Some Remarks on Premodern GDP

I have seen these misinterpretations more times than I can possibly count:

1. Confusion of premodern real GDP per capita with living standards. No such thing as a necessary lower bound on GDP to prevent widespread destitution exists. If everyone is a subsistence farmer with high labor productivity, but nobody sells or buys anything from anyone else, that’s a society with a GDP of $0, but with fairly high living standards historically. If the share of output which is not for sale is highly variable and is only weakly correlated with real GDP per capita between societies, real GDP per capita will often severely underestimate actual output per capita and generally be a poor measure of living standards. And if the statistical agencies did count household production in GDP, the world would be a lot different.

2. Confusion of premodern inequality with real GDP per capita. More luxury goods and services may simply be a result of a higher rate of exploitation by the elites of the commoners, rather than a higher real GDP per capita.

3. Confusion of premodern economic complexity with real GDP per capita. Living standards and real GDP per capita are absolutely not measures of economic complexity. High productivity due to gifts of nature is no substitute for high productivity due to the gifts of the human mind. Korea was at least as poor as Ghana back when Ghana became independent. That absolutely does not mean Korea’s economic complexity was even remotely comparable to that of Ghana (see also previous link). Likewise, in a Malthusian environment, increased population may result in simultaneously falling real GDP per capita as a result of falling per capita agricultural output, but rising economic complexity as a result of growing division of labor and easier ability to create goods and services with high fixed costs. Economic complexity is in many cases much more useful to analyze than real GDP per capita. Much easier to analyze, as well. You and I admire the extent of division of labor in the Empire of Rome much more than its per capita agricultural output.


The Ramesses III Sea Peoples Reliefs

Whenever you search Ramesses III Sea Peoples you ALWAYS get a depiction of the relief showing the Battle of the Delta. You never see a depiction of the relief showing the Battle of Djahay. I have sought here to remedy this.
The depiction of the Battle of Djahy:

The depiction of the Battle of the Delta:

From here.
Translations here.

The Myth of Desperation

One narrative that’s been floating around the lyin’ press throughout the past two years is that that Trump and Sanders voters were mainly driven by desperation -that one wouldn’t vote for a candidate of dramatic change if one was perfectly satisfied with one’s affairs.

Perhaps the perfect counterexample to that is the county in Michigan with the highest median household income and lowest poverty rate in the state -Livingston.

Livingston County is many things, but it ain’t desperate. It’s rich, very Republican -it went for John McCain with 55% of the vote in November 2008, and 61% of the vote for Mitt Romney in 2012- and is not the place where one would find out-of-work factory workers or coal miners discontented with their economic situation, because there aren’t much of them. And, during the 2016 primaries, the candidate there who got the most votes was Donald Trump. The candidate who got the second-most votes there was Bernie Sanders (indeed, Livingston County had a higher Bernie share in the Democratic primary than all the counties surrounding it). The candidate who got the third-most votes there was John Kasich -this county isn’t as socially conservative as the western part of the state. Nor did woke neocon Marco Rubio appeal there much -he got a lower share of the Republican vote there than in the rest of the state, and Rubio and Kasich’s vote share combined would not have sufficed to prevent Trump from winning it in the primary.

Now, before 2016, Michigan hadn’t had a real Democratic primary for ages. But it did have real Republican primaries in 1996, 2000, 2008, and 2012. And guess who won the vote in Livingston County (a solidly Republican county, it must be remembered) each time? Mitt Romney by double digits in 2012, Mitt Romney by double digits in 2008, George W. Bush by single digits in 2000 [most MI counties went for McCain at the time], and Bob Dole by double digits in 1996 (Buchanan did well in Lapeer and St. Clair, though, and nearly won the famous Macomb). Not Ron Paul. Not Mike Huckabee. Not Alan Keyes. Rich guy Mitt Romney and establishment candidate George W. Bush.

There are other examples of this. Nevada’s third congressional district. Long Island. In the general election only, Minnesota’s sixth and second congressional districts (though Trump did far worse than Rubio there in the caucuses, he did better than Romney there in the general election).

Now, yes, Trump and Sanders really did appeal more to those among the really desperate who are White, at least, relative to Ted Cruz and Hillary Clinton. The results of the 2016 primaries in the poorest non-Hispanic White majority congressional district in the country (KY-05) are enough to prove this. But that does not mean economic or social desperation was either a necessary or sufficient condition for Trump or Sanders support (many Whites in desperate rural areas in the South also voted for HRC in the primary).

