2
$\begingroup$

HI there: I have a one likert item I would like to analyze. I know that most scales require multiple likert items, but this is just an exploratory exercise of an existing survey. It's out of my control.

When I treat the likert categories (strongly support - support - oppose -strongly oppose) as a numeric variable (coding 0 to 1) and conduct a multiple regression, I get some predictors in my models as statistically s ignificant. When I try to create some predicted values based on different configurations of my indendent variables, however, some combinations of values create predicted values larger than 1, which is impossible given the original scale.

Incidentally, when I recode the four categories to two (support - oppose) and rerun it with a binomial logistic regression, the same predictors do not appear as statistically significant.

I have three questions: in the first scenario, is it meaningful and / or appropriate to rescale the predicted values so that the maximum predicted value is 1 and the minimum is 0? Or is that not appropriate?

Second question, can someone explain why the predictors would not be showing up as statistically significant when turning to logistic regression? Is there important variance at the poles of the DV (between strongly support and support) that is being lost when squishing them into binomial?

Third question, should I just analyze this as ordinal?

Thank you very much for your time.

I am conducting this on R 3.0.2. Simon

$\endgroup$
  • $\begingroup$ What method(s) would you use to "just analyze this as ordinal"? $\endgroup$ – whuber Apr 7 '14 at 18:16
  • $\begingroup$ Maybe something like this? ats.ucla.edu/stat/r/dae/ologit.htm, basically doing three separate logistic regressions. One for moving from strongly support to support; support to oppose; and one from oppose to strongly oppose? $\endgroup$ – spindoctor Apr 7 '14 at 18:20
  • 1
    $\begingroup$ Yes, that would be an excellent choice. Note that these are not truly separate logistic regressions: they are parallel regressions (as expressed by the proportional odds model). $\endgroup$ – whuber Apr 7 '14 at 18:28
  • $\begingroup$ Right, I actually knew that. I remember that term, parallel regressions. OK, thanks a lot! Does the ordinal package in R handles this stuff? $\endgroup$ – spindoctor Apr 7 '14 at 18:34
  • 1
    $\begingroup$ Why don't you read over the page you linked to in your comment? It does a wonderful job of describing how to perform, interpret, and check an ordinal logistic regression with R (using the MASS, HMisc, and GGPlot packages). $\endgroup$ – whuber Apr 7 '14 at 18:35

Your Answer

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

Browse other questions tagged or ask your own question.