Richard Hoenes wrote in news:[EMAIL PROTECTED]: > On Sat, 15 May 2004 19:42:42 GMT, David Winsemius > <[EMAIL PROTECTED]> wrote: > >>Richard Hoenes wrote in >>news:[EMAIL PROTECTED]: >> >>> >>>>> If it was the ICC he was pushing I wouldn't mind so much, but he >>>>> has insisted we include Bland & Altman's limits of agreement >>>>> (which is simply the mean difference +/- [1.96*stddev] which has >>>>> no signficance test), and he is now systematically having us >>>>> remove every other statistical test we've included in the paper. >>>>> The only other test left in the paper is the paired t-test and now >>>>> he wants a reference >>to >>>>> show it is valid to use. snipped stuff...
The paired t-test >>will be significant at the 95% level in exactly those situations when >>the mean is more than t(0.025,n-1) or t(0.975,n-1) standard deviations >>away from zero. t(0.025, n) will generally be greater than 1.96 but by >>not much. Reporting the confidence interval is more informative than >>merely reporting that the test was "significant", because it defines a >>range of plausible values for the difference. > > We did this when we were told to put in the Bland Altman since we > understood it more and, at least to us, it was more widely used. We > were told to take it out of the paper since we'll have the Bland > Altman intervals. I just looked at the Bland-Altman paper. In their example on p5 they did precisely a paired t-test with the use of the correct t-level based on the number of subjects. I don't know who corrupted that to the mean +/- 1.96sd's but I am sure Bland or Altman wouldn't tell you to use 1.96 in a study with small numbers. http://www.mbland.sghms.ac.uk/ba.htm Below is reference that is a direct response to your original request for a citation in support of using paired t-test for measuring concordance. But since Bland And Altman used the paired-t, too, I cannot see that it is really necessary. Just use page 5 of B-A. Other tests of agreement are also described. It describes the Lin concordance coefficient (AKA CCC, concordance correlation coefficient) which has features similar to a correlation coefficient but will be sensitive to either a difference in means or a difference in standard deviations. Unfortunately the confidence interval for this coefficient is given in another citation. Another possible problem arises because I found a citation saying that it was equivalent to the ICC. (Carrasco and Jover, Biometrics, v 59 n4) http://www3.oup.co.uk/eortho/hdb/Volume_22/Issue_03/pdf/220257.pdf > >>You could use Rosner's "Fundamentals of Biostatistics" or most basic >>stats books for this assertion. The accept-reject formalism and the >>confidence interval formalism have been shown to be equivalent, oh, >>about a half century ago. >> >>If you want to "bring back the Pearson's correlation", why don't you >>instead create a scatterplot of subject pre-post measures. snipped more stuff > The test we are reporting on is actually a series of vision tests with > different scales, so this would entail a large number of plots. We > were told not to include Bland Altman plots because of the large > number, so scatterplots wouldn't work. > You could report summary estimate on a large number of groups that coincide with the Bland-Altman framework by calculating the regression coefficients of X-Y against Mean(X,Y). If you centered the mean, you could also look at the intercept and s.e as a test of drift. The test for agreement would be looking at the slope of the line, its standard eror, and seeing if it was or wasn't plausibly different than zero. The usual caveats about insensitivity to curvilinear departures apply. This all assumes that there was no intervention between the measurements. This is a test for agreement where agreement should exist, right? No interventions or expectation of drift? -- David Winsemius If the statistics are boring, then you've got the wrong numbers. -Edward Tufte . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
