Estimating the Effects of Political Instability in Nascent Democracies

3.1 Outcome Variables

Our approach to measure the institutional quality of Slovenian municipalities and regions is per se constrained by the lack of observable and measurable characteristics of subnational institutional quality that may provide plausible insights in the variation of institutional quality across space and time. In a seminal contribution, Magnusson and Tarverdi (2020) highlight the problems arising from aggregate baseline indicators of institutional quality that reflect some information about quality as the underlying latent variable. Although aggregate indicators have been used extensively in cross-country empirical analysis of institutional quality and its economic effects, the aggregation method implicitly assumes that the errors linked with the baseline indicators are independently distributed, which is both questionable and subject to rigorous scholarly debate. They also propose the extension of the method with nonzero cluster dependence among the error terms.

Our approach to measure subnational institutional quality combines Magnusson and Tarverdi (2020) cluster dependence approach with the extraction of the residual component of institutional quality from the observable aggregate baseline indicators. The approach we propose hinges on the extraction of the latent variable component of institutional quality from the aggregate baseline level through the linear projection onto the local level using the set of predetermined and plausibly exogenous characteristics that are usually considered “fixed” effect and cannot be manipulated at the same time to minimize the measurement error. The proposed disaggregation scheme uses the ex-ante variation in predetermined geographic characteristics to extract the residual component of institutional quality from a higher level of aggregation. That said, disaggregating institutional quality scores can be readily projected to both regional and municipal (i.e., local) level of aggregation provided that sufficient variation in the predetermined characteristics is perceivable in the data. By disaggregating the institutional quality scores to the level, it should be emphasized that the level of institutional quality at the local and regional level per se cannot be recovered, which poses an intrinsic limitation of our approach. Nevertheless, the recovery of the residual component at the local level from a linear projection of predetermined characteristics is able to uncover the residual component, which indicates whether the existing set of institutions is either better or worse than the implied level plausibly expected in the potential control units with plausibly similar geographic characteristics. Since the predetermined physical geographic covariates can be easily compared across and municipalities, regions, and countries, the disaggregating allows for a straightforward comparison of residuals both across singular (i.e., within country) and multiple (i.e., across countries) sources of space and time for each dimension of institutional quality.

In a similar vein, our set of institutional quality variables is from Kaufmann et al. (2011) and comprises well-recognized and established baseline indicators of institutional quality, namely, (i) voice and accountability,^[2] (ii) rule of law,^[3] (iii) regulator quality,^[4] (iv) government effectiveness,^[5] and (v) control of corruption.^[6] Our set of predetermined characteristics exploited as a source of variation in institutional quality consists of region-level and municipal-level geographic covariates and, specifically, include latitude and longitude coordinates (in degrees), precipitation amount (in mm), duration of sunshine (in hours per year), mean annual temperature (in °C), landlocked binary indicator, and Köppen–Geiger climatic zone. These characteristics are collected from large-scale Beck et al. (2018) climate classification at 1 km resolution of topographically corrected climatic maps for the period 1980–2016. Using 1 km resolution of global map, we assemble hand-coded climatic characteristics for 2097 regions, provinces, or states from 121 countries for the period 1996–2020 and match it together with the Kaufmann et al. (2011) governance dataset that yields the potential donor pool with roughly 50,328 year-region/state/province-matched observations.

Since the institutional quality can be only partially observed at the subnational level given the measurement error and ex-ante constraints, our goal is to extract the residual component of baseline governance indicator from the aggregate level from the variation in predetermined subnational geographic covariates that are orthogonal to the institutional quality variable of interest. Residualizing the series of institutional quality to the subnational level allows us to examine whether projected institutional quality at the lower level of aggregation is higher or lower compared to the expected level implied by the multi-country and cross-regional comparison. If the underlying institutional quality residual recovered from the variation in geographic characteristics is positive, our interpretation is that the institutional quality at the subnational level is better than expected in geographically similar environment elsewhere. If the respective residual component is negative, then our interpretation is that the quality of institutions is worse than one can expect elsewhere with geographically similar characteristics. By contrast, zero residual may indicate the level of institutional quality as predicted by geographic endowments.

Suppose we observe a continuum of treated regions or municipalities indexed by i = 1,2, … N and a continuum of control units denoted by j = 1,2, … J. Since some of the geographic characteristics are time-invariant, we extract the institutional quality residual for kth baseline governance indicator through the canonical regression of the following form for each t = 1,2, … T:

where q denotes the institutional quality as a latent variable, η₀ is the level of institutional quality, X is the vector of geographic characteristics, α is the set of quality response coefficients, and ε is the canonical error term. From the canonical regression of institution on the predetermined physical geographic characteristics, the residual component for the full repeated cross section of treated regions and municipalities is recovered as the difference between the observed baseline indicators and its predicted counterpart implied from the variation in exogenous geographic characteristics:

where q j ϵ J , N o b s e r v e d is the observed level of institutional quality from the baseline indicator, and q i ϵ N ∪ j ϵ J p r e d i c t e d is the predicted level of institutional quality for ith municipality that belongs to jth region and for the jth region only. The general thrust of the residual component is that it can be used to determine spatial disparities and spreads of institutional quality over time based on the difference between the observed and predicted level of institutional quality. By the same analogy, the overall level of institutional quality can be recovered as the sum of the observed component at jth level and the exogenous residual component at ith level such that Q_i∈N = q_j∈J + e_{i∈N∪j∈J}.

