Essential aspects of data analysis in epidemiologic research have been reviewed elsewhere and are not specific to chemicals with short physiologic half lives.

However, for completeness of the proposed tiered evaluative system, these considerations are described here in brief. The overall analytic strategy in observational research depends on the main goal of the study. Generally, statistical models fall into two categories — predictive and explanatory (Shmueli, 2010). For predictive analysis, selection of variables into the model is data-driven and may differ from dataset to dataset. The goal of this approach is to maximize the CT99021 concentration model fit and a decision on whether to retain a particular covariate of interest is based on statistical tests and goodness-of-fit without a specified exposure of interest (Bellazzi and Zupan, 2008). In an explanatory (hypothesis testing) analysis, this approach may be inappropriate because it may wrongly eliminate potentially important

variables when the relationship between an outcome and a risk factor is confounded or may incorrectly retain variables that do not act as confounders (Kleinbaum and Klein, 2002). More importantly, for an explanatory model, which is focused on a pre-defined exposure–outcome association, inclusion and exclusion of control variables (confounders, mediators or effect modifiers) should be driven, at least in part, by a priori reasoning (Beran and Violato, 2010, Concato et al., 1993 and Hernan Ibrutinib ic50 et al., 2002). It is important to keep in mind that the results of observational studies are inevitably subject to uncertainty. This uncertainty may be attributable to various sources of unaccounted bias and to various data handling decisions and assumptions. The magnitude of uncertainty can be formally assessed through quantitative sensitivity analyses. The methods of addressing residual bias through sensitivity analyses are now well developed both in terms of basic theory (Greenland, 1996) and with respect to practical applications (Goodman et al., 2007, Lash and Fink, 2003 and Maldonado et al., second 2003). With respect

to sensitivity analyses of alternative decisions and assumptions, much can be learned from previous experience in economics, exposure assessment and quantitative risk analysis (Koornneef et al., 2010, Leamer, 1985 and Spiegelman, 2010). Tier 1 studies include those that clearly distinguish between causal and predictive models and demonstrate adequate consideration of extraneous factors with assessment of effect modification and adjustment for confounders. To qualify for Tier 1, a study should also perform formal sensitivity analyses. When consideration of extraneous factors is considered adequate and the model selection is appropriate, a study may still be considered incomplete without a sensitivity analysis. Those studies are placed in Tier 2.