Home » Politics & Society » Conservative Politics » Census Sensibility

Census Sensibility

© Frank Monaldo

National polling during the 1996 presidential elections provides an important lesson in the nature of statistics and the dangers inherent even in their ostensibly professional application. These lessons should make us less sanguine about the potential use of sampling for the official 2000 US Census.

Right up to Election Day, major national polls had President Bill Clinton ahead of Senator Robert Dole by double-digit percentage point leads. Leave aside the question of whether these polls had the self-fulfilling effect of making Clinton's re-election seem inevitable. Most major polls radically erred and predicted election results far outside the stated margins of error.

One major poll consistently found much smaller differences between support for Clinton and Dole, the Zogby poll by John Zogby. The table below shows the major poll predictions and the actual election outcome. Zogby's results mirrored the actual election tallies, giving Clinton an eight percent edge, far smaller than the 18 percent margin predicted by the CBS/New York Times poll.

Table: 1996 President Polls and Election Results. Values are in percentage. The last column is the margin between the Clinton and Dole percentages.

Poll	Clinton	Dole	Perot	Margin
CBS/NY Times	53	35	9	18
Pew Research	49	36	8	13
ABC News	51	39	9	12
Harris	51	39	9	12
NBC/WSJ	49	37	9	12
Reuters/Zogby	49	41	8	8
1996 Election	49	41	8	8

How is this possible? The stated margins of error were only a few percent. How could reputable, statistically savvy pollsters arrive at such radically different results? The reason is because the quoted margin of error, the sampling or random error is just one error source. All the polls suffered from about the same level of sampling error, but they labored under different levels of systematic error.

Given a finite, randomly selected sample of a population, the size of the sample determines how well that sample can characterize the population. These statistics are well understood. For example, let's say I want to know the mean height of the population. I can randomly select 100 people from the population and measure their heights. Sampling error tells me how close I can expect the mean height of 100 people in my sample to approximate the mean height of the entire population. Clearly, if I increase my sample size to 1,000 or 10,000 or 100,000, the mean height of my sample will more likely represent the mean height of the entire population.

The copyright of the article Census Sensibility in Conservative Politics is owned by . Permission to republish Census Sensibility in print or online must be granted by the author in writing.