One of the limitations of the repeated cross-sectional estimation of the residual component arises from the time-invariance of some geographic characteristics such as latitude and longitude or binary indicators of coastal access and landlocked status. Without the loss of generality, repeating the cross-sectional canonical regression may only shift the intercept η₀ upward or downward in each t = 1,2, … T, while the variation in Q_i∈N may be the same as in q_j∈J. Consequently, any attempt to isolate the effect of institutional quality from the observables and unobservables would be impossible per se since institutional differences would be treated as a set of fixed effects in panel-level analysis. It should be noted that the selection of the predetermined covariates used to construct the residualized level of institutional quality is driven by data availability. Although a much richer set of the covariates would reflect a more nuanced ex-post variation in institutional quality and further reduce noise-to-signal ratio, both cost and time of data collection are overly cumbersome. Nevertheless, we submit the battery of the predetermined covariates to further scrutiny and construct the residual series separately through a variety of combinations of X variables to tackle the robustness of the residualized series to the possible misspecification. We find that the correlation between the outcome variables using various sets of X variables is very high, with the sample correlation coefficient between +0.8 and 0.9 (i.e. p-value = 0.000). This implies that quantitatively and qualitatively similar results can be expected when different sets of predetermined covariates are used to build residualized institutional quality series.

To partially overcome these intrinsic limitations, we use Monte Carlo Markov Chain (MCMC) sampling algorithm for the probability distribution of the recovered residual component of institutional quality. Through the construction of Markov chain with the desired distribution as the equilibrium distribution, sample of target distribution can be obtained by recording the states from the chain (Jarner and Roberts 2007). We aim at obtaining a sequence of random samples from a probability distribution where random sampling is difficult and thus adopt the well-known Metropolis et al. (1953) and Hastings (1970) algorithm to approximate the true distribution of institutional quality at the subnational level. Randomly sequenced samples can be drawn from any probability distribution that has a known probability density for which a function must be known and proportional rather than exactly equal to its density and values. Metropolis–Hastings algorithm’s ability lies in generating a sequence of random sample in such way that as more samples are produced, the distribution of values more closely approximates the target distribution through an iterative procedure that produces a Markov chain. At each iteration, the algorithm selects a candidate for the next sample value from the current sample value. Provided that probability threshold is specified ex-ante, the candidate value of institutional quality is either accepted or rejected.

Let f(Q = e _iϵN ) be a function proportional to the target probability distribution denoted by P(Q = e _iϵN ). Our goal is to obtain a posterior parameter of institutional quality at the subnational level to be evaluated, π(Θ _i ) = p(Θ _i |Θ_i−1,X,Q) where a single parameter is sampled from one-dimensional distribution π(Θ). To generate samples from π(Θ), the algorithm specifies a recognizable proposed probability density denoted as q(Θ^g+1|Θ ^g ) with the density ratio requirement π ( Θ g + 1 ) π ( Θ g ) . In this respect, the algorithm is similar to the Gibbs sampler, but it draws a candidate value first that will be accepted or rejected based on the acceptance probability. For the sake of brevity, we adopt a two-stage sampling procedure:

• Step #1: Draw Θ^g+1 from the proposal density q(Θ^g+1|Θ ^g ).

• Step #2: Accept Θ^g+1 with probability γ(Θ^g+1|Θ ^g ) where: γ ( Θ g + 1 | Θ g ) = min ( π ( Θ g + 1 ) q ( Θ g + 1 | Θ g ) / π ( Θ g ) q ( Θ g | Θ g + 1 ) , 1 ) .

where the requirement to generate a sequence of random samples is that the values of the posterior parameter are drawn from the proposal distribution of the uniform random variable to evaluate the acceptance and rejection criteria. In the two designated steps, the algorithm decomposes the unrecognizable conditional distribution into recognizable distribution through the sequenced generation of candidate points and unrecognizable part from which the acceptance criteria is set. Such iterative procedure corrects the equilibrium distribution by allowing the functional form of the model to be nonanalytic. One notable advantage of the algorithm is that sampling can be conducted on specific tail of the distribution to analyze parameter restrictions imposed by the prior values. Since imposing prior values requires a complete information set and may be driven by subjective beliefs, our approach is to abstain from the imposition of priors. Instead, to enable the generation of random samples through the Markov chain, our approach is to impose a noninformative objective prior distribution on parameter space (Jeffreys 1946) with the density function proportional to the square root of the determinant from Fisher information matrix p ( Θ → ) ∝ det I ( Θ → ) . The key advantage of the objective noninformative prior is posed by the relatively nonthin tails of the distribution compared to its target counterpart, which ensures a straightforward and fast convergence of the algorithm.

By imposing objective noninformative prior to approximate the target distribution of institutional quality, we adopt an adaptive random-walk version of the Metropolis–Hastings algorithm where a candidate value of institutional quality for each region and municipality is drawn from a simple random-walk model Θ i ϵ N , t g + 1 = Θ i ϵ N , t g + ε i ϵ N , t where ε_iϵN,t = i.i.d ∼ (0,σ²) is the random component taken to be a symmetric density function with thick tails of t-distribution. Our choice of the density function is generic. As such, it ignores the structural features of the target density. Due to the symmetry in the proposal density functions, q(Θ^g+1|Θ ^g ) = q(Θ ^g |Θ^g+1), the algorithm to estimate the subnational institutional quality converges to the following:

• Step #1: Draw Θ^g+1 from the proposal density q(Θ^g+1|Θ ^g ).

• Step #2: Accept Θ^g+1 with probability γ(Θ^g+1,Θ ^g ) where: γ(Θ^g+1,Θ ^g ) = min[π(Θ^g+1)/π(Θ ^g )t,1].

