Clear Language, Clear Mind

May 2, 2019

FAQ for “Cognitive ability and political preferences in Denmark” Kirkegaard, Bjerrekær, Carl (2017)

Filed under: intelligence / IQ / cognitive ability,Political science — Emil O. W. Kirkegaard @ 04:38

Various critics of Noah Carl are being asked to produce something that shows they have read and understand his research, and they appear to be struggling. In an attempt to rectify this, we have Danish researcher Stine Møllegaard, a professor at Copenhagen University, making some criticisms of our study on Twitter. (Note that she subsequently deleted all of these comments.) Unfortunately, she does not seem that familiar with research in this area, since her points do not make a lot of sense. Stine is a moderately qualified critic by virtue of her own field, sociology, but does not appear to have published anything on psyhometrics and political attitudes (the topic of our paper)

Data

I had work to do. I am also very sceptical about the “danish data” used in multiple papers in OpenPsych, some of which Noah also co-authored. It requires quite a lot of digging to find the actual description of the data – and I find it curious. Have you read any of the papers?

A strange claim considering that all the data is public and so are the surveys given to people in the case of our study. The link to the data is given both on the OpenPsych website and in the paper PDF:

https://osf.io/xdpcq/files/

Unknown pollster

“1) It’s collected via a service I’ve never heard of (and I got quite some experience with quantitative data research in Denmark). This is not critical in itself, but strange..”

Not particularly strange. Has Stine heard of every internet pollster? In fact, when we looked for data collection, we reached out to multiple Danish pollsters. There was a huge price difference between the options because some suggested doing phone or face interviews (expensive! one estimated 100k DKK). We settled on the relatively unknown Survee because it relied on online collecting, which is similar to other highly utilized English language pollsters like MTurk or Prolific.

Representativeness

2) This service pays participants to participate – probably motivating some types of participants rather than others. How are basic demographic measures such as occupation, income, education, marital status related to the likelihood of being on such a site?

Saying that it is representative is a bit of a stretch, imo. This is further confirmed by the rather large number of participants the co-authors themselves note “… did not comply with instructions and filled out the questions seemingly at random.”

It is closely nationally representative on several metrics as we reported by comparing with data from Danish stats agency:

Because our sample was essentially a self-selected sub-set of another sample, it might be biased. As a check on representativeness, we calculated mean values of relevant variables for responders and non-responders from the original sample. As Table 1 indicates, responders were slightly younger, had slightly lower cognitive ability and were slightly less educated than non-responders. There was, therefore, some selection bias in responding to our survey. However, in all cases the differences were quite small (e.g. d = .23 for cognitive ability) and the subset was thus still fairly representative of the general population.

She continues:

3) The authors “openly stated the purpose of the study in the introduction of the survey” which might further have contributed to some selection into who would want to participate in such a survey and their answers.

Would she rather we don’t state the purpose of the study? It’s a rather odd criticism since most studies give a general introduction in the beginning of the survey. She presents no evidence that this would cause important selection bias.

Did responders understand the task?

4) The authors openly admit that their followup survey showed that a group of participants (size unknown) had misunderstood some of the questions, “however they may have been lying.”
🤔

There is ample research (see this random example) that representative surveys often have issues with responders not understanding what they are supposed to do. We decided to investigate this, but the standard method is to ignore it in studies. We went to extra lengths to ensure that participants understood the tasks by excluding data from those who didn’t.

The cognitive test

5) The authors use a measure of cognitive ability they themselves developed and validated – but on a completely different group: namely primary school students. I would be very surprised if primary school students are representative of 30-39 year old Danes in general.

The authors admit that “did not have other criteria variables than age and grade level to validate the test against”. A bit of a stretch to use it as a general measure of cognitive abilities.

Particularly a person (Noah) who has done research on intelligence should be more critical about how to measure cognitive ability.

Stine is not familiar with the test. This was in fact a Danish translation of the ICAR test, which has been well-validated on tens of thousands of people and used by numerous other researchers. We have used it not once, but multiple times in Danish samples (with middle schoolers, high school students, and adults), and there is no evidence it does not work as intended generally speaking. Obviously, such a short test (9 items) will have fairly low reliability compared to multiple hour long tests given in person, but that’s the reality of survey data: we can’t feasibly give everybody a full Wechsler assessment. For comparison, there are hundreds of studies using the 10-item vocabulary test found in the ANES and GSS survey datasets. Thus, her criticism is not on target and applies equally well to 100s of other studies by other researchers.

Funding

6) For a group of researchers publishing in their own journal in the spirit of open science, it is really curious that they cannot disclose who is funding their research; “This research was supported by two anonymous research contributions.”

Stine appears oblivious to the political bias of the field and the media. Suppose I got private funding from some person who was interested in immigration but not heavily involved in politics. Now, I could report this, and the social justice activists would then seek out that person and try to ruin their reputation by virtue of their association with me or the study. There is no legal mandate to report private funding, and no ethical mandate to do so. The only place really where funding sources should be reported are for research with commercial implications, which ours obviously does not have.

Controversial

I’m just saying… This is really not what I associate with well-conducted research. I would expect more of my students. Maybe I’m harsh? But when studying controversial questions, this is even more important, imo.

And here we get the truth: Stine has higher requirements for research she or other left-wingers don’t like: that’s what “controversial” means in this context. Selectively applied higher standards is the norm of political bias in science. There are various lines of evidence that show this, one can e.g. read the recently edited book by Crawford and Jussim (2017), or the long target article from 2015 (Duarte et al 2015).

Real Peer Review

OK. Well then he shouldn’t have a problem publishing these papers in actual peer-reviewed journals? As in _not_ reviewed by his friends, but researchers in the field?

OpenPsych is actually peer reviewed — no need for the motte and bailey. The reviewers on this particular paper are listed on the website: Robert L. Williams, Peter Frost, L. J. Zigerell. Zigerell is a professor of political science, Peter Frost is an anthropologist with numerous publications and a long running interest in IQ research, and Williams is retired but is well read on IQ research as evidenced by his publication of a paper in Intelligence previously, which is the highest impact journal of this field and run by Elsevier.

August 27, 2018

Political quiz results

Filed under: Political science — Tags: — Emil O. W. Kirkegaard @ 06:56

Since people keep asking about my results on such tests, and because my position has been misrepresented (e.g. by Les Temps).

There’s a lot of political tests online one can take, but some are more popular than others, and sensible. A popular choice is the libertarian-made, The World’s Smallest Political Quiz, a 10 item quiz based on the Nolan chart 2 dimensional representation of politics.

The RED DOT on the Chart shows where you fit on the political map.
Your PERSONAL Issues Score is 60%
Your ECONOMICS Issues Score is 30%
According to your answers, the political group that agrees with you most is…
Centrist
Centrist prefer a “middle ground” regarding government control of the economy and personal behavior. Depending on the issue, they sometimes favor government intervention and sometimes support individual freedom of choice. Centrists pride themselves on keeping an open mind, tend to oppose “political extremes,” and emphasize what they describe as “practical” solutions to problems.

My score on this have changed over the years, ironically, towards more leftism away from libertarianism. The changes concern my views on government interventions in markets, where I’ve become more open to the idea, mostly as a result of increasing disbelief in mainstream economics (e.g. due to reading Debunking Economics).

A similar test is that by IDR labs which has 36 questions instead of 10:

2.8% Right, 2.8% Liberal

So dead center.

Pew Research has a UC-centric test, Political Typology Quiz, 17 items, 1 dimensional output:

This overwhelmingly Republican group holds conservative attitudes across a wide range of issues, especially in their support for smaller government. Core Conservatives are deeply skeptical of the social safety net and favor lower tax rates on corporations and high-income individuals. While they are divided on whether immigrants do more to strengthen or burden the country, Core Conservatives offer far more positive views of immigrants than do Country First Conservatives. Core Conservatives are relatively upbeat about national conditions and a majority says that the United States “stands above” all other nations in the world.

This test had a lot of questions related to group differences (race, sex), which is probably why the right-leaning result happened.

PolitiScales is an 8-scale test based on 117 questions. It’s kinda too long.

Like the one above, I affirmed all the HBD questions hence the high score on ‘essentialism’. Justice questions concern basically regards for prisoners vs. regards for victims, in particular, relative focus on rehabilitation vs. punishment. Their framing were poor, so I mostly clicked agnostic option. Their questions about nationalism mostly concern stuff about immigrants, not so much the war-like national-expansionism type that many abhor. The questions on environment were somewhat unclear, so I usually clicked agnostic.

Another odd test is the 4 dimensional by 8values.

A neoliberal! I guess maybe if we change the definition to that given by Sam Bowman.

Then, there’s another 2 dimensional test called the Political Compass, which has 42 items, mostly US-centric:

Luckily, it produced quite similar results to the other 2, so at least there’s some inter-test reliability.

Bonus: views of my bubble

We have a private group for discussion of HBD matters and general freethought. People range from professors to random normal wage slaves who developed an interest. It’s pretty international, having a variety of Scandinavians (basically my Danish HBD cluster), US citizens (every race or so, think we’re still missing an Indian), various other Euros, Chinese Chinese and what various in betweens. I posted the 10 item test above, and this is the distribution of the group:

The summary stats are: n = 14, economic mean/sd 56/33, personal mean/sd 54/30. So it’s pretty dead center with a wide range of basically everything except it seems a hardcore anarcho-communist (bottom right). So I’m pretty happy with my bubble!


