MetaAnalysis Of The Final 1996 Pre-Election Presidential Campaign Polls:
The Odds Calculated Here Are The Odds Against A Chance Explanation Of Poll Prediction Failures.
| Poll | n | Clinton (pC) |
Dole (pD) |
Clinton-Dole (pC - pD) |
Std. Error of (pC - pD) |
Z-score |
P(Z) | Probability (2-tailed) |
Odds (1-p)/p |
Rounded Odds |
| CBS/NY Times | 1519 | 0.53 | 0.35 | 0.18 | 0.0236 | 3.9954 | 0.999968 | 0.000065 | 15479 | 15,000 to 1 |
| Pew Research | 1211 | 0.49 | 0.36 | 0.13 | 0.0262 | 1.6893 | 0.954423 | 0.091155 | 9.97 | 10 to 1 |
| ABC/Washington Post | 703 | 0.51 | 0.39 | 0.12 | 0.0355 | 0.9581 | 0.830986 | 0.338027 | 1.96 | 2 to 1 |
| Harris | 1339 | 0.51 | 0.39 | 0.12 | 0.0257 | 1.3358 | 0.909196 | 0.181607 | 4.51 | 5 to 1 |
| NBC/Wall St. Journal | 1020 | 0.49 | 0.37 | 0.12 | 0.0288 | 1.1890 | 0.882771 | 0.234458 | 3.27 | 3 to 1 |
| CNN/USA Today/Gallup | 1200 | 0.52 | 0.41 | 0.11 | 0.0277 | 0.8791 | 0.810326 | 0.379348 | 1.64 | 2 to 1 |
| Hotline/Battleground | 1000 | 0.45 | 0.36 | 0.09 | 0.0283 | 0.1497 | 0.559482 | 0.881037 | 0.14 | 1 to 7 |
| Reuters/Zogby | 1200 | 0.44 | 0.37 | 0.07 | 0.0259 | -0.5728 | 0.716617 | 0.566766 | 0.76 | 1 to 1 |
| Actual Vote | 96,211,883 | 0.493 | 0.407 | 0.0853 | ||||||
| Combination of Probabilities | 3.7138 | 0.999898 | 0.000204 | 4,896 | 4900 to 1 | |||||
| Note 1: The standard error of the difference between two categories of a multinomial distribution (here tabulated as the standard error of pC - pD) is derived in Kish, L. (Survey Sampling, 1965, Wiley, pp. 497-501). | ||||||||||
It is: STD.ERROR = SQRT((1 - (n/96,211,883))*(pC + pD - ((pC - pD)^2))/(n - 1)) | ||||||||||
| Note 2: The computation of a Z-score follows Snedecor, G.W. and Cochran, W.G. (Statistical Methods, 6th Ed., 1967, Iowa State University Press, pp. 211-212). | ||||||||||
It is: Z = (ABS((pC - pD) - 0.0853) - (1/(2*n)))/STD.ERROR | ||||||||||
| Note 3: Probabilities were combined by the weighting method described by Rosenthal, R. (Meta-Analytic Procedures for Social Research, Rev. Ed., 1991, Sage Publications, Ch. 5). | ||||||||||
Combined Z = SUMPRODUCT(n, Z)/SQRT(SUMSQ(n)). | ||||||||||
| Note 4: The P(Z) values associated with the two most extreme Z values were double checked against entries in the normal probability table in Abramowitz, M. and Stegun, I.A. (Eds.) (Handbook of Mathematical Functions, 1964, U.S. Government Printing Office, p. 972). | ||||||||||
© 1997 Gerald S. Wasserman |