk category, suggesting autocorrelation is present. Additional evidence of a MAUP is found in the normal levels of cause specific deaths at Yigo, when this is the closest village to Andersen Air Force Base (AFB). Using all available publically reported congenital deaths on Guam (available for the years 1991, 1993, and 1995) by village, there is an even lower rate of congenital causes of infant deaths among mothers residing in this village than was reported by the authors.³ In fact, adding all available data shows the village of Yigo would switch from “high risk” to “low risk” by the authors' definition of these terms. This evidence of MAUP points to the larger issue: artificially separating small scale contiguous locations leads to comparisons between populations and numerous exposures that are not independent, violating model assumptions.

Even if the authors had established independent observational units, the authors acknowledge the validity of their study is also threatened by confounding. They attempt to adjust for village level confounding, using aggregate measures, but the chosen numbers have remarkably low temporal granularity. Plus, over-parameterization happens quickly with only 19 observations. Thus, residual confounding is very likely as it is not possible to adjust for even well-established risk factors.⁴ More specifically, there is no adjustment for the other natural and industrial toxins on the island, maternal smoking, healthcare access, obesity, infectious diseases or the use of medications during pregnancy.

While there could be several sources of bias, the likeliest source of measurement bias comes from the questionable use of memory-based methods (MBM) to determine exposure. Because memories cannot be externally measured and because they are so inaccurate for dietary habits, memory based dietary assessment methods have been called pseudoscientific and inadmissible as scientific evidence.⁵ With so much controversy about near term MBM, the use of memories from several decades ago is not a scientifically appropriate basis for the claims that were made. Instead, all conclusions about the association should have used terms describing the measurement method. For instance, “spray area” should have been consistently referred to as “spray area identified by recall several decades after spraying.” It should also have been noted that recall was provided after village level outcome data was publicly reported for many of the years covered by the study.³

The authors further exacerbated the methodologic weakness by using several additional categories of scientifically inappropriate language throughout the paper. We recommend avoiding the use of the terms “increased” and “increases,” as these words can be used to describe a causal relationship.⁶ The authors also consistently over-generalize to unstudied populations, geographic areas, and time periods. Despite the low sample size, the authors excluded the births from residents of military installations. This exclusion was particularly questionable given the exposure in question is claimed to be military related and prior evidence has shown military members are an appropriate population for investigating spatial causation.^7,8 Approximately 20% of all births over the time period⁹ were not considered and 27% of the land mass,¹⁰ but the magnitude of these exclusions was not provided. All conclusions should have specified the populations and time periods covered by the study since it was not a randomly selected subsample. Plus, the land mass of the excluded installations was included in Figure 1, falsely indicating the residents on the military installations are a part of “high risk” areas when they were excluded from the study.

Additionally, the authors mishandled the precision of their data on both relative and absolute terms. For instance, Table 1 describes the birth rates for most villages with three significant digits even where case counts were less than 10 (ie, “pseudo-precision”). Even worse, the poor handling of precision was not constant across all villages, differentially exaggerating the influence of outlier observations. When case counts are low, it is expected there will be variance from the mean by chance alone until a large enough sample can be collected. For example, there were only 503 births observed in the village of Umatac, but yet the authors report the rate per 1000 births. Similarly, the regression model falsely assumed each rate was as precise as every other rate, even when sample sizes differed by >10,000. Looking at the rates in Table 1, the three villages with the fewest births were also showing cause-specific death rates furthest above or below the mean. Instead, each village should have been weighted by sample size using an “analytic weight” (ie, inverse variance weight) to account for the differing sample size in different villages. When assuming equivalent samples, the variance in cause specific death rates per 1000 births between villages is 2.3, but when accounting for sample size, the variance between villages falls to 1.1 (using Stata v13.0, StataCorp LP, College Station, TX). Thus, there is substantially less variance between villages than the authors claim. Imputation with the available data suggests this error substantially inflated the β coefficient from the linear regression model, but the degree of inflation cannot be certain because the authors did not disclose which villages were “exposed.” For these same reasons, the authors' claims about the proportion of total variance explained by their models are equally suspect.

Another aspect of language that would be better suited for an editorial was the authors' focus on the controversy about the name of the chemical formulation that Veteran remembers using. Specifically, the authors say the Veteran's “claims have been reported and confirmed” by the Daily Beast. Although the Daily Beast is not typically used as an authority for deductive reasoning, even this source did not report the same information as the current memories.¹¹ Specifically, the Daily Beast article says the spraying occurred around the perimeter of Andersen AFB and within the base without any mention of the 12 villages throughout the island.³ Additionally, the title of the Daily Beast article poses the story as a question rather than a scientific confirmation. Likewise, the authors incorrectly describe the Veterans Administration (VA) cases.^12,13 Unlike deductive standards, for legal determinations, doubt is to be resolved in favor of the Veteran.¹² Even by this standard, the VA did not assert Agent Orange (rather than another commercially available herbicide) was used in Guam.^12,13