Calculating Partisan Gerrymandering (Part III of a III-part series)

Transforming a percentage into a probability of victory is fairly easy. Convert a percentage into the log odds of the percentage, multiply that by some integer, and convert that back into a percentage.

By what integer should I multiply the log-odds(percentage)? The answer varies.

I first tried this out with Michigan’s presidential vote in 2012. Michigan is known, after all, to be a high-quality gerrymander on the federal level. The result, somewhat surprisingly, was that given low enough number the log-odds(percentage) is multiplied by [i.e., given high enough values of voter swinginess] it was the Democrats who were favored under that House map (i.e., the 2011 House map in MI was a dummymander), due to the safety of the Dem seats and the complete lack of safety of the Republicans’ seats (or so it appeared) that year.

The first row in the below table is the number the log-odds (percentage) was multiplied by to produce the estimated probabilities of victory in the below rows.

However, by the 2016 presidential election numbers, the Republicans became clearly favored due to the newfound safety of their seats and a newfound danger to the Dem seats:

Note: the two-party HRC percentage is listed as over 50% in the above table due to more Democratic districts having lower voter turnout, and each district being counted equally during averaging the vote.

The number one should multiply the log-odds percentage by remains to be debated with historical statistical evidence; but I would be surprised if it were not within the range of three to twenty.

How to actually measure partisan gerrymandering (Part II of a three-part series)

I. NC plan>>>PA plan
In the previous post, I attempted to show the tradeoffs available to a designer of a gerrymander with this graph (see previous post for an explanation):

This approach would work well in states in which the districts of the favored party are all roughly equal in partisanship, such as North Carolina.

(X-axis is the name of the district, the Y-axis is the two-party Democratic vote share on the presidential level in each district)

However, not all states wanted to create a bunch of Likely Favored Party districts that would give the incumbent party supermajorities in neutral years, but would risk the opposing party gaining every seat in the state in a wave year favorable to the opposing party. An example of this is Pennsylvania’s House districts (note Tom Marino of PA was selected to be the President’s Drug Czar just today, so this post is perfect timing).

Pennsylvania has, in my judgment three tossup districts, all of which are held by Republicans, two Lean R districts, six Likely R districts, and two Safe R districts (one of which is Marino’s). There is also one Lean D heavily Obama-Trump district (PA-17), which actually was close enough in 2012 as to be acceptable for a wiser (or more partisan) Republican state legislature to turn into a Lean R district even then under more aggressive lines, and actually went for Trump by its present boundaries by more than both the Lean R districts and, naturally, all three of the the tossups. The fact it has a Democratic representative now is a huge and inexcusable failure of the Republican state legislature in 2011. The current Pennsylvania plan is, in my judgment, a mess, much inferior to North Carolina’s, and unnecessarily creates opportunities for the Democrats where, by any Republican’s judgment, there should be none. It would be relatively easy to create a redistricting plan for Pennsylvania with four Safe D seats and the other fifteen Likely R. My judgment is that the risk of all seats going to the opposing party in a wave year is worth it, and is much superior to being subject to the whims of “moderates” who are afraid of alienating their district’s swing voters in a general election while the favored party is in power. But how does one judge plans such as PA’s, anyway?

II. How does one measure the utility of tossup seats?

Another question, obviously related to the above one, is how does one measure the utility to a party of redistricting a district from Likely Opposing Party to Lean Opposing Party. For example, here’s my proposed Republican gerrymander of Indiana (no county splits, resulting in some serious malapportionment in Marion, but that’s not important here). The 2016 U.S. Senate race is used as a guide:

IN-01: In green. Two-party vote for Bayh: 52.47%. Rating: Lean D, instead of the current Likely D. This is an Obama-Trump district. The right kind of Republican can definitely get elected here. Certainly it would not be left the only uncontested seat in Indiana, as it actually was.
IN-02: In dark blue. Two-party vote for Bayh: 44.32%. Rating: Likely R.
IN-03: In red. Two-party vote for Bayh: 42.14%. Rating: Likely R.
IN-04: In orange. Two-party vote for Bayh: 38.21%. Rating: Safe R.
IN-05: In light blue. Two-party vote for Bayh: 39.84%. Rating: Safe R.
IN-06: In white. Two-party vote for Bayh: 36.52%. Rating: Safe R.
IN-07: Marion (yellow). Two-party vote for Bayh: 61.87%. Rating: Safe D.
IN-08: In light yellow. Two-party vote for Bayh: 44.35%. Rating: Likely R.
IN-09: In purple. Two-party vote for Bayh: 39.18%. Rating: Safe R.

