¹ See, for instance, Borooah, Vani and van der Ploeg, Rick, ‘The Changing Criteria of Economic Issues in Performance and Popularity in British Polities’, Manchester School of Economics and Social Science, L (1982), 61–78CrossRefGoogle Scholar: Miller, W. L. and Mackie, M., ‘The Electoral Cycle and the Asymmetry of Government and Opposition Popularity’, Political Studies, XXI (1973), 263–79CrossRefGoogle Scholar; Pissarides, C., ‘British Government Popularity and Economic Performance’, Economic Journal, XC (1980), 569–81.CrossRefGoogle Scholar

² Norstrom, T., ‘The Popularity of Government: Lagged Response or Omitted Predictor’, paper presented at ECPR workshop, Barcelona, 03 19, p. 2.Google Scholar

³ See Whileley, Paul, ‘Inflation, Unemployment and Government Popularity’, Electoral Studies, III (1984), 3–25CrossRefGoogle Scholar, for a notable exception. Also note that whilst, like regression analysis, the procedure involves minimizing the sum of squared residuals such models, unlike regression analysis, are non-linear and require non-linear least squares estimation procedures.

⁴ McLeod, G., ‘Box Jenkins in Practice’, Gwilyn Jenkins & Partners, 1982.Google Scholar

⁵ Kernell, Samuel, ‘Presidential Popularity and Negative Voting’, American Political Science Review, LXXI (1977), 44–66.CrossRefGoogle Scholar

⁶ Enquiries of Gallup revealed that prior to the 1966 general election the political questions – voting intention, attitude towards the government's record, attitudes towards the prime minister, economic conditions, etc. – were asked in no set order. Following the 1966 election the question order was standardized with the first political question being that concerning attitude towards the government's record.

⁷ The essential characteristics of ARIMA modelling are that a series of data is regarded as being generated by either (or both) past values of the series itself (auto-regressive) or a combination of current and past values (α_t, α_t-1 etc.) of a random normal series (moving average). Therefore there is some process by which either or both of the past values of the series or current and past values of a random normal series are converted into the data series being modelled.

⁸ McCleary, R. and Hay, R. A., ‘Applied Time Series Analysis’ (Beverly Hills, Calif.: Sage, 1980).Google Scholar

⁹ Figures in parenthesis are Student t-values and all coefficients are significant at the 5 per cent level unless specified otherwise; RMS = Residual Mean Square; LBQ₁₅ = Ljung – Box Q Statistic for first fifteen residual auto-correlations. This may be compared with a chi-squared distribution having 15 – K degrees of freedom, where K is the number of estimated model parameters, to test whether or not residuals are random; = chi-squared value at 5 per cent level for 12 degrees of freedom; B = backward-shift operator such that BX _t = X _t-1. Note that if α_t = X _t-1 = X _t-BX _t = (1- B) X _t then X _t = α_t/(1 – B).

¹⁰ Analysing the series separately for pre-1966 election and post-1966 election data revealed that the only difference was in the level, not in the structure of the models.

¹¹ Incorporating a similar intervention variable produced an insignificant coefficient for this series.

¹² This is produced by correlating the series at increasingly positive and negative time lags.

¹³ McLeod, , ‘Box Jenkins in Practice’.Google Scholar

¹⁴ ‘Pre-whitening’ a series means that it is turned into a random (or ‘white-noise’) series. Prior to the CCF being calculated, the scries with which it is to be cross-correlated is then subjected to the same manipulation that induced randomness in the pre-whitened series.

¹⁵ Originally two variables were defined here, one when disapproval was above 60 per cent and one for when disapproval was below 40 per cent. However, the estimated coefficients were such that these two variables could be collapsed into one. There are thirty-three instances when (adjusted) disapproval fell below 40 per cent (fifteen during Conservative and eighteen during Labour governments) and thirty-four instances when it rose above 60 per cent (fifteen during Conservative and nineteen during Labour periods of office), accounting for 25 per cent of observations in total.

¹⁶ The multiplicative dummy, MCG, was included in zero-, first- and second-order moving average form; however, only the zero-order coefficient was significant.

¹⁷ The insignificant t-values on MCG only arose through introducing the noise model – in estimating the structural form of this variable without the ARIMA noise model the zero-order average coefficient only was significant and was thus retained. It becomes insignificant, though does not change its magnitude, when the noise model is added due to a slight correlation with the noise process. This was deemed not to be a significant problem. A similar argument exists for the t-value in the first-order coefficient of Dadj.

¹⁸ In order to investigate the election cycle the following procedure was adopted. Two variables were defined such that both took zero values except that one took the value 1 for the bth month prior to a general election and increased by 1 each month until reaching the value of b in the month prior to an election. The second took the value of c in the month following a general election and declined by 1 each month until reaching a value of zero (c + 1) months after an election. The initial value of b was fixed and the value of c varied until the maximum reduction in the sum of squared residuals was obtained. A new value of b was then used and, again, c was varied. The combination of b and c values resulting in the maximum reduction of the sums of squares failed to yield significant coefficients for either of the two variables.

¹⁹ Rather than incorporate ED, the model was estimated using the logit of the (adjusted) disapproval series log (Dadj/(1 – Dadj)), which has a similar form to the resulting relationship. However, whilst this represented an improvement, the use of ED was marginally superior and is retained here for ease of exposition. Statistically, it is preferable to work with logits of both the input and output series. Doing this did not alter the conclusions.

²⁰ This concurs with Hudson's results although the difference here is somewhat greater at around 5 percentage points compared to his value of 3 percentage points. This is primarily due to the inclusion of MCG which prevents CG from picking up both slope and intercept differences.

²¹ See Crewe, Ivor, ‘The Electorate: Partisan Dealignment Ten Years On’, West European Politics, VI (1983), 183–215CrossRefGoogle Scholar, in which it is shown that in 1983 there was a greater proportion of weak partisans amongst Labour identifiers than amongst Conservative identifiers.

²² Särlvik, Bo and Crewe, Ivor, Decade of Dealignment (Cambridge: Cambridge University Press, 1983).Google Scholar

²³ This concurs with Hudson's theoretical model but was not tested empirically by him.

²⁴ Whilst this value of about 0.62 was based on the percentage approving of the government, even when one takes into account the relationship between the approval and disapproval series

the estimates here remain somewhat different.