A routine review of the processes we are using to create weights for our data for the current poll has turned up a problem that affected the release on NJ voters attitudes toward Gov. Chris Christie and former Sec. of State Hillary Clinton as 2016 presidential candidates we put out on August 7 (see the corrected release here.) Fortunately, the error resulted in no change to our analysis of the results of the poll, but we do see some of our point estimates move up or down a small amount, primarily due to rounding up where we previously rounded down. We routinely round to the nearest percentage point, so if a point estimate is 39.4, we report 39%, while if it is at 39.6, we round up to 40%. We round like this because given a roughly 4-point margin of error, tenths of a percentage are inappropriate precision, especially because each tenth represents less than one person in our typical sample.
The somewhat technical details are these. The error crept in because of the change we reported in how we calculate our weights. Previously we only weighted based on demographics. Starting with EP193 we are now creating an initial “frame weight” that adjusts for the fact that we include both cell and landline phones in the sample, and thus some people have a greater likelihood of being chosen than others (those with both, versus those with only one or the other.)
The frame weight starts by examining the share of all cell and landline phones that are represented in our sample, and this is where the error occurred. Due to an inadvertent oversight, the baseline for the total number of cell phones and landline phones in New Jersey was incorrectly entered into the calculation. As a result the weight given to cells versus landline phones in our frame weight was incorrect. This error was perpetuated through the rest of the weighting process, throwing off our point estimates by very small, but sometimes noticeable, amounts.
Even though the effects are small and our analyses of the results remains unchanged, it is important that we note and correct this error. The corrected release now contains the following at the end.
Advisory: This release is a corrected version of the release of August 7, 2014. Due to an inadvertent error in calculating weights, some results were reported incorrectly. Most results were correct; while a few numbers used in the original release changed less than one percentage point. However, because of rounding to the nearest percentage point, some results moved up or down one or two points. Key changes include revising the Clinton-Christie match-up from 50%-40% to 51% to 40%. Clinton’s favorable rating should have been reported at 54% favorable instead of 53%, while Christie’s favorable rating should have been reported at 49% instead of 50%. In addition the reported weighted demographics of the sample have changed slightly, including increasing the rounded share of Democrats by 2 points to 33%, and decreasing the share of independents by 2 points to 48% and Republicans by one point to 19%. Women should have been reported as 54% of the sample instead of 52% and men 46% instead of 48%. These corrected demographics actually better represent NJ voters than did the original reports.
We are sorry about the error and are redoubling our efforts to ensure such errors do not occur again. This particular error only occurred in our latest poll and its first data release. The weighting has been corrected and subsequent releases will use the correct weights.
Director, Eagleton Center for Public Interest Polling