On 28 Apr 2002 13:25:06 -0700, [EMAIL PROTECTED] (F. Goldhammer) wrote: (quoting me, RU) > > - Does "internal consistency" as the measure of reliability > > make much sense for a time series? How tightly linked are > > the measures? If there is a high lagged correlation, then > > any 'internal consistency' is foolish to think about. Before > > you worry about that ratio, in fact, I think the nature of the > > time-series needs to be worked out. > > > > Is there an external criterion for this quotient, or is it > > intended to predict itself-in-the-future, and nothing more? > > > > Maybe a real-life example would impress me otherwise, > > but what comes to mind here, in the abstract, is that > > a) test-retest reliability would be more convincing; and > > b) even more than the usual cross-sectional report, > > your conclusions will be heavily tied to the character > > of the 'sample'; however well the quotient works, that > > depends on various variances remaining similar.
FG > > Thank you for your answers and comments on my problem. I'd like to go > into detail. > The 18 Items are measurements of concentration performance: > item1=concentration performance in the 1st testminute ... item18= > concentration performance in the 18th testminute. > The quotient [Mean(item7, ... item 18)/Mean(item1, ... item6)] is a > simple measure of the enduring concentration performance of a person. > I'd like to know a way to calculate the reliability (�true variance") > of this measure. I measured N=102 subjects just for once, so I can�t > calculate a retest-reliability. The implication of 'reliability' on a measure is that it should distinguish individuals, or individual performances. You want to distinguish the 'enduring concentration' of persons. I am not sure whether that means that you are expecting interesting changes in the 18 minutes, or not. The computation of that ratio makes me think that the change in performance is what is interesting. The correlations by themselves can't imply much about that. Is there a *trend* for a subject, and trends for separate persons, that account for variance, above and beyond what is accounted for by Personal means? The regression 'by person' (ID) is effectively a measure of the internal consistency -- it is not very interesting, I think. That would correspond to coefficient alpha, where the means and standard deviations are assumed to be equal at 18 measures. What ought to be interesting (as I understand it here) is the regression *trend* of scores across time: it improves for some, worsens for others. Whatever you get as a test for trends between persons is a measure of reliability. I don't know that I have seen it translated into a 'reliability coefficient'. - If I was presenting the data, I would (a) look for a published reliability coefficient devised on trends; or (b) point out exactly how Cronbach's alpha is computed from F, and show how these differences (presumably, as evidenced by their F ) are somewhat-comparably as strong. Or, not. > The lagged autocorrelation of the 18 occassions (18 testminutes): > lag0 1 > lag1 0,95 > lag2 0,89 > lag3 0,76 > lag4 0,58 > lag5 0,31 > Hope this helps you to give me further answers. Many thanks in advance r, r^2, r^4, r^8, r^16 - looks as if the lag-one correlation accounts for everything; or am I remembering that wrong? What -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
