2016 Taipei LY races revisited

It has been a while since I posted anything on this blog, but I actually have been mulling over a topic for a few months now. I want to write about the KMT chair election. Nope, just kidding. I want to write about the past rather than speculate about the future. I want to revisit the 2016 legislative elections. More specifically, I want to take another look at the DPP’s choice to cooperate with non-DPP candidates, especially in Taipei City. Did that strategy work well? Did DPP voters obediently vote for their party’s electoral ally? Did the non-DPP candidates bring in other voters that the DPP wouldn’t normally win? Should the DPP (or KMT) try this again in the future? I won’t provide definitive answers to any of these questions, but maybe we can unearth some suggestive evidence.

 

Caution: This section is about methodology. It is extremely boring. I suppose you could skip it, but I will judge you and think you are a bad person who probably believes all the fake news on the internet. If you can’t stand to read boring stuff, you aren’t educated. Also, what are you doing reading my blog.

Normally the way to attack this topic would be to look at individual-level vote choices from survey data. If we wanted to look at national-level vote choices, that is exactly what I would do. I’d go to the TEDS survey data and put together a table of how people voted in the presidential and district elections. Imagine if there were only two presidential candidates and two legislative candidates:

	LY: DPP	LY: KMT
Prez: DPP	p	1-p
Prez: KMT	1-q	q

Of the people who voted for the DPP presidential candidate, some percentage p also voted for the DPP legislative candidate, while the rest (1-p) defected to the KMT legislative candidate. Similarly, q of the KMT presidential voters were loyal while 1-q defected. We have two fairly large post-election surveys from TEDS for the 2016 election (one face-to-face, one telephone), so we could put together fairly good estimates for p and q.

However, I’m interested in Taipei 4, where the DPP supported PFP city councilor Huang Shan-shan. I’m also interested in Taipei 8, where the DPP supported independent (and former KMT and New Party) city councilor Lee Ching-yuan. For any given legislative district, we only have about 50 or 60 respondents from the two surveys combined. That isn’t enough to produce even a really bad estimate. National surveys simply aren’t much help; we would need a dedicated post-election survey for each legislative district. Unfortunately, I don’t have such data, and I am not aware that any such surveys were ever conducted. We will have to look elsewhere for evidence.

Instead of surveys, I look to the aggregate election results. The CEC has provided electoral returns from each precinct. In 2016, Taiwan had 15,582 precincts, so the average legislative district had 200 to 300 data points. The problem is that these are aggregated data, not individual-level data. These reports do not tell us how individual voters voted, which is what we really want to know.

This problem, which is known as ecological inference, has a long and not very distinguished history in political science. W.S. Robinson identified the basic problem in 1950. Aggregate data from the American states showed that states with higher levels of foreign born residents also had higher levels of literacy. In other words, the ecological inference should be that foreign-born residents had higher levels of literacy than native-born residents. However, it was well known that the opposite was actually true. The ecological inference was not just off by a little; the basic relationship was backward! From this, social scientists have learned to beware of what is called the ecological fallacy: inferring individual-level relationships from aggregate data can go horribly wrong.

Nonetheless, there are many instances in which all we have to work with is aggregate data. About twenty years ago, two of political science’s leading methodologists, Gary King (Harvard) and Chris Achen (then at Michigan, now at Princeton) published books about this problem. King’s book made the bigger splash. King put together a software program that would take the aggregate data, run thousands of simulations, and provide estimates for p and q. Unfortunately, there were a few limitations. King’s method worked well in a 2×2 case, but much less well with 2×3 or larger tables. His program might be able to manage 3×3, but as the number of rows or columns increased, the program often simply couldn’t converge to a solution. Second, his program didn’t handle covariates well. Imagine you wanted to control for the percentage of Hakka voters. His program could usually handle one covariate. However, if you wanted to control for the percentages of Hakkas, university graduates, farmers, and people who had moved into the district within the past five years, the program was likely to crash. I tried using King’s program a few times in grad school and gave up in frustration. (Another problem: Chinese language fonts make the interface go bananas, so it is hard for me to navigate.) That was probably fortunate for me. Not a whole lot of articles using his method were ever published, and I suspect many were rejected by reviewers who couldn’t be sure that they weren’t still running up against the ecological fallacy. At any rate, I thought I’d download the program and try some models for this blog post, but his software won’t run on my computer. Apparently, Gary King hasn’t bothered to update his software since 2003, which I take as a pretty good indication that he doesn’t consider it to have been a smashing success. But don’t feel too sorry for him, he’s got lots of other triumphs. Recently he has made headlines analyzing government funded posts on Chinese social media.