Update 23rd November 2018: Guardian Nolan-like test

The Guardian – How populist are you?

Well, nothing new there.

August 25, 2018

Data from NYT’s “An Extremely Detailed Map of the 2016 Election”

Filed under: Political science — Tags: , , — Emil O. W. Kirkegaard @ 23:38

NYT has a cool new dataviz called An Extremely Detailed Map of the 2016 Election.

Really, it’s just a more detailed version of their initial county election page which I previously scraped the data from. So I had a quick look to see if it was easy to get the data behind it. Seems the answer is ‘yes’. Here’s the files:

June 28, 2018

Country level indicators of altruism

Filed under: Immigration,Political science,Psychology — Tags: , — Emil O. W. Kirkegaard @ 07:47

A traditional approach to studying altruism has been to do some kind of economic game. E.g. seeing how nice people are in ultimatum games. One such study:

They tend to produce suspicious rankings:

Perhaps owning to terrible sampling, in terms of size and representativeness.

Objective, large data measures?

Surely, we can do better. How about proportion of GDP/GNI donated to foreign aid?

One probably wants to regress out country GDP from this, but this overcorrects to some degree because part of the reason why the altruistic ones donate so much is that they are also nicer to each other and this promotes country performance.

Friendliness to other-race immigrants (via Twitter):

I also suggested that one could measure pro-social behavior on online helping sites like StackExchange. Thousands of people take their time to answer others’ questions (contributors/altruists), but most people only read answers already given (leeches/lurkers). So one could take users per capita and regression out traffic per capita. One could also analyze the top users which more directly measures the , á la the method in my top mental sports players study.

(True to my people, my altruism score is high according to this metric.)

I predict such derived metrics would correlate highly with the above two altruism metrics, at least > 0.50.

July 10, 2017

Free trade: yay or not? (Review of Free Trade Doesn’t Work: What Should Replace It and Why)

Filed under: Book review,Government form,Political science — Tags: , — Emil O. W. Kirkegaard @ 00:44

Recently, I decided it was time for catching up on my to-read list. I try to read >=30 books a year, and I was behind, owing to spending a lot of time on company work. I also wanted to avoid reading too much of the same stuff. Two reasons. First, I want to avoid getting too much confirmation bias that inevitably happens from reading a lot of stuff that’s in high agreement with each other. Second, knowledge in general has strong diminishing returns. Knowing, say, 50% about physics is almost as practically useful as knowing 90%, But knowing 50% is a lot more useful than knowing 0%. Furthermore, there are diminishing returns to knowledge accumulation too because the material will inevitably cover some of the same stuff, meaning that you aren’t learning something new.

Taken together, I wanted to try reading something new to me. I decided on Big Politics, a topic I normally avoid because it’s full of feelings and the relevant data to decide the issues does not generally exist and in many cases could not even be realistically gathered if we were determined to do so.

I generally lean towards freedom on questions of policy, but I’m not a principled libertarian. What I have is a kind of libertarian default policy, which can be undermined by reasonable evidence that regulation/less freedom works better to further our collective goals. I’ve never really considered free trade, tariffs etc. (i.e. between-country trading) in detail so generally leaned towards free trade being good. On the other hand, macro economists — whose opinion people copy — tend to be not my cup of tea. Essentially basing their ideas on various mathematical models with totally unrealistic assumptions: everybody has only self-interest, consistent goals, perfect rationality, consistent time-preference, humans being substituteable (no individual differences), unrealistic beliefs in the causality of education in itself and so on.

So I looked around for a anti free trade book. There were several to pick from. I incidentally stumbled upon this article by the author of one such book: Free Trade Doesn’t Work by Ian Fletcher. It has reasonable reviews: 4.4/5 (n=69) on Amazon, 4.2/5 (n=59) on Goodreads. Good enough for me.

The book isn’t technical, but it gets the job done reasonably well. Since comparative advantage á la David Ricardo is the basic foundation for most claims about the benefits of free trade, naturally the author spends his time arguing against this argument. Because this argument is based on a lot of assumptions and some mathematical modeling, all one has to do is attack the assumptions. If they are shown to be very wrong, then the conclusion about free trade’s benefits won’t follow. This doesn’t establish that free trade is bad/protectionism is good, but it’s a start.

The criticism of free trade

Fletcher lists 7 dubious assumptions of free trade.

  1. Free trade is sustainable. One cannot keep a negative or positive trade balance forever. Keeping a negative one means importing more than one produces, which creates debt to foreigners. Not even keeping a positive one is always a good idea. If it’s based on selling off non-renewable property to foreigners such as natural resources. When they run out, there is nothing left for one’s ancestors.
  2. There are no externalities. These come in two kinds: negative and positive. The textbook negative externality are environmental pollution. If a country has lax environmental standards, one can import goods from it cheaply. However, this causes accumulated pollution in the production country. In the unlucky scenario where pollution is global in scope, it means that the world is essentially relying on the weakest environmental protection of any country. Because there are 200 countries or so in the world, probably some country or another will have misinformed, low protections. Free trade using this will cause global destruction of the environment.Positive externalities are the opposite: where some sectors of the economy produce spillover effects, like making it easy to break into another sector. For instance, production of materials used to manufacture computers makes it easier to break into the computer production sector because one can source locally. These factors are ignored in free trade. In the worst case, countries can get stuck in an agriculture sector. Agriculture doesn’t really lead easily to other sectors, but free trade will push a country towards it if wages are low and the climate warm. Ring any bells?
  3. Factors of production move easily between industries. The simplest example here are workers’ skills. While some workers are able to retrain, most are not or it is very costly. When the trade opportunities change, free trade forces a country to move its production into a new sector. However, since workers cannot adapt as fast as circumstances change, they will instead move into unemployment or underemployment.
  4. Trade does not raise income inequality. Comparative advantage, when it holds, implies that the economy as a whole will grow, not that it will grow equally. Fletcher gives a hypothetical example of importing clothing and exporting airplanes. Suppose we start with an economy that produces both. But then we find a trade partner with lower wagers that is able to produce clothes but not planes more efficiently. Great, so we start exporting planes and importing clothes. All good so far. But then, what’s the distribution of jobs required for producing planes and clothes? Maybe planes requires 3 high skill workers and 7 low skill, while clothes require 1 high skill and 9 low skill. So, by doing this, we’ve increased the demand for high skill workers and decreased that for low skill workers. Since workers can’t just change places (individual differences in basic traits + acquired skills), then this will cause lower wages or unemployment.
  5. Capital is not internationally mobile. Basically, capitalists won’t necessarily invest in creating jobs in your own country. Rather, if wages and free trade allows it, they will invest in other countries’ infrastructure. Fletches gives the example of British engineers building railroads in other countries instead of Britian. In 1914, 35% of British owned railroads were not in Britain. Furthermore, when capitalists move jobs to other countries, this will lower production costs (per free trade), which is good for consumers. But if they move too many jobs away, then it’s problematic too. Consumers and workers are the same people, and they can’t consume cheaper goods if they have no jobs, or can’t afford them if they have lower paying jobs. There is no theorem that says that these will necessarily balance out in favor of your country.
  6. Short-term efficiency causes long-term growth. Comparative advantage theory is a static model, about what would happen if things balanced out in an instant. It just so happens the world is not like that, things take time. Generally speaking, one wants growth, and change means positive change, be it skills, income or knowledge. Comparative advantage is about being the most efficient at what one currently does. If you work as a secretary, you don’t want to become the most efficient secretary. You want to build skills so you can move into a better job. Burkino Faso doesn’t want to be the most efficient Burkino Faso, it wants to become something like Denmark. Comparative advantage itself does not imply anything about how one would accomplish such goals.

    It goes back to Ricardo’s own example of wine and wool production. Britain produced wool and Portugal produced wine. Then they traded and everybody was happy. However, wine production did not spawn another useful sectors. Wine production has been essentially the same for hundreds or thousands of years. It has no node above it in the tech tree. Wool production, however, lead to mechanical treatment of wool, and then other mechanical parts and eventually a whole mechanical industry that lead to steam ships and trains. Lots of nodes in the tech tree above. If you’re a country, you want to move your production towards sectors with nodes in the tech tree above them.

    A personal example of this as applied to science. I have many co-authors. They almost inevitably want me to write the analysis because I’m so much better at it and much faster. This is the more effectively solution given the present distribution of abilities and results in us getting done with the study faster. Comparative advantage they say. However, as I keep telling them, if we keep splitting the work up like this, they will not learn to program, and this will have long-term consequences for their output. Programming ability is a force multiplier, allowing one person to do quickly what previously one person took a long time to do, many persons a long time, or was impossible before.

  7. Trade does not induce adverse productivity growth abroad. Trading with others might cause them to attain high growth rates, which changes their opportunity costs. These can become so high that they stop producing cheap products which you were previously importing. But now, you lost your own production in this sector, and it’s hard to start one up again. So you’ll have to keep importing more expensive products from the other country if they are necessary for you. If you had kept your own production on-going, then you would not have this problem because you would have kept the know-how in the country all along.

The prior

Seems pretty convincing to me, but we should be skeptical. The book itself has tons of footnotes: about 700 for the book or .about 2 per page. I did not generally check up on them, so maybe the references are not very convincing. Maybe they are like when Nisbett cites stuff.

