Semantic differential / Likert scale

[Serena Menny]

Hello everyone,

I have a variable that contains two items. One item is measured with a Likert Scale, and the second with a semantic differential. I want to test the effect of these items on the dependent variable, but I also want to examine the effect of the variable as a whole.

1. When testing the two items separately, I'm accustomed to interpreting Likert scale results. However, is there any specific interpretation method for the item measured with a semantic differential scale?
2. Can I aggregate the two items directly, or is there a preliminary step required?

Thank you,

[Daniel Schaefer]

It sounds like you are saying you have two measures (I would say two variables) of the same latent phenomenon. Is that correct?

1. With the semantic differential scale, you want to be careful about how the scale is represented numerically. Do values range from 0 to 1? -1 to 1? -10 to 10? If the scale values range from 0 to 1, then a regression coefficient will look very large because it represents the difference between someone at the far left of the scale and someone at the far right. If the scale ranges from -10 to 10, then a unit change represents a much smaller difference in opinion, so the coefficients will look much smaller.

2. If you are interested in the substantive effect of this phenomenon on some outcome, then I just want point out that these two items are probably tightly correlated and there likely isn't much to gain by aggregating them. If you want to aggregate them anyway to create an index, it depends on how you want to do that. For example, the simplest option would be to just add them together, but if the scales are widely different then one measure will tend to dominate the other. There are many technics to aggregate variables, but I'm most familiar with exploratory and confirmatory factor analysis as a set of techniques to extract the shared covariance. If you want to use either of those techniques, then yes, there is a fair amount of setup involved. If you just want to include both measures in the same model, that's fine if you only care about prediction, but if you want to explain the outcome it may lead to exaggerated coefficients and under-estimated standard errors because of the collinearity between your independent/predictor variables.