Yet another addendum to an addendum

This is another addendum to my recent post on the referendum. Kursat left the following comment to that post, which definitely deserves a reply:

I've been curious for a long time about how the votes of undecideds are distributed. I know most of the cases it's distributed according to relative frequencies which I don't find quite right since I am suspicious about randomness of the respondents of the survey. I just wonder whether KONDA has any statistical tool/model on that or they just religiously believe the respondents are random.

I personally believe if they are using relative frequencies to distribute, this will have a downward pressure on YES causing YES voters to be seem less than its actual value given my religiously belief of the data is nonrandom.

As I explained in the first addendum, KONDA tries very hard to ensure random sampling (they have a program that does random sampling at the street level and then go to the selected addresses even if they are in the middle of nowhere- all interviews are face-to-face by the way). Because of that, they distribute undecideds according to relative frequencies, as you guessed. However, even if the data is random, there is a strong argument that undecideds should not be distributed according to relative frequencies. My friends at Istanbul Analytics/ GlobalSource summarize the issues at stake at their report just before the referendum:

It is reasonable to conjecture that the undecideds will be distributed along a bell curve on Sunday, meaning that apportioning them (as is the current custom in the industry) pro-rata to YES and NO votes would not bias the results. There are two caveats. The first is an event that would significantly change the perception of the undecided voters. In our past comments on the referendum, we cited several such events. Of these, the most noteworthy is the backpedaling by the PKK-BDP-Ocalan Axis, which switched from impartiality to boycott. We assume that in large Kurdish cities poll security will be provided adequately, allowing the voter to express her preference freely in the ballot box. But, in rural areas and townships, the PKK/BDP pressure could work to the advantage of rejection. We make this conjecture, because poll data suggest that no-shows would have likely voted to ratify.

The second caveat is the observation, made by very reliable pollsters that the polarization of the nation and the mutual threats of retribution from the respective camps, coming on top of the growing fear of being wiretapped, have made many voters very conscious about expressing their views freely. These voters may be represented more than proportionally among the “undecideds” classification. We can’t estimate the magnitude of this effect, but would speculate that it works in favor of NO votes.

You could also argue that more potential "no" voters than "yes" voters decided not to go to the polls, as the so-called white Turks were still on vacation or preferred their Sunday brunch to their civic duty. Your friendly neighborhood economist, who was recently called a "grey" Turk by the editor-in-chief at the Daily News, did not even bother to look up where he was supposed to vote, as he is totally fed up with the small-town-merchant (kucuk kasaba tuccari in Turkish) mentality of the AKP and an opposition leader who is not even able to vote.

But at the end of the day, KONDA was right on target, so I guess  distributing according to relative frequencies was not such a bad idea after all. Maybe, the GlobalSource and white Turk effects canceled out:)

You could potentially come up with a statistical model of undecideds, but you'd have to ask specific questions on the survey towards that goal. Some food for thought for me until the general elections, as I am their Econ. consultant... 

As a side point, I am really not sure how KONDA managed to be so off-the-mark at the 2009 local elections. My only explanation is that they ended with a black swan in their sampling.