What about the author? Is he a well known economist, so that we can be reasonably sure he knows his stuff? Seems not. His page lists no academic publications. He says he was educated at Columbia and the University of Chicago, but not in what. Might be a ‘Doctor’ Laura case. She speaks as if she’s a doctor of psychology or psychotherapy (a really low bar to pass!), but actually she has a ph.d. in physiology. On diabetes. In rats. I am of course not very hostile to outsiders (my degree is in linguistics, bachelor only!), but not taking the prior into account would be foolish.

Searching for publications of his does reveal some. Mostly in obscure outlets (associated with the post-autistic movement). (Sounds familiar?). On the other hand, regular economists do cite his book not just for criticism.

So, I’m not wholly convinced yet, but will read more. Naturally, next up I decided to read something in the totally other direction: Democracy, the God that failed. An Austrian economics book.

February 2, 2017

Some simple models of US county voting outcomes

Filed under: Differential psychology/psychometrics,Political science,Sociology — Tags: , — Emil O. W. Kirkegaard @ 05:45

Woodley convinced me that these are of actual interest. As some of you may recall, I compiled a large county level (n≈3000) dataset some time ago, but didn’t use it for anything. I just thought it would be a cool dataset, but that results were not too interesting. Well, since someone did think these were important analyses enough to do a study using state level data on, I took a look at the superior dataset. The outcome variables are the fractions of votes for Democrats, Republicans as well as Libertarians and Greens for the 2016 election. Results for Dems and Reps are also available for 2012 and 2008. The 2016 data also has the various smaller candidates (e.g. this guy) but these were of little interest so I did not examine them.

Data sources

The cognitive ability (CA) score is from what used to be called the Global Report Card but which changed name to (mumble mumble), but one can find them here. I think these are actually from NAEP testing, but I’m not quite sure. They are some kind of scholastic testing, so not exactly standard IQ data, but good enough. The S — general socioeconomic factor (a fancy general social status metric) — is extracted from some varied 28 indicators, as detailed in this study. The voter outcome data comes from NYT’s map here. It took a bit of a scrape job to get them out, but I managed. The data are actually not the final counts, as I thought they were when I downloaded them, but they are very close to the finals, and so I didn’t bother updating them. I guess I should now that someone wants to publish this in some journal with my name on it! Demographic data was from the ACS.

Regressions

While one should use path models, I know that some will want to see the straight regressions. Regressions are basically path models where all the predictors are modeled as being causally independent and which cause the dependent outcome. Thus, one assumes that S is not caused by CA or by demographics. What this basically does is underestimate the variables which mainly work thru other variables (indirect effects).

What am I reporting? I report the standardized betas for the predictors, some model meta-data including cross-validated R2 (10 fold), and the etas. What are etas? These are the square root of the more common eta2., It’s an R2-type measure, i.e. about variance, so it is non-linear and not so easy to interpret correctly. Taking the square root puts it on the same scale as the correlation. The etas here are derived from the analysis of variance fit by stats::aov, which is passed to lsr::etaSquared function. This uses type 2 errors by default and I too like to live dangerously so I didn’t check method variance by trying the other methods. If you wonder what these are, you can find them explained here, here and here.

live dangerously

Etas have an advantage in comparison to the standardized betas which is that they make it possible to compare the importance of variables for the model’s overall explanatory power. Standardized betas do not allow for this because while they are standardized, a variable may be highly correlated with other variables such that it is redundant. Categorical variables may not have much variation. Being a type B may be associated with a large negative effect for some outcome, but if the dataset consists of 99% type A’s and 1% type B’s, variation in type will not explain much variation in the outcome. Etas take this into account.

Furthermore, categorical variables, such as state, are given a beta for n-1 of their levels (the last level is the reference level and thus has beta=0 using standard contrast coding). So if we have two categorical predictors, one with 5 levels and one with 10, we get a set of 4 betas vs. a set of 9 betas. This makes it hard to assess the relative importance of a categorical predictor compared to … any other predictor. Etas deal away with this problem because each categorical predictor is only assigned 1 eta value, just as every other variable is.

A problem with etas is that they are directionless (because based on eta2). However, we can look up the direction for the non-categorical variables using the betas. The categorical predictors of course do not have any consistent directions.

Choosing metrics for relative comparison of predictors is actually very difficult and I only used a simple method here because this is the only method I have implemented in my model summary function so far. I should implement the functions from the relaimpo package, but alas, I don’t have infinite time. So for now we will pretend that etas are totally fine for this purpose.

Here’s the results (6 long tables of numbers):

Democrats, 2016
Model coefficients
Estimate Std. Error CI.lower CI.upper
CA                     -0.0934      0.017   -0.127   -0.060
S                       0.1281      0.019    0.092    0.164
Black                   0.7662      0.017    0.733    0.800
Asian                   0.2719      0.012    0.248    0.296
Hispanic                0.3831      0.015    0.354    0.412
State: Alabama          0.0000         NA       NA       NA
State: Arizona          1.1032      0.165    0.779    1.427
State: Arkansas         0.4404      0.091    0.263    0.618
State: California       0.7013      0.107    0.491    0.912
State: Colorado         1.0043      0.101    0.807    1.202
State: Connecticut      1.6893      0.202    1.293    2.085
State: Delaware         1.0288      0.314    0.413    1.645
State: Florida          0.5031      0.095    0.318    0.689
State: Georgia         -0.0547      0.078   -0.208    0.099
State: Idaho            0.3316      0.108    0.121    0.543
State: Illinois         0.9745      0.088    0.802    1.147
State: Indiana          0.9326      0.090    0.756    1.109
State: Iowa             1.2884      0.090    1.113    1.464
State: Kansas           0.3538      0.089    0.180    0.528
State: Kentucky         0.7534      0.086    0.584    0.923
State: Louisiana       -0.1778      0.093   -0.361    0.005
State: Maine            2.1385      0.150    1.843    2.434
State: Maryland         0.8010      0.130    0.547    1.055
State: Massachusetts    2.5254      0.161    2.211    2.840
State: Michigan         1.3539      0.091    1.175    1.532
State: Minnesota        1.3040      0.092    1.123    1.485
State: Mississippi     -0.0042      0.089   -0.178    0.170
State: Missouri         0.6297      0.086    0.462    0.798
State: Montana          1.0762      0.102    0.876    1.277
State: Nebraska         0.3433      0.092    0.162    0.524
State: Nevada           0.2345      0.147   -0.054    0.523
State: New Hampshire    2.1916      0.183    1.833    2.550
State: New Jersey       0.8934      0.145    0.609    1.178
State: New Mexico       0.5739      0.125    0.328    0.820
State: New York         1.3975      0.098    1.205    1.590
State: North Carolina   0.7083      0.086    0.540    0.877
State: North Dakota     0.8483      0.104    0.644    1.053
State: Ohio             1.1032      0.092    0.924    1.283
State: Oklahoma         0.2687      0.092    0.088    0.449
State: Oregon           1.2490      0.115    1.024    1.474
State: Pennsylvania     1.1244      0.096    0.936    1.313
State: Rhode Island     2.2675      0.248    1.780    2.755
State: South Carolina   0.2430      0.102    0.043    0.443
State: South Dakota     1.0742      0.098    0.881    1.267
State: Tennessee        0.4817      0.087    0.310    0.653
State: Texas           -0.2096      0.085   -0.376   -0.044
State: Utah             0.2185      0.123   -0.022    0.459
State: Vermont          2.9258      0.159    2.613    3.238
State: Virginia         0.7378      0.083    0.576    0.900
State: Washington       1.3852      0.111    1.167    1.604
State: West Virginia    0.6874      0.100    0.491    0.884
State: Wisconsin        1.7170      0.095    1.530    1.904
State: Wyoming          0.2856      0.132    0.026    0.545

Model meta-data
 outcome    N   R2 R2-adj. R2-cv
 1 dem16_frac 3062 0.72    0.72  0.71

Etas from analysis of variance
 eta eta.part
 CA       0.052    0.098
 S        0.066    0.125
 Black    0.429    0.632
 Asian    0.216    0.380
 Hispanic 0.247    0.424
 State    0.474    0.670

 

