NEP raw data of exit polls released?

[davidgmills]

What is this?

http://www.icpsr.umich.edu/org/announce.html#nep


Mystery Pollster says that it is the raw data from the NEP exit polls.

Can anyone confirm?

[Igel]

1. Dunno.

But there's NEP exit poll data at http://www.ropercenter.uconn.edu/

Don't know if it's the raw data or just a complete set of the final data. (Although some data would invariably be raw: # of people interviewed/non-response at a given precinct.)

On edit: Need I add that it costs a few $?

[waz_nc]

2. You can get it at the Roper Center for $79

You get a CD that contains the national and state exit poll data for 2000 and 2004. You can download the questionnaires free.

http://www.ropercenter.uconn.edu/2004_presidential_election_polls.html

Here's what you get.

Each study includes:

ASCII data files for the national and all state polls
SPSS portable files for the national and all state polls
Crosstabular results released on election night for the national and all state polls
Complete documentation including the evaluation of the 2004 system of data collection produced by Edison and Mistofsky
Additional information contained on the CD:

National Election Exit Polls -- How Groups Voted in the Last 8 Elections.
Crosstabs from each Presidential Election Exit Poll from 1976 to 2004
Links to supplemental sources of information relevant to election analyses.
Presidential Approval Ratings from FDR to George W. Bush
Pre-election Bush/Kerry polls (horserace trends)
Current data tracking social, economic, foreign & domestic issues
Popular Vote/Voter Turnout

[jkd]

3. This looks like raw data

Each interviewee is weighted (lines 18-23 for Utah) This means that some respondents received 8 times as much weight as others depending on their demographics.

The data shows the characteristics of each selected voter, but it does not show how that data was interpolated to achieve the weighted value for each respondent. This data is significant and warrants study.

[jkd]

4. It's all here. This is great!

[Wilms]

5. But the raw data, or the raw data they needed to massage, first? n/t

[jkd]

17. Good question

The raw data are there, except the actual precinct is not identified. In Utah the weighted data reduced the raw exit poll numbers for Kerry by near 10%. If Mitofsky didn't adjust the formula on election day and stuck to the preconceived model, then he did a pretty accurate projection for Utah. The raw data was Kerry 29% and the weighted data gave Kerry just over 26%. The actual vote was 26% Kerry.

We still don't know the model that was used to weight each voter by their demographics. Each of the 828 respondents had to be weighted by the formula. If that model had been published before the election, then this data would provide more confidence in the process.

[Igel]

22. But that formula is Mitofsky's bread and butter.

Even if he copyrighted it or patented it, trivial changes would likely break it, and he'd never be able to prove that others stole it.

It's not the data ... it's the model.

[jkd]

25. That model would be in constant flux.

How closely it actually mirrored the election determines the accuracy of the model, and apparently the election results are used to tone it. That's why exit polls aren't that effective in detecting election fraud. They are part of the same process.

But, the within precinct error bears examination, if one could identify where the largest discrepancies occurred. They are not related to the poll itself, but how accurately the data was gathered and handled.

[Igel]

26. Yes, and while we can guess at the general shape of it,

the details are his.

And every year it's slightly wrong. I really have to wonder if he completely revises it each time?

[skids]

6. Another interesting set of data...

...since this thread will be crawling with people that know how to make use of data...

Check out what the NC SoS published -- voter registration, by precinct, with precise (I mean VERY precise) demographic information:

ftp://www.sboe.state.nc.us/enrs/voterstats11xx02xx2004.zip

And some of the other files on the page look pretty useful too.

http://www.sboe.state.nc.us/enrs/main_primary.asp?ED=11xx02xx2004&EL=GENERAL&YR=2004&CR=A

[southwood]

7. Individual precincts not identifiable!

"Although the data for each sampled precinct is designated by a code number, the precinct numbers in the data file do not correspond to the actual precinct number in any state. The data do no disclose the actual precincts sampled."

from:

http://www.mysterypollster.com/main/2005/01/nep_data_availa.html

[jkd]

8. Ask NEP for the actual precincts.

They have responded to some of my questions. Lines 10-12 give a precinct number. Just ask if they will convert these numbers to a county precinct number.

[southwood]

9. Individual precincts not identifiable (2)

Mystery Pollster:

"In response to an email query, an Edison Mitofsky spokesperson referred me to the following passage from the Code of Professional Ethics and Practices of the American Association for Public Opinion Research (AAPOR):

"Unless the respondent waives confidentiality for specified uses, we shall hold as privileged and confidential all information that might identify a respondent with his or her responses."

They feel that if they identify the polling locations it might be possible for a computer match to identify a small portion of actual individuals in the data. Some precincts are small enough that it would be possible to identify actual voters from their demographic data. They also feel that any effort to provide a precinct level estimate of actual vote or "within precinct error" would allow a user to identify the actual precinct and, theoretically at least, identify actual voters.

