Agent orange exposure modeling: fallacies and errors

In a 2009 paper in JESEE, Ginevan et al. questioned the validity of our model, also published in this journal, for estimating an index of exposure opportunity (EOI) to military herbicides used in Vietnam (1961–1971). We had already pointed out many of their conceptual and interpretational errors when the substance of their paper was presented to a 2007 Institute of Medicine meeting (they cite neither their 2007 presentation nor our response). While we regret not having replied to their published critique immediately, there is now urgency to correct the serious errors in Ginevan et al. because they have once again distributed their work to an Institute of Medicine (IOM) Agent Orange panel, and it is essential that these data not be used to further undermine efforts at studying the long-term consequences of exposure to military herbicides. To our surprise, upon further investigation of their paper, we have now found even more serious errors. The scores (E4) they use do not replicate the E4s actually produced by our software system. Ginevan et al. selected points on or near the spray paths of three missions to simulate deviation from the straight-line projections we use in our models. They do not describe the methods used to calculate E4. They then compound the E4 errors by interpreting both the scores and the model itself inappropriately. Ironically, the E4s for these points yield a strong confirmation of the utility of our model even for paths that deviate from straight lines. The Supplementary materials tabulate these data. The Ginevan et al. representation of our model seriously errs in many fundamental respects. Our overall model comprises two complementary sub-models. Sub-model 1 counts ‘‘hits’’ to a specific point from direct overflight by a leg of a spray mission, similar to models used by the CDC and others; sub-model 2 calculates an E4 measure of proximity in time and space, over a given time interval, from a given point, to all legs of all earlier spray missions that occurred, combining an inverse-distance measure with an environmental decay function. Ginevan et al. conflate these two models, purporting to assign an ‘‘area’’ to E4 when in fact E4 has no concept of area. They then discuss the supposed limitations in terms of 1-km and 5-km areas, that, simply put, have no meaning in the E4 model. Our EOI aggregates E4 scores into an area-under-the-curve measure. Herbicide concentration is continuously apportioned along the total multi-legged flight path (to reflect the careful calibration of spray nozzles). We operationalize this formula by application of a closed-ended solution. E4 also incorporates a firstorder exponential decay function to take environmental persistence into account. Our model uses log (E4). Ginevan et al. use raw E4 scores. A logtransform of the E4 scores produces values that vary p10% around the mean for each mission and supports our model. For epidemiological purposes, the raison d’être of the EOI, scores fall into appropriate ranked-exposure groupings, thus providing another confirmation of the robustness of our estimates, even for paths that deviate from the straight-line projections we used. The Supplementary materials reproduce a graph of the lognormally distributed aggregated EOI scores for a 1-month sample of over 140,000 soldiers in 815 military units. The combination of using both raw scores and scores that do not match the true E4s in their analyses lead Ginevan et al. down tortuous speculative paths. Their use of an incorrect E4 score of 60,791 for one point, when the true score is zero in our system, leads them into all sorts of wild speculation about how our system works. They guess at how we integrated inverse-square relationships. We used the analytic solution to the integral. They posit that we did not deal with dividing by zero when flight paths intersect the centroid. Of course we did, or our programs would not run. They ruminate that we used two significant figures for longitude and latitude calculations; we use five. They are troubled by the fact that the average E4 for the points on the flight path is slightly lower than the average E4 for points 0.2 km from the flight path. First, recall that many of their values are incorrect; second, their observations are an artifact of how they chose their points. Many of the 0.2 km points are close to the center of the flight path, while most of the ‘‘on the flight path’’ points are near the ends. E4 measures proximity as inverse-distance. Hence, the more of the flight path a point is near, the higher the score (1/D). A point at the middle is ‘‘near’’ the whole flight path; a point at the end is only ‘‘near’’ half of the path. Our E4 scores are calculated with respect to a receptor point at the centroid of grids in a customized geographical information system (GIS) we created, where former South Vietnam and parts of Laos are divided into 263,353 grids, spaced 0.011 longitude/ latitude apart (B1.2 km on a side). Much of the wobble observed in the raw numbers used in the Ginevan et al. analysis is a property of the granularity of the grid system and has no implications for exposure classification using log (E4). We carried out extensive sensitivity analysis for the IOM on just this point. At least half of Ginevan et al. is devoted to arguing against the appropriateness of an exposure opportunity model that relies on military data on spray history and troop locations. Both the National Academy of Sciences and the IOM have already carefully considered these arguments and issued a request for other models, to which we responded. We would love to take credit for creating an entirely new approach to exposure opportunity for epidemiology studies, but our models are similar to many others in widespread use. We set up a GIS using the 1974 NAS grid system model, albeit with a faster computer, modern database and GIS software, and vastly more complete and accurate spray and flight data that mathematically models full-flight paths rather than just relying on turning points given in the military spray records (HERBS file). Finally, our overall approach to exposure measurement is far more nuanced than only using the EOI for characterizing exposure to the herbicides, as the Ginevan et al. treatment would have the reader believe, and our approach has been evaluated, tested and endorsed by two separate IOM review committees. For example, while the three ‘‘isolated’’ missions selected by Ginevan et al. here appear to be isolated, in fact, they were located within Air Force defoliation targets in which tens of thousands of gallons had been previously sprayed. Soldiers were not simply dropped into little 1-km grids for their periods of service, but spent unwashed weeks on end in these areas during their tours of duty. Our models seek to take these factors into account. The Supplementary materials provide more information on the three missions in question. JESEE chose to publish this paper 4 years after it published a mathematical description of our model, without providing us any Journal of Exposure Science and Environmental Epidemiology (2014) 24, 444–446