Republicans, 2016
Model coefficients
Estimate Std. Error CI.lower CI.upper
CA                       0.101      0.018    0.065     0.14
S                       -0.172      0.019   -0.210    -0.13
Black                   -0.746      0.018   -0.781    -0.71
Asian                   -0.272      0.013   -0.297    -0.25
Hispanic                -0.386      0.016   -0.416    -0.36
State: Alabama           0.000         NA       NA       NA
State: Arizona          -1.253      0.173   -1.592    -0.91
State: Arkansas         -0.621      0.095   -0.806    -0.44
State: California       -0.845      0.112   -1.065    -0.62
State: Colorado         -1.264      0.105   -1.471    -1.06
State: Connecticut      -1.754      0.211   -2.167    -1.34
State: Delaware         -1.136      0.328   -1.779    -0.49
State: Florida          -0.531      0.099   -0.725    -0.34
State: Georgia           0.039      0.082   -0.122     0.20
State: Idaho            -0.940      0.112   -1.160    -0.72
State: Illinois         -1.113      0.092   -1.293    -0.93
State: Indiana          -1.052      0.094   -1.236    -0.87
State: Iowa             -1.409      0.094   -1.592    -1.23
State: Kansas           -0.523      0.093   -0.705    -0.34
State: Kentucky         -0.859      0.090   -1.036    -0.68
State: Louisiana         0.118      0.097   -0.073     0.31
State: Maine            -2.384      0.157   -2.692    -2.08
State: Maryland         -0.873      0.135   -1.138    -0.61
State: Massachusetts    -2.649      0.168   -2.978    -2.32
State: Michigan         -1.504      0.095   -1.690    -1.32
State: Minnesota        -1.554      0.096   -1.743    -1.37
State: Mississippi      -0.013      0.093   -0.194     0.17
State: Missouri         -0.726      0.090   -0.902    -0.55
State: Montana          -1.299      0.107   -1.508    -1.09
State: Nebraska         -0.445      0.096   -0.634    -0.26
State: Nevada           -0.537      0.154   -0.839    -0.24
State: New Hampshire    -2.275      0.191   -2.650    -1.90
State: New Jersey       -0.878      0.152   -1.175    -0.58
State: New Mexico       -1.087      0.131   -1.343    -0.83
State: New York         -1.516      0.103   -1.717    -1.32
State: North Carolina   -0.712      0.090   -0.888    -0.54
State: North Dakota     -1.111      0.109   -1.325    -0.90
State: Ohio             -1.207      0.096   -1.395    -1.02
State: Oklahoma         -0.406      0.096   -0.594    -0.22
State: Oregon           -1.500      0.120   -1.735    -1.27
State: Pennsylvania     -1.159      0.101   -1.357    -0.96
State: Rhode Island     -2.342      0.259   -2.851    -1.83
State: South Carolina   -0.334      0.107   -0.543    -0.12
State: South Dakota     -1.256      0.103   -1.457    -1.05
State: Tennessee        -0.538      0.091   -0.717    -0.36
State: Texas             0.189      0.088    0.016     0.36
State: Utah             -1.243      0.128   -1.493    -0.99
State: Vermont          -3.081      0.166   -3.407    -2.75
State: Virginia         -0.840      0.086   -1.010    -0.67
State: Washington       -1.572      0.116   -1.800    -1.34
State: West Virginia    -0.822      0.105   -1.028    -0.62
State: Wisconsin        -1.806      0.099   -2.001    -1.61
State: Wyoming          -0.525      0.138   -0.796    -0.25

Model meta-data
 outcome    N  R2 R2-adj. R2-cv
 1 rep16_frac 3062 0.7    0.69  0.68

Etas from analysis of variance
 eta eta.part
 CA       0.056     0.10
 S        0.089     0.16
 Black    0.418     0.61
 Asian    0.216     0.37
 Hispanic 0.248     0.41
 State    0.496     0.67

 

Greens, 2016
Model coefficients
Estimate Std. Error CI.lower CI.upper
CA                      -0.150      0.025   -0.198   -0.101
S                        0.076      0.027    0.023    0.128
Black                   -0.093      0.024   -0.140   -0.045
Asian                    0.097      0.017    0.063    0.131
Hispanic                -0.030      0.021   -0.072    0.012
State: Alabama           0.000         NA       NA       NA
State: Arizona           1.561      0.215    1.139    1.983
State: Arkansas          0.542      0.118    0.310    0.773
State: California        1.970      0.142    1.691    2.248
State: Colorado          1.466      0.133    1.206    1.726
State: Connecticut       1.942      0.263    1.426    2.458
State: Delaware          1.605      0.407    0.806    2.404
State: Florida           0.399      0.124    0.157    0.642
State: Idaho             0.990      0.141    0.712    1.267
State: Illinois          1.163      0.116    0.936    1.390
State: Iowa              0.403      0.118    0.171    0.635
State: Kansas            2.333      0.117    2.103    2.563
State: Kentucky          0.362      0.114    0.138    0.586
State: Louisiana         0.273      0.121    0.036    0.510
State: Maine             2.606      0.196    2.221    2.991
State: Maryland          1.284      0.169    0.952    1.615
State: Massachusetts     2.365      0.210    1.954    2.777
State: Michigan          1.129      0.120    0.895    1.364
State: Minnesota         1.077      0.122    0.837    1.316
State: Mississippi      -0.045      0.115   -0.271    0.181
State: Missouri          0.475      0.113    0.253    0.696
State: Montana           1.407      0.135    1.143    1.671
State: Nebraska          0.603      0.122    0.363    0.842
State: New Hampshire     0.944      0.239    0.477    1.412
State: New Jersey        0.895      0.190    0.522    1.268
State: New Mexico        1.037      0.164    0.715    1.359
State: New York          2.048      0.129    1.795    2.301
State: North Dakota      1.145      0.137    0.876    1.415
State: Ohio              0.700      0.121    0.463    0.937
State: Oregon            2.651      0.150    2.356    2.946
State: Pennsylvania      0.653      0.127    0.405    0.902
State: Rhode Island      1.675      0.323    1.041    2.308
State: South Carolina    0.301      0.133    0.041    0.560
State: Tennessee         0.200      0.115   -0.025    0.424
State: Texas             0.299      0.112    0.079    0.519
State: Utah              0.267      0.161   -0.048    0.582
State: Vermont           2.966      0.208    2.558    3.374
State: Virginia          0.475      0.108    0.264    0.687
State: Washington        1.454      0.146    1.167    1.740
State: West Virginia     1.012      0.132    0.753    1.271
State: Wisconsin         1.021      0.126    0.775    1.267
State: Wyoming           0.868      0.173    0.529    1.208

Model meta-data
 outcome    N   R2 R2-adj. R2-cv
 1 green16_frac 2556 0.54    0.53  0.54

Etas from analysis of variance
 eta eta.part
 CA       0.082    0.119
 S        0.039    0.057
 Black    0.052    0.076
 Asian    0.077    0.112
 Hispanic 0.019    0.028
 State    0.591    0.655

 

Libertarians, 2016
Model coefficients
Estimate Std. Error CI.lower CI.upper
CA                     -0.0038      0.017  -0.0373   0.0296
S                       0.3505      0.018   0.3146   0.3865
Black                  -0.0317      0.017  -0.0648   0.0015
Asian                   0.0143      0.012  -0.0091   0.0377
Hispanic                0.0738      0.015   0.0449   0.1026
State: Alabama          0.0000         NA       NA       NA
State: Arizona          1.2045      0.163   0.8846   1.5244
State: Arkansas         0.2946      0.089   0.1192   0.4700
State: California       0.6799      0.106   0.4719   0.8880
State: Colorado         1.0910      0.099   0.8962   1.2858
State: Connecticut      0.3227      0.199  -0.0684   0.7138
State: Delaware         0.7064      0.310   0.0984   1.3144
State: Florida          0.0761      0.093  -0.1071   0.2593
State: Georgia          0.3307      0.077   0.1789   0.4826
State: Idaho            0.6224      0.106   0.4142   0.8307
State: Illinois         1.2047      0.087   1.0347   1.3748
State: Indiana          1.6818      0.089   1.5075   1.8561
State: Iowa             0.4954      0.088   0.3221   0.6687
State: Kansas           0.9395      0.088   0.7679   1.1111
State: Kentucky         0.4006      0.085   0.2331   0.5680
State: Louisiana       -0.0810      0.092  -0.2614   0.0995
State: Maine            1.9222      0.149   1.6309   2.2136
State: Maryland         0.4547      0.128   0.2039   0.7056
State: Massachusetts    0.8918      0.159   0.5809   1.2028
State: Michigan         1.0787      0.090   0.9024   1.2549
State: Minnesota        0.7014      0.091   0.5227   0.8801
State: Mississippi     -0.1593      0.088  -0.3309   0.0123
State: Missouri         0.6371      0.085   0.4711   0.8032
State: Montana          1.7722      0.101   1.5742   1.9703
State: Nebraska         0.9864      0.091   0.8076   1.1652
State: Nevada           0.8487      0.145   0.5635   1.1339
State: New Hampshire    0.9316      0.181   0.5773   1.2860
State: New Jersey      -0.4833      0.143  -0.7645  -0.2021
State: New Mexico       4.4179      0.124   4.1754   4.6604
State: New York         0.7390      0.097   0.5490   0.9291
State: North Carolina   0.3557      0.085   0.1894   0.5220
State: North Dakota     2.0401      0.103   1.8380   2.2422
State: Ohio             0.6618      0.090   0.4846   0.8391
State: Oklahoma         1.7475      0.091   1.5691   1.9259
State: Oregon           1.8205      0.113   1.5986   2.0424
State: Pennsylvania     0.0888      0.095  -0.0976   0.2753
State: Rhode Island     0.5527      0.245   0.0718   1.0336
State: South Carolina   0.1063      0.101  -0.0913   0.3040
State: South Dakota     1.6102      0.097   1.4197   1.8008
State: Tennessee        0.2270      0.086   0.0577   0.3964
State: Texas            0.1331      0.084  -0.0307   0.2970
State: Utah            -0.0389      0.121  -0.2762   0.1983
State: Vermont          0.6762      0.157   0.3676   0.9847
State: Virginia         0.3101      0.082   0.1500   0.4701
State: Washington       1.1375      0.110   0.9218   1.3532
State: West Virginia    0.7544      0.099   0.5600   0.9488
State: Wisconsin        0.5313      0.094   0.3471   0.7155
State: Wyoming          1.5085      0.131   1.2524   1.7645

