This appendix discusses how to measure the significance of changes in public opinion over time. Two methodologies have been used in the public opinion literature. The time-series approach uses each survey result as a single point in a time series, and assesses the statistical significance of changes based on trends across surveys with the N equal to the number of separate national samples. This methodology was used to measure trends in partisanship in Chapter 2 , as well as some of the aggregate statistics on party change in Chapter 3 . It is equivalent to analyses of aggregate economic statistics or other national characteristics over time.
In contrast, the survey approach focuses on the raw survey data, measuring whether the differences in opinions over two or more surveys are statistically significant. Because these comparisons are based on large survey samples, the N for these comparisons is quite large—usually several thousand or more.
Employing these two techniques on the same cross-sectional data will yield dramatic=ally different estimates of statistical significance because of the large differences in the N used in calculating probability statistics. A brief example based on changes in American party identification can illustrate the effects of the two approaches. Between 1964 and 1970 the percentage of Americans identifying with a political party declined by a sizeable amount:
Using the time-series technique, one gets the following insignificant results:
b = -1.13, se = 0.38, sig = 0.099, N=4
In contrast, if one analyses the cumulative file of surveys from these four years, and calculates a regression model predicting identification with a party (coded 0,1) by year of interview, the following results are revealed:
b = -1.15, se=0.003, sig = 0.0000, N=5,925
The two methodologies thus give nearly identical estimates of the per annum changes because they are measuring the same trend over time. The time-series analysis is not statistically significant because of the small number of surveys upon which it is based. In contrast, the analysis based on individual data has a much greater N—being based upon individual interviews rather than timepoints—and thus the results are very highly significant.
Some recent research on comparative public opinion trends has used the time-series method (e.g. Klingemann and Fuchs 1995). If one had access to the raw survey data, we believe that it would be more appropriate to base the statistical analysis on the survey