One more Bolivia election fraud fraud thing

Following up on our post on their article on Bolivia election fraud fraud, Nicolás Idrobo, Dorothy Kronick, Francisco Rodríguez write:

The final OAS report on Bolivia presents a graph even worse than the rdplot you checked out in your post last November (which was from their preliminary report).

Here’s their main piece of evidence for a fraud treatment effect at 95% of the vote was counted:

In our paper, we replicate this graph using Stata’s lpoly (separately on each side of the cutoff), for which the default is local constant regression. (We don’t know for sure what they did because they won’t share their code.)

But simply switching to local linear regression eliminates the appearance of a jump at 95%:

Of course, this second graph isn’t the best way to look at these data, either (we do something more appropriate in our paper). But it does illustrate how inattention to the most basic boundary bias problem can have huge political consequences.

I definitely think the world would be better off had regression discontinuity never been invented. Also, I hate when people don’t share their code.

15 thoughts on “One more Bolivia election fraud fraud thing

  1. I used to work at the OAS Department of Electoral Cooperation and Observation. I am not surprised at these mistakes and I am even less surprised that they don’t publish their code.

    • As I recall, the only Democratic candidate who spoke out against what happened was Bernie Sanders, and he got flax for it. And, we should not forget the reaction to the overthrow of Zelaya in Honduras, on Clinton’s watch as Secretary of State. So, if it was not a coincidence, it was a bi-partisan non-coincidence.

    • Purely a coincidence.

      Just as the USA recognizing the new government in Venezuela within an hour or so hour after that semi-successful coup against Chávez back in 2002.

    • According to an LA Times article from January, Carlos Trujillo, the US ambassador to the OAS, ‘steered’ the OAS’s election-monitoring team to report widespread fraud. I doubt they needed much steering, however.

  2. > I definitely think the world would be better off had regression discontinuity never been invented.

    Clearly this is a question that should be empirically investigated and I know just the right technique. We need to plot the utility of the world against time and find whether there was a discontinuity at point where regression discontinuity was invented. I hear 1,000 degree polynomials are good models for fitting either side of the discontinuity.

  3. I don’t understand why you dislike regression discontinuity. The original version of RD you were just supposed to use a line. If people sticked to lines instead of N-degree cubic splines RD would be a fine tool for this kind of analysis.

    • Kg:

      I like RD when it’s done sensibly, but my problem with RD as a codified “thing” is that people often seem to have the attitude that they once they include the discontinuity variable in the regression, in some form, that they don’t need to worry about lack of balance or overlap in other pre-treatment variables. All the focus on the functional form for this one variable is a distraction from the big problem, which is when people just include THIS ONE VARIABLE. These are observational studies, and you have to worry about lack of balance and overlap. The fact that there happens to be zero overlap on one particular predictor does not get you off the hook regarding all the others. But people act as if it does.

  4. I’m curious what you think about the display of data in RD plots. These plots show the raw data points, but people often display only binned averages. I prefer to see the raw data because in situations like this, even if there were a legitimate discontinuity, the relative noise in the outcome makes it seem less consequential. Often I think that the summary statistic versions of these plots make it harder to evaluate the substantive significance of any discontinuities.

  5. Their analysis looks like a joke. Everyone knows that local regression error “explodes” at the boundaries of the x domain. Even with the RD approach I feel the error intervals would have made these differences not significant.

Leave a Reply to Ethan Steinberg Cancel reply

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