Escobari and Hoover forge a threesome and make things worse

This is the thirteenth article in a series of blog posts that examines Diego Escobari and Gary Hoover covering the 2019 presidential election in Bolivia. Their conclusions do not hold up to scrutiny, as we note in our Nickels Before Dimes report. Here, we expand on the different claims and conclusions made by Escobari and Hoover in their paper. Publication links: Part One, Part Two, Part Three, Part Four, Part Five, Part Six, Part Seven, Part Eight, Part Nine,Part Ten, Part ElevenAnd Part Twelve.

In the last post, we noted that – consistent with plausible benign interpretations of election results – relaxing the assumption of parallel trends helped explain Evo Morales’ victory in the first round of 2019. Even interpreting the unjustified increase in Morales’ support as “fraud”, the effect is very small. so that it is not politically significant. By offering the worst possible explanation, there was fraud in the 2019 elections, but it was unnecessary because Morales would have won anyway.

We also saw that Escobari and Hoover radically reformulated their analysis. They dropped the line of inquiry as the TSE advert “created a normal experience”. In its place, Escobari and Hoover postulated all benign explanations for non-parallel tendencies, arguing that the very existence of non-parallel tendencies serves as evidence of fraud in favor of Morales.

This shift in focus reeks of afterthought; It is totally out of place. In fact, Escobari and Hoover took another bite out of the apple with their “difference in difference in difference” models. The idea is that if there is a double difference in the MAS-CC voting margin, some of this may be a nice reflection of factors leading to a double difference in the voting margin for small parties.

There is some truth to this thinking, as voters chose Bolivia’s dice number (21F) in the 2016 referendum on Movimiento Tercer Sistema (MTS) more than 4:1 in urban areas compared to less than 2:1 in rural areas. However, this is the difference in difference in name only approach. The same 2016 results are used as a baseline for major and minor pairwise differences. So, in the “Triple” variation, the baselines cancel out entirely. What Escobari and Hoover actually offer is a dual variation of the major parties that use the minor difference as a baseline. This is evident in the results Table 1dropping all 2016 data from the analysis does not affect the estimate at all.

Table 1
Results of the “triple” divergence model

Complete data

2019 data only

Complete data

2019 data only

(1)

(2)

(3)

(4)

SHUTDOWN x TREAT x Y2019

16.26

(0.634)

16.26

(0.647)

off x treatment

*

16.26

(0.634)

*

16.26

(0.660)

Off x Y2019

-13.26

(0.595)

-13.25

(0.607)

close

13.77

(0.624)

0.511

(0.091)

2,616

(0.166)

-8,018

(0.329)

TREAT x Y2019

10.95

(0.253)

10.95

(0.258)

Treat

*

10.95

(0.253)

*

10.95

(0.263)

Y2016

0.101

(0.624)

0.090

(0.237)

Fixed

-3,173

(0.238)

-3,071

(0.036)

-1.383

(0.058)

-1,707

(0.129)

and fixed effects

Area

yes

yes

Notes

138,164

69102

138,164

69102

R2

0.037

0.061

0.755

0.555

* The reported estimate is too small

Sources: TSE accounts, OEP, and the author.

Table 1 shows a triple (twofold) difference of 16.26 percentage points, regardless of whether 2016 data are included, and regardless of whether geographic controls are included. Note that aside from the main results, the interpretations are not the same from column to column. For example, the closing coefficient of 13.77 in column 1 is the increase in net support for the referendum from early to late stations. In column 2, the closing coefficient is the increase in secondary limb margins from early to late stations.

See also that the off coefficient for column 2 estimates the increase in margin for MTS-21F from early to late stations. This is identical to the sum of the SHUTDOWN and SHUTDOWN x Y2019 coefficients for column 1. Of course, that’s 0.51 percentage point plus the estimated “triple” difference to 16.77 percentage points – exactly the original MAS-CC single difference. All results make sense if you simply cancel out the 2016 data.

may see in Figure 1 Residual double spreads with smaller tip margins as baseline.

Figure 1

“Triple” divergence reduces differences in differences: with or without geographical controls

Sources: TSE accounts, OEP, and the author.

Again, the small stocks that the minor parties got all claim to not have a significant trend compared to the major parties. That leaves a huge double digit difference of over 16 percentage points.

Now, Escobari and Hoover claim the three-way difference between them comes to just 2.9 percentage points. We are completely unable to reproduce similar results and believe this is a kind of mis-spec either in theory or practice.

they write:

This is the complete Three Differences model with a few additional, largely unrelated controls. We may use estimated coefficients to calculate the declared score

This is clearly not a variance-of-variance variance, as this would require eight terms, consistent with the estimation model. Now, there are at least two possible explanations for this. One explanation is that Escobari and Hoover simply planned to report their actual formula. Another explanation is that they mischaracterised BDDD as a trio.

so what BDDD If not the triple difference? To make the notation more compact, we rewrite it as

s = ([111]-[011]) – ([101]-[001]) – ([110]-[010])

The real difference will be in the difference in the difference

z = ([111]-[011]) – ([101]-[001]) – ([110]-[010]) + ([100]-[000])

in Table 2See how we count sIJ Based on the statistical model. The sum of the six terms in the formula is not equal BDDDbut for BDDDa1

Table 2
triple spread calculator

transactions across terminals

station specific

[111] =

a1

+ a2

+ a3

+ Dr1

+ Dr2

+ Dr3

+ bDDD

+XIJDr

+ nIJ

-[011] =

-a2

-a3

+ Dr3

-XIJDr

-nIJ

-[101] =

-a1

-a3

-Dr2

-XIJDr

-nIJ

[001] =

a3

+XIJDr

+ nIJ

-[110] =

-a1

-a2

-Dr1

-XIJDr

-nIJ

[010] =

a2

+XIJDr

+ nIJ

s =

-a1

+ bDDD

Fortunately, this works anyway, because SHUTDOWN is completely absorbed in station-level effects. Escobari and Hoover don’t actually appreciate it a1but simply assume it is zero.

Regardless, this three-difference approach—reduction to a difference in differences between major and minor parties—to fraud estimation is all the more questionable because the assumption of parallel trends is ephemeral. Although Escobari and Hoover again allow non-parallel directions, they also again point to the difference in directions as indicating fraud on Morales’ behalf. Even if these results could be reproduced, the interpretation would still be completely wrong as it was coming from difference-of-difference models.

Escobari and Hoover offer another scam-testing model: the regression stop. This is not the first application of regressive discontinuity to Bolivian election data. Although it has not been formally presented, this is in fact what Irfan Nooruddin contributed to the ill-fated OAS audit reports. Idrobo, Kronick, and Rodríguez address many of the shortcomings of the OAS approach in their paper. We also address conceptual issues with Noureddine’s approach in an appendix Our previous paper.

Rather than go into too much detail, we simply note this: To the extent that there is any meaningful discontinuity in the 2019 election data, we note an almost identical discontinuity in 2016. This is further confirmation that the results of late 2019 were predictable.

Figure 2

If there was a real outage in 2019, it has been going on since 2016


Sources: TSE accounts, OEP, and the author.

Related Posts

Leave a Reply

Your email address will not be published. Required fields are marked *