Author Topic: zero correlation  (Read 507 times)

Georg II

  • Newbie
  • *
  • Posts: 21
    • View Profile
zero correlation
« on: June 27, 2012, 04:59:07 AM »
Hello Ray,
In multidimensional models what happens if we force all correlations among dimensions to be zero?
How is this different from the (same analysis) when we allow the dimensions to correlate?

Thanks
Georg II

ConQuest

  • Global Moderator
  • Newbie
  • *****
  • Posts: 24
    • View Profile
    • Email
Re: zero correlation
« Reply #1 on: June 30, 2012, 03:22:30 AM »
It constrains the correlations to be zero, that is they are no longer free to be estimated but they are set to a zero value.

Georg II

  • Newbie
  • *
  • Posts: 21
    • View Profile
Re: zero correlation
« Reply #2 on: June 30, 2012, 07:15:16 AM »
Thanks Ray,
But constraining correlations to zero changes the difficulty and ability parameter estimates slightly and the deviance substantially. When dimensions are free to correlate fit is better (deviance is smaller) compared to when correlations are set to zero.

When and under what conditions should we constrain correlations to zero?

Thanks

ConQuest

  • Global Moderator
  • Newbie
  • *****
  • Posts: 24
    • View Profile
    • Email
Re: zero correlation
« Reply #3 on: July 04, 2012, 02:22:44 AM »
Allowing a correlation is a more general model, so it should fit better.  Fixing the correlations at zero has the same effect as running each of the scales separately.  If they are correlated you'll increase reliabilitly because you'll be drawing strength across dimensions.