The results are in from BMG’s exclusive statewide poll in next week’s special Senate election. Research 2000 interviewed 500 likely voters on Tuesday and Wednesday (and we do mean “interviewed” — Research 2000 does live interviews, unlike robo-pollsters Rasmussen and PPP). That means that our poll is the first (and so far only) one taken entirely after Monday’s final televised debate. Here’s what they came up with (margin of error is +/- 4%).
QUESTION: If the 2010 special election for U.S. Senate were held today, would you vote for Martha Coakley, the Democrat, Scott Brown, the Republican, or Joseph Kennedy, the Libertarian candidate?
ALL DEM REP IND Martha Coakley 49% 82% 7% 36% Scott Brown 41% 12% 85% 49% Joseph Kennedy 5% 1% 2% 11% Undecided 5% 5% 6% 4%
Particularly interesting in these numbers is the breakdown of unenrolled (independent) voters. Brown is ahead in that group 49-36; significant, to be sure, but not the overwhelming advantage suggested in the Rasmussen (71-23) and PPP (63-31) polls that came out recently.
The full set of crosstabs is here.
Some details for the wonkish:
The Research 2000 Massachusetts Poll was conducted from January 12 through January 13, 2010. A total of 500 likely voters who vote regularly in state elections were interviewed statewide by telephone.
Those interviewed were selected by the random variation of the last four digits of telephone numbers. A cross-section of exchanges was utilized in order to ensure an accurate reflection of the state. Quotas were assigned to reflect the voter registration of distribution by county.
The margin for error, according to standards customarily used by statisticians, is no more than plus or minus 4% percentage points. This means that there is a 95 percent probability that the “true” figure would fall within that range if the entire population were sampled. The margin for error is higher for any subgroup, such as for gender or party affiliation.
I assume the PAC paid for the poll?
<
p>hey, whose idea was it anyway for you guys to start a PAC?
<
p>Fantastic way to spend some money. Can build credibility over time and give you guys some juice.
<
p>Time will tell.
The PAC idea was from divine inspiration.
why billxi is still here.
I’ve never been called that before.
The dark angel of the blogosphere. Hey, that has kind of a nice ring to it.
Is there anything else available though?
Ernie knows.
<
p>Have the rights to “The Shadow” lapsed?
we paid for the poll out of our extravagant advertising revenues.
<
p>Nonetheless, the PAC is still an awesome idea! 🙂
Great move on the business side of things.
Let’s hope over time it will gain credibility, regardless of the BMG label.
<
p>So don’t let Chaley stop you from posting polls that hurt your candidates. I know he’ll want to.
Rasmussen has used “very likely voters” instead of “likely voters” to make the race seem closer. But then again, who votes on special elections? I don’t think pollsters have a good grasp on what group is voting, but “likely voters” in general had Coakley at 10 – 15. We don’t know the impact of identifying Scott Brown’s actual record, which has started this week.
<
p>Plus, we don’t yet know the full impact of the Tea Party and other national republican groups who have changed this race from outsider underdog Scott Brown, to a full on nutty Republican nation candidate.
There is no way to tell what “likely voters” means to any particular pollster unless they tell you in detail. It doesn’t have a standardized meaning.
but Rasmussen seems to be splitting groups in different categories. Not sure what questions he is asking to differentiate. But it seems odd that he has a “very like” category.
Nutty Republican nation candidate with at least two brain dead moves – demanding the Obama stay out of Massachusetts (didn’t Jesse Helms demand Clinton stay out of NC?) and sending his daughters out to complain about Coakely’s accurate ad. Maybe the girls didn’t know there dad actually did sponsor an amendment that did just what he was accused of.
… running straight up orthodox GOP conservative campaign will continue to attract enough extra-state support from the true believers to overcome the obvious deficiency that this isn’t how GOPers have won statewide in the past.
<
p>
… this is the correct link.
We don’t need a Nader here.