Unlike King, Chris Achen didn’t make any grandiose claims to have definitively solved the ecological inference problem. Achen’s book, co-authored with W. Phillips Shively (Minnesota), explored the old warhorse of ecological inference, the ecological regression, which is what I will use in this blog post. If you’ve taken any basic social science methodology course, you have probably studied ordinary least squares (OLS) regression. If you have a dependent variable Y and two independent variables X1 and X2, the normal equation is:

Y=b0 + b1X1 + b2X2 + e, where e is an error term (that we will henceforth ignore).

The normal model has a constant, b0. However, in an ecological regression, there is no constant. Effectively, the regression line is forced to go through the origin. This makes the equation:

Y = b1X1 + b2X2

In the above table, Y is the DPP’s district LY vote share, X1 is the DPP’s presidential vote share, and X2 is the KMT’s presidential vote share. B1 is an estimate of p, the percentage of DPP presidential voters who remained loyal and voted for the DPP legislative candidate. B2 is an estimate of (1-q), the percentage of KMT presidential voters who defected to the DPP in the legislative race.

Easy! Except it isn’t. Theoretically, b1 and b2 should never be less than zero or greater than one. Unfortunately, that happens quite frequently. You get results saying that 107% of DPP presidential voters voted for the DPP legislative candidate, and negative 11% of KMT presidential voters defected to the DPP legislative candidate. That doesn’t make any sense at all.

Achen and Shively suggest putting a squared term in the model to get a slightly less biased estimate. Unfortunately, this also makes it harder to interpret. Since the benefit is small and I’m looking for something quick and dirty, I’m going to just live with biased estimates.

Fortunately, Achen and Shively do tell us something about how the estimates are biased. First, loyalty rates are overestimated and defection rates are underrated. Second, if one election is more polarized than the other, estimates will commonly exceed the 0-1 range. Third, if there is a uniform swing, every ecological regression will have at least one estimate outside the 0-1 range (Achen and Shively 1995, 85). Let’s not worry too much about the second and third points, which simply suggest that it is nearly impossible to avoid having some estimates outside the 0-1 range. Unlike surveys, getting more data will not necessarily result in a better estimate. However, the first point is very important to keep in mind. The results I show you will probably overestimate the prevalence of straight-ticket voting for the KMT and DPP. When you see an estimate for straight-ticket voting, you should mentally insert “at most” in front of that number.

Achen and Shively have not solved the problem of the ecological fallacy. Strictly speaking, ecological regressions should be seen as evidence of aggregate-level trends rather than individual-level trends. As in most empirical inquiries, the best way to avoid falling victim to an unwarranted inference is a thorough knowledge of the facts and a robust theory explaining how the pieces fit together. It is usually a mistake to think the facts can speak for themselves, and this is more true than normal in the case of ecological inference.

 

[Digression: I first met Chris Achen about 20 years ago, when he spent a few months as a visitor at the Election Study Center. In fact, I have a vague memory of him teaching me how to do an ecological regression to estimate the incidence of split ticket voting in the 1997 Chiayi City mayoral election and the concurrent LY by-election. Chris is a wonderfully sunny and generous person; we quickly dubbed him 阿陳 (A-Chen). He has regularly returned to Taiwan over the past two decades, and he has trained several prominent Taiwanese political scientists. You might be most familiar with Hsu Yung-ming, who is now a New Power Party legislator. He is also one of the driving forces behind The Taiwan Voter, which will be the authoritative work on voting behavior in Taiwan for the next generation and the go-to source for anyone wishing to understand identity politics. (It will be published by University of Michigan Press later this year and is co-edited by Chris Achen and T.Y. Wang.) Chris is an awesome dude; everyone loves Chris.]

 

To recap the methodology section: I’m using ecological regressions to infer individual-level voting decisions from aggregate level data. This risks falling afoul of the ecological fallacy, and no reputable academic journal would publish such dodgy research. Nonetheless, the results are quite suggestive, and I believe they have some value. Moreover, we have fairly strong suspicions of how votes might have been distributed; we will want to see evidence that the results are more or less “reasonable.” Finally, remember that these results will overestimate straight-ticket voting and underestimate defections. When two people split their tickets in opposite directions, in the aggregate data it looks like they have both voted straight-ticket. There is a lot more churning under the surface. Proceed with caution. On the other hand, I wouldn’t have spent all this time working on this post if I didn’t think I wasn’t learning anything.

