On Tue, 26 May 2020 22:22:40 -0700 (PDT), Scarlett Brown

<***@gmail.com> wrote:

Observations and implied questions.

"Cronbach's alpha," from the Subject line, is a measure of

the within-scale similarity of items, for somehting that is

considered to have one "latent factor" or meaningful dimension.

It is typically measured at one point of time. It is computed

from variance terms, in several versions, and the usual version

is a transformation of the average correlation of the several items.

*Post by Scarlett Brown*im creating a three way repeated measures ANOVA.

I'm not sure how that is relevant. However, I do not see "3-way"

in what is described below. There are two time periods (as I read

it) and three variables.

I would probably be most interested in the alpha and correlations

among the three as measured at Pre. If Treatment has a strong

effect, the correlations are apt to change.

*Post by Scarlett Brown*I have two intervals, for before the condition is applied and after.

I have three dependent variables that have been reverse coded,

- if they are all scored in the same direction, that is not needed.

*Post by Scarlett Brown*and have checked the matrices to make sure theyre correct.

? "matrices" ? I would look at one correlation matrix.

I would look at Frequences and Crosstabs, to see that the

scores have been Reversed correctly, especially when there

are other results that need explaning.

*Post by Scarlett Brown*My scale has been collapsed down to a three point scale, ranging from positive to negative. -1, 0, 1.

I assume that means what I would write as, "My scaling on items

has been collapsed to three points...." I would reserve the bare term

"scale" for the composite score you may compute from the three

items.

You are throwing away detail by collapsing -- some reviewer

would demand justification for that. Collapsing to 3 points also

reduces the size of the correlations that you should expect, owing

to reduced precision. A correlation of 0.5 is nearing the limit of

reliability among dichotomous preference items, comparable to

0.7 (my guess here) for 4-point likert-type items.

*Post by Scarlett Brown*I do not understand why my data is weakly correlated as they are being tested on the same dimensions.

The possibilities for apparent data-error include data that were

recorded wrong, read wrong into SPSS, transformed wrong,

interpreted wrong. Scores that seem to share a dimension

judged by face value should be correlated, or be explained.

One thing that happens -- especially with over-collapsing of

scores -- is that you can get extremely skewed distributions.

Those can give you correlations beginners won't expect until

they have run into them, or considered technical computations.

For instance, when two raters Agree on 98 diagnoses out of 100,

the table (98, 1; 1, 0) will produce a negative r despite the 98%

"agreement." And if one rater is 100% Yes or No, you can't

compute a correlation at all.

Every reliability figure is a measure of a test IN some sample.

Occasionally, the samples are odd. Always, we should look at

and keep in mind what variability (including skewness) exists

and is available to be anaylzed.

--

Rich Ulrich