And conveys two chief advantages in estimating the latent institutional quality. First, the simplified algorithm has the ability to control the variance of the error term, which substantially reduces the measurement error. And second, the algorithm must be tuned by adjusting the variance of the error term to obtain an acceptable level of accepted draws. Given the size of our sample, we set the number of iterations to obtain a sequence of random samples at 12,500 for each outcome-year combination and set the acceptance rate at 25 %, which is in the conventional range between 20 and 40 %. At 12,500 iterations per outcome and year, the number of discarded observations through the burn-in amounts to 2500. Hence, for 24 years and 5 outcomes under consideration, the overall number of iterations performed in our analysis is around 1,500,000 million random sequences of samples. This allows us to construct the mean and median estimate of the posterior parameter for each region and municipality for each t = 1,2, … T. By inverting the test statistics, we also compute the upper and lower confidence bounds of the posterior parameter for each treated and nontreated region and for a full cross section of municipalities. One of the caveats arises from the stability and composition of the latent component over time. Estimating a repeated sequence of randomly generated samples through MCMC algorithms such as Metropolis–Hastings usually invokes a notable confluence of cyclical and deterministic components that may be inherent in the estimated series with noninformative priors. While the deterministic component captures long-term trend of the series, the cyclical component may capture short-run oscillations of the posterior parameter values from its long-run equilibrium. Suppose that the posterior parameters can be decomposed into the deterministic and cyclical component:

where τ captures he deterministic component and c represents the cyclical component of the parameter. Notice that the deterministic component captures the signal of the institutional quality parameter, whereas the latter captures the noise behind the latent quality trait. If the latent trait denotes short-term and temporary deviations from the long-run trend, Θ has a tendency to exhibit a stationary mean-reverting pattern since, in such circumstances, it is determined by c instead of τ. Hence, if c component prevails over time, the simulated latent quality trait will converge to the random-walk behavior. Conversely, if the latent quality trait exhibits little deviation from the cyclical component, the underlying posterior parameter Θ is dominated by τ. Our goal is to recover a smooth long-run trend component of the posterior parameter for ith municipality and jth regions, which is done by solving Hodrick and Prescott (1997) optimization problem:

where (Θ_i,j,t − τ_i,j,t) denotes the residual component of the latent quality trait in continuous time and satisfies the stability condition such that c t ∼ i . i . d ∼ N ( 0,1 ) . The parameter ϕ captures the speed of dynamic quality trait adjustment over time to parse out the cyclical component from its long-run counterpart. Ravn and Uhlig (2002) state that ϕ should vary by the fourth power of frequency observation ratio. For annual observations, this implies that ϕ = 6.25 (=1600/4⁴), which we adopt to smooth the series of latent institutional quality trait for each dimension.

3.2 Sample

To isolate the treatment effect of political instability, it is necessary to build a control sample of regions where the prevalence of political instability is imperceptible. This implies that regions from countries that have endured prolonged political instability such as Italy or Argentina that has persisted over time do not provide a plausible variation in the trajectory of institutional quality to estimate the effects of instability. To address these concerns, we leverage the variation in the estimated latent institutional quality of Slovenian regions and municipalities against their peers from countries with a high degree of political stability. To this end, we examine the differences in the political stability and absence of violence using Kaufmann et al. (2011) indicator of stability, which has been used extensively in the literature. In the next step, we build a distribution of political stability through the kernel density function and eliminate the region-level observations below the 80th percentile threshold of political instability score. Excluding these observations ensures that stable unit treatment value assumption (SUTVA) is reasonably met, while those observations where the influence of political instability is apparent are discarded from the sample. Figure 3 depicts the global distribution of political instability score in our treatment and control samples through a joint density function.

Under this particular criteria, the effective size of the donor pool to 246 regions from 16 countries^[7] for the period 1996–2020 yields a donor pool 5904 matched observations. The treatment sample consists of the 12 regions,^[8] 201 municipalities, and 11 urban municipalities for the same period. This yields a region-level treatment sample with 288 observations and a municipal-level treatment sample with 5088 observations. Hence, the combined size of treatment and control samples comprises 11,280 observations.

Figure 3:

Global distribution of political stability.

Figure 4 exhibits the trajectories of regional institutional quality estimated from the latent quality model using Metropolis–Hastings MCMC algorithm with the objective prior function. Each figure exhibits the mean score of institutional quality obtained from the repeated sequence of randomly generated samples together with the upper and lower bound of the 95 % confidence interval. The evidence readily uncovers the signs of deteriorating institutional quality after the onset of political instability in 2008. Declining institutional quality does not seem to be created equal across all respective dimensions. For instance, latent dynamic scores of quality suggest that voice and accountability, rule of law, and the ability to control corruption tend to deteriorate particularly fast. By contrast, the speed of decline in government effectiveness appears to be somewhat slower, while the trend of regulatory quality tends to improve around the year 2016 amidst persistent decline after 2008. The general thrust of these comparisons invariably advocates a pattern of accelerated deteriorating of subnational institutional quality in the midst of the rising political instability. In Figure 5, municipal-level latent scores are depicted for the same dimensions of institutional quality. In spite of the notable similarities of the trajectories of institutional quality, several differences are apparent. First, voice and accountability dimension tends to deteriorate relatively fast in the initial years of the shock, while it does not seem to decrease further. Second, rule of law and government effectiveness components tend to deteriorate much slower than at the regional level. Third, municipal regulatory quality tends to decrease substantially in the initial years of the shock, while it tends to recuperate strongly afterward and exceeds the preinstability level in the end-of-sample years. Fourth, the control of corruption tends to deteriorate even faster at the municipal level, which suggests that local communities may be even more prone to the rampant deterioration of the control of corruption. Compared to the regional level, municipalities have somewhat lesser drop in political accountability, rule of law, and government effectiveness. The estimated latent scores also advocate a somewhat better regulatory capacity of the municipalities amidst a temporary nature of instability shock. However, the deterioration in the ability to control corruption appears to be even stronger than at the regional level.

Figure 4:

Trajectories of regional institutional quality in Slovenia, 1996–2020.

Figure 5:

Trajectories of municipal institutional quality in Slovenia, 1996–2020.

