# Thread: Ordinal Regression: Test of Parallel Lines

1. ## Ordinal Regression: Test of Parallel Lines

Hi,

I am trying to fit a model by using ordinal regression. I have got two co-variates, and one factor with four groups. The dependent variable has 28 points (1.00, 1.25, 1.50... 7).
I use SPSS.

My research model fits nicely and the sig. of the Parameter Estimates of the IVs are quite good.
But; The Sig. of the Test of Parallel Lines is .000.

The Link Function I use is Complementary log-log. When I switch the Link Function to Probit and Negative log-log, my Test of Parallel Lines is fine, but the model don't fit any more.

I have tried regrouping the factors and redusing the number of desimals in the DV to the nearest half (1.00, 1.50, 2.00 ...7). Is there something else I can do, and how can I approach this a little more systematic?

Lotte (I'll be only to happy to send all test data if wanted)

2. Hi Lotte,

I don't know too much about ordinal regression, but 28 points for your DV sounds like rather a lot for this approach. It seems quite unsurprising that model slope coefficients would vary across response categories when there are so many response categories.

Sorry to go back to basics, but I'm wondering what happens when you fit a simple linear regression model? Is there a clear problem with the distribution of residuals that precludes this approach?

3. Hi,

Linear regression is not the answer as the DV and IV are not linear, the data is homoscedatic. Also I don't get significant results, and I am comparing respondents that are not randomly assigned to the groups.

No, ordinal regression is the answer. Also because it can handle small sample sizes.

New development: I have managed to pass the Test of Parallel lines by rounding the DV of to the neared whole digit.

New question: Am I allowed to do that?

Also, the significance of the different IV have suffered a bit (as the DV has become less sensitive, I suppose).
So another new question: Are there other ways of restructuring the scale of the DV?

F.Y.I: The original DV had two decimals because it was an average of four items measuring the same construct [ DV = ( V1 + V2 + V3 + V4 ) / 4 ]. The items measured an attitude on a seven point Likert scale.

All the best,
Lotte

4. Originally Posted by Lotte
Linear regression is not the answer as the DV and IV are not linear, the data is homoscedatic.
Hi Lotte,

What do you mean in saying that the DV and IV are not linear? Do you mean that their relationship appears to be nonlinear? How have you determined this? Homoscedasticity is a good thing - it is an assumption of OLS regression. Or do you mean that the data is not homoscedastic?

Originally Posted by Lotte
Also I don't get significant results, and I am comparing respondents that are not randomly assigned to the groups.
I would be very cautious about making decisions about which analysis to run based on whether or not you get significant results, this shouldn't really be a criterion for choice of analysis. Random assignment to groups within a factor variable is not an assumption of linear regression AFAIK (rather, it's a consideration with regards to internal validity).

Originally Posted by Lotte
No, ordinal regression is the answer. Also because it can handle small sample sizes.
That's interesting, I didn't know that ordinal regression handles small sample sizes better than OLS regression. Have you got any papers demonstrating this?

Originally Posted by Lotte
New development: I have managed to pass the Test of Parallel lines by rounding the DV of to the neared whole digit.

New question: Am I allowed to do that?
It'd be interesting to hear the thoughts of others on this, but my 2 cents would be that manually altering data values to obtain the result you want isn't a great way to go about things. In this case, the significance of the test of parallel lines and of the IV's will have been reduced simply because in rounding DV values you've removed some variability from the DV values. In general, removing variability (e.g. by dichotomising continuous variables) is a bad idea.

 Tweet

#### Posting Permissions

• You may not post new threads
• You may not post replies
• You may not post attachments
• You may not edit your posts