 

Substantive Results:

For each legislative district, I attempted to learn how the district candidates’ votes were produced by trying to figure out how the party list vote and the presidential vote translated into the district votes. For each important district candidate, I thus ran two regressions. The cases were precinct-level vote percentages. (Indigenous voters do not vote in normal legislative districts, so I only used precincts in which the number of eligible voters in the district election was at least 90% of the eligible voters in the presidential election.) For both regressions, the dependent variable was the vote share of the district candidate. In the first model, the independent variables were the vote shares of the party lists. However, the nine smallest parties kept producing crazy results, so I lumped them all together. Sorry. Next time win more votes. In the second model, the independent variables were the vote shares of the three presidential candidates. As you will see, the results of the two models generally tended to be consistent with each other, which should provide a small measure of confidence in the results.

 

Taipei 1

Taipei 1 is a good place to start since it was the only district in Taipei in which a KMT candidate faced a DPP candidate. The presidential and party list votes showed that this formerly blue district had turned light green:

Tsai	55.3	DPP	42.5
Chu	34.8	TSU	2.1
Soong	9.9	NPP	6.2
		KMT	27.2
		PFP	5.1
		New	6.5
		MKT	2.1
		F&H	2.2
		G/S	3.2
		Nine	3.0

There were five total candidates, but two got negligible votes and the third candidate got just a hair over 5%. It was fairly close to a straight KMT-DPP contest.

丁守中	Ting	KMT	82649	43.77%
吳思瑤	Wu	DPP	95951	50.81%	*
黃清原	Huang	台灣獨立黨	379	0.20%
王靜亞	Wang	MKT	9480	5.02%
吳忠錚	Wu	健保免費連線	348	0.18%

The overall totals are fairly similar across the three elections. However, this will not necessarily imply massive straight-ticket voting. If there are variations from precinct to precinct, the ecological regressions might show significant split-ticket voting. This could result if a candidate has a geographical stronghold, has done large and effective amounts of constituency service, or has some other cross-party appeal.

So what do the ecological regressions tell us? I ran regressions for the top three candidates:

	DPP	KMT	MKT
DPP	.985	.008	.008
TSU	1.050	.061	-.139
NPP	.712	.102	.173
KMT	-.085	1.071	.016
PFP	.031	.713	.240
New	.138	.928	-.063
MKT	.015	-.107	1.074
F&H	.023	.961	-.010
G/S	.827	.147	-.003
Nine	.327	.531	.116
.
Tsai	.949	.033	.016
Chu	-.041	1.050	-.009
Soong	-.026	.550	.446

These results show a fairly straightforward party to party fight. Of the voters who voted for the DPP list, 98.5% also voted for Wu Si-yao. Likewise, these results show that 107.1% of the people who voted for the KMT list voted for Ting Shou-chung. Technically speaking that isn’t possible, so we should probably state the results much more imprecisely: Almost all of the DPP and KMT list voters also voted for that party’s district candidate. Supporters of the two smaller parties in the green camp, the TSU and NPP, mostly followed suit, though the number is a bit lower for the NPP. On the blue side, New Party list voters mostly supported Ting. However, the fourth blue camp party, the MKT, had its own district candidate. The regression results show that MKT list voters overwhelmingly voted for the MKT district candidate. The PFP and MKT allied on the presidential ticket. However, in the district race, PFP voters mostly supported the KMT district candidate rather than the MKT candidate. Of the two non-aligned parties, the Green/SDP Alliance voters mostly opted for Wu, while the Faith & Hope League overwhelmingly supported Ting.

I was particularly heartened to see that latter result. F&H is led by conservative Christians whose main demand is to stop marriage equality. Ting Shou-chung has been one of the strongest voices against marriage equality. It makes perfect sense to see overwhelming F&H support for Ting, though the regression models, which know nothing of the fight over marriage equality, would not have been predisposed to produce this result. Ecological regressions CAN impart knowledge!

Looking at the district race from the presidential perspective, the story is roughly the same. Almost all Tsai voters supported Wu; almost all Chu voters opted for Ting; and Soong voters split roughly half and half between Ting and the MKT candidate.

Remember, these coefficients overestimate party loyalty, so there was probably a bit more split-ticket voting than these results suggest. Still, Taipei 1 was pretty darn straightforward. I couldn’t have asked for a better example to set the stage for the other seven Taipei districts.

 

Taipei 2

Taipei 2 is the DPP’s strongest district in Taipei, and it is the only one that the DPP won in 2012. This year, the KMT basically gave up on this district and did not nominate its own candidate. Instead, it nominated a New Party city councilor. Nominating a candidate from a more extreme party seems a somewhat dubious strategy to appeal to the median voter in a green-leaning district. For this post, an obvious question is whether KMT supporters and other more moderate blue camp supporters all lined up behind the New Party candidate.

