The Exploratory Factor Analysis (EFA) is a reduction technique better known to statisticians as Rotated (Principal) Component Analysis. The distinction is critical to statisticians as classical EFA carries more assumptions than standard market research data can usually justify.
EFA uses information about how a set of measurements are related to one another (known as the correlations between metrics). From that information, it can summarise the larger set of original measures into a smaller set of summary measures, known as factors or latent variables.
The primary purpose of EFA is ineffective if the original set of measurements are (virtually/completely) independent of each other. If there is no correlation between the variables, there can be no useful summary which reduces dimensionality.
However, the correlation between measurements is often an inherent feature of survey measurements, and it must be taken into account to minimise the risk of ‘double-counting’ the effect of (partly) duplicated measures. If not, the most frequently asked question (albeit in different guises) will inevitably appear to be the most important.
In applications, the set of original measures is often a battery of attitude statements or attribute ratings using an Agree-Disagree scale, Bi-polar differentials, Importance, Ratings out of 10 and many other forms. The battery may be ordered into topics (e.g. pricing, product features, image) or maybe randomly sequenced depending on the research design and objectives.
The output describes a range of possible factor solutions, each of which will have a different number of summary factors within it (e.g. a set of 30 rating scales could be summarised into, say, 6,7,8,9,10 or more elements). The fewer factors there are, the more substantial and more general their summary effect is. The more factors there are, the smaller and more specific, their effect. The researcher has the task of deciding, which, of a possible range of factor solutions, is the most likely to meet their needs.
The criteria for choosing a factor solution are usually a blend of art and science.
The science requires that there are enough factors to retain as much information as possible without there being too many, each very specific, to cope with in interpretation. In practice, this often means that researchers who wish to retain a clear understanding of the factors tend to work with a solution which has between 5 and 10 latent classes.
The art requires that the solution can be interpreted and is deemed capable of discriminating between either respondents or their views of products and services. It is possible, for science to confirm which model is most discriminating, but this can lead to data-led, rather than hypothesis-led choices. It is also likely that the most discriminating solution is always the largest, there are more factors from which to choose. The trade-off between art and science is an acquired skill.
With EFA, we are trying to find statements that are related to each other and hopefully come together in the topic areas discovered in any qualitative research. If EFA uncovers new topics, it is often wise to review the qualitative work to see if something is not recognised or not explored.
Please contact us if you would like any assistance in running exploratory factor analysis on your data.