Introduction
At first, the idea that coups might be “good for democracy” sounds paradoxical. Yet recent transitions in Ecuador in 2003, Bangladesh in 2009 and Niger in 2011 all quickly followed military coups, matching an earlier pattern in Panama, Portugal, Bolivia, Thailand, and elsewhere. Indeed, recent empirical work argues that coups within autocracy raise the ensuing likelihood of democratization (
Miller, 2012,
2016;
Thyne and Powell, 2016;
Varol, 2012). In particular,
Thyne and Powell (2016) find that both coups and coup attempts predict democratization from 1950–2008.
In a recent reconsideration of the evidence,
Derpanop-oulos et al. (2016) (hereafter, DFGW) dispute the idea that coups predict democratization. Applying a new model to
Thyne and Powell’s (2016) coup data, they find that “the association between coups and democratization is statistically insignificant” both during and after the Cold War (
Derpanopoulos et al., 2016: 2). DFGW should be commended for an innovative analysis of the varied effects of coups; in particular, I do not challenge their well-defended conclusions that coups are often followed by autocratic regime changes and increased repression.
This paper instead critiques DFGW’s findings regarding coups and democratization. I show that their modeling approach, which adds fixed effects for each autocratic regime spell, severely biases their estimate and effectively removes most recent coups from the sample. I then reanalyze DFGW’s data with similar models that do not suffer from this bias. In every case, I recover a significantly positive effect of coups and coup attempts on democratization.
Besides contributing to an important substantive debate, this paper identifies an instructive and surprising source of bias. Researchers frequently overlook the possibility that adding too many controls can produce bias. Yet methodologists recognize that controls can cause “post-treatment bias” and “collider bias” (
Pearl, 2009). Recent scholarship also cautions that unit fixed effects may exacerbate rather than reduce omitted variable bias (
Middleton et al., 2016). DFGW’s model is another example of over-fitting producing bias, which can serve as a useful caution for empirical scholars.
Background
Although they differ on the underlying mechanisms, both
Miller (2012) and
Thyne and Powell (2016) find that coups raise the likelihood of later democratization.
1Miller (2012) focuses on the five years after violent executive turnovers, demonstrating a stronger effect at higher average income. In related work,
Miller (2016) examines all democratic transitions from 1800–2010 and finds that 52 had a successful coup as a primary causal factor.
Thyne and Powell (2016) find that experiencing a coup or coup attempt predicts democratization in a window from the current year through the following two years. To critique this finding, DFGW retain the same coup data (from
Powell and Thyne, 2011) and post-coup window, but use a different measure of democratization (from
Geddes et al., 2014) and model (discussed below). Following
Marinov and Goemans (2014), DFGW also distinguish between coups during and after the Cold War.
Figure 1 presents a simple descriptive breakdown using DFGW’s data, graphing the annual likelihood of democratization by post-coup status and period. The results are remarkably strong – experiencing a coup shifts the chances for democratization from 1.0% to 5.1% during the Cold War and from 2.9% to 24.3% during the post-Cold War era. Surprisingly, DFGW claim to find no effect of coups in either period. The next section explains how their model biases their results.
Problems with DFGW’s model
In a sample of autocracies from 1950–2015, DFGW use a linear model to predict democratization based on whether a coup occurred in the current or previous two years.
2 Separate models test coup attempts. Each model controls for year fixed effects, a cubic polynomial of regime duration and the log of leader duration.
The most consequential and unusual choice, however, is DFGW’s addition of “regime-case fixed effects.” Rather than a fixed effect for each country, they include one for each of 285 autocratic regime spells. For instance, there are separate fixed effects for Cuba 1952–1959 and Cuba post-1959. The authors claim this leads to “a within-regime comparison of what follows a coup, while conditioning-out all differences between autocratic regimes” (
Derpanopoulos et al., 2016: 3). Although this choice stems from a good motivation, it unfortunately produces severe bias and effectively restricts the sample to a specific type of coup.
The core problem with regime-case fixed effects is that coups typically produce a new autocratic regime. In DFGW’s sample, of 149 regimes facing a coup, only 52 (34.9%) survive to the following year. Another 17 (11.4%) democratize the same year. Thus, the majority of coups (53.7%) produce a new autocratic regime with its own fixed effect.