4.1 Difference-In-Differences

Our goal is to estimate the contribution of political instability to the institutional quality consistently. To this end, we estimate a simple but compact difference-in-differences specification in the presence of unobserved effects:

where Q is the institutional quality in municipality i = 1,2, … N and region j = 1,2, … J at time t = 1,2, … T, X is the vector of structural covariates that vary systematically with the outcome variable, μ is the set of municipality- or region-level fixed effects which are, by default, unobserved to the econometrician, δ is the set of time-varying institutional quality shock common to all municipalities and regions, and ϵ denotes stochastic disturbances. Our key parameter of interest is α ˆ 1 , which represents the average treatment effect of political instability captured by the binary treatment variable (Post-Crackdown) for the full set of regions and municipalities in the post-treatment period 2008–2020. In the baseline differences-in-differences model of institutional quality, standard errors are adjusted for serially correlated stochastic disturbances and heteroskedastic distribution of random error variance using simple region- or municipality-level non-nested single-way clustering scheme. The key assumption that underlines the validity of the difference-in-differences estimate of α ˆ 1 concerns parallel trends of institutional quality between the treated municipalities and regions and their untreated peers. Against this backdrop, De Chaismartin and D’Haultfoeuille (2020) show that in the presence of parallel trends assumption, a linear panel-level regressions of such type estimate weighted sums of the treatment effect in each unit and time period. The weights used to compute the treatment effect of interest may be negative, which can be problematic if the treatment effect exhibits either spatial or temporal heterogeneity. For instance, treatment coefficient in our DiD regression may be negative, while the treatment effect may be positive. To address these concerns, they propose a two-way fixed-effects estimator of the treatment effect for the observed units that switch treatment only at the time they switch, which does not necessitates homogeneity assumption on the treatment effect. Hence, we adopt a two-way fixed-effects estimator and for a given time-fixed treatment in our model, we thus tackle and allow for the heterogeneity of the treatment effect in time.

4.2 Synthetic Control Estimates

In the absence of parallel trend assumption, difference-in-differences estimator may not isolate the treatment effect of interest. When parallel trend assumption fails, the average treatment effect may be tainted by preexisting trends or shock and, thus, it does not provide any evidence of whether the effect of political instability on institutional quality is significant or not. The hypothetical absence of parallel trend assumption can be partially addressed by adopting the synthetic control estimator to isolate the treatment effect of interest (Abadie 2021; Abadie et al. 2015; Ben-Michael et al. 2021; Billmeier and Nannicini 2013; Cattaneo et al. 2021; Powell 2022; Xu 2017). Suppose we observed J + 1 regions in the time period t = 1,2, … T where the entire set of regions is exposed to the political instability while {2, … J + 1} being directly unaffected. We assume that political instability takes place at time T₀ and lasts for the entire postintervention period, which implies that t < T₀ < T where T₀ + 1 indicates the first post-treatment period of instability. Our interest lies in estimating the effect of instability on institutional quality for the full set of treated regions. Let Y i , t N denote the outcome for ith region at time t in the hypothetical absence of political instability, and let Y i , t I denote the corresponding outcome for region i at time t that would be observed if ith unit were exposed to the shock posed by instability. By assuming that Y i , t N = Y i , t I for each jth region in the preintervention period t < T₀ + 1, our aim is to estimate the impact of instability on institutional quality for each post-treatment period:

where Y 1 , t N is by default unobserved to the econometrician and needs to be estimated. Without the loss of generality, our assumption is that Y 1 , t N follows a latent factor model, which allows the approximation of the counterfactual levels of institutional quality in the hypothetical absence of instability. Hence, for the combined pre- and postintervention period, we estimate the trajectory of institutional quality in the absence of instability through the following latent factor model:

where ϵ i , t ∼ N ( 0,1 ) transitory shocks under i.i.d. assumption, δ _t is the set of common temporal shocks that absorb time-fixed effects into the counterfactual trajectory, Z i ∈ R r represents the observed covariates unaffected by the shock, η t ∈ R r is the set of common unobserved factors, and μ i ∈ R r represents the set of unknown factor loadings. The key advantage of the latent factor model is posed by the ability to capture heterogeneous responses of the outcome variable to multiple unobserved factors and embed the time trends into the model. By estimating Y i , t N , the control group can be reweighted so that a synthetic version of the instability-affected region is matched with the donor pool on Z _i and some pre-T₀ {Y}, which immediately implies that μ _i will be matched automatically. Therefore, to approximate the missing counterfactual scenario, control group for the instability-affected region needs to be reweighted to ensure that the control group mimics the characteristics and attributes of the affected region inasmuch as possible. Let W = (w₂, … w_J+1) be a vector of weights such that w _j  ≥ 0∀j. For the full vector W, the outcome variable of the synthetic control group at time t is given by:

Where ∃W^* is such that the synthetic control group can be reasonably well matched with the treated region in the full pre-T₀ period wherein both ∑ j = 2 J + 1 w j * ⋅ Y j , t = Y 1 , t and ∑ j = 2 J + 1 w j * ⋅ Z j = Z 1 , t can be approximated. Provided that ∑ t = 1 T 0 λ t ′ λ t is nonsingular, then for all t > T₀, we have E [ Y 1 , t N − ∑ j = 2 J + 1 w j * ⋅ Y j , t ] → 0 as T₀ → ∞ or if T₀ is large relative to the scale of ϵ_i,t. The vector of weights W is built to minimize the distance in pre-T₀ outcomes and auxiliary covariates between the full set of treated regions and its respective control group such that ‖ X 1 − X 0 ′ W ‖ V subject to w _j  ≥ 0 and w₂+ … +w_J+1 = 1. To minimize the discrepancy in the attributes and characteristics of the treated regions prior to the political instability, we use a positive semi-definite matrix as a linear combination of preintervention outcomes and covariates through the following cross validation:

Which consists of two stages. In the training stage, the matrix V denotes the relative importance of covariates and pre-T₀ outcomes in explaining the institutional quality outcomes. In the validation stage, predictive weights of covariates and pretreatment outcomes are selected to minimize the root mean square prediction error to build W vector indicating which regions from the donor pool have similar covariate and outcome values and fall within the convex hull of characteristics and attributes of the treated regions but had no political instability shock. Under these conditions, the missing counterfactual scenario can be estimated through an approximately unbiased estimator of λ_1,t:

wherefrom it follows immediately that even though {μ} vector is unobservable to the econometrician, fitting the joint vector {Y,Z} provides the sufficient condition to match the observed realization of pre-T₀ {Y} with its counterpart in the donor pool. Hence, the difference between the observed trajectory of institutional quality and its synthetic peer tends to disappear E [ Y 1 , t N − ∑ j = 2 J + 1 w j * ⋅ Y j , t ] → 0 if Λ 1 = { Y 1,1 , … Y 1 , T 0 , Z 1 ′ } ϵ R 1 × ( T 0 + r ) falls within the convex hull of {Λ₂, … Λ_J+1}, which implies that the linear combination of the attributes and characteristics from the vector of covariates V in the training stage and vector of weights W can only include nonzero positive weight rates and thus, Rank({Λ₂, … Λ_J+1}) = T₀ + r condition holds. To estimate a plausible counterfactual scenario without preexisting trends in the outcome variables, we assume that the treated regions and their counterparts in the donor pool without political instability follow the same latent factor model over time. If preintervention time window is sufficiently large, reasonably unbiased estimate of Y i , t N is possible. Given that our time window spans more than 10 years, it is neither insufficiently low nor excessively large and enables us to track and reproduce the region-level trajectory of institutional quality is a reasonable manner to avoid biasing the estimate of λ ˆ 1 , t .

Since our model involves several treated units with a single-treatment year, we follow Cavallo et al. (2013), Acemoglu et al. (2016), and Gobillon and Magnac (2016) and extend the model setup toward multiple treated units. Under such setup, the vector of treatment effects of political instability { λ ˆ 1 , T 0 + 1 , … λ ˆ 1 , T 0 + k } is estimated for each region separately. Hence, a dynamic estimate of average treatment effect is computed by integrating over the treated units vector of λ ˆ 1 :

where the heterogeneity of treatment effect is further examined through the inspection of λ ˆ 1 , which allows us to build the distribution of treatment effects for each underlying outcome variable and compare its evolution before and after the political instability shock. Following Firpo and Possebom (2018), we modify root mean square prediction error from preintervention period to test sharp null hypothesis of no treatment effect whatsoever and invert the test statistics to estimate the 95 % confidence bounds and examine the precision of the estimated treatment effects as well as their respective significance and robustness.

5.1 Difference-In-Differences Estimates

Table 1 reports difference-in-differences (DiD) municipality-level estimates of the institutional quality effect of political instability. Panel A reports standard DiD estimates in the presence of municipality-fixed effect. Each specification confers the full set of both municipality-fixed effects and time-fixed effects to absorb both time-invariant heterogeneity and time-varying common institutional technology shocks into the model and eliminate them as the respective sources of confoundedness. The evidence points out a pervasive and substantial deterioration of institutional quality in response to the political instability. In particular, point estimates invariably suggest significant drop in institutional quality after the government crackdown in 2008. For instance, the estimates in column (1) suggest that the average treatment effect of political instability on voice and accountability dimension is around −0.13 and is statistically significant at 1 % (p-value = 0.000). This amounts to an average drop in municipal-level voice and accountability latent score by about one 10th of the standard deviation, which is appears to be large. That said, the advent of political instability in 2008 appears to have restrained the citizens’ access to participation in selecting the government and also deteriorated freedom of expression and association. The underlying post-treatment political instability variable accounts for around 20 % of the overall variation in voice and accountability dimension in the full sample, which appears to be large. Furthermore, point estimates in column (2) invariably indicate a substantial deterioration of the rule of law in response to the onset of political instability. Although the average treatment effect is about half as large as its voice and accountability counterpart in column (1), the point estimate is statistically significant and invariably suggests that rule of law score tends to decrease by 0.068 standard deviation in the post-treatment period while the underlying coefficient appears to be statistically significant at 1 %. By contrast, the estimated average treatment effect in column (3) indicates some improvement in government effectiveness in response to the political instability. The point estimate in column (3) shows that municipalities tend to have 0.058 standard deviation improvement in government effectiveness in the underlying post-treatment period. By contrast, the estimates in column (4) and column (5) uncover the evidence of rampant deterioration in regulatory quality as well as a substantial drop in the control of corruption. Both coefficients are statistically significant at 1 %, whereas the post-treatment effect of instability accounts for about 20 % of the overall variation in regulatory quality and control of corruption. In each specification, the null hypothesis of the absence of parallel trend assumption between the treated municipalities and its control group is rejected, which suggests that in the absence of instability, the differences in institutional quality between the treated municipalities and their control group is indeed constant over time.

