Atlas Google Consumer Survey polls megathread

[cinyc]

Is that a poll of Georgia?

You can always post images in the Atlas gallery.

[cinyc]

Do you need to know the ins and outs of polling in order to do this, or does Google do it all for you?

You probably need to know how to weigh the results, because Google's weighted universe is Internet Users, not RVs or LVs.  Other than that, GCS does everything but write the question.

[cinyc]

Hey, they changed the interface, removing some of the cross-tabs.  No more rural/suburban/urban tab.  The graphs look a little better, though.

But my poll got unweighted by Google all of a sudden.

[cinyc]

Thoughts on weighting:

1) Because they are one-question surveys, just polling the candidates makes the GCS poll a poll of adult Internet users, no matter how you weight them.

2)  Adding an option for "not registered in X state" and/or "not likely to vote in X state" is just a poor way to make them RV/LV polls.  In an ideal world, those options should probably be asked as a separate question to make sure respondents see the choice.  But that's expensive.  And including those options lowers the number of usable responses, increasing the MoE.

3) We can weight the one-question polls to RV/LV, but they're still polls of adult Internet users at heart.  At least Google's own weighting (when it works) takes into account geographic distribution, age and sex.

[cinyc]

How do we weight them, again?  Since it still has a ways to go to get to 333, I was thinking of providing pre-debate and post-debate numbers, along with an overall total.

The sample's probably too small to do that.  Even 333 is a somewhat small sample for a poll.  And, after weighting, the 333 is probably going to be even smaller than that, since Google doesn't include demographic data for all respondents.

How to weight these is an interesting question.  They're technically adult polls, and Google's weighting will give you the adult figure.  If you want to weight to RV or LV, start with amdcpus' spreadsheets here:

Template: https://docs.google.com/spreadsheets/d/10semu5xRZ-D6gG7J3d19asXVMnoCvh7G2RNuySOv_RI/edit?usp=sharing

(Obviously, copy it over to your own spreadsheet in Excel or Google Sheets, first.)

Get the 2012 RV/LV data for Tennessee from Census' website.  Or use my spreadsheet for almost all the info from all states here: https://docs.google.com/spreadsheets/d/11CzmVw3tIIDCDsIQ0QMlKDq45t6RqpKry8rJLwG_In4/edit?usp=sharing

If you want to do a RV weighting, you can get the data to put in amdcpus' amount and total columns (B&C) from Column E for your state of the "by age" tab on my spreadsheet.  You'll have to splt up the 45-64s into 45-54 and 55-64s.  Do that using the data from the Census bureau at the link at the top of my Splitting 45-64s tab in that spreadsheet.  Finally, you'll need to compute the male/female percentage to put in cells F2 and F10 of the template, respectively.  Do that by calculating the male/female percentage for TN in Column C of the By Sex/Race tab of my spreadsheet.  For RV, the percentage turnout by age column should be all 1.  

For a LV spreadsheet, you need to take the percent voted for each subgroup in Column K.  Assuming 65+ is the largest, the formula should be percentage of 18-24s voting/percentage of 65+s voting, in cells G3 and G10, percentage of 25-34s voting/percentage of 65+s voting in cells G4 and G11, etc. You also might want to replace the male/female percentages with the percentage that actually voted, instead of percentage registered.  You don't need to replace the data in columns B and C - the numbers in Column K will do that for you.

Finally, get the PERCENTAGE of age by sex for each candidate by clicking on the relevant boxes, and place that data in columns H, J, L, N, P, etc.  

When done, you can look at my SD spreadsheets to see if yours looks similar:
https://docs.google.com/spreadsheets/d/1GqiTvFhwr2uhPdGUx7jKSYPgdRi6nwedHLgkS8PK0DU/edit?usp=sharing

amdcpus' template should then work to weight the data.  It's a bit strange to weight adult data to RV or LV standards - but so be it.

I'll PM amdcpus to make sure my explanation is correct.

[cinyc]

Yeah, I probably won't be able to break it down weighted.  My results so far have very little in the way of 65+ voters, so it would produce some really odd results.  Fortunately for this, Southern states usually break on racial lines, not age lines, so having a predominantly 18-34 sample shouldn't be too terrible.  I won't weight it by gender yet, either, but I will post both the male and female numbers, if that sounds good.

That you have few 65+ voters is a bigger reason to weight it.  But Google will do it for you, if your universe is Adults, anyway.

I'm currently trying to simplify the weighting process by getting everything together in one spreadsheet.

[cinyc]

I think I made weighting even easier, based off of amdcpus' spreadsheet.  Just see the amdcpus Template revised tab of this spreadsheet.  All the information you need is in the other tabs of that spreadsheet or your survey:

https://docs.google.com/spreadsheets/d/11CzmVw3tIIDCDsIQ0QMlKDq45t6RqpKry8rJLwG_In4/edit?usp=sharing

[cinyc]

Looks like they're taking a day off on running mine, so I'll drop the current results just for fun.

