Likert Scales and Averages
Tired of being told again that you shouldn’t be calculating averages from your Likert scale data? Well, here’s your answer. It has been strongly argued that it is incorrect to use the measurement type of data (e.g. nominal, ordinal etc) to determine the statistical method appropriate to analyse it. Instead, the question it is designed to answer is what should determine the statistical analysis to be applied.
Some background
Likert scales have long been a very successful way to measure customer sentiment such as customer satisfaction, perceived value and so on. But along with their success has come much criticism. From a statistical perspective, the most significant criticism is that the data produced by Likert scales is, at best, ordinal and not continuous - as is assumed by the statistical analysis most commonly applied to it e.g. calculation of means, hypothesis tests etc. So technically, such analysis should not be carried out on Likert scale data.
Such strictness in classification, however, is at odds with reality. Consider this - when measuring customer satisfaction, the underlying construct (latent variable) is, in fact, continuous. However, it is captured in an ordinal form, because respondents simply don’t verbalise their level of satisfaction as, say, 67.25 per cent. The resulting data captured, although ordinal in type, actually describes a metric whose behaviour is best aligned with a continuous, normally distributed variable. It makes perfect sense then to apply statistical methods appropriate for numeric data.
Juster’s 11-point probability scale provides further validation for treating Likert data as continuous in its analysis. Juster argued that since customer ratings/intentions are just disguised probabilities, why not collect the probabilities themselves? This is possible according to Juster because everyone is familiar with the concept of chances out of 10. So whilst this 11-point scale looks similar to the usual ordinal ones, it becomes metric through the introduction of probabilities.
A final point to consider is that the data type can vary depending on the question it is used to answer. For example, tickets given to everyone who attended a reception can be treated as ‘nominal’ for determining ‘who won the door prize’ and ‘continuous (eg. ratio)’ for the purpose of estimating ‘how many people attended the function’.
Likert scales and averages?
Clearly then, the appropriate analytical techniques should be decided by the questions the data was collected to answer, not by some arbitrary, pre-determined, theoretically strict classification system of data types that is independent of the purpose for which the data itself was generated. This means that it is acceptable to calculate averages and other arithmetic on data collected using Likert scales. It is recommended, however, to have at least seven points on the scale.
