Was the Colombian plebiscite result foreseeable?

 Oct 10, 2016 

Today, a week after a result that many international observers fail to understand, Colombians have seen numerous analyses explaining the results. From my perspective, some of the best examples, all of them in Spanish, are those by Leopoldo Fergusson and Carlos Molina and the analysis made by the team of “La Silla Vacía”. This post will present some spatial analysis that might allow us to further understand the vote and will complement the results just quoted by exploring how the vote relates to other variables. I expect there will be two audiences for this post. For this reason, sections (i) to (v) of this post include an English summary of how the vote divides territorially and is related to other variables (which is covered in the Spanish posts listed above). Those already familiar with this material can skip ahead to the analysis on age and urban functionality, which is what I consider to be my original contribution. As a side note, I use urban functionality –an original variable that I have created– because I think it does a better job describing the urban rural divide in Colombia than the variables that are traditionally used.

Before I proceed with my analysis, I will share a couple more details in order to provide further context. First, while I was a supporter of the yes vote, I do not intend to use this analysis to land points for or against the results of the election. In fact, although I am disappointed with the results, I confess I am not entirely surprised. However, the purpose of this post is purely to contribute to our understanding of the plebiscite. Additionally, in this post I am exploring the relationship between variables, but by no means suggesting that these methods imply causal mechanisms. Another technical point that deserves clarification is that I am using municipalities indistinctively to refer to municipalities, districts, and non-municipalized areas. This is a minor detail many analysts omit but is relevant to explain missing data for some administrative units. 

Another purpose of this post is to make these data available for others to run their own analyses and draw their own conclusions. And here I must credit my friend, Christian J. Meyer, who is responsible for retrieving these data. The code that he created to extract data from the Colombian National Registry is available for all to see, use, and share at GitHub. Additional data for this analysis comes from data I have collected for my own research from the Colombia National Planning Department and Colombia’s National Statistics Department. You can find these data at http://www.jctaborda.com/data. Finally, this page was designed to be viewed on a desktop or laptop, so you might miss information if you view it on a mobile device. 

What others have shown, but I present in English:
(i) The General Trends

Let us start by analyzing the total results and how close they were. The difference between the yes and the no vote is only 53,894 votes out of 13,066,047 votes cast. How small is the difference? It is only 0.4 percent of the total votes. For the sake of comparison, 53,894 votes are less than the 86,243 blank ballots that were cast. It is also smaller than the 170,946 votes that were declared null because they were incorrectly marked.

This small difference between the yes and no vote totals can also be explained by the fact that there was so little time between when the plebiscite was announced to the actual voting day. Most people had to vote in the same polling stations where they registered two years ago (for those living in Colombia), or four years ago (for those living abroad). Having said that, and to bring the issues back into the context of the current election, only 31 out of 1,122 municipalities cast more than 53,894 votes.

The following map, which you should refer back to for the rest of this analysis, shows which vote won in each of the municipalities.

Map 1: Yes and no vote result per Municipality

Because the map might be a bit misleading due to the larger-size municipalities in the east and southeast of the country, it is important to clarify that the yes won in 577 of the municipalities, while the no won in 544.

The next map shows the results by including their relative relevance. Hence, it not only shows which vote won where, but it actually displays the number of votes cast for that option in that municipality.

Map 2: Vote size per Municipality

The darker shade of brown is we have similar number of votes for yes and no. Hence, when you see a blue circle around the brown dot, it means there were more no votes than yes votes. On the other hand, when you see a lighter shade of brown surrounding the brown circle it means there are more yes than no votes. Finally, to further facilitate understanding, I shaded the municipalities in the corresponding color of their outcome of the election.   

The following map shows how polarized the country is by showing places where each of the options won by more than a ten percent difference. As you can imagine in such a contested election, a ten percent difference is akin to a landslide.

Map 3: Yes or no win by more than 10% points per Municipality

Another critical trend in this election was abstention. The following map shows that low voting turnout was a constant in all municipalities.

Map 4: Abstention per Municipality

For such a significant election, low turnout was a common trend (especially in the Caribbean region), with most of the municipalities having more than 50% abstention. 

On the topic of abstention or electoral apathy, there are multiple explanations. Some argue that it happened because of voters’ fatigue. Others say Hurricane Mathew kept people away from the polls in the Caribbean region. Others say it was the result of an out-of-date electoral census. Still others claim there was a general feeling that no matter what happened, the yes vote was going to win. Whichever theory you ascribe to, the fact is that abstention or electoral apathy played a significant role in such a close election.

