On Nate Silver and 538’s 12.2% Clinton House Effect

cross-posted at CounterPunch

Nate Silver has taught us all a lot of things about how to accurately assess good and bad polling and how to forecast difficult to forecast political races in the United States accordingly. Unfortunately, he and his web juggernaut FiveThirtyEight dot com have forgotten several of their own lessons during this 2016 primary cycle.

And it shows.

It isn’t just that Silver and his main sidekick Harry Enten badly missed that Donald Trump’s early poll numbers were real (Enten famously said Trump had a better chance of playing basketball in the NBA than he did of securing the GOP nomination). But, outside the South, 538 has a 12.2% Clinton House effect on the Democratic side. In eighteen of the twenty-one races, 538’s projection or modelling has shown a Clinton bias.

It gets even worse if we ask about the three of the twenty-one. One of the three was Vermont. Not exactly something to write home about, having a relatively slight pro-Bernie bias in Vermont. If we suggest that calling a race within three points means you actually got it right, with an inevitable slight bias one way or another, then Ohio, at a 2.8% Sanders bias says more, perhaps, about an under performance there. Finally, there is voter suppressed Arizona that is starting to look like Florida 2000, and where it isn’t particularly easy to peg what FiveThirtyEight thought about it ahead of time anyway. Their main stance toward the Democratic race, most charitably interpreted, has been one of paternalistic boredom. It showed in Arizona. More on that in a bit. Since the only argument left on the Democratic side is Math! (think Jeb!): To The Numbers!

538 Bias Chart Final

You’ll notice methodologically that I had to account for places where there aren’t Poll Plus FiveThirtyEight projections. This was quite simple. In a piece entitled, Bernie Sanders Doesn’t Need Momentum – He Needs to Win These States, Nate Silver projected what each remaining state would look like if the race were tied nationally (If, if, IF “but it’s not” has been the constant refrain). So, if the race isn’t tied nationally, there is no way Sanders should be outperforming his projected numbers for those contests, right?

Now, of course, Math! isn’t the only argument right now on the Democratic side. There’s also the lingering White State! argument, but since Hawai’i and Alaska and Seattle and #BernieMadeMeWhite shut that down kind of definitively, there’s the new version of it: Caucus! So yes, redoing the numbers without the Western Caucuses this past week (in which case I’m also taking back Arizona), drops the 538 Clinton House Effect down to a more reasonable, but still measurable and statistically significant, 5.25% with thirteen out of fifteen races showing a pro-Clinton bias.

What if we toss out the caucus states altogether? Again, subtracting Arizona, 538 has a 5.7% Clinton House Effect in primary states outside the South. By now, loyalist to FiveThirtyEight and Clinton will begin shouting WHY OUTSIDE THE SOUTH. (I know this from pushing these numbers on Twitter a bit on Wednesday.) There are no races left in the South, unless you count Washington, D.C. and Maryland, which, for prognostication purposes, it is wise to do if you might, as I do, have the potential for a bit of a Sanders bias.

(Full disclosure, since I started projecting numbers race by race, I too have a pro-Clinton bias outside the South to the tune of 7.0%. I’ve had a Clinton bias in twelve of eighteen states including the Western swing last week, 2.2% in 7 of 12 states without it. In other words, I’ve tried to account for the possibility I’m doing math that makes me feel better as a Bernie supporter and have, generally speaking, undersold Sanders a bit. That said, in primaries alone outside the South, I have an 8 point Bernie bias, reducing only to 6.5% if I eliminate Arizona. That’s statistically significant too, and I’ll talk more about it next week when I project forward to the end of the race after Wisconsin. Basically, I did worst on March 15 when Sanders faced five tough races in large states in one day. The greatest danger zone left for him on the Primary Calendar is April 26. As for Sanders’ chances generally, if he survives April 26th, he could well end up losing by less than 50 delegates all told. In that case, his campaign team would likely excoriate themselves for letting Texas, Virginia, and Tennessee go to seed.)

All of this brings us to Nate Silver’s latest Sanders sneer piece. Wednesday he wrote a thing up entitled It’s Really Hard to Get Bernie Sanders 988 More Delegates. In it, he constantly reminded people that these weren’t his “straight faced” projections. They’re the “rose-colored-glass scenario.” It’s “a path-of-least implausibility.”  Indeed. Indeed. More mocking. But should people with raging and measurable pro-Clinton biases really take such a Tone! if they want to maintain credibility?

Let’s look at the remaining races if we take 538’s rose-colored scenario then compare it to Silver’s previous projection of what each state would look like if the race were tied nationally while adding 538’s bias to date per each region. Yes, regionalism matters everywhere in this contest, not just in states touching Vermont and the South. (Midwest 538 Clinton House Effect is 6.5% pro-Clinton, Northeast 3.9, West 23.3.) Now, these aren’t my actual projections at all, though I may consult the numbers in making mine as the race continues. But it’s a handy way to show that, no, it isn’t so hard to get Sanders to 988 delegates.
538 Sanders to 988

Again, this is just a handy tool for showing up the 538 farce in this election cycle. It isn’t my actual prediction. I plugged these numbers in going backward to Super Tuesday. Sometimes they’re deadly accurate (Illinois, off by 0.3, Maine by 0.9 and Oklahoma just right). Other times they’re laughable in a way that would have over projected Sanders significantly in delegate rich states like Ohio (Sanders by 21.3%!) or Arizona (Bernie Sanders +50!). The problem is that any projection outside the South is going to be wildly wrong if it is based in any way on 538’s boneheaded demographic modelling, which shows up in every Polls Plus projection and which they haven’t changed since July’s infamous White and Liberal piece. There, Silver predicted Sanders might win in Iowa and New Hampshire and nowhere else.

