Most polls (look at our poll page here) missed Joe Sestak's win over Arlen Specter last week by 5 to 7 percentage points.
At the National Journal, pollster Mark Blumenthal ponders the late-deciding voters and the undecideds (if they did vote at all, it was probably for Sestak), but says the biggest factor in the undercount was the difficulty of polling a low-turnout primary:
A survey conducted in early May by Franklin and Marshall University, for example, showed Specter leading by 9 points (38 percent to 29 percent) among all registered Democrats. But they found Sestak nominally ahead (38 percent to 36 percent) among a much smaller subgroup of likely primary voters.
And the tracking survey conducted by Muhlenberg University for the (Allentown) Morning Call showed the race finishing in a dead heat (44 percent to 44 percent) among all likely voters. However, when poll director Chris Borick narrowed the universe at my request to the subgroup who said they were "definitely going to vote" in the primary election, Sestak had a bigger advantage during the final week of tracking. For the last four days, Sestak led by 5 points (46 percent to 41 percent) among the definite voters.
. . . But the relatively consistent pattern in this case suggests that in low-turnout primaries, especially where vote preferences show consistent differences at various levels of turnout, pollsters would do better to report a range of results -- showing how vote preferences differ as the likely voter universe narrows -- than a single set of horse-race numbers.
Graphic: National Journal