Panel B reports heterogeneous DiD estimates with two-way fixed effects proposed by De Chaisemartin and D’Haultfoeuille (2020). The key advantage of two-way fixed-effects estimator is that it specifically addresses temporal heterogeneity of the effect. The panel reports the baseline institutional quality effect of instability in the contemporaneous year of the shock as well as the effect in 5-year time span together with the end-of-sample placebo coefficient to further tackle both the uniqueness and significance of the effect. The evidence suggests a reasonably strong deterioration in the voice and accountability score in the post-treatment period. The estimated set of post-treatment coefficients indicates zero contemporaneous effect and a relatively high negative effect of instability in the 5-year period. At the same time, the end-of-sample placebo coefficient is not statistically significantly different from zero, which implies that the estimated negative effect of political instability on voice and accountability is relatively unique and specific to the treated municipalities only. The evidence also suggests some improvement of government effectiveness in response to the political instability as indicated in column (3) and also reported earlier in Panel B. Conversely, point estimates in column (4) further uncover the evidence of the deterioration of regulatory quality. While the contemporaneous effect is not distinguishable from zero, the effect becomes negative in the first post-treatment year (−0.022, p-value = 0.000) and rises up to −0.216 (p-value = 0.000) in the fifth post-treatment year. Consistent with the baseline fixed-effects difference-in-differences estimate in column (5), we further uncover the evidence of the widespread deterioration in the control of corruption.

The deterioration is both immediate and appears to persist. The treatment effect coefficient in the contemporaneous year of instability is both negative and statistically significant and tends to widen in magnitude considerably until the fifth year of the post-treatment period while retaining a high degree of statistical significance (i.e., p-value = 0.000). By explicitly accommodating temporal heterogeneity into the underlying DiD model specification, the evidence further unravels and demonstrates debilitating effects of political instability on institutional quality, which does not seem to be temporary. Panel C re-examines the difference-in-differences effect through nearest-neighbor matching technique (Abadie and Imbens 2006) where the treatment effect of instability is obtained by the covariate distance matching between the treated municipalities and the control sample. The estimated coefficients on average treatment effects are negative and significant at 1 %, which confirms the institutional quality deterioration in response to the instability and suggests no evidence to support the notion of growth-enhancing effects of political instability. The drop in institutional quality appears to the largest for voice and accountability and control of corruption dimensions.

5.2 Synthetic Control Estimates

The local-level evidence based on difference-in-differences estimates of the institutional quality model specification indicates substantial deterioration in response to the political instability. Two limitations are inherent in the DiD approach. First, small size of the municipalities chiefly implies that local-level response of the latent institutional quality to political instability may be dominated by the response at the regional level. Second, if the parallel trend assumption is sensitive to the length of the pretreatment period or varying composition of control sample, estimating the counterfactual scenario through the application of synthetic control estimator may provide a more plausible evidence on the effect scope and size. Synthetic control estimator provides evidence on the average treatment effect in each respective year of the postintervention period, which seems to be more informative about the nature of the political instability shock.

Table 2 reports the balance of preinstability institutional quality outcomes and auxiliary covariates between the selected treated regions and their respective synthetic control groups for the three selected regions. Balancing the outcomes in pre-T₀ period and the auxiliary covariates indicates a reasonably good quality of the fit. Both the values of the outcomes and covariates in the preintervention period are closely aligned without major discrepancies. For instance, synthetically matched levels of outcome variables for the full set of treated regions are almost identical in the pretreatment period. The estimated root mean square prediction (RMSE) is low and appears to be within 2 % of the scale-adjusted error margin. At the same time, the matched control groups for the treated regions exhibit similar auxiliary characteristics, which imply that the synthetic control groups closely mimic both the dynamic and structural characteristics of the treated regions prior to the political instability shock. It should be noted that for some regions such as Littoral-Carso, the quality of the fit appears to be excellent (i.e., RMSE = 0.006) with evidence of few discrepancies compared to the synthetic control group in the pretreatment period.

Figure 6 reports the composition of synthetic control groups for the selected treated regions.^[9] In spite of the numerous similarities, several differences are perceptible in the composition of control groups between the treated regions. For instance, the voice and accountability trajectory of Littoral-Carso in the preinstability period is best reproduced as a convex combination of the implied attributes of Midland Ireland (24 %), Madeira (13 %), Burgenland (18 %), Aomori (14 %), Porto (11 %), Gaborone (11 %), Regiao de Leiria (4 %), Azores (4 %), and Kagawa (<1 %). Taken altogether, 41 % of the control group consists of Portuguese intermunicipal regions, followed by Irish regions (24 %), Austrian federal states (18 %), Japanese prefectures (15 %), and Botswanan districts, particularly Gaborone City (11 %). In a similar vein, the rule of law trajectory of Central Slovenia prior to the political instability is best synthesized as convex combination of institutional development and the auxiliary characteristics of Madeira (66 %), Saitama (13 %), Southwest Ireland (11 %), Tottori (7 %), and Gotland (3 %), respectively. Furthermore, the control of corruption trajectory of Drava region before the instability period appears to be best reproduced as linear combination of the implied characteristics of South West Ireland (76 %), Medio Tejo (15 %), and Ghanzi (10 %). In terms of further example, the composition of synthetic control group for Upper Carniola’s trajectory of regulatory quality is similar and appears to be dominated by Irish regions, followed by one Botswana’s district (Ngamiland) and one Portuguese intermunicipal region (Ave/Braga) where the combined total weight share of Irish regions is around 80 %.

Figure 6:

Composition of region-level synthetic control groups.