(ii) The Vote and the Victims

A common idea that trended after the results were public was that the most conflict-prone municipalities overwhelmingly voted to approve the peace agreement. However, the truth is that we can present data to both contradict or support this argument. There are different ways in which you can characterize violence from the conflict, but trying to identify an electoral trend from a particular conflict characterization is an oversimplification resulting from cherry-picking data. 

 

For example, when we look at data at the department level the NO vote won in Antioquia, Arauca, Caqueta, Casanare, Meta, and Norte de Santander. And these are some of the departments that suffered most from the conflict.

When we change our unit of analysis and zoom in on the municipalities that were most severely impacted by the conflict, on average they supported the yes vote. But again, this does not include all conflict-prone municipalities. This divide seems to be a bit more marked in those municipalities with conflict in the pacific coast and parts of the south, while the result does not hold for the Llanos, Catatumbo-Bari, Magdalena-Medio or the Caucan boot regions.

 

Map 5: Yes and no results by Department

Map 6: Yes and no results and Risk of Conflict

Map 7: Yes and no results and Risk of Conflict, with votes

Statistically speaking, the regions that have intermediate or high risk of violence according to the PARES score, which is one of the most reliable indicators to measure risk of conflict from FARC, were more prone to vote yes than those that were not affected by the conflict.

Perhaps, the idea of the victims supporting the agreement was magnified in the news, because most of the towns that experienced the harshest massacres and acts of violence by FARC voted to support the agreement (table 2).

However, this is anecdotal evidence. I do not wish to undermine any victims, but presenting this table and omitting other municipalities that suffered intensely from conflict, like those in the Llanos region, is disrespectful to all victims. Furthermore, neglecting to comment on voter turnout is also misleading. In the end, it is hard to draw any conclusion about how victims were represented or misrepresented by the vote, in part because we should not fall into the typical mistake of thinking all victims are the same.

This has been especially dramatic for the international audience since most of the foreign media picked up this generalization (see for example NY Times Op-Ed "The Victims of War are Sick of War"). This is a misleading generalization. The truth is it was a close vote, and victimization per municipality or risk of conflict do not seem to be the best estimator to predict the outcome. Using selective data to build a narrative, might be newsworthy but might not provide the most accurate understanding of a complicated election. 

 

(iii) The Vote and the Urban and Rural Divide

 

Another common trend was arguing that the yes vote was eminently rural and the no vote was eminently urban. This argument made by many analysts and pundits is factually inaccurate. The yes vote won in Bogota, Cali, Barranquilla, and Cartagena which are four of the five largest cities in the country. In fact, the yes vote in these cities represents 31% of the total votes for yes. 

 

Having said that, on average, municipalities with high levels of rurality tended to favor the yes.

 

The issue is that indicators of rurality that have been used to categorize municipalities are problematic - as I explain in detail in the final section of this entry - and this is why we should analyze the results between rurality and the yes vote with caution.

Map 8: Yes and no results and population living in rural areas

Map 9: Yes and no results and Rurality Index

Map Matrix 2: Yes and no results and their relation with Rurality Index

Map Matrix 1: Yes and no results and their relation with population living in rural areas

How to read a Map Matrix: To understand a Map Matrix you need to remember that it shows the relation between two variables. For example, in this case, it is the relationship between the Rurality Index (on the vertical axis) and the yes and no votes (on the horizontal axis).  

The top row of the map shows higher levels of rurality according to the Rurality Index, while the bottom row of the matrix shows the lower levels of rurality. Similarly, the columns represent different values of our yes or no vote. In this example, the left column is where the no vote won and the right column is where the yes won.

 

Taking into account this logic, the top right map shows municipalities that have high rurality where the yes vote won. In the opposite corner, the municipalities depicted in the bottom left corner of the matrix are those with low rurality where the no vote won.  

Please note the areas highlighted in yellow have no data for the variable. 

As with the victims’ analysis, it is impossible to draw definite conclusions about this relationship. The relationship between variables is not as clear with the cabecera/resto indicator (the standard urban/rural metric used by DANE), but is easier to see with the Rurality Index. When using regression analyses we observe that the divide is statistically significant for the Rurality Index, but not for the cabecera/resto indicator. However, we should not extrapolate from those correlations and assume this is the explanatory variable that determined the election outcome.

(iv) The Vote and Poverty

 

The last common trend shown in multiple analyses is a potential relationship between the results and poverty. Again, this relationship that is statistically significant. On average, municipalities with higher poverty indicators were more likely to vote for yes. As with the rural urban divide, we are assuming a correlation implies causation and might be extrapolating from very simplistic analyses.