GCS seems to run their polls in batches, so that you get responses at different times of the day.  The morning Internet audience isn't necessarily the same as the evening or work-time Internet audience.  So it's important not to get all the responses at 6AM, when early risers are answering the polls.

In my and amdcpus' experience, these polls should take 2-3 days to complete.

[cinyc]

Trump 58
Clinton 25
Johnson 9
Stein 4
De La Fuente 3
Smith 2

This definitely seems like an outlier, but it was similar without weighting for gender and age (54-26-12).  Johnson was doing far better in the first half of the poll, but fell off later (debate effect?).  I will be back soon with crosstabs!

What question did you ask?

[cinyc]

What candidate do you plan to vote for in Tennessee for President and Vice President of the United States

What were the choices?  Did you randomize them?

[cinyc]

What candidate do you plan to vote for in Tennessee for President and Vice President of the United States

What were the choices?  Did you randomize them?

The names of six candidates on the ballot and their VPs (randomized), with an "I am not registered to vote in Tennessee" option always at the end.  I recalculated all numbers as if the last option didn't exist (e.g. .365/.711 for Trump for 18-24s).

Thanks.  I ended up with 22% not registered in South Dakota in the raw tally of my GCS SD survey.  What was yours?  (I'm wondering what percentage of non-registereds Google usually picks up and how much it varies by state).

[cinyc]

Idaho poll just finished. Among 269 registered voters,
Trump: 50
Clinton: 27
McMullin: 12.5
Johnson: 7.5
Stein: 3

Thanks.

Same questions I asked ExtremeRepublican: What question did you ask, what were your choices, and if one option was "not registered in Idaho", what was the raw percentage that claimed that option?

[cinyc]

Idaho poll just finished. Among 269 registered voters,
Trump: 50
Clinton: 27
McMullin: 12.5
Johnson: 7.5
Stein: 3

Thanks.

Same questions I asked ExtremeRepublican: What question did you ask, what were your choices, and if one option was "not registered in Idaho", what was the raw percentage that claimed that option?

I asked the same question you did in your South Dakota poll, If the presidential election were held today, who would you vote for? Choices were Republican Donald Trump, Democrat Hillary Clinton, Independent Evan McMullin, Libertarian Gary Johnson, and Green Jill Stein. 20.2% said that they weren't registered. Also, did anybody else have a sample that really skewed male? The age numbers actually didn't seem too bad, but men heavily outweighed women.

My sample was heavily male, too.   It seems to be an issue with GCS.

[cinyc]

I don't think the third-party numbers are high for Alaska; the state has a long history of third parties doing much better than the national average.  Also, the current governor was elected as an Independent.  Here's a (partial?) list since 1968:

1968: George Wallace got 12.1% in Alaska, a little less than the 13.5% he got nationally but quite high for a non-Southern state (and Alaska is about as non-Southern as you can get!)
1972: John Schmitz of the American Party got 7.2; nationally he got 1.5.
1976: Roger Macbride (Libertarian) got 5.5, nationally 0.2
1980: John Anderson 7.0 (6.6 nationally); Ed Clark (LP) 11.7 (1.1)
1984: David Bergland (LP) 3.1 (0.2)
1988: Ron Paul (LP) 2.7 (0.5)
1992: Ross Perot (Reform) 28.4 (18.9)
1996: Perot 10.9 (8.4), Ralph Nader (Green) 3.1 (0.7)
2000: Nader 10.1 (2.7) Pat Buchanan (Reform) 1.8 (0.4)
2004: Nader 1.6 (0.4)
2008: Nader 1.2 (0.6)
2012: Gary Johnson (LP) 2.5 (1.0)

Not impossible then, although I'd still be surprised personally. I'd say the Johnson 15 is more likely than the Stein 6.

The third party numbers seem high in almost all the Atlas Google Consumer Surveys we've done, really.  It's an artifact of asking third parties by name, coupled with no real likely voting screen because we only get one question.

[cinyc]

My poll finished a few days ago, but the results will never load

The results should load.  Try logging into and out of your account, or using a different browser.

[cinyc]

I did a WA poll, and re-weighed found this, Clinton +10:

Clinton 47%
Trump 37%
Johnson 11%
Stein 5%

30-point gender gap

What exact question did you ask, what were the choices, and if you included a not registered/likely to vote/from WA option, what percentage of respondents chose that option?

[cinyc]

I did a WA poll, and re-weighed found this, Clinton +10:

Clinton 47%
Trump 37%
Johnson 11%
Stein 5%

30-point gender gap

What exact question did you ask, what were the choices, and if you included a not registered/likely to vote/from WA option, what percentage of respondents chose that option?

Question: "Washington State Voters: If the Presidential election were held today, who would you vote for?"

'Opt-out' question was "I do not plan to vote in Washington State" -- unconventional wording but I wanted to give it a try.  70/333 people chose that, so overall unweighted sample size of n=263 (MoE +/-6%).

Choices were my best effort to reflect the Washington ballot:

Hillary Clinton/Tim Kaine (Democrat)
Donald J Trump/Michael R Pence (Republican)
Gary Johnson/Bill Weld (Libertarian)
Jill Stein/Ajama Baraka (Green)

I excluded the two socialist options.

