Data Still Needs Theory: Collider Bias in Empirical Legal Research

Abstract

<p>Big data is characterised not only by the amount but also the kinds of information that can be created, stored, and processed. This explosion of data, accompanied by the capacity to analyse them, has catalyzed large n, quantitative approaches to the study of law and legal institutions.<br><br>But neither size nor quality guarantees the validity of causal inferences drawn from observational data. For example, although the inclusion of control variables can help isolate causal effects, not all variables are good controls. Bad controls are not harmless and can create the impression of a causal relationship where none exists. This spurious association is called collider bias.<br><br>We introduce the concept of collider bias and give motivated examples of how it can arise in empirical legal research. The selection of good controls requires knowledge and assumptions about causal structures. Theory and domain knowledge are essential for quantitative analysis, even in the era of big data.</p>