 |
|
| Wednesday 14th October 2009, 2.00pm - 5.00pm at Manchester Dental Education Centre (MANDEC), Higher
Cambridge Street, Manchester, M15 6FH (tea will be served about mid-afternoon) (building 41, entrance on corner facing building 35) Joint meeting with Manchester University's Biostats
Group Theme: Measurement Error Stijn Vansteelandt (The University of Ghent) While most scientists appreciate the value of a simple and robust intent-to-treat (ITT) analysis in clinical trials, many challenge the wisdom of relying solely on the ITT summary for major decisions facing a complex drug delivery system. In practice, drug exposure varies over time, as compliance with prescribed dosing regimens tends to vary widely within and between subjects. This affects the relevance of a measured ITT effect for future patient horizons. Research efforts have therefore focused on estimation of the causal effect of observed exposure patterns. In doing so, they silently assume that compliance was accurately measured. Unfortunately, there typically remains a margin of error in exposure measurement, even with today's highly sophisticated electronic monitors of drug intake. In this talk, I will discuss the impact of error-prone compliance measures on compliance-adjusted analyses of randomized clinical trials. I will propose a class of estimators of the causal effect of received treatment on response under linear structural mean models which acknowledge that exposure is incorrectly measured with known error mean and variance. In addition, I will present a new approach which allows correction for measurement error with unknown mean and variance in the presence of an instrumental variable, which is known to be uncorrelated with the error and not to modify the causal effect of interest. The methodology will be illustrated via simulations and a data application. Terence Iles and Jonathan Gillard, The University of Cardiff The presence of measurement error often obscures a linear relationship between two measurements - most undergraduate statisticians are familiar with simple least squares regression as a way of fitting straight lines to scattered data. In this talk we will concentrate on the effects of the presence of measurement error in both variables. Some slightly surprising results will be demonstrated; it turns out that the effect of measurement error on the appearance of the scatter plot depends on the distribution of the measured variables. The talk will include a discussion of the relevance of conditional expectation, and will describe approaches to the problems of prediction and the definition of appropriate residuals that might be used for diagnostic checking or the construction of reference intervals. Eva Batistatou, The University of Manchester Exposure measurement error can lead to substantial bias in assessing exposure effects which can be corrected if repeated exposure measurements (Ws), are available. A single-stage (1S) study design, in which response Y and repeated Ws are measured for all subjects, could be used to adjust for measurement error. For expensive exposure measures though, it is common to carry out repeated exposure measurements only for a sample of subjects, leading to a two-stage (2S) study which produces data ‘missing by design’. Several bias-correction methods of analysis have been proposed (i.e. regression calibration, SIMEX) in the medical literature, some of which can address missing data (eg using imputation methods). Here we will compare these methods in terms of bias and Root Mean Square Error - both for 1S and 2S designs - with recently proposed grouping bias-correction methods, in which means of exposure - across subjects grouped accordingly to similar exposure characteristics - are used as an instrumental variable in the analysis. If you would like to attend this free seminar please contact Wendy Lamb (+44 (0)161 275 5764): |
| Navigation:- External Links:- |