Figure 7 reports synthetic control estimates of the institutional quality effect of political instability. More specifically, the figure reports the quality effect of instability for each of the five outcomes by displaying the mean effect for the full post-treatment period alongside 95 % confidence bounds. Thus, the vertical axis exhibits the institutional quality gap between the treated region and its synthetic peer. The estimates indicate plausible evidence of the pervasive institutional quality breakdown induced by the political instability. The estimated effect of instability on voice and accountability appears to be temporary. While the respective trajectory decreases markedly in the early postinstability period, it tends to rebound back to the preinstability level in the end-of-sample years. This indicates a sharp temporary effect of instability instead of its permanency. We also find evidence of the pervasive deterioration of the rule of law in response to the instability. After 2008, the rule of law trajectory appears to improve modestly followed by a rapid drop, which tends to amplify in the end-of-sample year. Pointwise, the average treatment effect of political instability on the treated amounts to −0.2 basis points in the end-of-sample year, respectively. The evidence based on synthetic control analysis also uncovers substantial deterioration of government effectiveness in the postinstability years. In spite of somewhat markedly lower gap behind the synthetic peers after 2015, the upper and lower bound of the confidence interval are firmly below zero, indicating an arguably strong and pervasive effect of instability on the ability of the government to effectively formulate and implement sound policies. The estimated gaps of regulatory quality are similarly negative in the entire post-treatment period, indicating a large regulatory quality cost associated with political instability. Furthermore, the synthetic control estimates indicate a pervasive and rigorous deterioration of the control of corruption. While the gap between the treated regions and their synthetic peers appears to be both indiscernible and imperceptible in the pretreatment period, the average treatment effect on the treated with respect to corruption control is both immediate, large, and pervasive. The control of corruption trajectory of the treated regions deviates substantially from its synthetic benchmark in a considerable manner. The deterioration of the trajectory continues unabated and appears to be large. In the end-of-sample year, the average treatment effect of instability on the institutional quality amounts to −0.35 basis points and seems to be large and ubiquitous.

Figure 7:

Average treatment effect of political instability on institutional quality of Slovenian regions, 1996–2020.

In Table 3, we compare the contrasts between the pretreatment and post-treatment effects of political instability for each outcome variable. More specifically, in the presence of weak effects of instability, the pretreatment effects and post-treatment effects should be roughly equal with the difference being statistically indistinguishable from zero. However, if the political instability triggers a discernible shift in the trajectories of institutional quality, the differences in the estimated post-treatment and pretreatment effects should be both apparent and statistically significant. Three different implications emanate from these comparisons. First, Chow test statistics on the equality of coefficients before and after the instability shock indicates ample evidence of the structural break in each outcome, indicating a marked shift in the trajectories of institutional quality. The resulting p-values are uniformly within 1 % significance threshold, which implies that the political instability has indeed triggered a differential trend in the post-treatment quality trajectory compared to the synthetic benchmark (Spruk and Kovac 2020). Second, a simple test of parallel trend assumption in the spirit of Muralidharan and Prakash (2017) is performed to examine whether the trends in the gap between the observed and synthesized trajectories of institutional quality are parallel in the pretreatment period. It should be noted that our test concerns parallel trends in the trajectories of institutional quality gaps compared to the trajectories in the donor pool. Nonparallel trajectories’ trends may indicate either large predictive discrepancy between the treated regions and their synthetic peers, which may invoke the presence of policy shocks distinctive from political instability. In each model specification, we reject the null hypothesis of no parallel trends between the treated regions and their synthetic peers at 1 % significance threshold, which implies that institutional quality gap at the regional level is plausibly captured by the trends in the synthetic control group. Lastly, another test we conduct is to see if in the anticipation of political instability, the treatment and control group change their institutional quality behavior. In the presence of discernible anticipation effect, the estimated effects of political instability would be questionable at best. The evidence largely confirms the absence of the anticipation effects in each institutional quality outcome in the treatment and control group prior to the instability shock. The resulting p-value is high in each outcome even when the significance threshold is set at artificially elevated levels, which implies that the null hypothesis of no effect in treatment anticipation cannot be rejected. Taken together, the tests readily suggest that synthetic control estimates plausibly capture the effect of political instability on institutional quality. No rejection of the null hypothesis also implies that the resulting post-treatment gap in the institutional quality is unlikely to be driven by the alternative political or structural shocks. If the null hypothesis were rejected, the hypothetical possibility of other shocks influencing the institutional quality trajectory would become more feasible to discuss.

Figure 8 reports the overall composition of the synthetic control groups across the full set of treated regions for the full set of outcome variables. Instead of reporting the exhaustive composition for each region and outcome separately, the frequency distribution of donors with nonzero positive weight is reported, which enables a more straightforward comparison of the contrasts in the composition of synthetic control group across the outcome categories. The evidence suggests both notable similarities and marked contrasts in the composition of control group. For instance, in the estimated synthetic control specifications with voice and accountability latent outcome variable indicate Irish regions, Portuguese intermunicipal regions, and Japanese prefectures as the most dominant donors with nonzero weights for the full treated set. Similarly, in the rule of law specifications, the control groups are disproportionately dominated by Madeira and Azores, while with respect to government effectiveness, the control group is dominated by Azores followed by Japanese prefectures. In slight contrast, the control groups for regulatory quality specifications are characterized by the presence of Swedish counties such as Halland and Austrian federal states such as Salzburg and Burgenland. The presence of Australian states, particularly Tasmania, is also perceptible in the specifications estimating the effect of instability on the control of corruption. Taken together, this implies that preinstability latent institutional quality trajectories of Slovenian regions can be plausibly tracked and reproduced as a convex combination of attributes of regions from countries having markedly and substantially higher level of political stability after 2008, which are predominantly concentrated in the Western Europe, East Asia, North America, and Australasia.

Figure 8:

Frequency distribution of synthetic control groups.