I honestly have no idea. Joe is more right leaning, but I think some left leaning folk how would otherwise hold their nose and vote Coakley are choosing him as a protest vote.
N=500, MOE=4%, Coakley-Brown=8%
<
p>which means, within the 95% confidence interval, they might be tied.
<
p>A bigger N would have shrunk the MOE. A MOE of 3.5%, with the same percentage, would mean that there is a 95% probability that Coakley beats Brown (not 95% Pr that Coakley beats or ties).
<
p>Oh well. More N is more money. I understand that you guys are trying to get polls, and have limited resources. Still, N=500 is definitely on the low end of good polling… I wonder if you could make a deal with DailyKos, 538, etc to see if they would spring for an extra 100 each. That sort of thing.
500 is right in the sweet spot of good statewide polling. You’ll typically get 400-600 and going up to 1000 in a statewide poll is much less typical.
<
p>Yes, within the 95% confidence interval they might be tied, but each outcome in the 95% confidence interval is not equally likely.
Per pollster, at the time of this post:
<
p>Rasmussen 1000 LV
Mellman Group (D-DSCC) 800 LV
PPP (D) 774 LV
Boston Globe/UNH 554 LV
Rasmussen 500 LV
Suffolk 600 RV
WNEC 342 LV
Suffolk 500 RV
<
p>The mean: 633.75. Only one poll of the eight has fewer than 500, which seems to reinforce my claim that N=500 is definitely on the low end of good polling. Another two hundred (as I suggest) would put the number polled in 4th out of 9, the middle third.*
<
p>I didn’t claim the poll was invalid — just that polling a few more hundred would lead to more robust results (and a smaller MOE, which helps better understand reality). I think the other polls in MA reinforce my claim, and implicitly that you are lowballing a smidge.
<
p>
<
p> * of course, so would another one polled (501 would be outright 6th).
Mine is gone, too:
<
p>”it’s orchestrated and a reflection of the man”? [6.00 / 8] Replies: 0
posted by: melora on Sat Dec 05, 2009 at 19:29:18 PM EST
attached to: Coakley ties her own hands, misleads public in Wilkerson Case
So at least 20 are missing.
It hasn’t been hidden. Somebody chose to delete the entire string without notifying anyone. I hope this is the only comment string “cleaned out” the day before the primary. Personally, I thought it was useful to have all the Capuano people on the record promising to vote Coakley in the general when/if she’s the nominee.
I posted a diary for this, rather than comment further here.
<
p>It certainly is “odd”. Certainly BMG has been receiving more than usual traffic, so it could well be a technical issue. Nevertheless I, too, find the timing of the removal disturbing. I also share your hope that this only comment string where this is taken place.
Another few hundred respondents = another few hundred * 1+refusal rate calls
<
p>All for about a half point of MoE.
<
p>It really confuses me why people are so obsessed with the MoE. All it is is two standard deviations away from the mean number in the survey (the one that gets all the attention). You don’t expect something to be two standard deviations away with equal probability as something one standard deviation away or something close to the mean.
<
p>You could easily say 2/3 of the time, the election will be plus or minus 2, rather than 95% of the time it will be plus or minus 4. And if you really want to worry yourself, 97% of the time it will be plus or minus 6! There’s your dead heat!
I meant 99.7%, obviously.
As sco points out, 500 respondents is fine for state polling. I’d add that the Research 2000 poll BMG commissioned was done by live interviewers, and was not a robo-call like Rasmussen (where it is easier and cheaper to get a higher N).
<
p>Of course, if some of the criticism about robo-calls is correct (especially the problem of selection bias), then getting the higher N involves a trade-off by sacrificing accuracy of another sort. In that way, it’s not necessarily the smaller MoE that “helps better understand reality” — it’s the techniques and voter screens you use.