Model meta-data
 outcome    N   R2 R2-adj. R2-cv
 1 libert16_frac 3062 0.73    0.73  0.72
Etas from analysis of variance
 eta eta.part
 CA       0.0021   0.0041
 S        0.1810   0.3291
 Black    0.0177   0.0341
 Asian    0.0114   0.0219
 Hispanic 0.0475   0.0910
 State    0.6146   0.7637

 

Democrats, 2012
Model coefficients
Estimate Std. Error CI.lower CI.upper
CA                      -0.136      0.018   -0.172   -0.099
S                       -0.026      0.020   -0.065    0.013
Black                    0.646      0.018    0.611    0.682
Asian                    0.210      0.013    0.185    0.236
Hispanic                 0.319      0.016    0.288    0.350
State: Alabama           0.000         NA       NA       NA
State: Arizona           0.959      0.177    0.612    1.306
State: Arkansas          0.376      0.097    0.186    0.566
State: California        0.845      0.115    0.620    1.071
State: Colorado          1.324      0.108    1.113    1.534
State: Connecticut       2.204      0.216    1.781    2.628
State: Delaware          1.401      0.336    0.742    2.059
State: Florida           0.604      0.101    0.406    0.803
State: Georgia          -0.057      0.084   -0.221    0.108
State: Idaho             0.488      0.115    0.262    0.714
State: Illinois          1.424      0.094    1.240    1.608
State: Indiana           1.322      0.096    1.134    1.511
State: Iowa              2.040      0.096    1.853    2.228
State: Kansas            0.448      0.095    0.263    0.634
State: Kentucky          0.844      0.093    0.662    1.025
State: Louisiana        -0.248      0.100   -0.444   -0.053
State: Maine             2.602      0.161    2.286    2.918
State: Maryland          1.135      0.139    0.863    1.407
State: Massachusetts     2.750      0.172    2.413    3.087
State: Michigan          1.806      0.097    1.615    1.997
State: Minnesota         2.026      0.099    1.833    2.220
State: Mississippi      -0.059      0.095   -0.244    0.127
State: Missouri          0.973      0.092    0.793    1.152
State: Montana           1.290      0.109    1.075    1.504
State: Nebraska          0.644      0.099    0.451    0.838
State: Nevada            0.517      0.158    0.208    0.826
State: New Hampshire     2.567      0.196    2.183    2.951
State: New Jersey        1.401      0.155    1.097    1.706
State: New Mexico        0.835      0.134    0.572    1.097
State: New York          1.946      0.105    1.740    2.152
State: North Carolina    0.790      0.092    0.610    0.970
State: North Dakota      1.439      0.112    1.220    1.658
State: Ohio              1.662      0.098    1.470    1.854
State: Oklahoma          0.369      0.099    0.175    0.562
State: Oregon            1.403      0.123    1.163    1.644
State: Pennsylvania      1.479      0.103    1.277    1.681
State: Rhode Island      2.731      0.266    2.210    3.252
State: South Carolina    0.320      0.109    0.106    0.534
State: South Dakota      1.538      0.105    1.332    1.743
State: Tennessee         0.603      0.094    0.420    0.787
State: Texas            -0.236      0.091   -0.413   -0.058
State: Utah              0.032      0.131   -0.225    0.289
State: Vermont           3.380      0.170    3.046    3.714
State: Virginia          0.967      0.088    0.794    1.141
State: Washington        1.511      0.119    1.278    1.745
State: West Virginia     0.928      0.107    0.718    1.139
State: Wisconsin         2.270      0.102    2.070    2.469
State: Wyoming           0.439      0.141    0.162    0.717

Model meta-data
 outcome    N   R2 R2-adj. R2-cv
 1 dem12_frac 3063 0.68    0.68  0.67

Etas from analysis of variance
 eta eta.part
 CA       0.075    0.132
 S        0.013    0.024
 Black    0.363    0.542
 Asian    0.167    0.285
 Hispanic 0.205    0.343
 State    0.619    0.740

 

Republicans 2012
Model coefficients
Estimate Std. Error CI.lower CI.upper
CA                       0.138      0.019    0.102    0.175
S                        0.026      0.020   -0.013    0.066
Black                   -0.624      0.019   -0.661   -0.588
Asian                   -0.211      0.013   -0.237   -0.185
Hispanic                -0.310      0.016   -0.341   -0.278
State: Alabama           0.000         NA       NA       NA
State: Arizona          -1.016      0.179   -1.367   -0.665
State: Arkansas         -0.481      0.098   -0.674   -0.288
State: California       -0.973      0.117   -1.201   -0.744
State: Colorado         -1.411      0.109   -1.625   -1.197
State: Connecticut      -2.210      0.219   -2.640   -1.781
State: Delaware         -1.425      0.341   -2.092   -0.757
State: Florida          -0.596      0.103   -0.797   -0.395
State: Georgia           0.047      0.085   -0.120    0.214
State: Idaho            -0.582      0.117   -0.810   -0.353
State: Illinois         -1.469      0.095   -1.656   -1.282
State: Indiana          -1.378      0.098   -1.569   -1.187
State: Iowa             -2.066      0.097   -2.256   -1.875
State: Kansas           -0.506      0.096   -0.695   -0.318
State: Kentucky         -0.859      0.094   -1.043   -0.676
State: Louisiana         0.208      0.101    0.010    0.406
State: Maine            -2.698      0.163   -3.018   -2.378
State: Maryland         -1.211      0.140   -1.486   -0.935
State: Massachusetts    -2.796      0.174   -3.137   -2.455
State: Michigan         -1.792      0.099   -1.986   -1.599
State: Minnesota        -2.093      0.100   -2.290   -1.897
State: Mississippi       0.047      0.096   -0.142    0.235
State: Missouri         -1.027      0.093   -1.210   -0.845
State: Montana          -1.390      0.111   -1.608   -1.173
State: Nebraska         -0.696      0.100   -0.893   -0.500
State: Nevada           -0.643      0.160   -0.957   -0.330
State: New Hampshire    -2.579      0.198   -2.968   -2.190
State: New Jersey       -1.419      0.157   -1.728   -1.111
State: New Mexico       -1.009      0.136   -1.275   -0.743
State: New York         -1.987      0.106   -2.196   -1.779
State: North Carolina   -0.804      0.093   -0.987   -0.622
State: North Dakota     -1.529      0.113   -1.751   -1.307
State: Ohio             -1.717      0.099   -1.912   -1.523
State: Oklahoma         -0.272      0.100   -0.468   -0.076
State: Oregon           -1.559      0.124   -1.803   -1.315
State: Pennsylvania     -1.482      0.104   -1.687   -1.277
State: Rhode Island     -2.780      0.269   -3.308   -2.252
State: South Carolina   -0.346      0.111   -0.563   -0.129
State: South Dakota     -1.584      0.106   -1.792   -1.376
State: Tennessee        -0.612      0.095   -0.798   -0.426
State: Texas             0.228      0.092    0.048    0.408
State: Utah             -0.084      0.133   -0.345    0.176
State: Vermont          -3.447      0.173   -3.786   -3.108
State: Virginia         -1.010      0.090   -1.185   -0.834
State: Washington       -1.601      0.121   -1.838   -1.364
State: West Virginia    -0.989      0.109   -1.202   -0.775
State: Wisconsin        -2.262      0.103   -2.464   -2.059
State: Wyoming          -0.581      0.143   -0.862   -0.300

Model meta-data
 outcome    N   R2 R2-adj. R2-cv
 1 rep12_frac 3063 0.67    0.67  0.66

Etas from analysis of variance
 eta eta.part
 CA       0.077    0.133
 S        0.014    0.024
 Black    0.350    0.523
 Asian    0.168    0.282
 Hispanic 0.200    0.330
 State    0.626    0.739

 

