It is important to limit statistical testing of context–mechanism–outcome configurations (CMOCs) to those which are most plausible. This is because testing too many hypotheses will lead to some false positive conclusions. Qualitative research conducted within process evaluations is a useful way to inform refinement of CMOCs before they are tested using quantitative data. Process evaluations aim to examine intervention implementation and the mechanisms that arise from this. They involve a mixture of quantitative (for example, logbooks completed by intervention providers) and qualitative (for example, interviews or focus groups with recipients) research. Qualitative research can be useful in assessing and refining CMOCs because intervention providers and recipients will have insights into how intervention mechanisms might interact with context to generate outcomes. These insights might be explored directly (for example, by asking participants how they think the interventions works) or indirectly (for example, by asking participants about their experiences of an interventions, and the conditions and consequences of this). Sampling for such qualitative research should ensure that a diversity of different participant accounts is explored. Analyses of these accounts can draw on grounded theory approaches which aim to build or refine theory based on qualitative data.

From the later nineteenth century, a need was recognised for social data that covered a wider range of issues and that were also of more detailed kind than those that could obtained from ‘complete enumerations’ via national population censuses and registration systems. Initially, ‘partial studies’ in the form of monographs as produced by Le Play and his followers – essentially ethnographic case studies -- were seen as the way ahead. But those favouring this approach were unable to solve the problem of how to move from part to whole. Claims of the ‘typicality’ of monographs could never be substantiated. A different approach, that of sample surveys, was proposed by Kiaer, in a shift from typological to population thinking in data collection that paralleled that made by Galton in data analysis. Kiaer’s ‘purposive’ sampling was found to have serious flaws and Bowley, an economist but also an advocate of ‘modern statistical sociology’, proposed and applied the alternative of probabilistic or random sampling. Finally, in the 1930s, Neyman demonstrated the superiority of probabilistic sampling, which was then rapidly taken over by sociologists.