|
|
What are "completely standardized est... |
|
Message/Author |
|
|
hi there, first, apologies for starting a new thread, but i was not able to find a thread that seemed applicable. i have written a paper in which dichotomous (present/absent) diagnosis variables are being used to predict a latent variable. in response to this paper, an editor states: "For the binary dummy coded predictors (e.g., presence/absence of diagnoses), it would be more informative to report standardized coefficients (i.e., which reflect the SD unit change in the latent variable outcome given the presence of the diagnosis, as opposed to a standardized score increase in the diagnosis which would be reflected by a completely standardized estimate). " i *think* the editor is stating that for the diagnosis variables, what is labeled in mplus output as "std" would be more informative than the "stdxy" estimates. i have not been able to find a reference to "completely standardized estimate" in the mplus pages, so i wanted to check to make sure my interpretation makes sense. relatedly, if the editor is correct in this request, then it seems to me that other categorical predictors (e.g., a four-response-option measurement of education) should be reported similarly. does that seem reasonable? thanks in advance for any help on resolving this issue. best, tom |
|
|
I think you are seeing this correctly. When we teach we emphasize that you don't want to standardize wrt categorical independen variables, so for such variables you would std. |
|
|
thanks for the response. a follow-up question: if the predictors are ranked by stdxy, then the size of the coefficient seems to correspond pretty well to the level of statistical significance. however, if the predictors are ranked by the size of std, even predictors with the same response scale (0 and 1) don't show that much correspondence between the size of the coefficient and the level of significance. for example, a -.26 is strongly significant (stdxy = .05) whereas a .26 is not (stdxy = .02). it seems counter-intuitive to me and makes me wonder if something is wrong with the analysis. (so i'm sure a reviewer will have a similar issue.) does it perhaps mean that the .26 would be a stronger effect in theory but there is less confidence in it because there are fewer instances of "1" in the data? (the predictors do have different rates of people coded 1 in the data.) thanks again for any feedback you can give. best, tom |
|
|
Even though two covariates are binary the same slope value can have very different SE because as you say the proportion of people is different for the two. |
|
|
yup, that checks out with the output i have, and makes a lot of sense. thanks for the help! best, tom |
|
Back to top |
|
|