Can Variation in Practice Patterns be Used to Estimate a Causal Treatment Effect from Observational Data when Patient Treatment Assignment is Highly Confounded?
Laine Thomas, Duke University
Recent publications in medical journals adopt an intuitive, although unsubstantiated, approach to estimating causal treatment effects. Hernandez et al. (2010), an influential paper in JAMA, identifies survival benefit from early follow?up with a physician after hospitalization for heart failure. Early follow-up is targeted to the sickest individuals. Measured comorbidities are not adequate to capture this confounding. The researchers argue that unmeasured confounders are balanced across hospitals; selection is based on proximity rather than severity of illness. Additionally, hospitals are expected to differ in terms of policy regarding the importance of early follow?up. The idea is to exploit this variation by estimating and imputing the hospital?specific proportion treated for all patients within each hospital and evaluate the association between this variable and time?to?event. Although akin to instrumental variable methods, a methodological foundation for this specific strategy is needed. Does it estimate the effect of interest and under what assumptions? What sample size is desirable for hospitals? How does the power for detecting a treatment effect depend on the amount of true variation between sites? Should the relationship between hospital proportion treated and outcomes exhibit a dose effect (linearity)? Preliminary results will be presented.
Extended Approaches for Analyzing Longitudinal Data with Time-dependent Covariates
Yi Zhou, John Lefante, Shande Chen, and Janet Rice United Therapeutics Corporation
Quadratic Inference Functions (QIF) and estimating equation using conjugate gradient method (CGM) for fitting marginal model to longitudinal data show appealing feature in improving the efficiency without making assumption on the correlation structure. However, our simulations show that both methods produce biased and inefficient estimation on regression parameters when time dependent covariates are present. In this paper we extend both QIF and CGM methods for fitting marginal model to longitudinal data with time-dependent covariate. The idea is to restrict the moment conditions to the ones which are only valid to certain type of time-dependent covariate. Our simulations show that efficient estimation on regression parameters is achieved using extended approaches. Further we apply the extended approach to anthropometric screening data to evaluate the association between body mass index and morbidity in children in Philippine.
Plackett-Burman eXtended (PBX) Designs
M. B. Hardy Applied Research Associates, Inc. Raleigh, NC
PBX (Plackett-Burman eXtended, to estimate main effects in physics-based engineering models) uses Plackett-Burman designs to generate designs requiring a number runs within 4 of the number of parameters to be estimated. By equivocating PB columns with parameters of categorical variables, or multiple order terms of numeric variables, mixed categorical and numeric designs are possible, including higher order terms and interactions. D-optimality seeks to maximize the determinant of the information matrix X'X of a design, so these matrix calculations must be performed at each point in the candidate set. No such search and calculation is required for PBX, as the points are generated directly. In another approach, just as Box and Behnken combined two-level fractional factorial designs with incomplete block designs, two Plackett-Burman designs can be combined in similar fashion, producing similar but smaller designs. Axial points of length sqrt((k+1)/2) allow estimation of univariate cubic and quartic effects. Since each column has 50% 0s and 25% each +1s and -1s, multiplying each column, j, by sqrt(2)*stdev(Xj) and adding mean(Xj) prior to simulation, produces a "sample" of Y having roughly the first and second moments of Y.
