Finding Agreement With OVV

I found myself nodding along to much of OVV’s response to my criticism. In fact, a minor change to the correction they themselves propose gets us very similar estimates for the violent death rate in 2015.

Reading OVV’s response to my article on how to estimate Venezuela’s violent death rate, I was struck by the extent to which we agree on essential facts and opinions.

We agree, first, that the police (CICPC), the Ministry of the Interior and Justice, and the Prosecutor General’s office should publish their own counts of violent deaths, together with metadata that explains how those counts were constructed.

We agree, also, that, in the absence of such publications, we have to rely on estimates, and that both OVV’s number and mine are just that:estimates.

We agree about the discrepancy between our estimates of the violent death rate for 2015. This discrepancy does not stem from the method by which we generate predictions based on data available for prior years. Instead, it comes from the input data itself: In practice, if we were to apply our respective estimation techniques to the same data through 2013, we’d obtain similar estimates for 2014 and 2015.

Moreover, we apparently agree that the number OVV used for 2013 double-counted many violent deaths, creating the false impression of an increase in violence in that year—which then created artificially high estimates for 2014 and 2015. Their response acknowledges this mistake and proposes a correction:

After becoming aware of this situation, OVV repeated our calculations, separating the 17,962 cases reported by the journalist Ramírez—copied and used by OVV in the original prediction exercise—into cases of homicide and cases of police violence, using the historic average (1990–2010) of the relationship between the total number of cases of homicide and police violence, and the number of cases of police violence alone.

Just to be clear, let’s recall the nature of the problem with OVV’s original calculation. For the year 2013, instead of constructing their estimate by adding together homicides and victims of police violence, they added victims of police violence twice. That is, instead of adding A+B, they added A+B+B.

To correct, they subtract B (victims of police violence). After making this change, OVV obtained a revised estimate of 81 violent deaths per 100,000 in 2015 (about halfway between their original estimate of 90 and my estimate of 68.5).

But what does “subtract B” actually mean in this context?

To arrive at the B to subtract, OVV used an average of the percent of police killings and homicides that were police killings between 1990 and 2010 (i.e., B/(A+B)).

That would make sense if the percent were stable. But it isn’t. OVV’s own data from the CICPC—which, through 2010, is the same data published in Table 1 here—indicates that the proportion has been rising over time, from 11% in 1990 to 21% in 2010. Using an average, in that case, distorts their correction.

On the basis of their own data, then, OVV’s correction didn’t go far enough: by starting with Ramírez’s total number of homicide cases and police violence cases (A+B) and then subtracting an estimate of police violence cases “using the historic average (1990-2010)”, OVV under-corrected.

What if, instead, OVV had corrected by subtracting from Ramírez’s total their own estimate of the number of police violence cases in 2013?

To my mind, that would be the logical thing to do, although OVV disagrees.Instead, they subtract an estimate of B based on historic data (1990-2010)—a figure that is lower than their own 2013 estimate of B. Why? They don’t say.

If they’d subtracted their own 2013 estimate of B, and then applied their own model for creating forecasts, their resulting estimate for 2015 would have fallen between 70 and 75 per 100,000—within the prediction interval from the estimate I created using a different approach.

This constitutes major progress. OVV’s response acknowledges a flaw in the input data they used for 2013 and then proposes a way to correct that flaw. If, instead of their proposed correction, they had implemented a similar—and, to my mind, more natural—change, then, using their own model, voilà! Their estimate looks very much like mine, and like that of the police (CICPC).

What’s more, OVV and I agree not only about my critique of their analysis, but also about their critique of mine. In their response, they point out—as I noted in my original article—that both (a) public announcements by ministers as well as (b) Annual Reports from MPPRIJ have, at times, stated “homicide” numbers higher than those in my data from CICPC (at other times, these same sources have given lower numbers). How, OVV asks, could I put any stock in numbers lower than those acknowledged by the government?

Good question. The answer lies in those scare quotes around homicide. When the Minister of the Interior or the Prosecutor General goes on TV to announce that there were so many “homicides” in a given year, they don’t explain exactly what they mean by homicide: homicide in the colloquial sense, meaning deliberate killing of another person, which could include killings by police officers? Homicide in the CICPC sense, which excludes killings by police officers? Homicide in the legal sense, which would include some but not all killings by police officers? Without these definitions, it’s close to meaningless to compare, say, Nestor Reverol’s 16,072 “homicides” in 2012 to CICPC’s count of homicide cases.

In a second criticism of my analysis, OVV pointed out that I credit two different figures for the violent death rate in 2015: my estimate of 68.5 per 100,000; and CICPC’s own number, which was 73.5 per 100,000. Why, OVV quite reasonably asks, would I put more stock in my estimate than in the violent death rate created by the police agency itself?

The short answer is that I don’t; the CICPC number of 73.5 violent deaths per 100,000 may well sit closer to the “true” violent death rate than my estimate, and I would very much like to understand why CICPC registers more victims than the Ministry of Health (if you know, please write to me). One advantage of my estimate, for now, is that it can be used to describe trends over time (because there are comparable numbers for previous years), while CICPC only very recently began counting victims of violent death.

Regardless, CICPC’s violent death rate (73.5) falls within the confidence interval from my analysis (62, 75). OVV’s original figure (90) did not. In their response, as in their comments on the draft of my original article, they acknowledge that the 90 per 100,000 figure was the product of a mistake. We can, then, uncontroversially set that number aside, and move forward with the difficult task of measuring violence in Venezuela.