Map 10: Yes and no results and Unmet Basic Needs Index

Map Matrix 3: Yes and no results and their relation with Unmet Basic Needs Index

Map 11: Yes and no results and Multidimensional Poverty Index

Map Matrix 4: Yes and no results and their relation with Multidimensional Poverty Index

(v) The Vote, Santos, and Uribe (Zuluaga)

Perhaps the most interesting analysis is by Fergusson and Molina and looks at how the yes vote reflects the country divide between President Santos and former President Uribe, whose 2014 candidate was Oscar Ivan Zuluaga. The next maps summarize the correlation of the plebiscite results and the 2014 presidential election. For more on this analysis, I encourage you to go to their website. I hope this brief display of their results sparks your interest. 

Map 11: Yes and no results and 2014 Presidential election runoff

Map Matrix 5: Yes and no results and their relation with Santos and Zuluaga results in 2014

Age and urban functionality: the two variables that were left behind

To conclude this entry, I wish to contribute by analyzing two additional variables. One is age and the other one is urban functionality.

 

Age

I am convinced that age was another key divide between the yes and no camps. To my knowledge, no analyses have focused on age as an explanatory variable.  It is not an obvious analysis to conduct because Colombian voting data fails to include age of voters. However, population data from DANE do provide estimates on age groups per municipality. The hypothesis would be that municipalities where the leading age groups are older are more likely to vote no in comparison to municipalities that have younger populations. For all municipalities in Colombia the largest age group is between 27 and 59 years. However, in 23% of municipalities the second largest age group is citizens over 60 years. This hints that these municipalities are older, so by including this variable as a dummy we can try to understand how age affected the results.  This analysis can be seen in the following maps and table:

 

Map 12: Yes and no results and age groups per Municipality

I hope these simple results, that hint toward a relationship between older municipalities and a no vote, encourage future research on the topic. It would be especially interesting if future studies go beyond the standardized fact that older populations tend to be more conservative and additionally explore how different age groups might have forged their opinion based on their own lived experiences. 

 

Urban functionality

As already discussed, some analyses claim distance from large cities is a good predictor for the election results. However, in a country with such complex geography as Colombia, distance can be very misleading. Thirty kilometers might not seem very far, but if you add elevation or the lack of roads, those 30 kilometers become a long trip to the closest urban area. Similarly, it is incorrect to equate administrative features (being the capital of a state) to being part of the national urban system. This is particularly true for remote capitals that are not linked to the national road network, like Leticia, Inirida, or Mitú.

For this reason, using GIS network analysis and controlling by elevation, I calculated an urban functionality variable to evaluate the urban functionality of municipalities with respect to large metropolitan areas. The following maps and table share how this variable predicts the election results.

Map 13: Yes and no results and urban functionality per Municipality

In municipalities that have low urban functionality (the excluded group) the average support for the referendum was 55.8 %. Whereas, in municipalities with high urban functionalities, meaning those that are part of or close to – with proper road networks –  metropolitan areas, the average support for the referendum was 47.7 %. This result is both economically and statistically significant.  

 

Conclusion:

The objective of this post was to raise awareness to an international audience about some misleading messages that have arisen in the aftermath of the plebiscite. For example, stating that victims were betrayed by the rest of Colombians is factually inaccurate. It also leads to the inaccurate generalization that all victims are the same. Similarly, analyses that focus only on rurality or poverty fail by suggesting there is a key variable that distinguishes voters and explains the outcome of the election. If anything, the only variable that provides undisputed results is politics. The vote, and the country, is divided by support to President Santos or former President Uribe.

 

My analysis also shows that age might play a role in the divided vote, which is consistent with the political trend just depicted. The other relevant variable that determined the outcome of the plebiscite was abstention, which was particularly low even by Colombian standards. Whether this happened as a result of the inadequate electoral census, Hurricane Mathew, or the generalized belief that the yes victory was inevitable is irrelevant. 

Knowing these details, was it possible to predict the outcome of the plebiscite? I argue that the dynamics of previous elections provide valid hints. It was not surprising to see that there was a geographic and demographic divide. It was also predictable to see a disputed election in a country as polarized as Colombia. These results confirm a trend of political polarization, and how political communication and access to targeted information is leading to more polarized societies. The other determinant variable was the low voter turnout. The Colombian plebiscite is just another reminder that voting matters.  

 

Finally, an underlying intention of this post was to treat the data carefully. It is easy to show data of the result and bend it to strengthen our arguments. In fact, this is a worrying trend in the way we do politics. I believe you are entitled to have your own opinion, and you can choose to favor it over the other. But this does not entitle you to create narratives by cherry-picking the data that best fit your argument.