Filling the gap between individual and collective aspects of the policy process: multilevel analysis of policy learning

Abstract
Policymakers’ beliefs and preferences about a policy evolve over time. This process of policy learning is a multilevel notion. It is an individual-level, psychological process of beliefs and preferences change. At the same time, it has a collective dimension: given the influence that policymakers have on each other during the policy process, policy learning may not be reduced to a sum of individual processes, at the aggregate level. As policy learning is a crucial factor of policy change, the study of its determinants is central. A considerable amount of research has been published on policy learning but it has rarely been multilevel in approaches and methods. First, policy learning is most often measured at only one level, either at the individual level or at the aggregate level. Second, individual-level and aggregate-level conditions of policy learning are rarely investigated in combination. Most often, existing studies focus on a set of factors situated at the same level, for example several personality traits or a set of organizational factors. These two limitations have at least three drawbacks. First, as individual-level and aggregate-level policy learning are difficult to connect, methodologically speaking, the conceptual and theoretical discussion of this connection remains unclear. Second, it prevents to compare the amount of learning situated at the aggregate level – that is, common to the members of a same group unit – with the amount of learning situated at the individual level – that is, different among individuals even within a same group unit. Third, the relative influence of the factors of learning situated at different levels of analysis remains unknown as well. The present paper shows why and how multilevel analysis may be used to analyze policy learning and to solve those drawbacks. In particular, three tools associated with multilevel models are mobilized: the intraclass correlation coefficient, the Snijders & Bosker (1994)’s multilevel pseudo-R², and the Pfeffermann et al. (1998)’s Probability-Weighted Iterative Generalized Least Squares Estimators (PWIGLS). The paper does not only offer a systematic presentation of these multilevel tools but also their application to an illustrative study on policy learning among 318 policymakers from several organizations involved in the European liberalization process of two Belgian network industries: the rail sector and the electricity sector. In this illustrative study, the relative effectiveness of two measures of organizational interests is assessed for predicting policy learning: an individual-level measure based on the responses of policymakers to a survey and an aggregate-level measure based on the deduction of the researcher of the organizational interests of all policymakers pertaining to a same organization.