Democrats, 2008
Model coefficients
Estimate Std. Error CI.lower CI.upper
CA                      -0.118      0.020   -0.156  -0.0793
S                       -0.045      0.021   -0.086  -0.0038
Black                    0.609      0.019    0.571   0.6470
Asian                    0.215      0.014    0.188   0.2415
Hispanic                 0.280      0.017    0.247   0.3131
State: Alabama           0.000         NA       NA       NA
State: Arizona           1.005      0.187    0.639   1.3719
State: Arkansas          0.477      0.103    0.276   0.6776
State: California        1.054      0.122    0.815   1.2922
State: Colorado          1.512      0.114    1.289   1.7349
State: Connecticut       2.444      0.229    1.996   2.8920
State: Delaware          1.666      0.355    0.969   2.3626
State: Florida           0.700      0.107    0.491   0.9104
State: Georgia           0.014      0.089   -0.160   0.1883
State: Idaho             0.626      0.122    0.388   0.8650
State: Illinois          1.867      0.099    1.672   2.0617
State: Indiana           1.758      0.102    1.558   1.9573
State: Iowa              2.236      0.101    2.038   2.4347
State: Kansas            0.592      0.100    0.395   0.7887
State: Kentucky          1.068      0.098    0.876   1.2600
State: Louisiana        -0.253      0.105   -0.460  -0.0461
State: Maine             2.698      0.170    2.364   3.0317
State: Maryland          1.211      0.147    0.924   1.4988
State: Massachusetts     2.916      0.182    2.559   3.2718
State: Michigan          2.100      0.103    1.898   2.3017
State: Minnesota         2.170      0.104    1.966   2.3752
State: Mississippi      -0.057      0.100   -0.253   0.1401
State: Missouri          1.308      0.097    1.117   1.4978
State: Montana           1.603      0.116    1.376   1.8296
State: Nebraska          0.800      0.104    0.595   1.0050
State: Nevada            0.809      0.167    0.482   1.1356
State: New Hampshire     2.730      0.207    2.324   3.1360
State: New Jersey        1.424      0.164    1.102   1.7461
State: New Mexico        1.208      0.142    0.930   1.4857
State: New York          1.978      0.111    1.760   2.1954
State: North Carolina    0.923      0.097    0.732   1.1133
State: North Dakota      1.820      0.118    1.588   2.0512
State: Ohio              1.728      0.104    1.525   1.9308
State: Oklahoma          0.369      0.104    0.165   0.5737
State: Oregon            1.610      0.130    1.356   1.8644
State: Pennsylvania      1.727      0.109    1.513   1.9406
State: Rhode Island      2.849      0.281    2.298   3.3999
State: South Carolina    0.396      0.115    0.170   0.6227
State: South Dakota      1.819      0.111    1.602   2.0364
State: Tennessee         0.765      0.099    0.571   0.9589
State: Texas            -0.020      0.096   -0.208   0.1673
State: Utah              0.510      0.139    0.238   0.7816
State: Vermont           3.485      0.180    3.132   3.8387
State: Virginia          1.154      0.094    0.971   1.3373
State: Washington        1.662      0.126    1.415   1.9092
State: West Virginia     1.408      0.114    1.185   1.6309
State: Wisconsin         2.585      0.108    2.374   2.7964
State: Wyoming           0.669      0.150    0.375   0.9618

Model meta-data
 outcome    N   R2 R2-adj. R2-cv
 1 dem08_frac 3063 0.65    0.64  0.63

Etas from analysis of variance
 eta eta.part
 CA       0.065    0.109
 S        0.023    0.039
 Black    0.342    0.498
 Asian    0.171    0.276
 Hispanic 0.180    0.290
 State    0.650    0.738

 

Republicans, 2008
Model coefficients
Estimate Std. Error CI.lower CI.upper
CA                       0.118      0.020   0.0795    0.157
S                        0.055      0.021   0.0137    0.096
Black                   -0.586      0.019  -0.6237   -0.548
Asian                   -0.213      0.014  -0.2394   -0.186
Hispanic                -0.269      0.017  -0.3026   -0.236
State: Alabama           0.000         NA       NA       NA
State: Arizona          -1.026      0.188  -1.3945   -0.658
State: Arkansas         -0.612      0.103  -0.8139   -0.410
State: California       -1.160      0.122  -1.3994   -0.920
State: Colorado         -1.571      0.114  -1.7953   -1.347
State: Connecticut      -2.479      0.230  -2.9297   -2.029
State: Delaware         -1.684      0.357  -2.3848   -0.984
State: Florida          -0.698      0.108  -0.9085   -0.486
State: Georgia          -0.012      0.089  -0.1866    0.163
State: Idaho            -0.725      0.122  -0.9651   -0.485
State: Illinois         -1.905      0.100  -2.1011   -1.709
State: Indiana          -1.771      0.102  -1.9721   -1.571
State: Iowa             -2.277      0.102  -2.4770   -2.078
State: Kansas           -0.638      0.101  -0.8356   -0.440
State: Kentucky         -1.087      0.098  -1.2802   -0.894
State: Louisiana         0.201      0.106  -0.0064    0.409
State: Maine            -2.753      0.171  -3.0890   -2.418
State: Maryland         -1.276      0.147  -1.5644   -0.987
State: Massachusetts    -3.008      0.183  -3.3657   -2.649
State: Michigan         -2.146      0.104  -2.3493   -1.943
State: Minnesota        -2.269      0.105  -2.4747   -2.063
State: Mississippi       0.046      0.101  -0.1518    0.244
State: Missouri         -1.342      0.098  -1.5336   -1.151
State: Montana          -1.760      0.116  -1.9885   -1.532
State: Nebraska         -0.861      0.105  -1.0665   -0.655
State: Nevada           -0.984      0.168  -1.3127   -0.656
State: New Hampshire    -2.747      0.208  -3.1548   -2.339
State: New Jersey       -1.454      0.165  -1.7781   -1.130
State: New Mexico       -1.242      0.142  -1.5216   -0.963
State: New York         -2.012      0.112  -2.2308   -1.793
State: North Carolina   -0.927      0.098  -1.1188   -0.736
State: North Dakota     -1.899      0.119  -2.1321   -1.666
State: Ohio             -1.786      0.104  -1.9901   -1.582
State: Oklahoma         -0.273      0.105  -0.4785   -0.067
State: Oregon           -1.750      0.130  -2.0059   -1.495
State: Pennsylvania     -1.746      0.110  -1.9607   -1.531
State: Rhode Island     -2.922      0.283  -3.4759   -2.368
State: South Carolina   -0.429      0.116  -0.6565   -0.201
State: South Dakota     -1.893      0.111  -2.1112   -1.674
State: Tennessee        -0.788      0.099  -0.9826   -0.593
State: Texas             0.027      0.096  -0.1613    0.216
State: Utah             -0.652      0.139  -0.9248   -0.378
State: Vermont          -3.556      0.181  -3.9110   -3.200
State: Virginia         -1.176      0.094  -1.3603   -0.992
State: Washington       -1.738      0.127  -1.9868   -1.490
State: West Virginia    -1.450      0.114  -1.6740   -1.226
State: Wisconsin        -2.613      0.108  -2.8253   -2.401
State: Wyoming          -0.781      0.150  -1.0759   -0.486

Model meta-data
 outcome    N   R2 R2-adj. R2-cv
 1 rep08_frac 3063 0.64    0.64  0.63

Etas from analysis of variance
 eta eta.part
 CA       0.065    0.109
 S        0.028    0.047
 Black    0.329    0.482
 Asian    0.169    0.272
 Hispanic 0.174    0.279
 State    0.661    0.741

Summary & interpretation

In general, the models performed fairly well, the mean cross-validated R2 was 65% (54% to 72%). The best way to summarize the findings for the predictors would be to aggregate/meta-analyze the results. I’m too busy to do that now, so we will just look at the non-state predictors presented in less space:

     CA     S Black Asian Hispanic       group
1 -0.09  0.13  0.77  0.27     0.38   fit_dem16
2  0.10 -0.17 -0.75 -0.27    -0.39   fit_rep16
3 -0.15  0.08 -0.09  0.10    -0.03 fit_green16
4  0.00  0.35 -0.03  0.01     0.07 fit_liber16
5 -0.14 -0.03  0.65  0.21     0.32   fit_dem12
6  0.14  0.03 -0.62 -0.21    -0.31   fit_rep12
7 -0.12 -0.04  0.61  0.21     0.28   fit_dem08
8  0.12  0.06 -0.59 -0.21    -0.27   fit_rep08

So, for predicting democrat votes, we can see that the betas are all negative for CA: -.09, -.14 and -.12. All else equal, smarter counties voted less for democrats, whether it was Clinton or Obama. S is weird. The beta for 2016 was .13 but it was -.03 and -.04 for 2012 and 2008! A sign change and it’s not a chance finding because the use of a dataset with n≈3,000 gives us a lot of precision, and none of these did actually have CIs that even overlapped zero. So for 2016 this gives us the odd situation where the highly correlated CA and S variables (r = .71) have reverse signs for the outcome: -.09 and .13. Smarter counties voted less for democrats, but those higher in S voted more for democrats — all else equal. That wasn’t so in the Obama elections where CA and S had the same directions. As for demographics, the situation is not surprising: non-Whites like democrats, a lot. We knew this from simpler statistics showing that Blacks vote 93% for Obama. The curious finding here is that this was not just due to the lower CA and S for Blacks or Hispanics. The Black effect was even stronger for the 2016 election than the Obama ones, which is somewhat curious. The general idea seems to be that minorities like to vote for their own candidates, but it seems not to be the case for these data. Or there’s some annoying confound, like turnout %. Hispanics are voting increasingly for democrats (betas: .28 to .32 to .38) and Asians too, maybe (.21 to .21 to .27). The republican results are not so interesting because they are essentially the opposite of the democrat results (for non-2016, they are necessarily the opposite because NYT did away with the third party votes).

Results for the two smaller parties are somewhat interesting. Greens showed the same mismatch in directionality for CA and S, just with reversed beta strengths (-.15 and .08; CA stronger, reverse for democrats). Interestingly for libertarians, there was no effect of CA, but a large one of S (.35). Given the generally positive correlations between libertarian preferences and CA, this is somewhat surprising (see this and this). Perhaps more interestingly, demographics had little to no effect on preferences for libertarians. This was also true to a bit lesser extent for greens.

The relative importance of variables can be glanced from the etas. Most of the models’ validity is due to state-level effects (whatever these represent) and demographics, mostly % Blacks. The mean eta for State was .59 (range: .47 to .66), and for Black .29 (.02 to .43). The small values for Black are from the third parties which, as we saw, were not a thing that Blacks cared much about as a group once controlled for CA and S. CA and S themselves had mean etas of .06. As such, cognitive ability and social inequality were not particularly important for explaining the election outcomes at the county level.

