Deepwater Horizon Supporting Stmt B

B. Collection of Information Employing Statistical Methods If statistical methods will not be used to select respondents and item 17 on Form 83-I is checked “No” use this section to describe data collection procedures.

The target population for the Worker Health Survey comprises all workers who responded to the Deepwater Horizon oil spill event, and our intent is to include all of these workers on the roster. The survey will be administered to the entire roster, minus those who meet exclusion criteria, by contacting them using the phone numbers they provided on the roster. Since all workers are to be included, there is no sampling of workers using statistical methods.

Every effort is being made to include all response workers on the roster. However, it is unlikely to be one hundred percent complete due both to refusals and logistical issues. While it will be possible to assess and, to some extent, adjust for nonresponse on the part of rostered workers, uncertainty about basic parameters of the population of interest (including size and age, race, sex, and job/response role distribution) makes assessing noncoverage or undercoverage more problematic. Therefore, results from the statistical analysis of survey data will, strictly speaking, be applicable only to the population of workers comprised by the roster.

We will determine non-response bias by comparing the distribution of variables available on the roster form—including demographics, work location, and type of worker (BP, contractor, government worker, volunteer, or other)—among respondents and non-respondents to the survey. The detection of significant differences between the two groups based on appropriate statistical tests (Chi-square tests for categorical and t-tests for numeric variables) will be taken as an indication that systematic differences exist between responders and non-responders. We intend to generalize the survey results to all rostered workers, and should any non-response bias become apparent based on these analyses, we will qualitatively describe the limits of generalizability. Additionally, if necessary, we may adjust for non-response by using post-stratification weighting to fit the distribution of roster variables available among respondents to their original distributions in the full roster.

Generalization of results beyond the roster to the population of all Deepwater Horizon response workers will require the assumption that the roster is representative of the larger responder population. The reasonableness of this assumption will depend on the completeness of the roster and the likelihood that certain categories of workers are preferentially included, neither of which can be directly assessed empirically. Therefore, our principal arguments in favor of the assumption of representativeness will be the completeness of the roster, assessed on the basis of a comparison of the number of rostered workers with best estimates from the Incident Command of the actual number of responders, and the lack of any other available data.