Oboy, here I am posting not one but two posts in the same day. Here is something interesting - so interesting that I am actually posting for the second time today - something for the election year:
I attended Dr Karandikar's talk at The Hindu centre, the day before yesterday...
He spoke on the science behind opinion polls and it was very interesting to know that statistics can be used so powerfully. With a sample of about 4000 from each region you go on to predict the outcome at the national level... fine..
I did raise a question though - whether the model throws up variations in the regions. For example if Tamil Nadu is a region, the historical behaviour of voters in Kanyakumari has differed from that of South Chennai voters (keeping fingers crossed) - He replied that they were only looking at the final outcome at the national level and not so much at regional variations - that would require (micro statistics is what he said) a much more complex and costly survey involving more time etc...
But I am still left with the question... in a region with inherent patchiness, clumpiness call it what you may - small domains of peculiar but persitent convictions and bias - does it even make sense to take a random sample and ask them about national outcomes? I mean you go to region like south chennai and pick a sample that has been randomised not with respect to the biases in south chennai but with respect to general ideas of community etc, is it likely to yield correct results - you might pick a sample that is more characteristic of Tiruchi or Kanyakumari...
Is there any meaning in not doing a survey employing micro statistics (I assume that is like a micro canonical ensemble)... Dim memories of tackling Kerson Huang flash in the corner of my mind.
Well said. There is an inherent weakness in almost all random sampling methods. In small as well as large samples, which lies on the basic principles of law of averages, various statistical methods like standard curve and skew distribution are used. Even with regard to basic measures like mean, median and standard deviation, there is a margin of error say from - 2 to + 2 in large samples.
ReplyDeleteAgain, the error lies in another direction also.
Voters, most if not all, do not speak for whom they would be voting for. So, their responses to environmental conditions are taken and projected to be applied as their voting pattern or behaviour.
For instance, a question like this ," Do you agree that the Tamils in Sri Lanka are quite aggrieved ? "
a. very strongly. b. yes. c. somewhat. d. no .
A voter's answer as "a" may project him to be a voter of Vai.ko. whereas an answer "c" may lead to his being projected to belong to the congress party.
The questions are to be worded in such a way as to elicit the emotional responses of the voters.
In any case, the sum total of all the replies of all the answers in an election year, or the mean thereof, may vary very fast, say from today to the date of election.
That is how, last time, almost all the opinion polls missed the goalpost.
subbu thatha.