We are rapidly descending into “African swallow carrying coconuts” level of debate.
Read it yourself. Unless you have some magic quotes up your sleeve, you’re wrong.
I think the graph at the top of the wiki article is most intuitive - the “peak” is the result that comes out as the headline of the poll, as it is the most likely (i.e. highest probability) of the possibilities, and as the numbers get further from the peak, they become less and less likely, as per the graph. Hopefully that makes sense when you look at the graph.
Is your position that with a given reported poll result (say, 40% +/- 5%) that the actual underlying distribution of voters in the population is equally likely to be 35%, 40%, or 45%? Is it your position that 40% is not the most likely reality within that range, based on this information?
I just want to make sure we’re actually disagreeing before I take the time.
Richard Parker, nobody is saying that all outcomes within the MOE are equally likely, so I’m not sure whom you’re arguing against there. What CarnalK said was that there is a 95% chance that Biden is within the range of 13-25% and Warren and Sanders are within the range of 14-26%. Which, yes, is exactly what the MOE means.
I said, “You also seem to think that MOE means any result in that range is equally likely. That is also incorrect. The most likely reality is the one reported by the sample. In only a small percentage of cases would the underlying reality be at the MOE margin (by chance as opposed to some other driver of inaccuracy unrelated to MOE, of which, again, there are many).”
He said that wasn’t true.
So it certainly looks like he disagrees. No?
That was an earlier post. And the problem I had with it was his use of “accurate.” The MOE is about the likelihood that your sample has diverged by the population by mere chance. It does not account for all the other possible reasons that your sample does not reflect the population. So in the context of a discussion of those other sources of inaccuracy it is quite wrong to say that " there is a 95% chance that Biden is within the range of 13-25%." That isn’t what MOE is.
No offense …or maybe some offense … but this is a pretty idiotic discussion.
RTF has done us the huge favor of grinding out some decent national poll aggregation, which we should all be appreciative of. The top bunch:
That aggregation falls not so far out of the range of 95% confidence, the MOE, of this single high quality so far outlier poll. It does not contradict the aggregated results at all. And it is true that taking this poll in isolation, one would make a better bet saying the true number is the center of the MOE than one of its farther points.
This seems like a basic Bayesian priors circumstance to me. The aggregation of other high quality polls is priors.
OK, there are really only two ways worth noting that the sample doesn’t reflect the population. One is sampling error, which has been the subject of most of the discussion. The other is coverage error, which is when your sampling frame doesn’t reflect the population as well as it should. But the combination of coverage error and differential response rates by different race/sex/age categories that make it into the sample is usually accounted for in the weighting, by breaking up the population into cells based on the above factors (and any other relevant demographic factors), which really just leaves you with the sampling error.
Of course, in a small-N poll such as Monmouth’s, you can’t have very many weighting cells because the more cells, the fewer respondents in each one, which gives you big sampling errors in the individual cells. So small N doesn’t just give you a big sampling error overall, it also degrades what you can do to rectify coverage error.
But even if it had a huge sample size, it’s still just one poll. Even a poll by an excellent pollster can be an outlier due to sampling error, because there isn’t a damned thing the world’s best pollster can do about sampling error. I’m not a big fan of the RCP polling average (that’s why I made my own) since they act as if all pollsters are equally good, which they’re not. But even their average is way better than focusing on a single poll.
Maybe I can split the baby here. Richard Parker and Andy are both correct in what they are pointing out as regards MOE in a single poll. But DSeid is also right about other polls providing Bayesian priors. Therefore it is far more likely that this poll is understating Biden’s strength than that it is overstating it. If OTOH we see another poll with similar numbers, then we may have to quickly shift our perspective and consider that the poll was the first “canary in the coal mine” for him.
One other thing that deserves mentioning here is that pollsters are sometimes guilty of “clustering”: being reluctant to report results that are outliers from what other pollsters find. So if nothing else, Monmouth deserves credit for not falling victim to this temptation (unless they actually found that Biden was several points behind and fudged the numbers to make it less of an extreme outlier).
I don’t want Biden to win the nomination. But I’m not so sure if this poll is “nice”, because I also really, really don’t want Sanders or Warren to win the nomination. To me, a nice poll would be one showing all three of them cratering and a more electable candidate surging. Bullock, Beto, Booker, or Castro would be fine. Harris has many skills and would also be better than the top three, even if she took a big hit in my eyes for the Mike Brown “murder” nonsense.
Heck, I had declared my home state senator Amy Klobuchar was dead to me after the revelations about how poorly she treats her employees; but at this point, if she were the only other contender besides those three I would gladly grab ahold of her like a drowning man on a piece of driftwood that floats by. So that makes six candidates running who would IMO be clearly better than the top three. I’m not being picky!