Other notes:

  • Analyses were unweighted. I reasoned that we are here thinking of the counties as the units of interest, and so we should weigh them equally, not give more weight to the larger counties. We would do that if we were interested in modeling the national outcome itself or persons inside counties.
  • For the Green’s analysis, n≈2,500. Why is n only about 2,500 instead of 3,000? Because the Greens did not run in all states, and so these have missing data. Perhaps one should impute these values, maybe with 0%, maybe with estimated values.

January 11, 2017

Cross-national data for social desirability bias in polling

Filed under: Political science — Tags: , , — Emil O. W. Kirkegaard @ 21:16

This is just a brief post summarizing some stuff because I don’t have time to do more right now.

Noah Carl asks me:

Can you point me to some evidence (e.g., a blog post) on “PC-bias” in polling in recent years?

It is hard to establish biased polling. One has to:

  1. Obtain datasets of polling data
  2. Forecast the election using the data
  3. Compare the forecast to the election results

What exactly is the bias we are talking about?

In general, when one asks people questions, they don’t always answer truthfully. There is a large literature (review) on why people lie on surveys which applies equally to surveys of political intentions (=polls). Why do people lie on surveys? Generally, its when the response would be embarrassing, unpopular or illegal — in short, what is socially desirable. The stress is on socially, because sometimes what is popular in the media to talk about is sometimes very different from what people really like/dislike. I equate the media with socially for the time being because the media reflects the collective thinking, and humans are sensitive to what (they think) others think about them. The context matters too: people lie more when they have to answer to another human (face to face, phone) than to a machine (internet).

So we expect people to lie on polls when one or more of the above are true. We expect the most lying when voters are interviewed face-to-face and when they intend to vote for parties/politicians that are very unpopular in the media. Compared with the general population, the media is pushed to the left to some degree. I reviewed polls of journalists etc. (in Danish), but generally the situation looks something like this:

This shows the journalists preferences for the Norwegian parliament election in 2016. The numbers are the number of seats, which is 169, so each seat is ~.59%. The last election was in 2013, but one can find recent polls to compare with.

norway-polls

Doing some quick arithmetic, we see that the big center-left party (Ap) gets some 38.5% among journalists and about 35% for general population. Much the same. However, the big center-right party (H) gets about 25% among voters, but only 13.6% among journalists. In general, the two blocks are about equally large in parliament and in general elections. But when we look at the journalist cake chart, it has a very large left-wing lead: 12+24+65+18=119 seats, out of 169 is 70%. The immigrant critical party, Frp, gets about 15% in the general polls, but ~0% among journalists. Talk about the under-representation of minorities! Alas, it is the wrong kind — European and with more men.

A simple way to estimate how much bias to expect for each party is to calculate the difference between the general population polls/elections for a party and the journalist polls. Another way is to rank the parties on a left-right spectrum (even if this doesn’t make much sense) and then expect more bias from the parties the furthest to the right.

(I am aware that the general population polls show some bias — according to the thesis of this post — and thus are somewhat problematic to use to compare to the journalist polls, which probably also show some bias. I don’t know the details of how they were done. I ignore this problem for the being.)

Show me the data

This kind of bias has been investigated in different countries and even re-discovered a number of times. For this reason, it has a number of names in the local languages. In the UK it’s called the shy Tory factor. Before writing this post, I searched a bit and one can find a number of okay sources on the topic covering the UK (Tory, UKIP, SNP), US (Trump) and Canada. In no particular order:

  • http://www.nature.com/news/the-polling-crisis-how-to-tell-what-people-really-think-1.20815
  • https://fivethirtyeight.com/features/shy-voters-probably-arent-why-the-polls-missed-trump/
  • http://www.threehundredeight.com/2015/05/are-conservative-parties-under.html
  • https://techcrunch.com/2016/10/19/the-perils-of-polling-in-a-brexit-and-donald-trump-world/
  • https://ballotpedia.org/Shy_Elephant_Factor
  • http://eprints.ncrm.ac.uk/3789/1/Report_final_revised.pdf
  • http://www.jstor.org/stable/3078789?seq=1#page_scan_tab_contents
  • http://www.sciencedirect.com/science/article/pii/S0261379415002231

The report above is probably the most detailed on the UK.

But the amount of data is not really sufficient. One should dig up more data, for more countries, more parties and more elections. Sweden is a particularly interesting case given that the media in this country is perhaps the most biased. This should be reflected in especially strong polling bias against the parties the media doesn’t like, which is primarily the immigrant critical party (Sweden Democrats, SD). There is a website that monitors SD in the polls and also has the actual election results. The result looks like this:

sweden-sd-polls

The yellow line is a moving average of the polls, the red dots are the election results. So, out of 3 elections, SD was underestimated 3 times. By how much? Here we have to eye-ball the chart, but something like: 1, 0.5, 2%points. So, small effects even in the most extreme country.

Denmark?

Erik Gahner has made a github repo for polling data from Denmark 2010-present. The last elections were in 2011 and 2015, so that gives us 2 datapoints. Instead, I was even more lazy and used the polls listed on Wikipedia for the 2015 election. Plotting the data with a loess looks like this:

danish_polls_2015

The squares at the end show the actual results (not used in the fitting process). Looking at the plot, one can see the large gain for blue, which is O, the nationalist party. The same one polls sometimes fail to find just a single journalist who would vote for. If we take the model predictions and subtract the actual result, we get this:

> preds - d[1, parties]
    V    A    O    B   F    Ø    I    C    K     Å
1 1.1 -1.4 -3.1 0.34 1.5 0.97 0.11 0.34 0.18 0.021

So, there was a 3.1%point error for the nationalist party. Note that there was also a moderately high underestimation of A, the large center-left party, and about equal loss for V, the large center-right party.

Turns out that someone already did this kind of analysis for Denmark, data 1990-2015. Results:

danish-polls-bias

Again, the bias for DF (nationalist) is the largest, but there is a also a substantial bias for S (large center-left). Not sure why this is.

Where to now?

Collect data like these for all countries one can find, particularly those with growing nationalist parties (so pretty much all European countries). Fit the same kind of model to all the data (methodological consistency), make the predictions, compare with results. Aggregate results. The effect size will not be large, so expect a lot of noise.

Wikipedia is helpful for gathering data:

  • https://en.wikipedia.org/wiki/European_Parliament_election,_2014_(United_Kingdom)#Opinion_polls
  • https://en.wikipedia.org/wiki/Opinion_polling_for_the_2010_United_Kingdom_general_election
  • https://en.wikipedia.org/wiki/Opinion_polling_for_the_2015_United_Kingdom_general_election
  • https://en.wikipedia.org/wiki/Norwegian_parliamentary_election,_2013#Opinion_polls
  • https://en.wikipedia.org/wiki/Swiss_federal_election,_2015#Opinion_polls
  • etc. Just search “wiki opinion polling [country]”

If someone were to make a large, clean dataset of this kind of data, the analysis would be fairly trivial to carry out.

November 15, 2016

Crowdsourcing a dataset: Multiparty democracies, election thresholds and running for parliament

Filed under: Political science — Tags: , , , — Emil O. W. Kirkegaard @ 09:30

I’ve decided to re-do my old and not too well done 2013 study. It goes like this, for multiparty democracies:

  • Countries differ in the number of houses: 1 vs. 2.
  • Countries differ in the number of seats in these houses.
  • Usually, to get a party on the ballot, one must first collect X signatures. This number X varies.
  • To get a seat in parliament, usually there’s a threshold so that parties below that do not get in, even if they would have gotten >0 seats. This threshold varies.
  • These variations likely have a substantial effect on the party-diversity of parliament.

Apparently, there’s very little academic research on how works, so I collected some data years ago. I found a correlation of about -.50 between election thresholds and number of parties in parliament. Looks like this.

old

This was all done using sketchy Wikipedia data, so it must be redone. I also failed to realize the number of seats in parliament was a confounding factor.

There’s a spreadsheet here.

So I’m looking for collaborators who can fill in data for countries. I’ve prefilled the sheet with Wikipedia data. But more is needed for a reliable dataset:

  • Find the laws for each country that state the rules for 1) election threshold, and 2) number of required signatures to get on the ballot.
  • Check numbers and update datafile and Wikipedia accordingly.

It’s not easy to collect this data because of the many languages, so it’s a crowdsourcing project. Finding the laws and exact § even if one speaks the language often isn’t easy.

In many cases, judgments must be made because voting systems vary quite a bit. There’s also the question of how to deal with independents. I counted them as 1 party before, but perhaps that’s the wrong approach. Many countries use two-set systems where the national parliament is made up of members or delegates from regional parliaments. Complicated…

If you want to help, send me an email or send me your email on Twitter. Note which languages you want to cover.

So far, have people covering:

  • Scandinavian (including Finland).
  • German
  • Italian

So, need everything else.

September 6, 2016

Cognitive ability and political preferences in Denmark: knitr edition

http://rpubs.com/EmilOWK/208757

Analysis of the data collected so far, presented side by side with R code. It will be expanded into a proper paper and submitted to OQSPS soonish. Soonish here meaning when Noah gets around to do it!