Here are the presidential and party list results:

Tsai	61.6	DPP	48.6
Chu	28.7	TSU	2.6
Soong	9.7	NPP	6.3
		KMT	23.4
		PFP	5.5
		New	4.9
		MKT	1.4
		F&H	1.8
		G/S	2.7
		Nine	2.8

The district race had seven candidates, but the top two candidates took over 95% of the total votes:

王銘宗	Wang	IND	1342	0.74%
陳民乾	Chen	台灣獨立黨	865	0.47%
吳俊德	Wu	F&H	3550	1.96%
林幸蓉	Lin	健保免費連線	1561	0.86%
潘懷宗	Pan	New	65967	36.42%
姚文智	Yao	DPP	107366	59.29%	*
陳建斌	Chen	自由台灣黨	433	0.23%

The ecological regressions:

	DPP	New	Faith
DPP	1.012	-.014	-.004
TSU	1.284	-.532	.067
NPP	.681	.209	.045
KMT	-.056	1.058	-.010
PFP	.228	.512	.037
New	-.062	1.033	.029
MKT	.236	.328	.237
F&H	.030	.510	.622
G/S	.457	.574	.083
Nine	.446	.551	-.044
.
Tsai	.979	.000	-.002
Chu	-.094	1.055	.050
Soong	.178	.629	.068

The regression coefficients for party lists are a little far out of the bounds, especially for some for some of the smaller parties. -53.2% of TSU voters supposedly voted for Pan Huai-tzong, and the coefficients for the three equations don’t add up to anything near one for the NPP, PFP, or MKT. Maybe we shouldn’t take these results too seriously. Instead, let’s look at the regression coefficients for the presidential race. With only a 3×3 table, these bottom coefficients place fewer demands on the data. The bottom set of coefficients show another fairly straightforward green-blue contest, in which both sides mostly stayed loyal. The major exception is with Soong voters, of whom 17.8% voted for Yao Wen-chih. So KMT supporters did not seem to defect, but a fair proportion of Soong voters did. I would be cautious about attributing this to an extremist New Party candidate, though. As we go through different districts in the country, you will see that coefficients for Soong voters vary significantly from district to district, and 17.8% is a bit high but not terribly extreme.

 

Taipei 3

Now we are getting to the messier and more interesting districts. In Taipei 3, Wayne Chiang Wan-an defeated the KMT incumbent Lo Shu-lei in the primary, thus avenging his father’s own defeat at Lo’s hands in the primary four years prior. The DPP made a mess of this district. It originally nominated city councilor Liang Wen-chieh, but the self-appointed moral authority Lin Yi-hsiung protested and Liang was forced to withdraw. For months, it was not clear who the main green camp candidate would be. The DPP finally agreed to throw its support behind Billy Pan, a physician running as an independent who styled himself as a second Ko Wen-je. (Ko, however, did not seem so interested in supporting Pan.) This was a contentious decision in the DPP. Pan was strongly supported by the Hsieh faction, but he had less support from the rest of the party. Many green camp voters were also attracted to Social Democrat Lee Yan-jung. However, she ran a lackluster campaign. Until the last few weeks, she seemed more interested in her ideals than in actually winning votes, leading her to do things such as print fewer leaflets (such nasty and unenvironmental waste!).

The district, which had previously been solidly (though not overwhelmingly) blue turned a light shade of green. Here are the presidential and party list results:

Tsai	52.1	DPP	39.7
Chu	37.5	TSU	2.1
Soong	10.4	NPP	6.0
		KMT	28.3
		PFP	5.7
		New	8.4
		MKT	1.2
		F&H	2.2
		G/S	3.5
		Nine	2.9

The district race had ten candidates, but only three mattered:

高士恩	Kao	大愛憲改聯盟	541	0.28%
潘建志	Pan	IND	73797	38.41%
林新凱	Lin	台灣獨立黨	794	0.41%
趙燕傑	Chao	IND	352	0.18%
邱正浩	Chiu	和平鴿聯盟黨	450	0.23%
陳科引	Chen	IND	1448	0.75%
李晏榕	Lee	G/S	23706	12.34%
李成嶽	Lee	軍公教聯盟黨	721	0.37%
黃麗香	Huang	IND	607	0.31%
蔣萬安	Chiang	KMT	89673	46.68%	*

Chiang Wan-an won the district, putting the KMT’s first family back into the political fray. However, he did not get a majority; the combined totals of Pan and Lee would have defeated Chiang. Was Chiang’s victory due to the green camp’s failure to concentrate its support on one candidate?