(...)

Like it or not, the released data are all we are likely to see."

http://www.mysterypollster.com/main/2005/01/nep_data_availa.html

[althecat]

10. Releasing individual WPE numbers for the 1400 precincts...

... would also enable someone else to do some actual modelling of whether the election was stolen.

I.E.
- the election must have been stolen in some precincts not all
- the WPE outliers are a good place to start looking
- while each individual WPE is not a good place to look as the samples are too small.. the distribution of WPE's by state/voting technology is where someone who wants to seriously look at Election Fraud might logically start

[althecat]

11. A further point.... You can issue WPE data by precincts without...

.. compromising privacy. I.E. we do not need to know where the exact precints are - just what state they are in and what voting technology was used.

Secondly... if the problem is the small precincts they could leave those out. What I would be looking for in vote fraud analysis would be occurring in the big precincts not the small ones.

[southwood]

13. Identifying precincts

WPE per precinct is not released (in combination with exit poll results per precinct) because that would make it possible to calculate the actual election outcome ("vote count") and therefore to identify the actual precinct.
(this attempts to reply to your first post)

Also, it is not just voting technology that may have swung the election, there has probably been e.g. large scale "voter targeting" (purging, pre-challenging) based on extensive data mining, producing provisional ballots that were subsequently rejected. This effect would also show up in differences between exit polls and election outcome.

[althecat]

14. This is stretching things quite a mile or three...

Unless the WPE was produced with an accuracy of three significant figures or more I doubt what you are saying would be possible to identify the precincts - though even then it would require an enormous amount of work. That said it may be the BS explanation.

A combination of bumping the smallest precincts would resolve this too.

[southwood]

15. Well...

.. I would have expected raw data i.e. numbers (not percentages) in terms of exit poll results per precinct. If the WPE is given in 0.1 % accuracy that would get me pretty close to the actual election outcome (in a 500 voter precinct 0.1 % would mean 0.5 vote.) But, I must admit, I'm not much of a mathematician.

Nevertheless, it does seem no less theoretical then deriving the individual preferences of an individual voter from the raw data of a precinct level exit poll.

Basically, not publishing the raw precinct level data looks like a CYA effort, making it impossible for anyone to check on the Edison/Mitofsky outcomes, or to finger interesting "hot spots" for that matter.

Must sign off now - different time zone here.

edit: typo

[althecat]

16. AGREED...

"Basically, not publishing the raw precinct level data looks like a CYA effort, making it impossible for anyone to check on the Edison/Mitofsky outcomes, or to finger interesting "hot spots" for that matter."

AGREED...

[southwood]

12. Precincts not identifiable (3) - calls for outrage!

This means two things.

1. As mystery pollster points out in the post linked above, it will not be possible to independently check the infamous "withing precinct error" (or its potential determinants such as the Reluctant Bush Respondent theory, the weather, type of voting machine or "irregularities".)

2. It will not be possible to identify specific precincts with large disparities between exit poll outcomes and election results, i.e. the sort of places that merit further investigation. A.k.a. the hot spots of "irregularities".

The reason given for not identifying polled precincts is very unsatisfying and theoretical, and very much against the public interest of finding out whether election outcome actually matched voter intent. I'm surprised that this way of making public exit poll results has not yet caused public outrage.

[Igel]

23. It does and it doesn't.

The precinct I worked at averaged less than 70 voters on non-presidential years, maybe 120 on presidential election years. You'd want to survey at least 20 or 30 of those people, maybe more. So that would be every 4 people. Black, hispanic, Asian, Russian ... not many of each.

Find out how "the" 60 year male Asian voted, go to the publically available poll book ... find Shinichi Shibatani, check the birth date and registration (dem/repub) in the book. For that precinct in Rochester you could probably identify 20 or 30 people from the exit poll data. For smaller precincts, it might be even a higher percentage. For large districts, not the same problem.

Poll data isn't good for the one thing I thought it'd be useful for ... checking turnout. Dang.

[ClintCooper2003]

18. Um, NO, THIS IS NOT THE RAW DATA! LOOK AT THE DATES!

They all say Nov. 3rd on them. These are the CORRECTED data!!!

[ClintCooper2003]

19. And isn't it kind of funny how all of the exit polls now match the results

EXACTLY???? They tweaked the exit polls. Don't trust these. Use your head.

[skids]

20. While the data could be fudged...

...you are looking at the wrong files. Look at the "data", not the
crosstabs. That's what you call raw results. The crosstabs also have absolute numbers of data in them, but they are just for giving a rough overview.

[skids]

21. WE CAN BEAT THIS. How to figure out the precincts' real numbers.

...at least in North Carolina...

