Comparing methods of measurement: Extending the LoA by regression

Abstract

Method comparison studies are usually analyzed by computing limits of agreement (LoA). If only one measurement by each method is taken on each person, and the difference across the range is not constant, it has been suggested (Stat. Methods Med. Res. 1999; 8:136-160) to regress the differences on the averages and use the resulting equation to construct LoA. LoA can be converted to a prediction formula for one method given a measurement by the other. The meaning of the regression of differences on means is clarified in the framework of a proper model and prediction equations linking one method to another are devised. The performance of this model based method is evaluated against the simple approach proposed earlier and against the Deming regression.

Evaluation of the relationship between capillary and venous plasma glucose concentrations obtained by the HemoCue Glucose 201+ system during an oral glucose tolerance test

“…Mean capillary and venous glucose concentrations obtained during the OGTT are shown in Table I According to Carstensen [15], equations were calculated for conversions between the two measurement methods. The equations, conversion lines and 95% p.i.…”

“…Altman, in which differences between paired measurements are plotted against the mean of each pair [14]. Conversion equations were derived according to the method described for not constant differences [15]. To study the agreement between categories of glucose tolerance obtained by either capillary or venous glucose measurements a cross [16].…”

The Bland-Altman Method Should Not Be Used in Regression Cross-Validation Studies

“…Our concern is with its inappropriate use in regression validation studies: as it is biased, there is no purpose for using the BA method in regression validation. There is some debate regarding A B the advantages of the BA method for measurement agreement and other types of studies (Aspden, 2005;Atkinson & Nevill, 1998Batterham, 2004;Carstensen, 2010;Hopkins, 2000Hopkins, , 2004Hopkins, Marshall, Batterham, & Hanin, 2009;Krouwer, 2008;Twomey, 2006). Investigators who are considering using the BA method to evaluate their data should familiarize themselves with this literature.…”

Visual acuity and contrast sensitivity screening with a new iPad application

“…We checked the hypotheses of constant differences and constant standard deviations by means of a Deming regression (function DA.reg) [31]. If the corresponding p values for both hypotheses were significant (p < 0.05), conversion equations were employed on the plot, and mean differences (MDs) or LoAs were represented considering linear correlations (function BA.plot, parameters dif.type = ''lin", sd.type = ''lin").…”