I have made significant advances to the fields of economics, epidemiology, and statistics.
I developed a new standard error estimator <PDF> based upon the Census Bureau's random-groups method. A conspicuous problem with the Census Bureau's methodology was that it yielded different standard error estimates for the "yes" and "no" response for the same binomial variable, when both standard error estimates should have been identical. If most respondents answered a dichotomous variable one way and a few answered the other way, the standard error estimate was considerably higher for the response with the most respondents. Moreover, when 100 percent of respondents answered a question the same way, the standard error was not zero, but was quite high. Reporting average design effects which were weighted by the number of respondents that reported particular characteristics magnified the problem. My modified estimator, which was based on ratio estimation, resolved these problems.
At the Census Bureau, I also conducted original research on the impact on the reliability of Census estimates assuming various scenarios (i.e., varying undercount rates and different imputation methods, such as nearest observation, previous observation, hot-deck, etc.) for incorporating the undercount adjustment into sample data products. In addition, I created a SAS program for determining confidence intervals for medians <PDF>.
While a statistician with the National Animal Health Monitoring System (NAHMS) of the United States Department of Agriculture (USDA), I introduced innovative methods to the weight-adjustment procedures, including raking for weight adjustment <PDF>, and propensity analysis <PDF> (applying logistic regression analysis to identify variables associated with non-response) to develop poststrata for nonresponse adjustment.
I have performed a considerable number of epidemiologic analyses. I applied time-series analysis to data on Rinderpest, and accurately forecasted a subsequent outbreak in East Africa (Veterinary Record, 2003, 152(21):641-647). I used logistic regression to analyze epidemiologic risk factors associated with:
My thesis work in Agricultural and Resource Economics, for which I developed a Cobb-Douglas production function to study returns to scale in the swine industry, led to four journal articles (Invest Agr: Prod y San Anim, 1999, 14:71-84; Prev Vet Med, 1998; 34:147-159; Ind Farming, 2000; 50:32-34,40; and Ind J Anim Sci, 2001; 71(10):995-997). I completed further analysis to examine economies of scale in the production of swine manure ( Arq Bras Med Vet e Zootecnia, 2000; 52:285-294).
To study economic interactions between feeding rates and stocking densities in intensive catfish production, I developed a Just-Pope simulation model (which simultaneously characterizes the mean and variance of production, and separates the output variance versus input interactions from the expected output versus input interactions) from NAHMS Catfish '97 data I used the mean model (J World Aquaculture Soc <PDF>, 2000, 21(4):491-502) to determine production stages, to develop a profit function (from which I derived profit-maximizing input intensities), and to examine marginal rates of technical substitution between inputs; and the variance equation (Aquaculture Int <PDF>, 2006, 14(5):415-419) to identify systematic impacts on yield variance. Typically, more risk-averse producers have an augmented concern with profit variance and prefer to apply input at higher rates if they are variance-reducing (i.e., risk-reducing) inputs.
I used Mishan's normalization procedure to perform a cost-benefit analysis for catfish farmers using recirculating systems (Ind J Anim Sci <PDF>, 1999; 69:204-206). I applied economic-welfare analysis to measure changes in producer and consumer surplus to determine who would ultimately pay the price for Hazard Analysis and Critical Control Points (J Appl Anim Res, 2000; 17:197-200).
While a statistician with the New Brunswick Laboratory, I became very familiar with the International Organization for Standardization (ISO) guidelines for computing and expressing measurement uncertainty. I found the GUM Workbench <PDF> to be an especially useful tool for performing uncertainty analyses quickly. While uncertainty analyses for nuclear measurements are quite rigorous, other fields (such as economics) have tended to ignore uncertainty altogether, even when offering rather broad claims based upon seemingly complex and sophisticated-looking calculations. To address this gap, I discussed the problem and applied the GUM Workbench to an economic-welfare analysis of Bovine-Leukosis virus in US dairy cows (Agricultural Economics<PDF>, 2006, 35:363-372).
I identified and corrected a serious blunder that epidemiologic researchers had been repeating for two decades in the computation of population attributable fractions (i.e., the fraction of disease that could have been prevented by eliminating exposure to specific risk factors) (J Dairy Res <PDF>, 2005, 72:1-11). This correction vastly simplified the calculation of uncertainties in measurements of the economic impacts of reduced milk production associated with epidemiologic risk factors for Johne's disease on dairy operations. I subsequently addressed the welfare effects of reduced milk production associated with Johne's disease on Johne's-positive versus Johne's-negative dairy operations (J Dairy Res <PDF>, 2006, 73:1-7).
I found extreme flaws in an article that purported to measure economic impacts associated with the use of recombinant bovine somatrotropin, which is used to increase milk production in dairy cows (see AgBioForum <PDF>, 5(2):25-27). I offered an improved analysis, and analyzed changes over time in the welfare effects of recombinant bovine somatotropin (Int J Dairy Tech <PDF>, 2007,60(3):157-164).
At the U.S. Department of Defense, I identified severe problems in the statistical methods followed for personnel surveys (Armed Forces & Society <PDF>, 2010,36(3):558-570). Briefly, in terms of the sampling:
As anticipated, survey-response rates were quite low, most notably in strata that comprised spouses of lower-ranking service members; hence response bias is probably high enough to raise concerns about the value of the data. In an effort to correct survey estimates for response bias, weight adjustments occurred in three stages: a logistic-regression model to adjust for unknown eligibility; a logistic-regression model to adjust for survey completion among eligible respondents; and finally, a post-stratification adjustment to force specific survey estimates to match certain known population totals. Briefly, problems with the weight adjustments were as follows:
Finally, sampling strata were collapsed together to form new "variance strata" for variance estimation, which effected a downward bias in the variance estimates, in order to make the survey results appear more precise and accurate than they actually were. The variance estimates that resulted from the creation of the "variance strata" were inappropriate because the variance estimates did not reflect the actual sampling design. The margins of error presented with DMDC survey results grossly misrepresent the actual uncertainty associated with the estimates.