Cluster sampling bias in government-sponsored evaluations: a correlational study of employment and welfare pilots in England

For pilot or experimental employment programme results to apply beyond their test bed, researchers must select ‘clusters’ (i.e. the job centres delivering the new intervention) that are reasonably representative of the whole territory. More specifically, this requirement must account for conditions that could artificially inflate the effect of a programme, such as the fluidity of the local labour market or the performance of the local job centre. Failure to achieve representativeness results in Cluster Sampling Bias (CSB). This paper makes three contributions to the literature. Theoretically, it approaches the notion of CSB as a human behaviour. It offers a comprehensive theory, whereby researchers with limited resources and conflicting priorities tend to oversample ‘effect-enhancing’ clusters when piloting a new intervention. Methodologically, it advocates for a ‘narrow and deep’ scope, as opposed to the ‘wide and shallow’ scope, which has prevailed so far. The PILOT-2 dataset was developed to test this idea. Empirically, it provides evidence on the prevalence of CSB. In conditions similar to the PILOT-2 case study, investigators (1) do not sample clusters with a view to maximise generalisability; (2) do not oversample ‘effect-enhancing’ clusters; (3) consistently oversample some clusters, including those with higher-than-average client caseloads; and (4) report their sampling decisions in an inconsistent and generally poor manner. In conclusion, although CSB is prevalent, it is still unclear whether it is intentional and meant to mislead stakeholders about the expected effect of the intervention or due to higher-level constraints or other considerations.