How to treat differing likert scales

[Caitlin Hopkins]

Hi all,
I am using panel data from British Election Survey to perform a fixed effect OLS regression looking at how party contact affects a persons political preferences (e.g. probability of voting for certain political party).
The main independent variable is a dummy variable for party contact but the rest of the independent variables are likert scales for the different questionnaire answers. However lots of these scales are not the same, some are 0 to 10, other 1 to 5 and some 1 to 7 etc. Would it be correct to leave them as they are as the unit increases is what will be picked up by the regression? or should I normalise or standardise the variables?
many thanks
Caitlin

[Joseph Coveney]

Your coarsest item is five categories. Set up a measurement submodel with all of the Likert scale items using sem. Your predictor would be in the structural submodel. See

help sem

for further details, and for options to assess measurement invariance and so on between party contact groups if that's desired.

If someone among your research team wants to be persnickety, then use gsem with an ologit or oprobit link function.

[Joseph Coveney]

I'm sorry, I misunderstood: all of the variables you're referring to are predictors. I would find their regression coefficients more interpretable leaving the variables as they are, but maybe that's just me.

[Phil Bromiley]

There has been substantial discussion on this listserv about standardizing variables. When variables have a meaningful scale (e.g., $1 to $3) , standardizing seems to make it harder to interpret what is going on. However, your Likert scales don't really have that kind of meaningful scale. On a Likert scale, a 2 is bigger than a 1, but we wouldn't want to say it is twice the size.

Whether you find it easier to understand standardized or not standardized I suspect is a matter of taste. Note, you are assuming that these are not simply ordinal scales while they really are just ordinal scales. That is, you are assuming 2 is twice a 1 when it really isn't.

However, a lot of research has been done treating such scales as if they are ratio. You might see how this is generally handled in your field.

[Joseph Coveney]

And not only what Phil said, but even if the measurement scale is completely ratio, you don't really know whether four days' exposure will have double the influence on the outcome variable as two days'. You'd need to explore these kinds of possibility with, say, polynomial terms or nonlinear transformations.

If this ratio-versus-ordinal business presents too much of a quandry, you can always treat the Likert scale items as factor variables. Then you can use contrast and its orthogonal polynomial contrast operators (p. and q.) if you want to explore what I mention above.