	IND	KMT	G/S
DPP	.867	.000	.125
TSU	1.041	-.080	.002
NPP	.466	.155	.360
KMT	.097	.878	.010
PFP	-.362	.915	.236
New	-.061	1.214	-.089
MKT	-.376	.570	.587
F&H	-.004	.612	.386
G/S	-.024	.264	.792
Nine	-.212	.932	.019
.
Tsai	.783	.020	.181
Chu	-.049	1.023	.044
Soong	-.052	.702	.123

The ecological regressions suggest that green camp supporters were not enthralled with Pan. Only 87% of DPP list voters supported Pan. 87% may sounds high, but in normal districts this figure was generally close to 100%. 87% is actually a disaster. Pan’s support was even lower among NPP voters, who gave almost as much support to Lee as to Pan. This ambivalent green camp support for the designated DPP ally is also evident in the presidential candidate regressions, where only 78% of Tsai voters supported Pan. In short, Pan failed to consolidate the green camp vote.

But wait, there’s more. Pan was an independent. When Ko Wen-je ran for mayor, we heard endless arguments that blue camp voters might be able to defect to Ko because he was not a DPP member. Did Pan eat into the blue camp vote? Or was Chiang able to consolidate all the blue votes?

The evidence here is mixed. The top set of regressions have some numbers that suggest there may have been significant numbers of voters who crossed the normal boundaries. Chiang only has 88% of KMT list voters, while nearly 10% of them voted for Pan. However, the bottom set of ecological regressions seems to indicate that Pan got almost no support from Chu or Soong voters. My guess is that the bottom numbers are probably closer to the actual result.

Both sets of regressions indicate that the blue camp lost some votes to the Social Democratic candidate. In the party list regressions, Lee took 24% of the PFP votes and 59% of the MKT votes. In the presidential regressions, she won 12% of Soong’s votes.

Overall, it looks to me as though Lee took votes from both camps, but she took significantly more from the green side. In the alternate world in which the DPP nominated Liang Wen-chieh and he consolidated almost all of the green camp support, it would have been a much closer race. The Chiang family had better remember to send a Christmas card to Lin Yi-hsiung.

 

Taipei 4

Taipei 4 might be the most interesting race of all; the rest of this post was mostly a by-product of a crazy idea that maybe I could figure out what happened in Taipei 4. In past years, this district had been nearly hopeless for the DPP. Several DPP city councilors were eager to run, but the party leaders refused to nominate any of them. Instead, Tsai Ing-wen doggedly insisted on cooperating with PFP city councilor Huang Shan-shan. I will not delve into the various strategic angles of this decision at this time. For this post, we simply note that it was a bit controversial for the DPP to collaborate with a PFP stalwart, even as the PFP chair ran against Tsai in the presidential race. Many traditional DPP supporters were unhappy at being instructed to vote for Huang Shan-shan. Without a major “true green” candidate in the race, minor party candidates volunteered to fill the void. TSU, NPP, and Social Democrat candidates all picked up significant numbers of votes. The top two candidates finished within 2% of each other, so a minor shift from any of the three minor candidates to Huang would have delivered the seat to her. Of course, as a PFP candidate, Huang presumably tapped into pools of support that normal green candidates could never hope to win.

At the national level, in 2016 the district shifted from a deep blue district to a tossup district.

Tsai	51.3	DPP	38.0
Chu	37.3	TSU	1.8
Soong	11.5	NPP	6.8
		KMT	28.9
		PFP	6.2
		New	7.8
		MKT	1.5
		F&H	2.4
		G/S	3.3
		Nine	3.3

The district race was extremely fragmented, with five relevant candidates:

何偉	Ho	IND	2497	1.16%
陳尚志	Chen	G/S	10278	4.78%
黃珊珊	Huang	PFP	85600	39.86%
李岳峰	Lee	和平鴿聯盟黨	251	0.11%
陳兆銘	Chen	台灣獨立黨	568	0.26%
李彥秀	Lee	KMT	89612	41.73%	*
蕭亞譚	Hsiao	TSU	13648	6.35%
林少馳	Lin	NPP	12246	5.70%

Here are the ecological regressions:

	PFP	KMT	TSU	NPP	G/S
DPP	.627	.121	.128	.077	.033
TSU	-.080	.760	.270	-.074	.123
NPP	.399	.081	.082	.228	.142
KMT	-.124	1.106	-.021	.005	.043
PFP	.721	.377	-.041	.009	-.056
New	.736	.288	.010	-.008	-.048
MKT	.048	.047	.187	.353	.134
F&H	.750	.035	.202	-.131	.099
G/S	1.152	-.929	.134	.219	.397
Nine	.339	.538	-.001	.082	.001
.
Tsai	.501	.200	.131	.091	.061
Chu	.202	.739	.004	.007	.050
Soong	.573	.359	-.048	.065	-.019

