Using repeated measures data to analyse reciprocal effects: the case of Economic Perceptions and Economic Values

Disentangling causal relationships between abstract concepts is the ‘holy grail’ of empirical social science. The enterprise is, however, stymied by the seemingly intractable limitations of observational studies which constitute the data source for the vast majority of social scientific research programmes. Without random allocation to ‘treatment’ and control groups, we are simply unable to discount alternative models that might account equally well for observed covariances between dependent and independent variables, yet do not posit a causal relationship between them. The primary limitations to valid causal inference with observational data relate to the issues of unobserved variable bias and reciprocal causality. The former pertains when X and Y are conditionally independent, given a third variable (or vector of variables), Z, which has not been included in the model. The latter arises when both X and Y exert a direct causal influence on each other over time. In this paper, we fit a series of models which attempt to estimate the reciprocal relationship between retrospective perceptions of the macro economy (EP) and ‘left-right’ economic values (EV). EP is measured by three five point items tapping ‘past-year’ perceptions of unemployment rate, inflation rate and the general standard of living. EV is measured by the six item scale developed by Heath et al (1993). We hypothesise a possible reciprocal relationship between these variables on the grounds that economic perceptions have been shown to be partially endogenous to political preferences (Evans 1997) and because it is likely that beliefs about how the economy should be managed will be influenced by macro-economic performance. The models we present are deliberately selected to demonstrate the importance of two issues in model specification: correcting for stochastic and systematic errors in the measurement of concepts; and specifying an appropriate covariance structure between error terms in both the measurement and structural components of a model. Measurement error in independent variables results in attenuated effect sizes (Bollen 1989). Failing to adequately account for covariances between error terms can bias estimates of auto-correlation and crosslag parameters (Williams and Podsakoff 1989). This is not an esoteric statistical point; failing to account for these errors and their covariance structures can lead to seriously flawed causal inference. Our data source for all analyses is the 1992-1997 British Election Panel Study, with an analytical sample size of 1640. Table 1 presents a summary of the cross-lag and stability parameters estimated in each model.