For the current (actual) Indiana map, see here and here -click on the districts for info about them.

HRC got 47.50% of the two-party vote in my IN-01 and Obama in 2012 got 54.70%. In real life, HRC got 56.57% of the two-party vote in the actual IN-01 in 2016 and Obama in 2012 got 62.01%. Obviously, my redistricting plan significantly improved the GOP’s position in Indiana’s first district while not reducing the utility of the rest of the GOP-held seats for the GOP. The reason my IN-01, which has very similar political demographics to the actual PA-17, is more acceptable in Indiana is because Indiana is a more Republican state. Thus, it would be an improvement over the present plan. But how does one measure that? How does one display that using this curve?

I do not see a way for the above curve to be useful in cases in which districts within each side’s are of heterogeneous partisanship.

After thinking about it, the best advice I can give in regards to measuring gerrymandering is to multiply each seat by the probability of victory of the favored party. The maximization of these seat equivalents should be the measure of the worth of a gerrymander. Relative to the actual Indiana House map, my redistricting plan increases the chance of a GOP victory in IN-01 by more than it increases the risk to the Republican-held districts. I thus consider it a better gerrymander for the GOP than the actual GOP plan.

Measuring the probability of victory of the favored party in each district should be fairly simple (I will do this in the third post of this series), and historical data should be used for this purpose. Of course, the most important variable for determining the probability of victory of a party in a district is the district’s presidential vote. If it’s lower for the favored party, the probability of victory for that party is, all else being equal, lower. Historical swing data for congressional districts for the past few cycles should also be brought into account.

So, the above is how to measure a partisan gerrymander.

III. A fair map

Of course, for every villain -the partisan gerrymander- there necessarily has to be a hero to compare it to -the fair map. Of course, the question now becomes: what is a fair map? Certainly, a fair map cannot be a map in which each district is representative of the state in an identical fashion. Otherwise, Massachusetts’s map would be called a fair map, and Tennessee’s map would be considered as biased toward the Democrats. Obviously, neither is the case. Ideally, a fair map should have its median seat (now it enters into play) be representative of the state, though this is obviously not important (as I’ve shown in the previous post, comparing the median seat to the state is not an important measure of gerrymandering).

Perhaps this could be a fair map for a 60% Democratic state:

Of course, when each district becomes 10% more Democratic, the state as a whole won’t become 10% more Democratic, because districts 1 and 2 in the above graph cannot get any more Democratic. Of course, Dem chance of victory can also be used in place of Dem vote share in the y-axis of the above graph.

IV. Issues with the Princeton Gerrymander Tests

The Princeton Election Consortium developed three tests for gerrymandering. I don’t think the current ones are especially valid or useful.

The first test is right out; Texas is penalized because of Will Hurd’s narrow victory, despite the fact he’s obviously going down in 2018. A streak of luck in closely-contested tossup races is obviously no sign of a gerrymander, it’s just a sign of a side’s better campaigning.

The second test is also barely useful; it only measures the partisanship of one seat, the median seat, and, as the authors of the site admit, isn’t very useful at all in safe states.

The third test is a votes-seats curve. To which extent the specific votes-seats curve used is accurate is debatable.

The biggest problem with the tests is that they use the House vote instead of the presidential vote. The problem with this is that this makes no sense due to candidate heterogeneity. Collin Peterson ain’t Keith Ellison. The presidential vote should be used instead of the House vote to account for such differences.

Explaining partisan gerrymandering

In the United States, both state legislative and congressional districts are designed by politicians. These politicians, especially in the Republican-controlled states (this has only been true since 2010 or so) tend to design districts to give a clear and consistent advantage to their party. The canonical example of this is North Carolina (current [smoother] district lines):

The x-axis indicates the district, the y-axis shows the two-party Democratic presidential vote. Given the highly sorted party system currently existing in the United States, the presidential vote functions as a very good proxy for House candidate vote and is more appropriate here than House vote since the candidates are the same in every district. As one can readily see, in North Carolina, Democrats are packed into only three out of thirteen congressional districts, even though they won over 48% of the two-party presidential vote in both 2016 and 2012. Only in 2011 were the lines redrawn to favor the Republicans (and they will continue to favor the Republicans for a long time), before, they favored the Democrats since the 1890s.

