I am trying to conduct an ordinal regression, but I am really struggling. When I was reading about how to interpret the results, I found out that the value of 'zero' is important in this type of regression. For example, for the thresholds, the model assumes that all independent variables are zero. However, I have not recoded my Likert questions so that the lowest answer has a value of zero. Is this the intention?
Should I recode the lowest value of Likert scale answers to zero (ordinal regression)?
Many thanks in advance for helping me out!
Recoding your Likert Scale so the lowest value is zero (e.g., 1 to 7 -> 0 to 6), or really any linear transformation to any variable (e.g., 1 to 7 -> -3 to +3; 1 to 7 to 0% to 100%), should not change the results of your regression. In particular, it won’t change the p-values, standardized betas, or effect sizes. It could change the beta (if, for example, you convert from 1 to 7 -> 0% to 100%), but simply subtracting 1 won’t even change that (because you didn’t change the slope). The only thing that changes is the constant. If feel the easiest way to think about regressions like high school algebra. Afterall, dependent_variable = beta1* var1 + beta2 * var2 + beta3 * var3 + constant is really just a more elaborate version of y=mx+b. So you can think of the constant in a regression like the y-intercept in y=mx+b; it’s the value of your dependent variable when all your independent variables are 0. In this respect, it is weird to have your Likert scale range from 1 to 7 since you would never have 0 as the value of that variable. But usually we don’t bother interpreting the constant. Rather than recoding your variable, you might plug your neutral values into your regression equation (e.g., 4 for a 1 to 7 Likert) and then you can calculate directly what your dependent variable would be under this assumptions. You could even plug & chug over and over again for different possibilities (e.g., neutral, mean, mean plus/minus 2 sd, min, max) and that can help you interpret each of your participant’s actual scores. Hope this is somewhat helpful Marieke. ~ Kevin
I suppose the answer may depend on the software package you are using, but as @KevinGrobman notes, it shouldn't be necessary to recode your data.
In R's ordinal package, for example, the dependent variable doesn't need to be numeric at all. That is you could use: "strongly disagree", "agree", etc., as long as you tell the software what order the categories follow.
Dear Dr Lauder,
I totally agree with Kevin and Salvatore. If you are using an ordinal scale, the rank is important, not the number itself. As long as bigger is expressed as a bigger number (e.g. strongly agree would be considered bigger than agree etc.), you would just need to make sure that "strongly agree" is assigned a bigger number than "agree".
A recoding of your Likert scale would therefore not be necessary. Let me give you the following example:
the following rank order 1 - 2 - 3 - 4 - 5 - 6 - 7 is equivalent to this one:
0 - 15 - 1000 - 1001 - 250000 - 250001 - 1000000
In an ordinal scale, the numbers per se do not matter as long as the rank order is not changed.
Yours sincerely,
Rainer
Transforming different Likert scales to a common scale
http://www-01.ibm.com/support/docview.wss?uid=swg21482329
- Badges
- Science topic
- Similar topics
- Philosophy
- Esthetics
- Interpretation