On the Meaning of Causal Generalisations in Policy-oriented Economic Research

Abstract

Current philosophical accounts of causation suggest that the same causal assertion can have different meanings. Yet, in actual social-scientific practice, the possible meanings of some causal generalisations intended to support policy prescriptions are not always spelled out. In line with a standard referentialist approach to semantics, we propose and elaborate on four questions to systematically elucidate the meaning of causal generalisations. The analysis can be useful to a host of agents, including social scientists, policy-makers, and philosophers aiming at being socially relevant. To illustrate our proposal, we analyse the complexities related to the meaning of causal generalisations in the context of a concrete case of economic research which is explicitly intended to guide public policy, namely, the OECD research on the causes of unemployment.

Acknowledgements

We want to thank Lorenzo Casini, Kevin Hoover, Caterina Marchionni, Julian Reiss, Attilia Ruzzene, and two anonymous referees of this journal for very helpful comments and suggestions. We are also grateful to the audiences in sessions at CitS Workshop ‘Causality and Intervention’ (Paris, June 2011), SPSP Conference (Exeter, June 2011), and TINT-POS Seminar (Helsinki, September 2011), where we have presented earlier versions of this article. The work of François Claveau on this project was supported by the Social Sciences and Humanities Research Council of Canada (fellowship 767-2009-0001).

Notes

1 For some insiders' recognitions of the tremendous impact of the OECD study on the expert community, see Freeman (2005), 131–132; Blanchard (2006), 51–52; Boeri and van Ours (2008), 1–2. For a more recent and revised perspective on unemployment by the OECD, see OECD (2006).

2 The OECD did not stop using causal generalisations after 1994. Recent generalisations formulated in OECD reports include ‘short-time work schemes helped preserve permanent jobs during the economic downturn’ (OECD 2010, 68).

3 There are some exceptions to a referentialist (sometimes called ‘truth-conditional’ or ‘representationalist’) approach to the meaning of causal claims: see, for example, Williamson (2005), ch. 9; Spohn (2006); Beebee (2007); Reiss (2011, 2012). These alternative accounts can be understood as variants of an inferentialist semantics—an approach that locates the meaning of a statement in the role it plays in the inferential practices of the utterers. For the distinction between referentialist and inferentialist semantics, see Peregrin (2012) and references therein. The present article is the first part of a broader project on the semantics of causal generalisations, in which we are developing an inferentialist analysis as well.

4 Following a strong trend in the philosophy of causation (Spirtes, Glymour, and Scheines 1993; Pearl 2000; Hitchcock 2001a; Hoover 2001; Woodward 2003; Hausman 2005), and in conformity with general usage in economics, upper-case italics (X and Y) are variables, and lower-case italics (x and y) represent specific values of these variables.

5 The causal variables in other generalisations present in OECD research have the same multidimensional character. For instance, if one considers the purported cause ‘short-time work schemes' (see note 2 above), one finds that recent discussions about short-time work schemes actually decompose it into 14 dimensions (which are then regrouped into four main families of features), see OECD (2010), annex 1.A1.

6 The concepts discussed in what follows are meant to be illustrative in relation to the OECD example, yet they are not exhaustive. Many additional causal concepts can be characterised using alternative criteria to the ones we refer to in this article (Pearl 2000).

7 The notion of a ‘net effect’ is also sometimes called ‘total effect’ (Pearl 2000, 151–152, 164; Pearl 2001) or ‘total cause’ (Woodward 2003, 50–51).

8 Hitchcock's (2001b) ‘component effect along a causal route’ is essentially the same concept that Woodward (2003, 50, 57) calls ‘contributing cause’, and similar to what Pearl (2001) defines as ‘path-specific effect’.

9 Some methodological accounts on the use of regression analysis as the primary tool for causal inference in empirical economics are Pearl (2000), ch. 5; Hoover (2001), ch. 7; Morgan and Winship (2007), ch. 5; Angrist and Pischke (2009).

10 This sentence in fact refers to two assumptions common to regression analysis, ‘linearity’ and ‘homogeneity’, yet the focus in this section is only on the consequences of the former. The consequences of the latter on establishing causal effects for policy purposes are extremely significant, since (as it will be further discussed in section 5) in heterogeneous populations it is not straightforward whether generalisations that are true for the population as a whole would also be true for specific individual units of the population.

11 Note that, for expository purposes, we assume here causal sufficiency (see section 3.2). The statement would be more convoluted if we were to consider other cases covering alternative combinations of sufficiency and necessity.

12 Note that many authors (Morgan and Winship 2007, sec. 2.4; Imbens and Wooldridge 2009, sec. 2.3) reduce SUTVA to the condition about the stability of a treatment effect to the treatment assignment of other units. Yet there are some exceptions (Heckman 2005, 35–38).

13 The notion of ideal intervention that is typically used in the manipulationist account of causality rules out the variability in the treatment assignment mechanism by the very definition of an ‘ideal’ intervention, see Woodward (1996) and (2003), ch. 6. As argued by Reiss (2007), ch. 10, and Cartwright (2007), ch. 16, the notion of an ‘ideal intervention’ is employed at the cost of making causal claims less directly relevant to policy, since in real life ‘ideal interventions' are seldom feasible. Indeed, even if one warrants the truth of a causal claim thanks to an ideal intervention, it can still be the case that the effect will not follow from a real-world implementation of the cause simply because this manipulation fails to be of the ideal type.

14 Notice that Rubin (1986) seems to be of the opinion that SUTVA is a necessary condition for a causal claim to be meaningful.