There is exactly one coefficient in this table that looks normal. 111% of KMT voters supported Lee Yen-hsiu. Ok, that isn’t quite logical, but the idea the almost all of the KMT party list voters also voted for the KMT district candidate is the most normal thing about this table. Lee didn’t get a lot of the other blue camp votes, though. Nearly three-fourths of PFP and New Party voters opted for Huang. On the other hand, Lee did get 12% of DPP list voters. In addition to that 12%, a further quarter of DPP list voters plumped for one of the three minor party candidates, leaving a mere 63% for Huang. A very, very large number of DPP voters were unwilling to support a PFP nominee, regardless of Tsai Ing-wen’s entreaties.

It’s important to note that there are a lot of really strange results on this table that are probably simply wrong. If we take the numbers literally, most TSU list voters supported the KMT candidate (even though there was a TSU candidate in the race); the models can’t figure out what happened with nearly half of the MKT voters; and negative 92% the Green/Social Democrat voters supported Lee Yen-hsiu. We are probably demanding too much of the data.

Nonetheless, the regressions from the presidential election largely echo those from the party list election. Tsai’s voters were fragmented in the legislative race, with half voting for Huang and the other half splitting their votes among the other four candidates. Three fourths of Chu’s voters supported Lee, but 20% opted for Huang. Among Soong voters, more supported Huang than Lee.

What can we conclude? The DPP strategy partly succeeded and partly failed. Their ally took quite a few blue votes. If Huang had been able to add all the green votes to her blue votes, she would have won. However, the green camp clearly did not unite behind Huang. Hordes of green voters could not stomach supporting a candidate from the other side of the political spectrum. Judged narrowly by the district election outcome, the DPP’s alliance strategy was a failure. Not only did it not win, it also angered and frustrated a large number of sympathizers.

 

Taipei 5

Taipei 5 was one of those districts that the DPP erroneously deemed as “difficult.” It had lost by a lot in 2008, and by a fair amount in 2012. Nonetheless, this was a district that the DPP could hope to compete in with a mildly favorable partisan swing. In the tidal wave of 2016, this district turned a clear green.

Tsai	53.4	DPP	41.3
Chu	36.3	TSU	2.2
Soong	10.4	NPP	5.9
		KMT	28.3
		PFP	5.7
		New	7.1
		MKT	1.4
		F&H	1.9
		G/S	2.8
		Nine	3.3

The district race was essentially a two man race, featuring two of the most different people imaginable. On the on hand, the KMT featured a long-term incumbent, Lin Yu-fang. Lin was straight-laced, professional, conservative, and a committed Chinese nationalist. Facing him was Freddy, the death-metal rocker. In a result that almost certainly still bewilders Lin Yu-fang, Freddy won:

林郁方	Lin	KMT	76079	45.58%
李家幸	Lee	台灣獨立黨	885	0.53%
黃福卿	Huang	IND	587	0.35%
洪顯政	Hung	大愛憲改聯盟	478	0.28%
龔偉綸	Kung	IND	1710	1.02%
林昶佐	Lim	NPP	82650	49.52%	*
尤瑞敏	You	樹黨	4506	2.69%

How did Freddy win? Did the NPP open up some new constituency, or was it simply a DPP campaign clad in yellow? The ecological regressions:

	NPP	KMT	Tree
DPP	.971	-.014	.018
TSU	.845	.242	-.130
NPP	.953	.048	.057
KMT	-.038	1.021	.023
PFP	.116	.765	.042
New	-.043	1.037	.008
MKT	-.261	.935	.163
F&H	.308	.682	-.037
G/S	.749	.046	.146
Nine	.064	.611	.123
.
Tsai	.936	.019	.020
Chu	-.056	1.030	.021
Soong	.152	.699	.085

For the most part, this looks like a classic blue-green contest. Lin Yu-fang got almost all of the KMT, New Party, and MKT votes, while Freddy soaked up almost all of the DPP, TSU, and NPP votes. The only number that might be a bit surprising is the 11% of PFP voters who opted for Freddy. In doing this exercise, I’ve noticed again and again that there seems to be a minority of both NPP and PFP voters who will support the other one. During the campaign, I also remember social movement activists who seemed to differentiate between the PFP and the other blue parties. I suspect that there was a pool of voters that liked both the PFP and NPP (and perhaps also the Social Democrats). If so, it would make sense that this pool was prone to split its support among those parties.