We have to deal with actual poll numbers, right? Not just 538’s fevered projections when they aren’t dealing with what used to be their bread and butter: properly critiquing, weighing, and averaging real poll numbers. So, let’s look at the poll numbers Nate Silver used to mock the idea that Sanders has any chance left. He specifically fluffed bad polling for Sanders in delegate rich states Pennsylvania, New York, and California on Wednesday.

What’s more rose-colored? Thinking Sanders might squeak out a four point win or that he is down 48 in New York?  Ah, so you might just roll with Emerson Polling. Emerson has been wrong by double digits in 5 of the 6 races it has polled on the Dem side this year. The one race it got kinda right, I guess, was Iowa. It missed there by 8 points.

How about Pennsylvania? Down 25, right? Mmm.

That Franklin and Marshall poll is one of those senior citizens only specials, like the horrible math that miscalled Michigan so badly. 2% 18-24-year-olds, 18% total 18-45-year-olds. 41% 65 and older, 68% total 55 and older.This is exactly like, say, including Alan Grayson’s internet poll which he used to decided how he’d vote as a superdelegate while insisting that it is actually predictive of what the results should have looked like in Florida.

It continues: 22.5% 18-44-year-olds in the poll, as unweighted, in California putting Sanders down 7. The average per exit polling for 18-44-year-olds is closer to 45% in most states; it was 35% even in Florida. However they weighted things, it didn’t move Clinton down much: 54% unweighted cross tab to 47% topline. You certainly couldn’t measure the Latino vote accurately with those unweighted numbers. 18-44-year-olds make up a whopping 63.1% of eligible California Latina and Latino voters, with 18-29-year-olds alone at 35.6%, according to Pew’s in-depth research on Latino voting for 2016. Yup. You read those numbers right 35.6%, 18-29 (cough, cough, Nevada entrance poll anyone?).

The otherwise well-done, and well-weighted USC/LA Times poll has Sanders down 11, but also likely misses out on this key demographic in California. While it worked hard to talk to LOTS of Latinx respondents in Spanish and English (501), then actually weighted their numbers downward, it simply couldn’t get an accurate read on Latino 18-29-year-olds because there are only 234 18-29-year-old respondents across all races for the entire poll. 178, or nearly all of them, would have to be Latino to get an accurate read for the ethnicity that is now the plurality in California.

Perhaps most laughable of all, and the state we will know about most quickly, is Wisconsin. There we have a triple special: (1) Emerson Polling said Sanders would lose by 6 on the basis of (2) an 8% millennial sample size and (3) before new numbers could come out, a FiveThirtyEight contributor waxed eloquent to Newsweek about Clinton’s 84% chance of winning in Wisconsin. Sanders needs to win by 16% to meet the rose-colored glass scenario. What if he wins by double that number?

I could go on, but maybe you get the picture. Inaccurate sampling of age demographics or of racialized people, as well as bad urban-suburban-rural splits, has been the achilles heel of polling in Michigan, as well as in a few other places it’s been badly off in the Dem race so far, including in South Carolina. It’s more starry-eyed to think you can poll half the percentage of youngsters that came out in Florida in Pennsylvania and get an accurate reading than it is to think Bernie might do kinda alright in California the way he did in Hawaii and Washington (ALERT: PEOPLE WILL BE YELLING “CAUCUS!” AT THIS LINE. Psst, I’m not saying he’ll win California by 45). Now, let’s look at something a bit more astounding from the USC Dornsife/LA Times poll since, after all, we are still two months and a bit out from actual voting in California, if we get that far. Buried in the weeds is this gem of a chart: Bernie actually has a higher favorability rating than Clinton amongst all, yes all means all, racialized groups.

The race hasn’t anywhere near begun in earnest in California, but Sanders seems to be doing okay with his supposed greatest weakness. Maybe, just maybe, winning by 15 in a West Coast state is a bit of a modest target for Sanders. You never know. It is why we let voters, rather than insult spewing number crunchers and their hordes of Clintonite followers yelling #ITSMATH, actually determine who governs.

Or that’s the idea outside of Arizona anyway.

Update: just as I was first publishing this at goodgawdanotherblog on Thursday, Quinnipiac published a poll with Sanders down just twelve in New York. Without delving into their details, this makes sense to me. Three weeks out, and campaigning there like it’s a gubenatorial race, a twelve point loss is probably Bernie’s floor there. A four point win, as in 538’s “rose-colored” assessment is likely near his realistic ceiling. Also, two polls of Wisconsin are now out with Sanders +4 and +6 in Wisconsin. In the PPP poll, Sanders leads among African Americans 51-40. 538 has moved Sanders to a 57% chance of winning.


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s