To see how this produces bias, consider
Figure 2. If coups predict democratization, then we should see several cases where a coup is followed quickly by democratization. This is pictured in panel (a): A coup occurs in Year 2 and the
post-coup variable (pictured in grey) covers Years 2 to 4, when democratization occurs. However, the democratization will be ignored if we include regime-case fixed effects and the coup produces a new autocratic regime. With these fixed effects, variables are tested
relative to the regime average. For Regime B, however, there is no variation in post-coup status. Thus, the
post-coup variable is measured as 0 and Regime B’s observations have no leverage on its coefficient. This follows from the same reason that independent variables that are constant within country cannot be tested in the presence of country fixed effects. Further, the effect of coups will be biased towards 0 because
post-coup in Year 2 does not lead to democratization that year.
The problem is worse for cases that democratize just after the two-year window, as shown in panel (b). Here, when Regime B democratizes in Year 5, this spuriously counts as a negative effect of coups on democratization. This is because a lower-than-regime-average value of post-coup in Year 5 coincides with democratization, generating a negative correlation. In fact, every case in which a coup produces a new autocratic regime will generate either no effect or a negatively biased effect of coups on democratization.
What could generate a positive effect in DFGW’s model? There are two possibilities. First, a coup could precede democratization in the same year, which occurs in 17 cases. However, five of these cases are ignored because every earlier year in the regime is also post-coup. Second, a coup in autocracy could lead to democratization without first generating a new autocratic regime, which occurs in only four cases. All are coups that shuffled leadership within a military junta. In total, 37 cases of democratization occurred within two years of a coup, but the regime-case fixed effects effectively remove 21 of them. In sum, DFGW’s estimates combine negatively biased results for regime-changing coups with results for a specific sub-sample of coups. Clearly, this cannot accurately represent whether coups predict democratization.
A reanalysis
I now reanalyze DFGW’s data, using models that hew closely to DFGW’s design but are not susceptible to the same bias. Specifically, I keep all elements of DFGW’s model, but alter the regime-case fixed effects. As in DFGW, I separately test coups during and post-Cold War era and cluster standard errors by the regime spell or its equivalent.
Figure 3 displays the results for successful coups (top) and attempted coups (bottom). I consider six total models. The first model replicates DFGW, confirming a null effect on democratization.
3 The second model is closest in spirit to DFGW’s original: It includes regime-case fixed effects, but extends the regime by two years following a coup. Thus, if a coup originally ended a regime in Year
t, the regime is redefined to end in Year
t + 2. By including the post-coup period in the same regime, this negates the source of bias. Further, it directly compares the post-coup likelihood of democratization to the preceding autocratic regime. At the same time, this design still finely controls for differences across regimes. In fact, the regime fixed effect is identical to DFGW’s original for 94.8% of the sample.
In the third model, regimes are considered continuing unless they fail without a coup. Again, this avoids the bias from coups producing regime failures. The fourth model uses country fixed effects, replicating a result that DFGW show in their online appendix. However, DFGW argue this is unsatisfactory as different regime types within the same country may have different propensities for liberalization. The fifth model therefore includes a distinct fixed effect for each autocratic regime type in each country (192 in total). The regime types are broken down into monarchies, military, party-based, and personalist dictatorships, using
Geddes et al. (2014). For instance, El Salvador gets separate fixed effects for its period as a military dictatorship and as a party-based dictatorship. Finally, the sixth model includes a separate country fixed effect for each 20-year period.
As seen in
Figure 3, post-Cold War era coups and coup attempts are significantly positive (
p < 0.05) for democratization in every altered model. Further, the effect sizes are quite large. Since this is a linear model, the coefficients represent shifts in probability. Thus, a post-Cold War era coup raises the likelihood of transition by 16–26%, consistent with
Figure 1. Cold War results are less robust, but two of five coup models are significant at the 0.1 level.
In further analysis, I extended the post-coup window from two previous years up to four and found similar results for successful coups, although coup attempts are no longer significant. I also ran the models removing coups and attempts within democracies from the
post-coup definition and found even stronger results
4 (see Figures A1 and A2 in the online appendix).