We have a metric dependent variable (impairment, scale 0-40), and a larger number of symptoms of a mental disorder, each on a scale of 0 (nonexistant) - 3 (very severe).

We try to find out whether symptoms are associated with different levels of impairment. We test this in two steps. First, we use a linear regression, predicting impairment by individual symptoms, in which symptoms explain about 45% of the variance of impairment. Second, we use the LMG metric of the R package RELAIMPO (Grömping, 2006; "relative importance") to allocate unique R^2 shares to each symptom. The result is that individual symptoms vary drastically in their explained variance of impairment. The estimate range between 1% and 20% explained.

This fits nicely with our hypothesis:

Some symptoms are more impairing than others, because they are different things (and not simply interchangeable indicators of the same latent). Impairment depends on the nature of symptoms (thus, sum-scores of symptoms might obfuscate information).

However, a reviewer put forward an alternative hypothesis:

Impairment depends on the severity of the symptoms, irrespective of the nature of symptoms. This hypothesis would state that all symptoms that explain a lot of variance are symptoms that have a high mean, and that symptoms that explain moderate amounts of variance are average severe symptoms, and that symptoms that explain only little variance are only slightly debilitating.

I don't quite know how to test this idea, and would very much appreciate input.

(EDIT: A colleague of mine suggested to test a null hypothesis model that uses the score sum only (i.e. forcing all betas to be identical) against the model with individual betas (13 degree of freedom test), but frankly, I don't understand what that would do and how it would help).


1 Answer 1


Maybe I am missing something but why not just look at the relationship between importance (from your previous output) and average severity of each symptom? You could do a regression with DV = importance and IV = average score.

  • $\begingroup$ Peter, I used a correlation instead of a regression, but I guess the idea is basically the same: find shared variance. However, the data matrix is tiny (14 symptoms x 2 rows, 1 for importance, 1 for average score), I hope that is no problem? The correlation is .45, and I am not sure about the interpretation: "the importance of a symptom on impairment is positively associated to the mean, but the mean does not entirely explain importance?" Would you recommend a regression over a correlation? $\endgroup$
    – Torvon
    Mar 28, 2013 at 16:28
  • $\begingroup$ They will give similar meanings; I thought of regression because you clearly have one dependent variable. $\endgroup$
    – Peter Flom
    Mar 28, 2013 at 16:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.