Thanks.  Unless you weighted the respondents without sex or age data, the weighted sample is probably smaller than just excluding the opt-outs, unfortunately.  Unfortunately, GCS can't impute demographic information for everyone.

[cinyc]

Why did you not include Castle, btw?

Castle's not the only minor candidate on the Washington ballot.  Alyson Kennedy/Osborne Hart (Socialist Workers) and Gloria Estela La Riva/Eugene Puryear (Socialism and Liberation) are also on it.

Adding the more minor candidates by name leads to poll clutter, which sometimes leads to more inaccurate results.

[cinyc]

Any suggestions on that?  I considered re-weighing on the assumption that the unknown group had a comparable gender gap, and a comparable oversampling of males, which came to Clinton +12 or so.  I also tried just separating that group and adjusting only known-gender people.  Got Clinton +10 rounded.  Not terribly happy with either option.

I've been ignoring those who we don't have demographic information for when weighting.  I haven't tried to readjust making assumptions, because whatever assumptions are made are probably wrong.  Usually, that subsample is only 10% voters or so, anyway, so they shouldn't move the margin much.  

In my SD poll, there were 63/500 without gender info, and a plurality chose the "not registered in SD" option (33.3% versus 22.1% overall).  There were 82 without age info, and an even higher percentage, 35.2% chose not registered.  So they're probably more likely to be out-of-staters, too.

Google's own weighting for Internet users ignores those without imputed age/sex data.

Out of curiosity, what were the raw (non-Google weighted) numbers for Washington?

[cinyc]

Trump leading by 3.9 in MT.

Is this poll complete or still in the field?

[cinyc]

Atlas Google Consumer Surveys so far:

State	Sponsor	Margin	Weighted to	Dates
Utah	amdcpus	Trump +6	Likely Voters	Oct 12-13
South Dakota	cinyc	Trump +17	Registered Voters	Oct 15-17
Tennessee	ExtremeRepublican	Trump+33	Registered Voters	Oct 18-20
Idaho	anthony1691	Trump +23	Unweighted	Oct 18-21
Georgia	rafta_rafta	Trump +11	Unweighted	?
Utah	Speed of Sound	Trump +3	Likely Voters	Oct 19-22
Alaska	Moderate Pennsylvanian	Trump+5	Likely Voters	Oct 19-22
Washington	Alcon	Clinton +10	Likely Voters	Oct 21-23
Maine	cinyc	Clinton +14	Likely Voters	Oct 26-30

I think I captured them all.  Please help me fill in the question marks.

[cinyc]

Also, here's the link to the survey for anyone interested. https://surveys.google.com/reporting/survey?survey=gdpj672ouix5gvzihxobgaf6jq

Thanks.  I get Trump+20, weighted for RVs (227 with demographic info):
Trump 48%
Clinton 28%
McMullin 14%
Johnson 7%
Stein 3%

I also get Trump+22 for LVs, but you didn't ask that question.

[cinyc]

I was going to run one of Florida or Iowa, but it didn't give me the $50 off coupon.

Unfortunately, the coupon offer seems to be over.

I will probably do another poll this week or next.  Any suggestions?  I prefer doing polls for rarely-polled states.  I'm thinking maybe Nebraska.  If the sample size is large enough, we might be able to guestimate NE-02.

[cinyc]

I was going to run one of Florida or Iowa, but it didn't give me the $50 off coupon.

Unfortunately, the coupon offer seems to be over.

I will probably do another poll this week or next.  Any suggestions?  I prefer doing polls for rarely-polled states.  I'm thinking maybe Nebraska.  If the sample size is large enough, we might be able to guestimate NE-02.

Yeah, I think Minnesota would be interesting.  We really don't have enough polling there.

The problem with Minnesota and the larger states is that doing a Google Consumer Survey poll for them largely duplicates what GCS does themselves.  The sample sizes of their 50-state national subsample is large enough that me doing a 333/500 or even 667-respondent poll would be largely duplicative but for ensured geographic weighting.  Plus, chances are Survey USA and maybe even Mason Dixon will re-poll Minnesota for their media partners in the next two weeks.  

Oregon has had 8 polls partially or all in October so far.   Indiana would be interesting, but, according to 538, 7 polls have been conducted partially or all in October (despite the state's ban on robo-pollsters), including GCS's own survey with 717 respondents.

In contrast, Nebraska has only been polled by 2 pollsters in October (GCS and CVoter) and GCS's sample size was only in the 200s.  That's why I'd like to stick to a smallish 3-7 EV state where there's little chance of getting a good poll.  We've already done Alaska and Utah, so those are out.

[cinyc]

^ The problem is also that these polls seem to be sh-t, but I'm glad you guys have the money to throw around.

A good number of these polls were done for free after the now-defunct $50 off your first survey coupon offer. And I don't think the results we've gotten so far are that far off from reality.  They at least seem plausible.