To assess causal direction in the valence model, Evans and Chzhen adopt a cross-lag specification with panel data. This is essentially the Granger setup, where a dependent variable is regressed on the lagged dependent variable and the lagged independent variable, estimated in both directions, first for party preference with performance as a predictor and then for performance with party preference as a predictor. The approach is not ideal, as it does not account for contemporaneous effects, which is important. Consider that voters most likely do not go into the voting booth asking themselves “So, how was the economy last year?”Footnote ¹ That said, it is a conservative approach, biased toward the null hypothesis, and so positive results are more noteworthy than negative ones. Evans and Chzhen are well aware of and acknowledge this limitation.

Their analysis examines the relationship between Labour performance assessments and preferences over a three-wave British Election Study (BES) panel leading up to the 2010 UK general election. The variables do not exactly match what Clarke et al. have done, though it is worth noting that the valence models in their different publications do not exactly match each other either, and what Evans and Chzhen do is very much in the spirit of those models. What does their reanalysis show? To begin with, they demonstrate that party preferences influence performance assessments at each point in time. This clearly contrasts with Clarke et al.’s implicit assumption that performance assessments are exogenous to party preferences. That said, Evans and Chzhen do find evidence that performance assessments influence preferences, but only in the middle of the election cycle. The effect is not present in their analysis of preferences at the end of campaign. Based on their analysis, Evans and Chzhen conclude that “evaluations of party performance appear to have limited merit as simplifying heuristics.”

Now, the results do not mean that individual-level valence assessments do not matter for the vote, just that they do not matter controlling for prior party preferences. We already have considered that performance assessments have contemporaneous effects that are not captured in the cross-lag models. Perhaps, more important is the fact that Evans and Chzhen showed that earlier valence assessments influenced party preferences, and that they in turn powerfully predict final, pre-election support. The impact of valence can be both direct and indirect, after all, and it ultimately may be that individual-level variation in performance assessments are increasingly reflected in preferences over the long campaign (also see Erikson and Wlezien Reference Evans and Chzhen2012). All of this said, there is no escaping the conclusion that the Clarke et al. models almost certainly overstate the influence of valence on voting.

Clarke and company (Whiteley et al. Reference Whiteley, Clarke, Sanders and Stewart2016) respond. They challenge Evans and Chzhen on various fronts. Some of what they claim is not right, most importantly that Evans and Chzhen do not test the valence model. (See Evans and Chzhen’s (Reference Evans and Andersen2016b) reply to Whiteley et al. for details about this and other issues.) Some of what Whiteley et al. claim is on the mark, including the limitations of the cross-lag approach that we have already discussed. They do not propose an alternative individual-level specification, however, and instead shift focus from the individual to the aggregate level. Why they change the level of analysis is not perfectly clear, but it seemingly is as Kramer (Reference Ladner and Wlezien1983) would have it; focusing specifically on economic voting, he argued that individual-level variation in economic perceptions at particular points in time are not meaningful causes of vote preference, and that the aggregate level is the more appropriate level of analysis. In effect, the individual-level effects of valence could take the form of an intercept at each point in time (see Wlezien et al. Reference Wlezien, Franklin and Twiggs1997).

Whiteley et al. analyze the dynamic relationships between indicators of aggregate Labour performance assessments and vote intentions. The results of their vector error correction models contrast with what Evans and Chzhen found: in Whiteley et al.’s analysis, vote intentions do not Granger-cause the public’s performance assessments but those assessments do predict vote intentions, and powerfully so (also see Green and Jennings Reference Kayser and Wlezien2012). These aggregate-level effects are not surprising given the now huge body of work on economic perceptions (and conditions) and government approval and election outcomes (for summaries, see Lewis-Beck and Stegmaier Reference Lewis-Beck and Stegmaier2000 and Lewis-Beck and Whitten Reference van der Brug, van der Eijk and Franklin2013). Whiteley et al. are right about not losing sight of the aggregate relationship in the face of the individual-level results. Yet, this does not gainsay what Evans and Chzhen have argued and shown about Clarke et al.’s individual-level analyses. The estimates of valence effects based on those analyses pretty clearly overstate the true effects, and seemingly not by a little.

There is reason to suppose that the problems with the individual-level analyses of valence have implications for the aggregate level, even if these may not be as troublesome as some might suppose, including Evans and Chzhen, who fret about the “ecological fallacy.” Consider that we know from research on economic effects that valence does matter for political judgments and voting (see the reviews referenced above). Most of this work shows that objective conditions influence aggregate government approval and election outcomes, though there is individual-level support as well (e.g., van der Brug et al. Reference Campbell, Converse, Miller and Stokes2007; Kayser and Wlezien Reference Kramer2011). Other work shows that economic perceptions track objective economic conditions over time, so they are informative, if imperfectly, about the real economy (Erikson et al. Reference Erikson, MacKuen and Stimson2002; Erikson and Wlezien Reference Evans and Chzhen2012; also see Reference WlezienWlezien 2015). This tends to support Whiteley et al.’s dynamic aggregate-level analyses of performance assessments. The problem is that research (listed above) demonstrates that economic perceptions are not useful indicators of real individual-level differences, as Evans and Chzhen posit and show for valence assessments. Their analysis highlights the need to incorporate objective measures, at the least to identify non-economic indicators to complement the plentiful economic ones. Doing this may not be easy but seems necessary in order to make real progress on the impact of valence on political behavior.