Training

Ordered scales with five or more response categories are commonly used in surveys. There is an ongoing debate among researchers as to whether such scales should be considered ordinal or interval. IBM SPSS Statistics contains procedures capable of handling such variables under either assumption. When in doubt about the measurement scale, some researchers run their analyses using two separate methods – each making different assumptions about the scale of measurement. If the results agree, the researcher has greater confidence in the conclusion.

There are both practical and theoretical reasons for treating ordinal scales as interval for purposes of analysis. Practical reasons include:

• General Linear Model (GLM) techniques are more powerful than categorical data analysis methods and the results can be presented more concisely. Summarizing the difference between two groups in terms of their mean response is simpler than having to describe a large table.
• Strong relationships will not be distorted no matter what technique is used; weak or non-existent relationships will also not be distorted no matter what technique is used.
• The techniques have worked well in practice through years of experience on all types of data.

While the last point is generally true, the logic of this rationale doesn’t always hold when variables have odd distributions. Consider Figure 1:

Figure 1: Two dissimilar response scale distributions.

Variable 1 has a bimodal distribution, with the greatest number of responses at the endpoints of 1 and 5 on the scale. Variable 2 has a symmetric distribution, about as bellshaped as can be managed for a five point scale, with a mode of 3. But the mean for both distributions is 3, even though these are clearly very different response patterns. GLM techniques may or may not distinguish between these two sets of responses, but categorical variable techniques invariably will do so. Thus, the mean is a poor measure of central tendency for variables with odd or highly skewed distributions. This example illustrates the importance of reviewing variable distributions before using GLM methods, especially for typical Likert scales.

The theoretical argument for considering response scales to be interval is more compelling. Consider a typical measure of overall customer satisfaction on a five-point scale. In theory, you conceptualize customer satisfaction as being a latent continuous variable. It is interval as you conceptualize it, not ordinal. The method you use to measure it is crude — a five-point scale — but that will simply introduce error in the responses. The response scale itself does not alter the fact that you are measuring a continuous quantity — overall customer satisfaction. By this reasoning, the use of GLM techniques with response scales is justified so long as you proceed with caution.

Below are additional recommendations:

• Use whatever technique is most appropriate for your situation. If you believe that discussing mean differences on a Likert scale will be most meaningful for you and most appropriate for the goals of your survey, then use GLM techniques. If you believe that discussing percentage differences will be more useful, use categorical methods instead.
• Avoid using GLM techniques on scales with fewer than five points.
• Whenever possible, use both GLM and categorical (nonparametric) methods to study the same relationship.
• There are sophisticated techniques available to “correct” for the non-interval nature of a scale (such as optimal scaling regression). Use these when they are available and if you are confident in interpreting and understanding their output.