AK-Lake Research (D): Trump +1

[Alcon]

Boos for using the awful term "statistical tie".

^

Burn it in a fire!

[Alcon]

Was this an internal poll? Because 538 just entered it into their tracker

Yes.  Not sure of their policy, but although a Dem firm, Lake is reputable and I doubt they'd massage their data for their client.  The bigger problem with an internal like this is that it's likely to be released selectively, i.e., only when it shows a good result.  But 538's staff may feel like they have enough information about past polls from Lake that this is less of a concern.  Not sure.

[Alcon]

Was this an internal poll? Because 538 just entered it into their tracker

Yes.  Not sure of their policy, but although a Dem firm, Lake is reputable and I doubt they'd massage their data for their client.  

http://elections.huffingtonpost.com/pollster/polls/lake-research-d-24267
Clinton 50 Trump 37 on April 2016 in Georgia.

http://politics.blog.ajc.com/2016/07/25/georgia-democrats-polling-shows-state-is-in-play-in-november/
Clinton 41
Trump 40

Just from their Georgia results, I would prefer a more traditional pollster to believe #BattlegroundAlaska.

That is a weird poll result, but it was during the primary, and Lake has a long-term record that's pretty good.  I agree we shouldn't enter internals, though, because many are skewed and the releases are so selective.

[Alcon]

Did we really enter an internal into the database?

mds32 likes to enter every single poll into the database, but I mean you literally added the MN-Breitbart/Gravis poll,so...

Which wasn't an internal...

I'm with him on this.  There are some bad polls out there, and polls with house effects, but the internals we get are so damn cherry-picked that there's almost no way to adjust for the way they skew averages.  Also, there's the nasty problem of credibly determining which public polls should be excluded.  Short of fraud, I say throw them all in and let things average out.  If they appear to skew one way, adjust for the house effect.  If they're crappy, consider them less the next time around.  But otherwise we get into subjective territory where crazy people start discounting CNN polls because THE MEDIA!11, etc.  Gravis is a pretty bad poll historically, and seems to have a GOP-leaning house effect, and I think it's better to adjust for that than throw them out selectively.

[Alcon]

Did we really enter an internal into the database?

mds32 likes to enter every single poll into the database, but I mean you literally added the MN-Breitbart/Gravis poll,so...

Which wasn't an internal...

I'm with him on this.  There are some bad polls out there, and polls with house effects, but the internals we get are so damn cherry-picked that there's almost no way to adjust for the way they skew averages.  Also, there's the nasty problem of credibly determining which public polls should be excluded.  Short of fraud, I say throw them all in and let things average out.  If they appear to skew one way, adjust for the house effect.  If they're crappy, consider them less the next time around.  But otherwise we get into subjective territory where crazy people start discounting CNN polls because THE MEDIA!11, etc.  Gravis is a pretty bad poll historically, and seems to have a GOP-leaning house effect, and I think it's better to adjust for that than throw them out selectively.

I think the argument might rather be that a poll conducted for Breitbart is basically an internal for the Trump campaign. Which isn't totally unreasonable.

exactamundo.

I know that's the argument, but I'm saying that the biggest problem with internal polls is not their bias, but rather that they are not always released (so you can't just adjust for a known house effect -- they cherrypick certain results to release) or are basically engineered to generate a specific result (so they can't be compared apples-to-apples).  Even for crappy, skewed polls like Gravis, those things aren't generally a problem.

You could argue that Gravis is doing one of those things -- presumably the latter -- and thus it's not sufficient to merely adjust for house effect, because they're actually changing their methodology from poll-to-poll in a way that defeats that adjustment.  But there isn't direct evidence of that.  And, in absence of direct evidence, you get to a place where you start having to make subjective calls based on the existence of conflicts of interest, etc.  That gets you into territory where you're having to explain why you do that with Gravis/Breitbart polls, but not CNN polls or Fox News polls.

Do I think it's relatively likely that Gravis/Breitbart polls are dubiously manipulated?  Yes.  Do I have any direct evidence they are?  Beyond a Republican house effect -- which happens with plenty of honest pollsters -- no...they mostly just seem crappy.  Can I say with high certainty that they probably are manipulated?  No.

And, in light of that, I'd rather throw them on the data poll (adjusting for R house effect, and down-weighting because Gravis sucks) than opening up Pandora's box.

That said, Gravis literally sends me spam emails I never opted into, so I'm not 100% convinced they're even a legitimate business enterprise, so...I feel ya on some level.

[Alcon]

I feel like we're talking past each other. The poll is conducted for Breitbart. The founder of Breitbart is Trump's campaign manager. So one could argue that a poll conducted for Breitbart is a poll conducted for the Trump campaign and is an internal. Then all the arguments you brought would apply.

Not quite, though.  I know it seems like I'm splitting hairs here, but I'm not.  Bear with me...

When a poll is commissioned out of the campaign budget, it's expected to serve some use for the campaign -- to test messaging, to provide actionable information, or to drive a media narrative.  That third reason is the only reason internals are released.  That means that even a well-conducted internal poll will only be released when it helps a certain narrative.  That also means that poorly-conducted, skewed internal polls are more likely to be released, since they're manipulated to conform with the narrative.  Taken together, the sample of internals you get is skewed both by manipulation and cherry-picking.  There is no reason for them not to be.

The Gravis/Breitbart polls are different in a few mitigating ways.  First, presumably all of them are released, whether it's good or bad for Trump, because they're released on a regular schedule.  That takes away the cherry-picking problem.  The second is that the media, at least in theory, has use for releasing unmanipulated polls.  Fox News viewers may "want" a certain outcome, but Fox News has other incentives to avoid manipulated or fabricated polls.  Internal polls do not.  You may think that Breitbart, by nature of why it's popular, has a lot less incentive than Fox News does.  I agree.  But Breitbart at least theoretically has such an incentive, like CNN or Fox; campaigns have absolutely no such incentive.

I agree with you that I'm incredibly skeptical of how much incentive Breitbart has to not manipulate.  But a lot of people are skeptical that CNN and Fox News aren't manipulating numbers to fake a competitive race...and it's true that they have an incentive to.  Dismissing Gravis/Breitbart opens up a slippery slope of objective interpretations about who's bought and sold and who's not.  I think that's reasonable for us to do in our personal evaluation of the polls, but I don't think it's a realistic question for a site like Atlas or FiveThirtyEight to adjudicate.

Sorry, I hope I'm not coming across as a pedantic ass (like usual!) but data curation is something I think a lot about for work, and I do think this is the only realistic policy Atlas could take.