I can’t see much pressure in that direction at this point. Just for starters, you’ve got the two weekly polls, YouGov and Morning Consult, showing very different numbers for the three main contenders. That should provide plenty of cover for any pollster that might be hesitant to show differing numbers of their own.
We can argue all day and night about who’s ‘electable’ and who isn’t, so I’ll skip past that. But it’s clear that the presence of two dozen candidates in the race is causing prospective voters to throw up their hands in a way that hasn’t happened in the past when there have been six, eight, or even ten candidates running. Most voters aren’t even looking past the top four or five, if even that far.
And given your clear preference for a more ‘establishment’ candidate - especially Booker or Bullock or Klobuchar - it’s really going to have to be Biden that stumbles in order to give them an opening. Booker, for instance, has done quite well in the debates, but what good has it done him? He’s still polling ~3%. So it’s hard to see anyone on your list, save Harris perhaps, breaking through without Biden dropping in the polls enough to make voters look around for an alternative on the moderate/establishment side of things.
Well, you ARE being picky, it’s just that to a limited extent, your desperation is overcoming your pickiness.
For the most part, you’re being the guy at the table who decides nothing on the menu looks good, and grouses to his companions that they should have gone to an entirely different restaurant, one that had stuff like Booker and Klobuchar and Bullock and Beto on the menu.
Anyway, two new polls have been released this morning - Quinnipiac and Suffolk universities, both of which have Biden at 32%. The Biden cratering is apparently taking its sweet time. 
The good news out of Quinnipiac, which is one of the debate-qualifying polls (Suffolk isn’t), is that Steyer got 0% in their poll, so he’s still one poll shy of qualifying for the next debate.
Today’s the last day for a qualifying poll to be released, so I wouldn’t be surprised to see a bunch of polls being released today, including a few by qualifying organizations*. If Steyer gets one more qualifying poll that has him at or above 2%, there will be 11 qualifiers, and they’ll break the September debate into two debates as the first two were. If not, there will be just one debate with 10 participants, and about damn time.
(The only others who are close to qualifying are Gabbard and Williamson, but they each need two more qualifying polls, which seems unlikely with less than a day to go.)
*The poll sponsor has to be one of:
TV/wires: AP, ABC, NBC, CBS, CNN, Fox News, NPR
Print: NYT, WaPo, WSJ, USA Today, Des Moines Register
Universities: Monmouth, Quinnipiac, UNH, Winthrop
I hope Steyer doesn’t get that 2%. What a pointless candidacy (and not just him!).
At this point we’re pretty far in the weeds, and not disagreeing much, but I think you’ll find my list in Post 2458 to be more accurate. Stuff like how much money the pollster spends on following up in order to avoid non-response bias really does seem to affect accuracy. Stuff like how and which demographic variables the pollster weighs really does seem to matter. This is why aggregation is so critical.
Another way of getting at this would be to look at historical accuracy of different polling houses. Some are more accurate than others, consistently over multiple elections, and by more than a few percentage points. This suggests that polling methodology and other considerations really do matter, at least as much as MOE.
Yes, of course. It is wrong to say that a single poll is meaningless. It is even more wrong to say that a single poll shows a dramatic change in the race. The better analysis is that this should very mildly move your estimation of Biden’s chances downward.
That’s nice to see for those of us who want polls to reflect who might win the nomination.
In fhe weeds we are, but I was responding specifically to Post 2466, about “all the other possible reasons that your sample does not reflect the population.” That’s pretty much limited to coverage error and sampling error.
Now there are many possible sources of nonsampling error that come into play after your sample is selected. Several of which you list in post #2458 and here:
These affect how your estimates reflect the population, or fail to.
My apologies if I’m being overmuch of a stat nerd. But sampling is what I do, so the distinction between how things can go wrong with the sample you send out into the field, and what can go wrong after you send it out, is a bright line in my day-to-day work. The latter is someone else’s fault. 
We are in full agreement. drops Mission accomplished banner
As for the state of the race, it seems to me pretty interesting that the betting markets have Warren as the favorite when that is not reflected in the polling. I wonder what people make of that? Is it nothing more than the fact that Biden looks Jeb Bush-y?
Betting markets are a poor second place for judging the state of the race. Warren has been moving her numbers up well with surprisingly strong campaigning. That makes her a great underdog bet even if she is still less likely to win than Biden.
I’m assuming that it is based in part on the thought that once one of either Warren or Sanders drops out, that their supporters will in the main stampede to the other. And that Sanders is more likely to drop out at this point than Warren.
But personally I wouldn’t place a money bet based on that hypothesis ;).
As I mentioned earlier (in this thread, I think??) the Morning Consult polling doesn’t show Warren as a Sanders supporter’s main second choice, nor the reverse.
Huh? If you think Biden is more likely to win that Warren, don’t you buy his shares now which are valued significantly less than Warren’s?