 

Taipei 6

This is the forgotten race of Taipei. Taipei 6 has always been a deep-blue stronghold, and the DPP was fully justified in looking for an ally here. They opted to support Fan Yun, the well-known sociologist and activist. Unfortunately, she turned out to be an uninspired candidate. It turns out that elections and social movements are completely different animals.

Tsai	47.0	DPP	34.2
Chu	42.9	TSU	2.0
Soong	10.1	NPP	5.5
		KMT	30.3
		PFP	5.2
		New	11.1
		MKT	1.6
		F&H	2.9
		G/S	3.9
		Nine	3.2

This is the most highly educated district in Taiwan, and, perhaps not coincidentally, it always draws large numbers of useless candidates. This time there were twelve candidates, but only two got more than 4%. That isn’t to say this was a classic one-on-one race; the two top candidates combined for only 81% of the votes.

陳家宏	Chen	樹黨	1986	1.23%
范雲	Fan	G/S	56766	35.35%
龎維良	Pang	IND	2963	1.84%
趙衍慶	Chao	IND	3212	2.00%
林珍妤	Lin	台灣獨立黨	1328	0.82%
周芳如	Chou	IND	4120	2.56%
蔣慰慈	Chiang	IND	271	0.16%
鄭村棋	Cheng	人民民主陣線	4927	3.06%
曾獻瑩	Tseng	Faith	4826	3.00%
蔣乃辛	Chiang	KMT	74015	46.09%	*
古文發	Ku	大愛憲改聯盟	184	0.11%
吳旭智	Wu	MKT	5962	3.71%

Chiang Nai-hsin, the KMT incumbent, was held under 50%, but he won by a comfortable margin since Fan Yun only managed 35%. Did she fail to consolidate all the potential green camp voters? Was her poor showing the result of the same process as in Taipei 4, where DPP voters refused to obediently follow their party leaders’ instructions? Here are the ecological regressions for the top five candidates:

	G/S	KMT	MKT	Faith	Cheng
DPP	.857	-.038	.008	.018	.046
TSU	.733	.180	.012	.105	-.050
NPP	.542	.337	-.175	-.046	.061
KMT	-.066	.998	.066	-.019	-.006
PFP	-.017	.744	.026	.007	.153
New	.136	.828	.006	.012	.086
MKT	-1.172	.662	1.312	-.014	-.002
F&H	.367	-.074	-.028	.842	-.002
G/S	.963	-.034	-.045	.005	-.052
Nine	-.287	.361	.124	.122	-.027
.
Tsai	.788	.017	.004	.032	.034
Chu	.018	.871	.036	.036	.027
Soong	-.238	.782	.195	-.006	.029

These results indicate that both of the two main candidates had some difficulties consolidating potential votes, though it was nowhere near the scale of Taipei 4. Chiang was somewhat more successful, winning almost all of the KMT votes and 87% of Chu’s voters. Most DPP voters went along with Fan, but not all. 86% voted for her, and 79% of Tsai’s supporters also did so. Perhaps the most interesting numbers are from the NPP supporters. Since the NPP and Social Democrats were both born out of the Sunflower Movement, I expected that NPP voters would enthusiastically support Fan Yun. The regression coefficients say otherwise, indicating that she only won 54% of NPP list voters.

 

Taipei 7

Taipei 7 nearly turned out to be one of the biggest shocks of election night. This had always been a deep, deep blue district. I questioned the DPP’s decision not to nominate its own candidate in many districts, but not this one. Defeating incumbent Alex Fai with an ordinary DPP candidate seemed rather hopeless to me, and the DPP needed to try something new. It turned to Yang Shih-chiu, an erstwhile KMT city councilor. Yang had dropped out of the KMT to run as an independent, but he had longstanding blue credentials. The DPP’s hope was that Yang could keep some of his blue voters and that they could deliver all of their green voters. This was a reasonable strategy, though I had very little reason to believe it would succeed.

Shockingly, this deep blue district experienced such a big partisan swing that that the district seat came into play.

Tsai	49.6	DPP	37.2
Chu	39.6	TSU	1.9
Soong	10.8	NPP	6.0
		KMT	29.9
		PFP	5.6
		New	8.9
		MKT	1.5
		F&H	2.2
		G/S	3.4
		Nine	3.4

The district race turned out to be very close. This was even more surprising since there was a G/S candidate available to soak up any protest votes from green camp supporters who couldn’t stomach voting for Yang:

林文傑	Lin	IND	1063	0.64%
呂欣潔	Lu	G/S	17747	10.73%
詹益正	Chan	IND	625	0.37%
蘇承英	Su	和平鴿聯盟黨	689	0.41%
范揚律	Fan	大愛憲改聯盟	231	0.13%
林芷芬	Lin	台灣獨立黨	588	0.35%
費鴻泰	Fai	KMT	74455	45.04%	*
楊實秋	Yang	IND	69882	42.28%