Given that not all gerrymanders are created equal, several ways have been proposed to measure this phenomenon.

I. Insufficiency of commonly used methods

One way has been to compare the difficulty of recapturing the majority of the seats relative to winning the majority of the two-party vote in a state by looking at the difference between the presidential vote of the median district and the statewide presidential vote. However, one of the most gerrymandered states by this measure is Tennessee, which is a 60%+ Republican state with two Democratic seats (out of nine total)! Massachusetts (a 60%+ Democratic state with the same number of seats as TN) does not have even a single Republican seat! Yet, nobody can seriously call Massachusetts gerrymandered against the Republicans. Tennessee’s creation of more Democratic seats almost necessitates a higher difference between the median seat and the state due to more Democratic voters having to be taken away from the state’s median seat into the Democratic-held seats. Thus, the median district approach cannot be used as a serious way to measure gerrymandering, as it only looks at one district-the median one.

Another way to measure gerrymandering has been some kind of way of comparing share of seats won by v. share votes cast for a party. This is also a very flawed method.

The problem with these approaches is that they cannot distinguish between this (super-weak Democratic gerrymander in an evenly tied state with ten districts):

and this (strong Democratic gerrymander in an evenly tied state with ten districts):

But pretends there are giant differences between this (super-weak Democratic gerrymander in an evenly tied state with ten districts):

and this (super-weak Republican gerrymander):

As you can see, the problem with this approach is that the most gerrymandered maps by this measure will inevitably be dummymanders -that is, maps in which the party drawing the districts is so thinly spread out, if the popular vote shifts uniformly to the opposing party just a little, the map will look identical to a gerrymander designed by the opposing party.

In any case, any gerrymander has to be judged on two criteria: seat maximization and safety. There is a direct trade-off between the two (as thus, for an evenly tied state in which any and all district boundaries are permitted):

Notice that the curve is bowed in. However, in practice, the difference between a two point and a ten point presidential win margin is worth much more in terms of a House member’s win probability than that between a thirty point and fifty point presidential margin. Given this, the top and the bottom portions of the y-axis should be compressed and the middle expanded. With the axis like this, the curve would be bowed out, and the point of most correct gerrymander be placed at the outermost point of the curve.

Such curves should be designed for every state in the union with a reasonable number of House districts to test for gerrymanders there. Generally, the optimal average win margin for a favored party (that is, one that does not waste votes, but still keeps seats reasonably safe) is probably around ten points for an evenly tied state. Any good gerrymander should be directly on the possibilities frontier (as my strong D gerrymander graph with red bars is), not within it (as my super-weak D gerrymander graph with red bars is).

Ideally, a gerrymander should have

  1. Zero seats flipping on the presidential level between elections outside wave years (on the logic that it is better to have a bird in the hand than two in the bush)
  2. Total unity between the House member’s party and the party of the presidential candidate that wins the district (the state that is easily farthest away from fulfilling this ideal is Minnesota; the closest states to this ideal are, as far as I can tell from a quick glance, Maine, Missouri, and North Carolina).
  3. Maximized number of seats given a state’s partisan lean

Goals 1. and 3. are obviously inconsistent if a state hugely changes partisanship between elections.

Pennsylvania failed all the criteria in 2016 (it had split districts both ways, seats flipped in presidential vote between 2012 and 2016, and obviously it didn’t maximize GOP seats in 2016, and probably not in 2012, either), but it’s a pretty clear GOP gerrymander regardless. Wisconsin blatantly failed all the criteria in 2016 and probably failed the third criterion (though not the second) in 2012, though it was a very clear pro-GOP gerrymander in 2012. Michigan and Ohio satisfied all the criteria in 2012, but did not satisfy the last criterion in 2016, due to the state changing partisan lean between those years and there being obvious Democratic seats in both states which could be removed in 2016. North Carolina clearly satisfied all the criteria in 2016, but had some disunity between House member’s and district presidential candidate victor’s party in 2012. Texas satisfied none of the criteria in either 2012 or 2016.