1
$\begingroup$

I keep reading that multiple regression is "just-identified" (df = 0) when viewed as a path analysis. If I'm using unstandardized variables, I believe this. For example, with five variables—4 predictors (IVs) and 1 response (DV)—there are 5(5 + 1)/2 = 15 variances and covariances available to me as data. Then the model needs to estimate 4 variances for exogenous variables, 6 covariances between exogenous variables, 4 paths from IVs to DV, and 1 error variance for a total of 4 + 6 + 4 + 1 = 15 parameters. Okay so far.

Now suppose I want to use standardized variables. Because all variances are now 1, I only have 10 correlations to use as data. On the parameter side of things, I no longer have to estimate 4 variances for the exogenous variables (those are set to 1), but I still have the 6 correlations between exogenous variables, the 4 paths from IVs to DV and the 1 error variance. But now I count 6 + 4 + 1 = 11 parameters. Ruh, roh.

What am I missing here?

$\endgroup$
1
$\begingroup$

I think you no longer have the error variance.

If you know the four regression paths and you know the correlations between predictors, you know $R^2$. If you know $R^2$, and you know the variance of the outcome (it's 1.00), you know the error variance and so this does not need to be estimated.

$\endgroup$
2
  • $\begingroup$ That makes sense. I'll have to continue to think about how this works for more complicated path analysis, but I imagine a similar principle applied. $\endgroup$ – Sean Raleigh Jun 1 '16 at 5:29
  • $\begingroup$ Yep, same basic idea. It's sometimes tricky to work out what's an estimated parameter (which costs you a df) and what's not (which doesn't). Sometimes I think I've got it right, but when I run the model the df aren't what I expect. $\endgroup$ – Jeremy Miles Jun 1 '16 at 15:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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