Google Consumer Surveys 50-state breakdown

[cinyc]

So this... thing has coughed up some more numbers, featuring these gems:

Illinois C+4
Kansas C+19
Utah: 21-15-11 T-C-L
Colorado: C+19
Indiana: C+6
Arkansas: T+4
NM to the left of Delaware, but to the right of Kansas
Oklahoma to the left of Texas
Idaho to the left of Georgia
Trump under 20 in three states, but NOT DC.

Something's probably wrong with the IP locating in Kansas.  No way the state is C+19.  I'm almost tempted to run a Kansas poll with a "not from Kansas" option just to prove it - but I won't.

[cinyc]

So this... thing has coughed up some more numbers, featuring these gems:

Illinois C+4
Kansas C+19
Utah: 21-15-11 T-C-L
Colorado: C+19
Indiana: C+6
Arkansas: T+4
NM to the left of Delaware, but to the right of Kansas
Oklahoma to the left of Texas
Idaho to the left of Georgia
Trump under 20 in three states, but NOT DC.

Something's probably wrong with the IP locating in Kansas.  No way the state is C+19.  I'm almost tempted to run a Kansas poll with a "not from Kansas" option just to prove it - but I won't.

We should crowdfund a Google Consumer Survey for Kansas where the only question is "Are you from Kansas? Yes/No/Not Sure"

That would be interesting.  I would probably phrase it "Are you currently in Kansas?", and drop the Not Sure option.  GCS would probably make us add a "I prefer not to say" option, though.

Thinking about it, an actual state poll might not cause as many problems as a U.S. poll due to weighting.  My South Dakota poll was pretty evenly geographically distributed, at least in theory.  Now that all these IP addresses in the Washington Post Article resolve to a lake in Wichita, a statewide poll will probably only include so many of them before moving on to other parts of the state.  The Wichita sample might get skewed, but the rest of the state should be okay.

A nationwide poll (which is what the GCS 50-state surveys ultimately are based on) might not have as fine a geographical distribution - just being from "Kansas" might be good enough for a national poll, instead of them further breaking down Kansas into other areas.  And if more people with IP addresses from "Kansas" actually aren't from the state, their whole Kansas subsample would be off.

[cinyc]

Google Consumer Survey's 10/20 results (click on the image for a larger view):



GCS seems to weigh their 50-state polls the same way they weigh the polls we've been conducting - by Internet users.  That's a bit bizarre for a poll of likely voters, but it is what it is.  As an experiment, I extracted the raw numbers and reweighted them according to the method amdcpus and I have been using to weight our GCS LV polls.  The Trump-Clinton margin raw data was, on average, 1.08 points more favorable to Trump, probably because GCS samples tend to be male-heavy.  The Trump-Clinton margin using the amdcpus method was, on average, 1.6 points more favorable to Trump, likely because younger voters are less likely to vote Trump, but get more heavily weighted in the GCS Internet User method.  Three states - Indiana, Pennsylvania and Maine - flipped to Trump as a result of the amdcpus method.

I didn't do this exercise to question the GCS polls.  As I've always said, polls are a snapshot of the electorate at a given point in time using the methodology that the particular pollster uses.  No more.  No less.  But how polls are weighted does matter.  Thus, any pollsters' polls are sometimes more useful in determining trends than anything else.

[cinyc]

The realignment!!

Or, only using the right column:

293-245

Your maps have Kansas incorrectly colored.  GCS has Clinton up there, no matter how you weight them.

[cinyc]

I made a systemic computational error with the LV margin.  (Fortunately, this is only an issue with this dataset, and not what I've weighted before).  Basically, I had the wrong formula in the female rows, and I didn't include the Female 18-24s while wrongly weighting the other Female age categories.  Fixing that issue lowered the average LV margin difference versus GCS Internet weighting to Trump +1.56.  Still a difference, but not as stark.  The same states flipped to Trump, but only barely.

[cinyc]

Why are 50 state polls always complete junk?

Partly because they often have very small sample sizes in many states, and partly because any pollster crazy enough to conduct such a type of poll must be a bad pollster in the first place.

They also don't weight the polls for the electorate in all 50 states, just to Internet users nationally.

[cinyc]

10/20-10/24 update: Gaze upon thy glory and tremble in terror

By Google standards, that's downright reasonable! They're improving.

Alaska should be Atlas red and New Hampshire Atlas blue.  With whopping sample sizes of 86 and 56, what could possibly go wrong?

[cinyc]

Google Consumer Survey's 10/25 Results:



This time, the margin in Google's model is only off the margin in the amdcpus LV model by an average 0.84 points.  The raw, unweighted results are off the Google model by 1.06 points.

[cinyc]

The 11/1 Iteration:

[cinyc]

Iowa and Deleware for Arizona, huh?

Delaware is the one state that flips if weighted by likely voter demographics.  Google itself has it going to Clinton - barely.  But the sample size is really small, 94.

[cinyc]

Just checking, we realize NH went from Clinton +12 to Trump +9 to Clinton +31 in the first column, right? Geeeeeeezzzzzzzzzzzzz

You mean over the last 3 releases, right?

That's bound to happen when you literally have polls of 70 people.  It's not a big enough sample to get realistic results.

[cinyc]

Just checking, we realize NH went from Clinton +12 to Trump +9 to Clinton +31 in the first column, right? Geeeeeeezzzzzzzzzzzzz

You mean over the last 3 releases, right?

That's bound to happen when you literally have polls of 70 people.  It's not a big enough sample to get realistic results.

Then why even post it?

Ask Google Surveys.  They've chosen to release polls with 70 respondents, and 538 is including them in their models.  I only post what Google Surveys has reported.  The sample size is included as a column for a reason.

[cinyc]

Just checking, we realize NH went from Clinton +12 to Trump +9 to Clinton +31 in the first column, right? Geeeeeeezzzzzzzzzzzzz

You mean over the last 3 releases, right?

That's bound to happen when you literally have polls of 70 people.  It's not a big enough sample to get realistic results.

Then why even post it?

Ask Google Surveys.  They've chosen to release polls with 70 respondents, and 538 is including them in their models.  I only post what Google Surveys has reported.  The sample size is included as a column for a reason.

Yeah, but if you think they're that worthless (I agree with you, as should anyone that isn't a discredited fraud), nobody's forcing you to post them...

They're not all worthless.  Most state polls have sample sizes that are sufficient for a decent poll.  It's largely the polls in the 3-5 EV states that are sometimes dubious due to low sample sizes... and Kansas.  I still suspect Google Surveys has a problem resolving all the non-Kansan IP addresses that are geographically resolved to a lake near Wichita, Kansas.

The Google Survey polls Atlas users have commissioned in the 3-5 EV states are at least plausible because the sample size is larger.

[cinyc]

Did this update again yet?

I don't see an update on FiveThirtyEight or Google's own websites.  I expect their final polls on Monday or Tuesday.  They usually take 5 days.

[cinyc]

Did this update again yet?

I don't see an update on FiveThirtyEight or Google's own websites.  I expect their final polls on Monday or Tuesday.  They usually take 5 days.

The last update was on the 1st.

That puts them on track to possibly update their numbers on Monday.  It's usually been 5-6 days between 50-state poll releases.

[cinyc]

The final Google Survey 50 State Poll Results.  Some of these are a real mess.  Salt to taste:

[cinyc]

Google Survey Reported:


Raw:



LV Weighted: