Introduction
In “Are coups good for democracy?” (
Derpanopoulos et al., 2016), we look at the political consequences of coups in autocracies. We find that, though democracies are sometimes established in the wake of coups, more often new authoritarian regimes emerge. Further, we show that coups frequently increase state-led violence, particularly following coups that lead to new autocratic regimes.
In a recent critique of our study,
Miller (2016) challenges one particular finding: that coups rarely lead to democratization. He argues that our use of a regime-case fixed-effects (FE) model to estimate the impact of coups on the likelihood of transitions to democracy drops a number of key observations from the sample, biasing our estimates. He emphasizes that when a country FE model is used instead, there is a positive relationship.
In our original analysis, we find that there is no association between post-Cold War coups and democratization in a regime-case FE model, but show in the online appendix to our published results (Figure A-2) that there is a positive association between post-Cold War coups and democratic transition when using a country FE model.
1Miller (2016) re-reports and expands on the country FE model of democratic transitions that we report in the appendix to our original analysis. In fact, the conclusions of
Miller (2016) are identical to those drawn from one set of tests we report in the original appendix. We thank the editors of
Research and Politics, however, for allowing us to further the discussion of the points articulated in
Miller (2016).
In the analysis that follows, we first offer background on the key issues that underlie Miller’s critique. We then re-state Miller’s critique in a slightly different way and demonstrate visually the differences between regime-case and country FE models in this application. In doing so, we focus on the post-Cold War period because Miller’s only point of disagreement about the consequences of coups pertains to the relationship between post-1989 coups and democratic transition.
Background
Miller’s (2016) main argument is that a country FE estimator is preferable to a regime-case FE estimator when modeling the effect of coups on the chance of democratic transition. Yet, all modeling choices have advantages and disadvantages, and the “optimal” approach is up to scholars to assess. It is for this reason that we presented the results of both in our study (the regime-case FE model in the main text and the country FE model in the appendix). Here, however, we provide a more in-depth discussion of the logic underlying both approaches and their strengths and drawbacks in light of the
Miller (2016) critique.
A FE estimator isolates the over-time variation in the data, allowing inferences that are not contaminated by unobserved unit (or cross-sectional) heterogeneity. When estimating FE models, researchers must determine the cross-sectional unit over which to isolate the time trends. In our original paper, we reported results from two sets of models that make different assumptions about the appropriate cross-sectional unit: the country and the regime-case.
A country FE model isolates variation over time within countries; that is, it uses information about movement over time in the independent and dependent variables to identify the estimated parameter of interest (in this application, the marginal effect of a recent coup on the probability of transition to a new autocratic regime or to democracy). This approach accounts for time-invariant heterogeneity across countries, such as colonial history and geographic endowments. When a country contains more than one autocratic regime – for example in the Democratic Republic of Congo (DRC), Guinea, or Thailand – a country FE estimator compares periods of no recent coup with periods of recent coup within the same country to inform an estimate of how recent coups influence regime instability.
In this application, a regime-case represents a continuous spell of years under the rule of a single autocratic regime. In some instances, regime-cases are preceded and succeeded by periods of democratic rule. For example, coups in Thailand upended democracy in 1991, 2006, and 2014, each time ushering in a new autocratic regime. In other instances, we see successive autocratic regime-cases during the period of analysis. Examples include the DRC (former Zaire), where a rebellion led by Laurent Kabila toppled the long-standing Mobutu regime in 1997, and Guinea, where a junior officer coup led by Dadis Camara ousted the Conté regime in 2008.
A regime-case FE model therefore enables a within-regime comparison of what follows a coup, while conditioning out all differences between autocratic regimes (e.g. level of development, how the regime seized power, the types of leaders the regime selects to rule). Thus, periods of
no recent coup are compared with periods of
recent coup within the same regime-case. Critically, this approach takes into account autocratic regime type, such as whether the incumbent regime is ruled by the military. This is crucial because of strong empirical evidence that coups are more frequent in military dictatorships and that military dictatorships are more likely, all else equal, to democratize (
Geddes et al., 2014).
Employing a regime-case fixed effects estimator implies that: (1) the baseline likelihood of a democratic transition and coup varies by the regime as well as by country; and (2) short-lived regimes may not have sufficient variation in the key explanatory variable (
recent coup) to inform the estimated parameter of interest. In this application, the latter feature of the regime-case fixed effect estimator has the effect, as
Miller (2016: 2) notes, of “effectively remov[ing] most recent coups from the sample.” While Miller focuses on the fact that the regime-case FE estimator drops some cases where the short-lived autocratic regime is always in a state of having had a recent coup, the substantive benefit of this design choice is that the baseline likelihoods of democratization and coups are not constrained to be the same for all regimes within the same country.
2
The issue that Miller raises regarding the bias of the regime-case FE modeling approach is largely not present when employing a country FE estimator in this application because, when there are multiple regimes within the post-Cold War period, at least one of them has a period where the
recent coup indicator is equal to 0.
3 There are, therefore, trade-offs in choosing between regime-case FE and country FE estimators. Our published analysis included both types of tests for interested readers to compare the differing results for democratic transition.
One way to summarize the
Miller (2016) critique is as follows: in some short-lived autocracies, the main explanatory variable (
recent coup) does not vary over time within the regime-case. For example, if a coup occurs in time
but the regime only lasts one year, then
recent coup will have a value of 1 for the entire period of regime rule. A regime-case FE estimator does not draw support from these cases to estimate the parameter of interest because it isolates the “within” variation, and these regimes have no “within-case” variation in the
recent coup indicator.
Consider a second way to illustrate the same point in a model specification that separates the “within” and “between” variation in the explanatory variables
4
In this equation,
is the mean value of the main explanatory variable and
can be either the country or the regime-case. Thus
is the deviation from the unit mean (i.e. the “within” variation) and
is the unit mean (i.e. the “between” variation). In our application, the parameter of interest is
, which is the estimate of the marginal effect of recent coup (
) on the probability of democratic transition.
The concern that Miller raises with respect to fixed effects models can be illustrated by noting that, for cases in which there is no “within” variation – those where the regime is always in a state of recent coup – the mean level of coups,
is 1. Thus
=1 for all
and
=1, while
=0 in these cases.
Note that, if there is more than one autocratic regime per country, then
is not necessarily equal to
. Thus while the regime-case mean for
recent coup may be 1, the country mean for the same variable may be less than 1.
5Figure 1 shows the distribution of regime-case
s and country
s for the post-1989 sample. The left plot shows the entire sample, while the right plot shows only regimes with at least one successful coup during this period. The gray bars represent the regime-case
s for
recent coup, while the un-shaded bars with blue outlines represent the country
s for
recent coup. The right plot shows that there are substantially more regime-case
s equal to 1 than country
s. The regimes where the regime-case
s equal 1 are effectively dropped from the estimating sample for lack of within-case variation in the
recent coup indicator.
Table 1 lists the 16 cases of short-lived autocratic regimes where the regime-case mean for
recent coup is always equal to 1. Half of these cases, or eight in total, democratize after the regime collapses. Miller’s argument is that the regime-case fixed effects model does not draw support from these sixteen cases to estimate the main parameter of interest. While we agree with Miller’s larger point that regime-case FE in this application effectively remove short-lived regimes from the analysis, we note that many of these short-lived dictatorships dropped from the analysis are coded as dictatorships
precisely because a recent coup toppled an incumbent democracy. The cases highlighted in gray are those in which the historical coup event that makes the indicator of
recent coup equal to 1 is a coup that ended a democracy. One practical effect of using a regime-case estimator that does not draw support from short-lived autocracies is that this approach does not count the “democratic reversal” coups in
Table 1 as potential candidates for coups that enhance democracy.
More to Miller’s point, there are four cases where recent coups in short-lived dictatorships are followed by democratic transition: Egypt 2011, Guinea-Bissau 2002, Haiti 1988, and Niger 2010. In a regime-case FE model, these cases do not inform the estimate of interest; however, in a country FE model they are included because the country
is less than 1. Thus the key difference, in this application, between a regime-case FE estimator and a country FE estimator, is that the latter allows those four ‘positive’ cases to inform the estimate of interest.
Remarks
We acknowledge that there are trade-offs in choosing between different estimators (regime-case FE and country FE in this application). In our original analysis, we reported the results from both FE estimators for both of our dependent variables: transition to democracy and transition to new autocracy. Results from both FE estimators yielded consistent results for coups and transitions to new autocratic regimes. This gives us confidence that these results are not dependent on the choice of cross-sectional ‘unit’. The tests for transitions to democracy, however, were mixed.
The larger point of our original study and the main point we emphasize here is that coups in dictatorships increase the risk of a new autocracy taking power; indeed, most coups in dictatorships are followed by new autocratic regimes. This finding is not dependent on the time period under analysis (1950–1989 vs. 1990–2015). The (average) positive correlation between post-1989 coups and democratic transition is less robust than prior studies have suggested, as we showed in our replication of
Thyne and Powell (2016).
We do not anticipate that our analysis will be the last word on this issue because our results – and all results using less than 30 years of data (post 1989) – are based on only a handful of instances when coups in dictatorships coincide with subsequent democratic transition. As more coups occur in the real world, additional data may alter our conclusions.
Therefore, we urge policy-makers to digest evidence for the “good coup” hypothesis with some skepticism. We also encourage researchers to dig deeper in terms of better understanding why some post-Cold War coups precipitate transitions to democracy, while other coups simply unleash new forms of authoritarianism, often accompanied by increased state-led violence (
Derpanopoulos et al., 2016). Gaining a more solid grasp of why it is that coups can lead to such disparate outcomes seems like an important step for moving forward.