Someone should do a 2020 Google Consumer Survey poll

[cinyc]

Sorry, it kept being weird when I tried to add more candidates so I could only do the number that I did. Coupons are available if people want to make their own for free. This didn't cost me anything.

Where did you find the coupons?

The coupon offer was off during the late stages of the primaries, but is now back:  
https://surveys.google.com/offer/view_survey_offer

It is for $50 off your first survey, which, for a national survey, should buy 500 respondents.  Unfortunately, it is available for only those who haven't conducted a Google Survey before.

I think Google Surveys limits you to 7 potential candidates/options when doing a one-question poll.  If you include more than 7, they split the sample up, IIRC.  That really wouldn't work.  So, if you include a "I am not registered to vote/will not vote in the Democratic primary" option, you're down to 6 named candidates or 5 named candidates and a write-in Other.

Before you dismiss these, I'm pretty sure my GCS poll wound up being the most accurate Tennessee poll this cycle.

Perhaps, but most of our GCS polls were way off.  My Maine poll was a disaster.  Then again, so was UNH's poll.  IIRC, my late-in-the-field New Mexico poll wasn't terrible if you weighted those without demographic info at 1 - which is what we probably should do in the future.

I'm in the process of figuring out the 2016 state RV/LV weights, assuming Census asked the same voting questions they did in November 2012 and 2014.  It's probably too early to use a 2020 RV/LV screen, anyway.

[cinyc]

I am doing a 2020 survey right and I made the mistake of including Other as an option and 67% of the responses I have gotten so far are Other (Trump supporters?).  I do however have multiple gmail accounts so I'll try another one when I'm done with this one.

Did you do other with a write-in option to specify?  You can separately count those write-in responses, but it's a pain.

[cinyc]

Detailed crosstabs are obviously not too useful for a sample like that, but is there anything that jumps out at you re: demographic differences between those who picked Warren and those who picked someone else?

Most filters didn't change the order much, but there were some filters that provided interesting results. Limiting answers to just the Northeast brought Booker into almost a tie with Warren. Likewise, Booker did better when answers were limited to voters in Urban areas. Kamala Harris did better in the West but not by much. Warren most strongly performed in the West and Midwest. There were very few differences when looking at gender. The sample size is probably too small to really get anything else out of that.

I'm curious having never conducted a national poll - for national polls, does Google Surveys break down the results only by region or also by state?  And am I right in assuming you still get age data?

[cinyc]

This was really fun and I have a 2nd email, so I've started a Trump/Cuomo general election match-up since I think he'd be one of the weakest Democrats could offer as a candidate. I would do more Democrats, but multiple question surveys are very expensive.


Edit: These initial Cuomo vs. Trump results would be pretty surprising if they hold up.

Unless someone has a huge lead, they sometimes don't.  Depending on when the poll was put into the field, Google Surveys sometimes seems to ask its Android app users questions first, which skews the early results younger.  Thus, the early sample isn't necessarily representative of the country/state as a whole after the news/other site results are added in.

It is really interesting to watch the results come in and track how each type of geography/response type/age group/gender polls, though.  Almost addicting.  As you probably know, you can download the results in real time and track subsamples from batch to batch.

[cinyc]

Well, this is interesting:

If the 2020 Presidential Election were held today, who would you vote for given the following candidates?

Donald Trump (Republican Party) - 29%

Andrew Cuomo (Democratic Party) - 17%

Undecided/Unknown - 38%

Would Not Vote - 16%


If you factor out those who would not vote, Trump leads Cuomo 34%-21%, with 45% Undecided. I think I'm going to go without an Undecided option next time and see what happens.

Here's how they perform under particular filters, with non-voters factored out:

Men: Trump leads 37%-19%
Women: Trump leads 32%-22%

Midwest: Trump leads 32%-20%
Northeast: Cuomo leads 34%-26%
South: Trump leads 39%-15%
West: Trump leads 37%-20%

Your crosstabs sound plausible, except perhaps for West.  What states does Google Surveys consider West?  California must make up a good portion of its population, no?  What do they classify Texas as?  South, I hope?

[cinyc]

Google emailed me seperate results that were reweighted by age and gender:

Undecided 44.1%
Warren 31.2%
Trump 18.4%
Petersen 1.7%

Not registered 4.5%

Note that Google's reweight is to the Internet population, not registered voters, so it skews younger than what you'd get if you weighted for RVs or LVs.  The 2016 CPS RV/LV weights aren't available yet.  But your actual sample was too male-heavy, so the Warren percentage should go up a bit in either case.

[cinyc]

Wow, this looks really good, Castro. Would be nice if someone could do Montana and poll the special election, the Senate race and maybe the 2020 gubernatorial race.

For 500 responses, this would cost $700 with the coupon. Beyond what I imagine Atlasian users want to spend on this.

SurveyMonkey does multi-question polls cheaper ($1 per respondent for a national poll, I think, so $500 for up to 10 or 15 questions), but polls for states cost more and may not even be available for a small state like Montana.  SurveyMonkey will also allow you to only poll RVs as an option, usually for free, which is a slight improvement over Google Surveys.  Survey Monkey one-question polls are way more expensive than Google Surveys' - the same $1 per respondent, IIRC.

Three one-question Google Survey Montana 500-respondent polls would cost $225 sans coupon, which is cheaper than doing a multi-question poll.  BTW, Google Surveys at least used to give a small coupon with the final results e-mail if you polled again within 2 weeks - something I unfortunately missed when conducting my primary polls until it was too late.

[cinyc]

Though yeah, 500 responses instead for 50/27 would be much more useful if I can figure out how to do that

Was this a multiple question survey or multiple 50-perso surveys?

[cinyc]

You paid for a multiple question survey? WOW!

A 50-respondent multiple-question Google Surveys national poll costs $50.  It's $1.00 per response, instead of 10 cents per response.  (State multiple-question polls are $1.50 per response.)  A 50-respondent poll is likely nowhere near statistically significant, of course.  

We'd probably have to pool our resources to conduct a statistically significant multiple question poll.  It's too expensive.