Response to Greckhamer et al (2018): Studying configurations with qualitative comparative analysis

Gross calibration

I agree that QCA fuzzy set analysis permits fine-grained attribute distinctions or calibrations. However, the majority of QCA studies, including most I have referenced, did not avail themselves of such distinctions, and are especially challenged when considering attributes with interval or ordinal properties (e.g. size, ownership concentration). For many studies, the ‘qualitative thresholds’ for such attributes would be difficult to ascertain a priori.

Sample size and clustering versus QCA objectives

Certainly, QCA can be used with substantial samples and in fact I have referenced numerous large sample QCA studies. Unfortunately, these studies have tended towards gross calibration and have made little use of rich qualitative data – a key potential strength of QCA.

My larger point was that clustering and QCA are to be used for different purposes, the former for early inductive research and pattern recognition (not merely ‘to determine which cases are more similar to each other’ (p. 5)), the latter for more theoretically directed exploration and testing of causal relationships associated with particular outcomes.

As I have suggested, there is an exploratory role for clustering and other ‘Q-techniques’, mostly as an aid in the early detection of common organizational types or other patterns, especially when there is too little theoretical guidance to be had in the early stages of classification or typology generation. In those circumstances, one is trying to locate and characterize the densest parts of a data space consisting of many subjects, ideally described in some detail along the variables needed to encapsulate a research question. By using different clustering techniques, it is often possible to detect whether a relatively small number of richly characterized and internally homogeneous clusters account for a large percentage of a sample (Miller and Friesen, 1984). As with QCA, it is important to specify the rationales for sample and variable selection and the use of particular clustering algorithms.

Having identified some promising clusters, hold-out samples should be used to establish the robustness and predictive significance of the clusters (see Miller, 2017; Miller and Friesen, 1984). Subsequently, one reverts to qualitative data and perhaps where warranted QCA – the former to help discover themes, better understand processes and generate hypotheses and the latter to establish associations with particular outcomes.

Of course, like clustering, QCA has its own subjectivity and reliability challenges, including those relating to inappropriate causal inference (see the detailed critiques and commentaries by Collier (2014), Hug (2013), Lucas and Szatrowski (2014), Ragin (2014) and Seawright (2014). In addition, the very significant conceptual and methodological demands of QCA are made very clear by Greckhamer et al. (2018), confirming the selective situations for which the method can be used (e.g. due to important limitations on the number of included conditions, required calibration and rationales). Indeed, QCA relies on a good deal of prior knowledge about a research domain – a condition that is by no means universal. ¹ That is less true of clustering approaches.

In short, with clustering, the intent is exploratory – to find common, richly characterized types and important distinctions among them, and then to use these as starting points for subsequent qualitative investigation, theory generation and hypothesis testing. These techniques are not intended to establish causality or to determine whether qualities are complements or substitutes with regard to particular outcomes.

More generally, there is a difference between a method’s potential and the ways in which it has been most commonly used, perhaps because of the demands of the method itself. A portion of my essay was directed towards the challenges associated with how QCA and clustering methods have been used in previous studies, often departing from their ideal deployment. Both my own and Greckhamer et al.’s (2018) essays have attempted to address these challenges.

Finally, there are alternative ways of discovering and investigating configurations that go beyond clustering and QCA methods; some of these are theoretically driven typologies, others a product of purely inductive qualitative research and others still operationalize configuration as a quality or degree of alignment among organizational properties (Miller, 1996). At the end of the day, it is the quality of the empirical and conceptual outcomes, and the theoretical and practical insights these yield that matter the most.