Andrew Gelman wrote today about some erroneous U.S. Governor approval ratings, noting that the ratings for Janet Napolitano sum to 108%.

In fact most of these ratings do not sum to 100%. I prepared a clean CSV file of the ratings, making use of R's XML library and the readHTMLTable function. The ratings data file is here approval.csv.

> approval <- read.csv("http://biostatmatt.com/csv/approval.csv")
> table( (approval$Excellent + approval$Good + approval$Fair + approval$Poor) - 100 )

-5 -4 -3 -2 -1  0  1  8
 2  3  8 12 13  6  1  1

There is so much variability here, we could start to think about the sampling distribution, and the factors contributing to variability.

I don't know much about this survey, or about survey conventions when reporting percentages. But I know my advisors wouldn't have let me report percentages like this. Is it common to report percentages that sum to less that 100% when there is nonresponse? Or are these typos too? Also, where are the ratings for Hawaii, Idaho, Indiana, and Wyoming? (I knew learning the alphabetical state song in elementary school would be useful some day)