Escobari and Hoover faux a triple distinction and make issues worse

That is the thirteenth in a collection of weblog posts masking a report by Diego Escobari and Gary Hoover on the 2019 presidential election in Bolivia. Their conclusions don’t stand as much as scrutiny, as we observe in our report nickels earlier than dimes. Right here we increase on numerous claims and conclusions that Escobari and Hoover make of their article. Hyperlinks to posts: half One, Second half, Half three, half 4, half 5, half six, half seven, half eight, half 9,half ten, half elevenand half twelve.

within the final contributionwe discovered that – according to believable benign explanations of the election outcomes – the relief of the parallel tendencies assumption helped clarify Evo Morales’ first-round victory in 2019. Even when one interprets the in any other case inexplicable surge in help for Morales as a “cheat,” the impact is just too small to be of any political significance. With the worst interpretation conceivable, there was dishonest within the 2019 election, however it was pointless as a result of Morales would have gained anyway.

We have now additionally seen that Escobari and Hoover then radically reformulated their evaluation. They deserted the road of inquiry the place the TSE announcement “created a pure experiment.” Of their place, Escobari and Hoover took apart any benign explanations for non-parallel tendencies and argued that the existence of non-parallel tendencies itself serves as proof of fraud in Morales’ favour.

This shift in focus smacks of post-hoc reasoning; it’s very misplaced. Certainly, Escobari and Hoover chew the apple as soon as extra with “difference-in-difference-in-difference” fashions. The thought is that if there’s a double distinction in MAS-CC voting margin, a few of it might be a benign reflection of things that create a double distinction in smaller celebration voting margins.

This reasoning has some validity, as voters within the 2016 referendum on Movimiento Tercer Sistema (MTS) voted greater than 4:1 in favor of Bolivia Cube No (21F) in city areas, in comparison with lower than 2:1 in rural areas . Nonetheless, this can be a difference-in-difference-in-difference strategy in title solely. The identical 2016 outcomes are used because the baseline for each the big and small celebration double variations. With the “triple” distinction, the baselines cancel one another out. What Escobari and Hoover truly supply is a double distinction for the large events that makes use of the distinction between the smaller events as a foundation. That is evident within the outcomes of Desk 1omitting all 2016 knowledge from the evaluation has no impact on the estimate.

Desk 1
Outcomes of the “triple” distinction mannequin

Full date

Date 2019 solely

Full date

Date 2019 solely

(1)

(2)

(3)

(4)

SHUTDOWN x TREAT x Y2019

16.26

(0.634)

16.26

(0.647)

SWITCH OFF x TREAT

*

16.26

(0.634)

*

16.26

(0.660)

SHUTDOWN for 2019

-1:26 p.m

(0.595)

-1.25 p.m

(0.607)

DECOMMISSION

13.77

(0.624)

0.511

(0.091)

2,616

(0.166)

-8,018

(0.329)

TREATMENT for 2019

10.95

(0.253)

10.95

(0.258)

TO 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)

Mounted Results

district

Sure

Sure

observations

138,164

69.102

138,164

69.102

R2

0.037

0.061

0.755

0.555

* The given estimate is negligible

Sources: TSE, OEP and personal calculations.

Desk 1 reveals a triple (double) distinction of 16.26 share factors no matter whether or not 2016 knowledge is included and no matter whether or not geographic controls are included. Observe that other than the primary outcomes, the interpretations will not be the identical from one column to the following. For instance, the SHUTDOWN coefficient of 13.77 in column 1 is the rise in internet help for the referendum from early to late stations. In column 2, the SHUTDOWN coefficient is the rise in minor celebration margins from early to late stations.

Additionally see that the SHUTDOWN coefficient of Column 2 estimates the rise in MTS-21F margin from early to late stations. That is an identical to the sum of the SHUTDOWN and SHUTDOWN x Y2019 coefficients from Column 1. In fact, that 0.51 share level plus the estimated “triple” distinction equals 16.77 share factors—precisely our authentic single distinction for MAS-CC. The outcomes all make sense if the 2016 knowledge merely cancels out.

We are able to see inside numbers 1 the remaining double variations with the margins of the smaller celebration as a baseline.

numbers 1

The “triple” distinction boils right down to the distinction in distinction: with and with out geographic controls

Sources: TSE, OEP and personal calculations.

Right here, too, the small proportions of the smaller events nearly imply that there may be no main development in comparison with the big events. This leaves a big double distinction of greater than 16 share factors.

Now Escobari and Hoover declare their triple distinction is simply 2.9 share factors. We’re completely unable to breed comparable outcomes and imagine that is some form of misspecification in concept or follow.

You write:

That is the total triple distinction mannequin with some further, largely irrelevant controls. We are able to use the estimated coefficients to calculate their declared consequence

Clearly this isn’t a difference-in-difference-in-difference as it could require eight phrases that match the estimation mannequin. There are at the very least two potential explanations for this. One clarification is that Escobari and Hoover merely received their precise system improper. One other clarification is that they have been mischaracterized βDDD as a triple distinction.

I do know what’s βDDD if not a triple distinction? To make the notation a bit extra compact, we rewrite this as

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

An actual difference-in-difference-in-difference could be

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

In Desk 2let’s examine easy methods to calculate Yij primarily based on the statistical mannequin. The six phrases within the system don’t add up βDDDhowever to βDDDa1

Desk 2
Calculation of the “triple” distinction

Inter-station coefficients

station particular

[111] =

a1

+ α2

3

1

2

3

DDD

+Xijδ

ij

-[011] =

2

3

3

-Xijδ

ij

-[101] =

1

3

2

-Xijδ

ij

[001] =

a3

+Xijδ

ij

-[110] =

1

2

1

-Xijδ

ij

[010] =

a2

+Xijδ

ij

Y =

1

DDD

Fortunately it really works anyway, as a result of SHUTDOWN goes utterly into the station-level results. Escobari and Hoover do not actually admire a1however merely assumes that it’s zero.

Regardless, this triple-differences strategy – which boils right down to a distinction within the distinction between massive and small events – for estimating fraud is much more questionable, as the belief of parallel tendencies is much more short-lived. Whereas Escobari and Hoover once more admit non-parallel tendencies, in addition they reiterate the variations in tendencies that point out fraud in favor of Morales. Even when these outcomes might be reproduced, the interpretation continues to be simply as flawed because it got here from the difference-in-difference fashions.

Escobari and Hoover supply one other mannequin for testing for fraud: regression discontinuity. This isn’t the primary utility of the regression discontinuity to the Bolivian election knowledge. Though much less formally introduced, that is truly what Irfan Nooruddin contributed to the discredited OAS audit stories. Idrobo, Kronick, and Rodríguez tackle lots of the shortcomings of the OAS strategy of their article. We additionally tackle conceptual issues with Nooruddin’s strategy within the appendix of our earlier paper.

Somewhat than go into element, we merely observe the next: to the extent that there’s a important discontinuity within the 2019 election knowledge, we observe a virtually an identical discontinuity in 2016. That is further affirmation that the outcomes from Ende have been foreseeable in 2019.

numbers 2

If there’s a important discontinuity in 2019, it has been since 2016


Sources: TSE, OEP and personal calculations.