Pollster Franken v. Coleman
<
p>Number: 55
Mean: 688
Median: 625
StDev: 263
Range: 300-1572
Number <500: 5
Number =500: 14
Number >500: 36
Number 400-600: 22
Number <400: 2
Number >600: 31
<
p>Based on Minnesota’s 55 polls for the US Senate, in a state which seems to be somewhat similar to MA in population, 400-600 doesn’t seem to be normal. 500-700 does however, and includes the mean as a not too important bonus.
<
p>Number <500: 5
Number 500-700: 35
Number >700: 15
<
p>I didn’t cherrypick this particular sample — it was just the first Senate race of 2008 I could think of which was competitive enough to get lots of polls. Of course, the expectation of close polling may encourage higher samples, precisely to reduce the chance that there is a “tie” within the MOE.
<
p>
<
p>Think it’s fair to say that 400-600 is too low; that 500-700 is probably a more appropriate range for describing what is typical?
“I think one of the molars on this pony might have a cavity. Thanks anyway, mister.”
My interest is in:
1. Getting good polling, and
2. Being creative on how to get it.
<
p>To address 1, I point out that N=500 is low. You disagreed. I came back with some data that reinforces my claim.
<
p>To address 2, I provided what I think is a creative way of getting ponies with good dental hygene.
<
p>
<
p>Instead of an academic discussion about a good sample size — which is what I expected from you — you come back and employ sure a lot of words to say I’m ungrateful, completely ignoring my critique of your 400-600 claim.
<
p>Bah.
I had a dentist appointment.
<
p>Yes, a bigger N is better, but you get really diminishing returns as N increases. The margin of error going from 500 to 1000 moves from roughly 4 to roughly 3. If you’re going to use MoE as your standard of what makes a poll good, you’re not really making that much of a difference.
<
p>They used to do statewide polls at N=400 almost exclusively as it’s on the low end of numbers you can get useful results from, but now as these things have gotten cheaper, I concede that is no longer the case — particularly as robopolling has become more accepted by the mainstream. Shame on me for becoming an old man while I wasn’t looking.
<
p>I would argue that there’s nothing mathematically wrong with a sample of 500 to get top-line numbers. If you want low MoEs in your crosstabs, however, you’ll want to go a little higher. That’s where the larger sample size starts to pay off because suddenly you have statistically significant subsamples work off of.
I had my teeth cleaned today at 10am. Lots of scraping. Strange coincidence.
<
p>It’s true that there are diminishing returns. I didn’t want to nit-pick, but the actual MOE of this poll is more like +/-4.3% and to get it down to +/-4.0% you’d need N=600.
<
p>
<
p>Indeed — for p = .5 it goes from 4.38% to 3.10%.
<
p>
<
p>Depending on which cross-tabs are interesting, N=800 gets you gender and unenrolled, N=1100 (roughly) gets you D and N=3000ish gets you GOP (only 11.6% of population).
<
p>
<
p>Final question for The Editors: how does pricing work? What would it have cost for an additional N=100? Is it like a base fee plus an additional cost per N+1?
Not going to get into that here. Email me if you want.
And you’ve doubled your work for what? A whole point and a quarter of MoE?
<
p>Question wording and order have more of an effect than that. And once you start talking about the effect and design of likely voter screens, it’s basically noise.
All the MoE gains in the world pale in comparison to the fact that we really have no idea who’s going to drag themselves out to the polls the day after a long weekend in the middle of January.
<
p>The population we’re trying to get a representative sample of does not even exist until election day. This is a problem with every political poll, but it is more acute in this election than it would be in November.
because if you don’t believe the crosstabs, you can re-weigh the poll to your preference. Larger N means larger N for crosstabs which means your re-weighing will be more robust.
<
p>This particular poll has N=92 GOP (18%). If you think that the GOP will make up 30% of the turnout, you can re-weigh… but the problem is that your magnifying the proportion which is based on a sample of only 92… you’re amplifying the error.
<
p>As you know, a larger N means a larger N for the crosstabs, which means the reallocation is more robust.