Here are the ecological regressions for the top three:

	IND	KMT	G/S
DPP	.832	-.015	.161
TSU	1.162	-.004	-.152
NPP	.971	-.230	.368
KMT	-.072	1.062	-.019
PFP	.160	.544	.201
New	.258	.845	-.046
MKT	.306	.636	-.027
F&H	.612	.496	-.044
G/S	.138	.100	.686
Nine	-.008	.678	.145
.
Tsai	.789	.009	.182
Chu	.024	.985	-.007
Soong	.200	.518	.186

The picture these numbers paint is fairly clear. Fai consolidated almost all of the KMT votes, but Yang ate away at rest of the blue camp vote pool. If DPP voters had unified behind Yang, he would have won. However, they were unable to do this, and about 16% of DPP voters protested by voting for the G/S candidate. Probably the very same qualities that enabled Yang to win a number of blue votes were also the qualities that prevented him from getting all the green votes.

A quick note: Alex Fai has been a prominent opponent of marriage equality, so you might expect him to win all the Faith & Hope votes. However, Yang Shih-chiu is a Christian, something I know because most of his campaign ads featured a cross somewhere in the logo. It is thus reasonable to see that they split the F&H vote.

 

Taipei 8

Taipei 8 was almost exactly the same story as Taipei 7. This was a hopeless district for the DPP, so they cooperated with a KMT city councilor who had left the party to run for legislator. In this case, they cooperated with independent Lee Ching-yuan, who had previously been a member of the KMT, PFP, and New Party. As in Taipei 7, there was also a G/S candidate ready to soak up protest votes.

The big difference between Taipei 7 and Taipei 8 is that the latter is even deeper blue. There would be no election night suspense in this district. Even with the 2016 DPP tidal wave, this remained a solidly blue district.

Tsai	44.2	DPP	31.5
Chu	44.3	TSU	1.7
Soong	11.4	NPP	6.0
		KMT	32.1
		PFP	5.9
		New	10.4
		MKT	1.4
		F&H	3.2
		G/S	4.0
		Nine	3.8

The KMT incumbent easily won the district race:

李慶元	Lee	IND	60459	35.79%
賴樹聲	Lai	IND	1071	0.63%
苗博雅	Miao	G/S	21084	12.48%
賴士葆	Lai	KMT	83931	49.69%	*
陳如聖	Chen	台灣工黨	808	0.47%
方景鈞	Fang	IND	1540	0.91%

The ecological regressions:

	IND	KMT	G/S
DPP	.956	-.054	.077
TSU	.029	.302	.615
NPP	-.036	.420	.540
KMT	.124	.921	-.045
PFP	.584	.370	.021
New	-.372	1.220	.120
MKT	-.037	.650	.362
F&H	.015	.867	.113
G/S	.586	-.437	.884
Nine	-.004	.530	.366
.
Tsai	.828	-.022	.171
Chu	-.088	1.020	.066
Soong	.274	.477	.169

There are some differences from Taipei 7, but I think the overall story is roughly the same. In these models, the top set of regressions don’t seem quite right to me. I’m a little skeptical that almost zero TSU or NPP supporters voted for Lee Ching-yuan, but 30-40% of them voted for the KMT candidate. I think it may be wiser to look to the bottom set of results. Nearly all the Chu voters stayed loyal to the KMT, but roughly a quarter of Soong voters shifted to Lee. However, Lee could not win all the Tsai voters, as 17% of them defected to the G/S candidate.

 

Conclusions

We observed several patterns. (1) When the KMT or DPP nominated its own candidate, that candidate was generally able to consolidate the entire KMT or DPP vote as well as most of the allied small party vote. (2) When the KMT or DPP supported a candidate from an allied party (New Party in Taipei 2, NPP in Taipei 5), there wasn’t much difference from the cases in which they nominated their own candidates. (3) When the DPP supported an independent who had formerly been part of the blue camp, that candidate was able to win some blue votes. However, roughly a fifth of DPP voters refused to go along with the strategy, thus negating those gains. (4) When the DPP supported Huang Shan-shan, who was still a member of a party actively competing with and usually opposed to the DPP, there was a widespread rebellion within the party. Only about half of Tsai’s voters voted for Huang. The losses from unhappy green voters probably outweighed the gains from luring over blue voters.

Please remember that ecological regressions are a very shaky method. These results are not definitive in any way, and there is a possibility that they are entirely wrong. I don’t think they are generally misleading, but I might be wrong.