A Gravis Fraud
Editor’s Note: This article was submitted by dawolf, one of our regular readers and commenters. We always welcome your contributions.
Another Editor’s Note: I have published an update on this.
Much has been written of late regarding whether polls need to be “unskewed” due to incorrect Democratic/Republican registration. But what if there are results being made up simply out of whole cloth?
One pollster that arrived on the scene recently is Gravis Marketing who has shown a strong Republican house bias this cycle. A recent investigation of Gravis by Democratic Underground made it clear that there were questions surrounding the company.
This is not the first time that a pollster has been accused of publishing “poll results” that do not correspond to actual polled individuals. Nate Silver at FiveThirtyEight made such an accusation of the now-defunct Research 2000 on June 29, 2010.
However, questions such as those raised by Democratic Underground are not evidence of guilt. So is there any real evidence that the work they are committing is fraudulent?
Here is a recent poll they claim to have conducted is currently available, with crosstabs. The executive summary states
On September 29th – 30th, 2012, Gravis Marketing, a non-partisan research firm, conducted a survey of 914 likely voters in the state of Florida regarding their likely vote for a given presidential candidate, senate candidate, and other social and economic issues important to voters.
Nine hundred fourteen likely voters. This number is key. It means one likely voter is 0.1094% of the total. If a likely voter didn’t answer one particular question, this percentage goes up, not down, so for any question the smallest possible result, to three decimal places, is 0.109%.
That leads us to be able to recognize many impossible results, such as the one circled in the table to the right.
Since 0.07 is less than 0.109, it would require just under two-thirds of a person to be unsure of whom to support.
Here is a summary of their impossible results.
Do you need more proof? Lets go back to that first impossible example given above. We can go back to the raw numbers by multiplying out from the percentage, to indicate how many people must have given a particular answer in order to achieve the supplied crosstab results. This gives the table to the right:
Not a single answer has an integer number of people. Most are too far away from an integer to have possibly been accounted for as a rounding error.
Clearly, Gravis Marketing should not be trusted, and their polls should not be used in any manner. They are currently being included in the poll aggregation being conducted by Nate Silver at FiveThirtyEight, and by RealClearPolitics. Hopefully this will be rectified soon.