In this paper we assess properties of commonly used estimates of total effects of obesity on mortality and identify consequences of these properties for inferences. We argue that standard estimates have important shortcomings that at best limit the reach of inferences and at worst lead to misleading conclusions. Although some of these limitations are routinely acknowledged, rarely is their use accompanied by careful scrutiny of their weaknesses, let alone by a quantitative assessment of their sensitivity to violations of some stringent assumptions on which they are based. In this paper we develop an integrated framework based on a multistate hazard model to describe properties of the simpler standard estimates, identify conditions under which their performance is best, and define the nature of biases and interpretational ambiguities that emerge when empirical conditions depart more than modestly from optimal ones. In particular, we show formally that estimates from limited panel data and two-state hazard models with obesity as a covariate, the workhorse in this area, produce estimates that are difficult to interpret and compare across studies and, in some cases, biased. Finally, we propose a simple procedure that can be employed when the use of conventional two-state models is risky and illustrate its application to an a empirical case.
I10: Health: General
C33: Multiple or Simultaneous Equation Models: Models with Panel Data; Longitudinal Data; Spatial Time Series
D81: Criteria for Decision-Making under Risk and Uncertainty