<
p>This is great because… if you don’t like the turnout sample that R2K is using, you can use their polling responses and reweigh to what you think it ought to be.
<
p>
<
p>I really think that Internet + blogging + open source calls out for larger N because it allows for re-weighing. It frees individuals from being restricted to just the top line of the poll… the “source” is open for us to tweak. A larger N allows for those tweaks far more robust.
it’s important not to compare apples and oranges here with the sample sizes. The Pollster list includes some surveys done with live interviewers as well as those done with robo-calls (Rasmussen, SurveyUSA). It would make sense to compare the N for all of the live surveys only. If you did, I’m certain you would find a lower mean, probably in the 500s.
lots of n=400 polls. The 400-500 range seems pretty normal to me. Rasmussen and PPP can do much bigger numbers because they’re robo-pollers. You pays your money and you takes your chances.
<
p>And yes, the result of a tie is barely within the margin of error. So what. The numbers are the numbers. If they had polled 800, maybe it would’ve come out 49-42, in which case, same result.
and I don’t mean to sound ungrateful. But:
<
p>
<
p>See above. I’m not sure what “pretty normal” to an expert in law and opera means with respect to sample size in state polling. The sample of polls in MA for this race shows that N=500 is on the low end of things.
<
p>
<
p>Indeed. There’s little evidence (per Nate Silver and others) that robopolling has less reliable results. In any case, if paying more money improves ones chances, perhaps it’s worth it.
<
p>
<
p>No, not same result. We’d have a higher confidence that the race was tighter, instead of what we have now (a lower confidence that the race is less tight). Not at all the same results.
<
p>
<
p>I must say, I’m a bit bummed you didn’t bite on my suggestion. Since you’re using the same polling outfit as kos and since he’s really big on polling, I think he might go for the idea of adding to the “N” of state blog polls. Sort of a partnership thing. To my knowledge 538 doesn’t do this sort of thing (commission their own polls), but they’ve gotten big enough that Silver et al might be willing to subsidize some polls if they feel there is a lack of good polls for that particular race. Getting multiple blogs with different audiences to “chip in” and help raise the N for polls (or the frequency of polls) seems like a pretty reality based partnership idea to me…
Most errors in polling are not due to statistical error, but due to sampling error. Sampling error is also much harder to measure.
<
p>To clarify, the 95% confidence interval assumes a uniformly random samples of voters. However, that’s almost impossible to get, so pollsters use various weights and so forth to simulate a random sample. Mistakes in this process are what causes polls to be far more inaccurate than any error caused by statistical variance.
Interesting–key is still, or course, getting out the vote.
<
p>This is kind of what I was getting at. Kos is even going to use the same firm (contract with R2K), so I’m not sure why he’d expect different results.
<
p>My question for cos, others: is it better to have two N=500 polls within a few days of each other at this point in the race, or better to have one N=1000 poll? This isn’t a hypothetical — methinks that if the case is made that a single N=1000 poll is better (for cross tabs, lower MOE, whatever reason), then there is a real opportunity for state and national blogs to combine resources to get a better final product.
<
p>Thoughts?
One out of every twenty polls (on average) is going to be outside the ‘true’ 95% confidence interval anyway. Better to have two polls that either reinforce each other or are wildly different than a single poll with a large N.
I don’t know that the kos poll will ask the same questions, but it will almost certainly be done with the same technique, since it’s the same polling firm.
<
p>Maybe I’ll ask the question this way: if you have a poll with N=1000 and you split (randomly) it into two N=500 polls, are you better off? Isn’t that the same question, given that we’re controlling for polling method (same firm)?
<
p>To control for everything — question, method, firm, etc: better to have two N=500 polls, one N=1000 poll, or (c) doesn’t matter because the N=1000 poll is two N=500 polls?!?!
If I reach into a bag of different colored marbles and grab a handful, I’m not always going to get the same handful every time, even though I’m using the same technique (stick hand in bag, grab as many as will fit).
