EDITOR’S PROOF
176 J.X. Eguia323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368social safety net, if some agents prefer a larger safety net to help needy immigrants
when immigration policy is lax, while other agents prefer a smaller safety net to not
spend money on immigrants when immigration policy is lax, then we can redraw
the axes to make the preferences of one group of agents separable, but in doing
so, the preferences of the other group of agents remain non-separable. In very non-
technical terms, agents have non-separable preferences if their indifference curves
are tilted; if all agents have curves equally tilted, we can tilt the whole map to return
to a standard model over newly defined dimensions.
If, on the contrary, different agents have preferences tilted in different directions,
we cannot correct this problem by tilting the whole map. We need instead to intro-
duce parameters to accommodate the correlation across issues. This is a consider-
able setback, similar to the problem of agents who assign different relative weights
to the various dimensions -but more damaging, because we need more parameters
to fix it. In order to accurately represent the preferences of agents who disagree on
the weights they assign to the different dimensions we need to add one parameter
per dimension per agent or group of agents who disagree on these weights, for a
maximum of(K− 1 )(N− 1 )new parameters if there areNagents andKdimen-
sions. In order to represent the preferences of agents who disagree on the correlation
in preferences between issues, we must add one correlation parameter per possible
pair of issues and per agent or group of agents who disagree, for a maximum of
K(K− 1 )
2 Nnew parameters.
While violations of separability do not affect classic results on the instability
of simple majority rule as long as preferences are smooth (Plott 1967 ; McKelvey
1979 ), they affect how we can interpret and use common spatial models. Consider
the structured-induced equilibrium theory (Shepsle and Weingast 1981 ), which pro-
poses that the instability is solved by choosing policy dimension by dimension. In
the standard structured-induced equilibrium theory, the order in which the legisla-
ture considers the various policy dimensions is irrelevant, because preferences are
separable. With non-separable preferences, the order in which each policy dimen-
sion is considered affects the chosen policy outcome. For a second example, con-
sider the ideal point estimation literature (Poole and Rosenthal 1985 ; Clinton et al.
2004 ): if preferences are not separable, estimating the ideal point of each legislator
is not enough to predict vote choice.5 Discussion
Theoretical and empirical work questions not only the standard assumption of Eu-
clidean utility functions in multidimensional spatial models, but the more general
assumptions of separable, convex and/or smooth preferences.
Standard spatial models suffer from limitations that I have not considered here.
For instance, an increasing body of literature argues that we must add a candidate
valence term to capture the actual preferences of voters about candidates. Valence
is any quality that all voters agree is good, and makes the candidate who possesses