Due to the surprising results, we should probably do a follow-up replication using new subjects. These results are hard to believe in the light of earlier findings from other countries based on larger samples etc.

Summary for the less math inclined:

  • Despite virtually every other published study, we did not find much correlation between preferences for personal and economic freedom and cognitive ability: r’s about .05.
  • Self-rated preference for personal freedom correlated very poorly with measured preferences, r≈.10, but for economic freedom, the correlation was alright r≈.50. This means one can use the latter as an okay proxy, but not the first. People don’t seem to have an idea of how important they actually think personal freedoms are relative to other people.
  • There was little to none correlation between the political axes, r≈.10. This means that one cannot summarize people’s political preferences using only 1 dimension such as the popular left-wing axis.
  • Using people’s self-rated agreement with parties, one can estimate the political positions of the parties. Doing so using people’s self-rated preferences recreated the familiar left-right economic axes, but not the personal freedom axes, since all parties were about equally in favor of personal freedoms. However, using measured preferences, the left-economic parties were found to be less in favor of personal freedoms than the right-economic parties. The result is that the party political dimensions are highly correlated (about r≈.80) and so one can summarize the parties using a 1-dimensional model.

May 29, 2016

Personal freedom and cognitive ability: OKCupid dataset replication

Noah Carl has been investigating the relationship between cognitive ability and political opinions aside from the usual confused 1-axis left-right model. Specifically, looking at the economic freedom and personal freedom axes (á la this test). He did this in two datasets so far, covering the UK and the US:

OKCupid does not have that many questions on economic freedom (but there are some), but it has a lot of questions on personal freedom. I identified 18 in my search of which 17 had a sufficient sample size.

Questions

The questions are the following:

q175 q218 q219 q340 q341 q868 q7204 q13054 q30455 q36357 q45158 q49714 q52503 q54450 q55344 q61830 q82734 q85583 q91207

For each question, the order was changed so that later options were more in favor of freedom. If necessary, some options were removed (recoded to NA), such as those refusing to say how they would vote (for assisted suicide/q45168). In the case of child limits/q49714, more freedom was coded as anyone not answering “Yes”. And so on for other questions.

Correlations

The latent correlations (estimated Pearson correlations if the variables had been measured as normally distributed continuous variables) were:

CA flag burning freedom of religion prostitution child limits child test gay marriage cigarettes smoking bars cannabis illegal drugs illegal drugs2 illegal drugs religious weapons motorcycles seatbelts mandatory voting assisted suicide
CA 1.00 0.40 0.12 0.28 0.06 0.14 0.35 0.12 -0.04 0.24 0.20 0.23 0.02 0.05 -0.11 0.12 0.01 0.31
flag burning 0.40 1.00 -0.04 0.43 0.08 0.12 0.43 0.16 -0.03 0.46 0.45 0.44 0.14 0.16 -0.10 0.32 -0.16 0.40
freedom of religion 0.12 -0.04 1.00 0.12 0.19 0.05 0.00 0.04 0.08 -0.06 0.01 0.07 0.10 0.04 0.21 -0.03 -0.06 0.02
prostitution 0.28 0.43 0.12 1.00 0.02 0.01 0.46 0.25 0.25 0.56 0.50 0.49 0.18 -0.12 -0.06 0.36 0.03 0.58
child limits 0.06 0.08 0.19 0.02 1.00 0.59 -0.19 0.14 -0.07 0.06 0.05 -0.13 0.00 -0.06 0.11 0.01 0.12 -0.21
child test 0.14 0.12 0.05 0.01 0.59 1.00 -0.22 0.12 -0.07 -0.05 0.15 -0.05 -0.08 -0.01 0.19 0.09 0.13 -0.24
gay marriage 0.35 0.43 0.00 0.46 -0.19 -0.22 1.00 0.08 -0.14 0.58 0.31 0.47 0.46 0.13 -0.62 -0.06 0.00 0.73
cigarettes 0.12 0.16 0.04 0.25 0.14 0.12 0.08 1.00 0.55 0.41 0.41 0.38 0.40 -0.14 0.11 0.27 0.10 0.32
smoking bars -0.04 -0.03 0.08 0.25 -0.07 -0.07 -0.14 0.55 1.00 0.13 0.41 0.31 0.09 -0.27 0.55 0.48 0.04 0.16
cannabis 0.24 0.46 -0.06 0.56 0.06 -0.05 0.58 0.41 0.13 1.00 0.78 0.78 0.64 -0.13 -0.50 0.27 0.17 0.68
illegal drugs 0.20 0.45 0.01 0.50 0.05 0.15 0.31 0.41 0.41 0.78 1.00 0.54 0.38 -0.08 -0.04 0.34 0.05 0.47
illegal drugs2 0.23 0.44 0.07 0.49 -0.13 -0.05 0.47 0.38 0.31 0.78 0.54 1.00 0.49 -0.12 -0.19 0.21 -0.04 0.52
illegal drugs religious 0.02 0.14 0.10 0.18 0.00 -0.08 0.46 0.40 0.09 0.64 0.38 0.49 1.00 -0.19 -0.50 0.14 0.05 0.45
weapons 0.05 0.16 0.04 -0.12 -0.06 -0.01 0.13 -0.14 -0.27 -0.13 -0.08 -0.12 -0.19 1.00 -0.07 -0.39 -0.15 -0.03
motorcycles -0.11 -0.10 0.21 -0.06 0.11 0.19 -0.62 0.11 0.55 -0.50 -0.04 -0.19 -0.50 -0.07 1.00 0.11 -0.45 -0.50
seatbelts 0.12 0.32 -0.03 0.36 0.01 0.09 -0.06 0.27 0.48 0.27 0.34 0.21 0.14 -0.39 0.11 1.00 0.34 0.28
mandatory voting 0.01 -0.16 -0.06 0.03 0.12 0.13 0.00 0.10 0.04 0.17 0.05 -0.04 0.05 -0.15 -0.45 0.34 1.00 0.18
assisted suicide 0.31 0.40 0.02 0.58 -0.21 -0.24 0.73 0.32 0.16 0.68 0.47 0.52 0.45 -0.03 -0.50 0.28 0.18 1.00

 

CA = cognitive ability.

The distribution of correlations among the personal freedom measures was this:

intercors

So, it leaned towards positive but not overwhelmingly so. People are not that consistent on the personal freedom axis. The negative correlations mostly come from the motorcycle question: apparently many people who want to ban motorcycles also support e.g. gay marriage (r = -.62).

The mean intercorrelations by item were:

Variable Mean cor with others
flag burning 0.20
freedom of religion 0.05
prostitution 0.25
child limits 0.04
child test 0.05
gay marriage 0.15
cigarettes 0.23
smoking bars 0.15
cannabis 0.30
illegal drugs 0.29
illegal drugs2 0.26
illegal drugs religious 0.17
weapons -0.09
motorcycles -0.11
seatbelts 0.17
mandatory voting 0.02
assisted suicide 0.24

 

Item response theory factor analysis

Altho some (mostly te Nijenhuis and his co-authors, and previously myself) have been analyzing item-level data using classical test theory measures, this approach is inappropriate because Pearson correlations are influenced by the difficulty of the items (proportion who gets them right).

The proper method is to use item response theory factor analysis (I use the version in the psych package, irt.fa). The item plot was this:

item_info

So, we see that many items were not very good measures (Y axis = discrimination ≈ factor loading, but not standardized to max of 1) of the assumed single underlying factor. This is in part because the factor structure is more complex than a single factor. Most items were too far on the left side, meaning that they could not distinguish well between the people on the right side (various sorts of libertarians, presumably). One would have to include more extreme questions, such as perhaps getting rid of FDA approval for drugs (the hardcore libertarians will say that the market solves this too, assuming perfect rationality…).

Still, suppose we ignore the factor structure problem and calculate factor scores anyway, then they look like this:

pf_dist

And again we see the problem: both ceiling and floor effects. Some people wanted to ban everything (left end) and many wanted to ban nothing (right end). One would have to introduce more questions to remove these effects, if that is even possible. Alternatively, some of the scores may be due to missing data for these persons. Most persons did not have data for all questions, so the scoring function tried to estimate scores the best it could from the available data.

Still, we can correlate the scores with cognitive ability (based on up to 14 items, see the OKCupid dataset release paper), and we get this:

CA_PF CA_PF_2

There is indeed a positive correlation of .24 (N≈50k, but many cases are estimated with missing data). Obviously, the distribution is not normal, so Pearson correlation is a somewhat biased measure. Spearman’s correlation gave the same result, however. The strength of the relationship is around what previous studies found (.20 to .30).

The second plot is a more fancy plot which shows the density at each area. Lighter = more persons. So we see that the highest density of points is in the top-right quadrant explaining the positive correlation.

Jensen’s method

It is possible to use Jensen’s method (method of correlated vectors) using item response theory results, but one must use the discrimination scores (from the item plot) and the latent correlations (from the first table). If one does, one gets this:

jensens_method_CA_PF

And the result is positive as expected. We can also see that there are only two things that smarter people support banning more than less smart people: smoking in bars and motorcycles. The second one sounds pretty odd to me.

So, hopefully, someone will write this into a real paper and submit to ODP or OQSPS. At least, if I’m a co-author. If not, then submit to Intelligence or PAID, I guess.

Older Posts »

Powered by WordPress