The NEP polled at a little under 1% of all precincts. The chances of noticing an exit poller working a precinct at the same polling place as you proably brings brings the odds up to 1/100 that any given person knows one of the precincts where the NEP polled.

Out of thousands of DUers and other activists, we should be able to identify some of the precincts (we have to be careful that we know that the pollsters were NEP as opposed to some other local exit poll.) Then it becomes an excercise in elimination to map these precincts to the ones in the NEP file.

As an experiment, I took the detailed voter registration information for NC which I posted a link to above, and wrote a program to eliminate precincts that did not have as many voters of an age/sex/race/party registered as the exit poll interviewed. I chose only those categories because the answers are reliable -- for example I did not filter on "hispanic" voters because the ethnicity question is optional on registrations.

Now, admittedly, my program needs a little work. The below results are probably better than a more rigorous treatment would provide. Though, my guess though is that the best results mostly hinged on rare birds like "black female republicans age 26-40" and should hold up.

Out of 2835 precincts in NC, the possible NC NEP precincts were limited in this fashion:

Number of possible matches for NEP precinct  34: "1415"
Number of possible matches for NEP precinct  56: "496"
Number of possible matches for NEP precinct  50: "1620"
Number of possible matches for NEP precinct  57: "504"
Number of possible matches for NEP precinct  24: "1678"
Number of possible matches for NEP precinct  77: "1290"
Number of possible matches for NEP precinct  12: "1540"
Number of possible matches for NEP precinct  82: "1912"
Number of possible matches for NEP precinct  14: "1556"
Number of possible matches for NEP precinct  18: "261"
Number of possible matches for NEP precinct   5: "1677"
Number of possible matches for NEP precinct  16: "2135"
Number of possible matches for NEP precinct  81: "1566"
Number of possible matches for NEP precinct  53: "1597"
Number of possible matches for NEP precinct  88: "1464"
Number of possible matches for NEP precinct  92: "2070"
Number of possible matches for NEP precinct  46: "408"
Number of possible matches for NEP precinct  85: "1504"
Number of possible matches for NEP precinct  64: "2410"
Number of possible matches for NEP precinct  17: "712"
Number of possible matches for NEP precinct  68: "1698"
Number of possible matches for NEP precinct  94: "1928"
Number of possible matches for NEP precinct  31: "2232"
Number of possible matches for NEP precinct  27: "2156"
Number of possible matches for NEP precinct  48: "1538"
Number of possible matches for NEP precinct  37: "1937"
Number of possible matches for NEP precinct   4: "1481"
Number of possible matches for NEP precinct   9: "123"
Number of possible matches for NEP precinct  25: "961"
Number of possible matches for NEP precinct  96: "1493"
Number of possible matches for NEP precinct  60: "760"
Number of possible matches for NEP precinct  39: "1513"
Number of possible matches for NEP precinct  26: "2557"
Number of possible matches for NEP precinct  30: "1695"
Number of possible matches for NEP precinct  97: "1525"
Number of possible matches for NEP precinct  90: "1648"
Number of possible matches for NEP precinct  19: "351"
Number of possible matches for NEP precinct  69: "245"
Number of possible matches for NEP precinct  75: "2152"
Number of possible matches for NEP precinct  10: "1960"

Now, divide the above numbers by 13. Why? Because the NEP data recorded the congressional district number of each survey. There are 13 districts in NC. The reason I didn't have my program filter that is because we do not know for sure yet that the number in the NEP file correctly corresponds with the official NC congressional district number. It probably does (NC districts 1 to 13 had races in them) but we shouldn't take that for granted. It should be possible to figure out which districts are which with a bit of effort, if not.

Working with a list of *all* precincts of course is infinitely inferior to working with a list of precincts where we know the NEP was at the polling place. We probably won't be able to get them all, but we CAN figure out a good number. The number of NEP precincts in each congressional district ranges for 3 to 6. If we did know all the NEP polling locations...

What we need is to collect trustworthy data from people who ran into exit pollers. Not my specialty. Plus I have other work for USCountVotes to be doing. Though I will keep an eye out for detailed registration data files from other states.

My program is available for any Perl coders who would like a few code fragments to jump-start their own analysis. PM me or wait a day for the SourceForge mirrors to catch up and look here for a file named nc_nep.pl.

http://cvs.sourceforge.net/viewcvs.py/uscvprogs/project/one_offs/

[southwood]

24. Miraculous! n/t

[skids]

27. Has anyone figured out WGT vs WGTN vs XSWGT?

There are three weights in the data files.

WGT in the state file -- that's easy enough, the weight for the statewide survey.

XSWGT in the "ALL" file... what's this for?

WGTN in the "national" file... does this include the flat statewide rate mentioned in the doc, or not?