<
p>To answer your question about which is better two N=500 or one N=1000, with the assumption that they’re both taken by the same firm on the same day, I would say it depends on what you’re after. If you want to break down gender/age/region/etc, then the larger one is a better bet since your subset N’s will be bigger. If you don’t care about that, then having two polls could either self-reinforce or show you that there might be something wrong (that is, non-representative) with your sample.
<
p>But that’s just my opinion. An old man like me has been conditioned to think of these things as prohibitively expensive, so large sample sizes strike me as extravagant.
At least in terms of tracking. We can see what effect Coakley’s going on the offense or Ayla’s whining has had.
<
p>It’s probably quite a bit better to get two N=500 polls than one N=1000 poll, especially at the late stages of a race like this one. Theoretically, if you took two identically worded and screened polls at the same time, both including a sample of 500 respondents, 95% of the time the numbers of each candidate will be within 4% of the other poll — but much more likely, within $2 of each other (or closer). Essentially, this is what you’re doing when you do a N=1000 poll.
<
p>Having two polls — even at N=500 — a few days apart is more valuable, however, because this is the stage of the race that preferences change quickly (as voters get tuned in). I’d rather have two smaller polls at different times than a larger poll capturing one moment of time.
Our friends at RMG are working overtime these days. Within an hour of this post going up, they had one by Mike Rossettie “debunking” Research 2000 as a respectable polling organization, claiming that R2000 routinely shows Dems doing better than they should.
<
p>Of course, the post is bullshit — a brilliant exercise in cherry-picking results to fit your theory, and ignoring others. Responding in detail is frankly not worth my time. But here are a couple of things to consider.
<
p>First, of the 14 polls that Rossettie chose for his post, R2000’s result in 10 of them was within the margin of error of what actually happened in the election. Guess what, Mike — they call them “margins of error” for a reason. If a poll result is within the margin of error of the actual result, the poll was right. So you can toss those 10 (in several of which, by the way, R2000 had the R overperforming) right out the window.
<
p>Second, of the remaining four, in one of them (NM-Sen) R2000 showed the Dem doing worse than he actually did. And of the other three, sure, R2000’s result was outside the MoE of the actual result. Doesn’t mean they were “wrong,” since things often change quickly before an election. And even if they were wrong, so what? It happens to all pollsters — again, they call it a 95% level of confidence (rather than a 100% level) for a reason. And by the way, Rossettie of course doesn’t mention that R2000 was one of the very few pollsters in the country to have accurately predicted what was going to happen in another highly-contested, teabagger-infested special election, namely, NY-23. (Amusingly, in that race, PPP had an enormous 1,754 likely voter poll that showed Hoffman up 17 points. Oops.)
<
p>Methinks Mr. Rossettie doesn’t actually know very much about how polling works. I recommend to him our excellent summary, from which I suspect he will learn something if he bothers to read it.
I’ve heard a lot of buzz from people who were briefly enamored of Mr. Brown but have since come to their senses. Given the media scrutiny of his inconsistencies and, shall we dare say it, hypocrisy, I suspect the flirtation has reached its zenith.
Yeah, she’s ahead by 8 points, but this is the first poll that puts her under 50%.
but shows that nobody should feel good about winning this race.
<
p>I still think Coakley wins by a 15 point margin, but that’s only if the Dems work hard through Tuesday and there isn’t a major gaffe/expose. Either of those assumptions fail and it’ll be a squeaker.
but this is a great use of PAC resources 😉
<
p>I loves me some polls, though unfortunately they don’t come cheap, but they’re well worth it (not mention both fun and newsworthy for BMG).
It’s to support candidates. News on that soon … perhaps after Tuesday.
Just incidentally. 🙂
. . . but, a new Suffolk poll has Brown up by 4.
<
p>http://www.bostonherald.com/ne…
<
p>Gulp.
<
p>We got more work to do . . . and I don’t have a good feeling about this . . .