Date:         Thu, 10 Jan 2008 14:24:29 -0600
Reply-To:     Mary <mlhoward@avalon.net>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Mary <mlhoward@AVALON.NET>
Subject:      Re: Slightly OT: New Hampshire Polling Fiasco
Comments: To: Peter Flom <peterflomconsulting@mindspring.com>
Content-Type: text/plain; charset="UTF-8"

I'm starting to get it.

Phone call fatigue is another huge issue.  There are many "fake" polls that really are push-polling to get people to turn against one candidate or for another; event the legitimate campaign polls make up some fake name rather than saying that they represent a campaign.  So when the REAL pollsters call on the day or two before the election, people have just had it with phone calls altogether.   I had easily 50 phone calls on the Sunday before the Iowa election, and usually I love surveys!

-Mary
  ----- Original Message ----- 
  From: Peter Flom 
  To: Mary ; SAS-L@LISTSERV.UGA.EDU 
  Sent: Thursday, January 10, 2008 2:15 PM
  Subject: Re: Slightly OT: New Hampshire Polling Fiasco


  Mary <mlhoward@AVALON.NET> wrote
  >Peter, 
  >
  >I guess I really don't understand the difference- could one not compare the demographics of those who responded to the census characterisitics of the state from which the sample was given, or from actual voting demographics from previous years, to determine if the non-response was biased towards certain groups, and thus weight the results accordingly?
  >

  OK....let's take two examples

  1) Sampling error.  Let's say that, after completing the survey, you realize that 20% of your sample is Black, but that you think 30% of the voters will be Black (based on previous years).  You can then weight responses from Black people more heavily.   That assumes that the reason they are not answering has something to do with their being Black.

  2) Non=response due to partisanship, though, is different.  If people are not responding because they don't care who wins, then how do you account for that?  It *may* be that this lack of caring is related to demographics.....but may not

  In missing data research, it's the difference between missing at random and nonignorable nonresponse.  That is, is the reason for missingness related to the DV, after accounting for the IVs? If not, then you have MAR; but if it is related then you have nonignorable error.  MAR is relatively easy to deal with; NINR is pretty hard

  Peter