I guess if symptom A and symptom B are in the same nature, than my design matrix (see at the end below) would be good right? But then would there be issues regarding whether the symptom A and symptom B (counts) are correlated or not? Obsessive-compulsive symptoms are one type of anxiety symptoms, which are commonly comorbid with depressive symptoms. If I am just comparing anxiety symptoms with depressive symptoms, according to what you responded, my design matrix below would work fine, am I correct? But then whether the fact that the anxiety symptoms and depressed symptoms are statistically correlated with each other become an issue? However, different subsets of the anxiety symptoms (e.g., obsessive-compulsive symptoms, or phobic symptoms, etc) are not statistically correlated with the depressive symptoms. Will you please help let me know if, regardless of the definition of the nature of the symptoms that I am intended to comparing, whether my design matrix at the bottom would work for a direct comparison between two sets of symptoms' relationship with FA?
Regarding the alternative approach you mentioned, I wonder what is the difference between contrast [0 0 0 1 0] vs. [0 0 0 -1 0] and how I am to determine the direction of the interaction, i.e., whether the slope of FA with symptom A is greater or smaller than the slope of FA with symptom B? This is not F-test right? Thank you so much for your help!
Date: Wed, 9 Aug 2017 10:22:00 -0300
From: "Anderson M. Winkler" <[log in to unmask]>
Subject: Re: Help with GLM for testing Correlation Difference in a single
This design has issues I'm afraid, because we cannot add or subtract
regression coefficients that have a different meaning (e.g., we cannot use
+1 and -1 in the contrast for the two sets of symptoms, because these are
of a different nature, like apples and pencils).
It seems you'd like to see:
1) Whether the number of symptoms of set A is correlated with FA (main
effect of symptom A count);
2) Whether the number of symptoms of set B is correlated with FA (main
effect of symptom B count);
3) Whether there is a difference in the slopes of the two above comparisons
To do this, make the design as follows:
EV1: intercept, as you already have.
EV2: Symptom A count, as you already have.
EV3: Symptom B count, as you already have.
EV4: Product of EV2 and EV3, for an interaction EV.
EV5: Covariate, as you already have.
The contrasts are then:
C1: [0 1 0 0 0]
C2: [0 0 1 0 0]
C3: [0 0 0 1 0]
You can add 3 more contrasts to test the opposite effect (the same as
above, but with -1 instead of 1).
Hope this helps.
All the best,
On 7 August 2017 at 20:45, Yuwen Hung <[log in to unmask]> wrote:
> Dear Anderson,
> Sorry that I thought I answered your questions:
> - Are Symptoms A and B: (a) actual symptoms that can be present, absent,
> graded into some scale, or (b) are these *sets* of symptoms (hence the
> count)? This would clarify why in your design you have, in addition to
> count, there are two binary regressors.
> =>They are symptom count of different sets of symptoms, not the actual
> symptoms. For example, symptom A is the number of depressed symptoms, and
> symptom B is the number of obsessive-compulsive symptoms. These two
> measures are not correlated with each other, at statistical level or at
> individual checklist content.
> Please disregard my previous questions, now my question is can I use the
> following matrix and contrasts to test the differences of correlations
> (slopes) on FA with symptom A versus FA with symptom B, while controlling
> for other covariates? The scores in the following example will be
> SubjectID EV1 EV2 EV3 EV4
> AllSub SymA SymB Covariate
> Subject1 1 2 1 21
> Subject2 1 0 3 25
> Subject3 1 6 2 27
> Subject4 1 1 7 45
> Subject5 1 5 0 18
> Subject6 1 0 1 35
> Subject7 1 3 2 40
> Subject8 1 2 0 22
> Subject9 1 9 7 28
> Subject10 1 1 4 37
> Title Ev1 Ev2 EV3 EV4
> C1 SlopeA>SlopeB 0 1 -1 0
> C2 SlopeB>SlopeA 0 -1 1 0
> Thank you again for your help,