Jan Ivanouw posted on Tuesday, March 29, 2016 - 5:19 am
I have a problem with how to formulate known reliability/measurement error for observed measures
If for example from earlier research it was known that in the User Guide ex 3.1 x1 is measured with a scale reliability of .7 and x3 is measured with a reliability of .6. How would this be represented in the input file?
See Slide 44 of the Topic 1 course handout on the website.
Jan Ivanouw posted on Tuesday, March 29, 2016 - 2:51 pm
Thank you for your answer. I have a problem, though.
From slide 44 on topic 1 I made the following modification for the input file ex3.1 assuming reliabilities .7 and .6 for x1 and x3:
DATA: FILE IS ex3.1.dat; VARIABLE: NAMES ARE y1 x1 x3 ; MODEL: f1 BY x1@1; x1@.3282; ! (1-.7)*1.094 (sample variance of x1) f3 BY x3@1; x3@.3828; !(1-.6)*0.957 (sample variance of x3) y1 ON f1 f3; f1 WITH f3@0; OUTPUT: stand;
In User Guide ex 3.1 x1 and x3 are presumed to be measured perfect (without measurement error). This gives the regression coefficients (StdXY) of 969 (.653) og .649 (.409) for x1 and x3 on y1. However, when introducing measurement error as in the input file above, the regressions coefficients are 1.385 (.787) og 1.082 (.532). They are assumed to be lower due to attenuation, but in fact they are higher. What is wrong here?
Jan Ivanouw posted on Tuesday, March 29, 2016 - 2:57 pm
Error: In the previous post I naturally cited the regression coefficients for y1 on x1 and x3, not the inverse.
Jan Ivanouw posted on Wednesday, April 13, 2016 - 7:04 am
Hi Maybe my error message was misunderstood. I still have the problem described in my post of 29 March, that my model with measurement error based on slide 44 gives higher regression coefficients then the model withoug measurement error.
As the reverse should be true, that a model without measurement error should give higher regression coefficients than a model with error, there must be an error somewhere.
Did I misspecify the model in my post of 29 March? Or is there another explanation?