As for the Environmental Protection Agency (EPA) and the Agency for Toxic Substances and Disease Registry (ATSDR), both agencies maintain websites providing public information about dioxins on Andersen Air Force Base (AFB) on the island of Guam.^14,15 While dioxins are the principal toxin within Agent Orange, the presence of dioxins does not necessarily signal the presence of Agent Orange. More specifically, it is known that pesticides were present at the Andersen superfund site¹⁴ and commercially available pesticides are a source of dioxin exposure.¹⁶ For instance, historic use of pesticides in Japan released an estimated 460 kg of dioxin toxic equivalents (TEQ) as compared to estimates between 130 to 366 kg of dioxin TEQ from spraying defoliants in the larger country of Vietnam.¹⁷ Additionally, the presence of dioxin on Andersen AFB does not mean it originated from military activity. Specifically, the EPA's website describes how, “Most soil contamination problems at Andersen are either the result of nearby industrial activity, or the result of materials being placed into scattered dump-sites.”¹⁴

The larger public health question is why would it matter if dioxins came from Agent Orange or commercially available pesticides? While the public might falsely perceive risk based on the name of the formulation rather than seeking a comprehensive, scientifically-designed hazard assessment, the authors are still responsible for appropriately communicating their findings. Although the risk of Agent Orange is related to the fact that it contained dioxin, there are several contextual factors that have historically made the perception of Agent Orange different than other sources of dioxin. First, Agent Orange was sprayed from aircraft with the intent of defoliating jungle within a warfare context. As a result, Agent Orange is associated with presumptive exposure because the methods of application and the warfare setting would be expected to cause incidental exposures. With this unique context, it is not surprising that Agent Orange is more strongly associated with congenital malformations within Vietnam than elsewhere.¹⁸ Thus, perceptions of Agent Orange toxicity could be inflated if Agent Orange was a surrogate marker for exposure to other combat related phenomena within Vietnam that are more difficult to measure. By the same logic, the hazard assessment for Agent Orange would be different if it were used in a peaceful location for purpose of weed control using a ground based application method, in the same manner as commercial pesticides containing dioxin. In this scenario, it would not be valid to make equivalent presumptions about village-wide exposure based on the name alone. Had the authors considered the larger picture of the hazard, they could have described the hazard assessments that have already been completed including the ATSDR identified sites with dioxin soil contamination within Andersen AFB. These sites are not expected to have previously caused hazardous exposures to the public in the past, present, or future.¹⁵

Although the distinction between Agent Orange and dioxins generally wouldn't impact hazard assessment, the distinction could be consequential for other reasons. Recalling a chemical formulation that conflicts with a documented comprehensive DoD assessment¹⁹ draws more attention to the pitfalls of memory-based methods of exposure assessment.⁵ Local and federal environmental protection agencies might already have reason to discount the memories because several decades lapsed before they were reported and because prior reported descriptions of spraying locations were incomplete.¹¹ Thus, the authors' selective focus on a toxic formulation that historical records show was not sprayed on the island unnecessarily detracts from the Veteran's memories about particular locations where he sprayed.

Nevertheless, we hope the new information will be taken for what it is. Given the prior detection of dioxins and pesticides, the Veteran's memories about spraying a chemical on 12 villages in Guam is something the public should know more about. It is unclear why spraying operations would occur in civilian areas of firmly controlled US territory or why the prior reports^3,20 of these memories were incomplete. Using an editorial could have expedited further public disclosure of these additional memories and provided a platform to call for further investigation. Instead, the authors published a research paper indicating they have access to information about 12 additional villages, but they do not disclose the village names, cardinal directions, or any other identifying details of the locations to enable soil testing.

Finally, framing the paper as if the study was deductive adds additional ethical considerations.²¹ Had the authors opted for an editorial, then it would make sense to cover anecdotal claims from popular media as a relevant part of a larger narrative. On the other hand, if the authors attempted to systematically collect memories from an identifiable human for purpose of generalizable knowledge, then the authors represent the Veteran as a human subject and the description of the review board approval is missing. If memories were not collected systematically, the authors should have disclosed this methodologic weakness. Similarly, the justification for describing identifiable medical information about the Veteran and Granddaughter was not clear and informed consent for this disclosure was not documented within the paper.²¹

Overall, the rates of congenital causes of infant death available thus far are reassuring. First, residence in the island of Guam for the studied population appears safe for fetal development: cumulative rates appear slightly lower throughout the island than found in the United States overall.¹ Additionally, rigorous risk assessment from Andersen AFB shows no evidence of past, present, or future hazard to the public from dioxin soil contamination and a synthesis of all available evidence suggests congenital causes of infant death occur at the same rate among those living near Andersen AFB in the village of Yigo as they do on the rest of the island. Still, public health authorities can always strive to lower the rates of congenital deaths through a comprehensive approach. Adequate folate for all women of childbearing age (as available in a prenatal vitamin) can lower the risk of congenital malformations² and providers should counsel women about the developmental risks of prescription medications, even if pregnancy is not being planned. Once pregnant, women should be offered routine obstetric care to optimize health outcomes for both mother and child.² Ultimately, appropriate public health practice depends on rigorous science, proper implementation of well-established interventions, and proper collaboration between federal and local authorities.

Conflict of Interest

None of the authors identify any conflict of interest.