5.3 Placebo Analysis

The question behind our estimates that still remains unanswered is whether the estimated effects of political instability are driven by chance. To assess the significance of the estimated institutional quality gaps, we rely on the multiple-event placebo analysis proposed by Cavallo et al. (2013) based on Abadie et al. (2010) and further extended by Galiani and Quistorff (2017). In the likely absence of anticipated treatment and presence of structural break reported in Table 3, our approach is to perform a series of placebo analyses by applying the synthetic control estimator to the battery of regions in the donor pool that never underwent the treatment itself and thus never experience large-scale political instability opposed to Slovenia. If the placebo analysis creates institutional quality gaps with similar size to the ones obtained for Slovenian regions, then it is not likely that our analysis conveys evidence of the significant effect of political instability because, under such circumstances, the declining institutional quality conditions are absorbed by the general trend perceptible everywhere. By contrast, if the in-space placebo analysis uncovers evidence of unusually large institutional quality gaps of treated regions compared to the regions that did not undergo widespread political instability, then the notion of significant evidence of the negative effect of instability on institutional quality becomes more plausible. To quantitatively evaluate and assess the significance of the estimates, the synthetic control estimator is iteratively applied to every region in the donor pool, which effectively reassigns the treatment and shifts the treated regions into the donor. In the next step, the estimated effect for each outcome and each quasi-treated region is estimated through the placebo run, which yields a distribution of placebo effects.

Without the loss of generality, we follow Cavallo et al. (2013) and adopt a multi-treatment placebo framework. Let g ∈ {1, … ,G} be an index of treated regions and let J represent the number of regions that never undergo the treatment. For each treated unit, we estimate the effect of instability in the first post-treatment year denoted by λ ˆ g , 1 , T 0 + 1 . Across the full treatment set, the average effect is given by λ ‾ = G − 1 ( ∑ g = 1 G λ ˆ g ) . For each treatment g, we generate the unrestricted set of placebo effects, λ g , t P l a c e b o = { λ ˆ j , t : j ≠ 1 } where each region enters the treatment at the same time as region g. By averaging λ g , t P l a c e b o over the instability treatment to obtain λ g , t P l a c e b o ‾ partially removes the noise from the estimates. This is constructed from all possible averages where a single placebo coefficient set is taken from each λ ˆ g , t P l a c e b o . Notice that there are N P l a c e b o ‾ = ∏ g = 1 J J g possible averages. To keep the placebo analysis parsimonious, we restrict the donor pool by match quality and discard placebos that do not match well to prevent the distribution from being tainted by the placebos with extreme rarity of negative instability effect. If K g m denotes the number of controls that match as well as the treated unit g for the same time period, then N P l a c e b o ‾ m = ∏ g = 1 G J g m . Furthermore, if b is the selection index where a single placebo effect is drawn from each treatment placebo set, let λ P l a c e b o ‾ ( i ) the mean of the placebo selection. Inference on the instability effect with multiple treated regions is then conducted through the following two-sided p-value:

where we perform to basic checks on the internal validity of synthetic control estimator in approximating the counterfactual scenario. In the first step, we examine whether the treated regions attributes and characteristics fall within the convex hull of the control units. In the second step, the distribution of pretreatment RMSPEs is modified by discarding the units that have the pretreatment RMSE at least twice as high as that of the treated regions because such regions cannot be matched appropriately. Although this is not a lenient criterion, discarding the control units that fail to match well with the treated regions in pre-T₀ period partially avoids overestimation of the p-value based on placebos where the relative rarity of obtaining artificially large effect with poor quality of pretreatment match is neither trivial nor low. For a total of 248 regions in the donor pool, 12 multiple-treated regions and five outcome variables, the estimated number of placebos is very large for each outcome (i.e., 1.297e+28). This implies that instead of standard treatment permutation traditionally used in smaller samples, our placebo analysis fully relies on the random sampling algorithm to obtain the sequence of p-values associated with the institutional quality effect of instability.

Table 4 presents the parametric analysis of the placebo effects through the application of difference-in-differences analysis of the estimated gaps compared to the full set of placebos. The underlying parameter of interest is the post-treatment indicators of the change in the estimated gap in comparison with the placebos. The evidence suggests that the estimated institutional quality gaps induced by political instability appear to be both large and statistically significant. For each of post-treatment gap coefficient, 95 % confidence bounds are computed. Each specification also conveys the full set of region-fixed effects and time-fixed effects to control for the confounding influence of unobserved heterogeneity bias and common time-varying shocks. The estimated gap coefficients are statistically significant at either 5 % or 1 % threshold, respectively. The largest magnitudes of the gap coefficients are found for the regulatory quality and control of corruption dimensions, while the smallest ones are found for government effectiveness and voice and accountability dimensions.

Figure 9 reports the intertemporal distribution of p-values associated with the institutional quality effect of political instability. The evidence indicates the embedded presence of the institutional quality breakdown triggered by the rising instability. The associated p-values are both low and relatively stable over time with two distinctive observable characteristics. First, the p-values linked with the effect of instability on voice and accountability, regulatory quality, and control of corruption are within 10 % significance threshold immediately from the initial post-treatment year onward. Without the loss of generality, this implies that political instability has triggered a bold and substantial institutional quality breakdown through the decline of liberal democracy, deteriorating quality of regulation and significantly weakened control of corruption. The negative effect of political instability on these particular quality dimensions is both immediate and apparent. And second, the p-values on the rule of law and government effectiveness dimensions appear to be initially high and gradually become low, falling well within the 10 % threshold. This implies that the negative effect of instability does not materialize immediately but tends to loom large over time in a more gradual fashion. Hence, it can be implied that the deterioration of the rule of law and government effectiveness in response to political instability appears to be time-contingent and substantially more gradual compared to the immediate negative effects on corruption, regulatory quality, and voice and accountability. Taken altogether, the evidence offers ample support for the notion of political instability leading to the institutional quality breakdown with a heterogeneous pattern across the rule of law and government effectiveness on one hand and corruption–regulation–accountability dimensions on the other hand.

Figure 9:

Lead-specific p-values of the institutional quality effects of political instability.

5.4 Discussion

Perhaps the most obvious question concerning the validity of our results concerns the mechanisms through which either gradual or permanent deterioration in the specific institutional quality indicator has materialized over time. Although a more exhaustive analysis would harness transmission mechanisms in a more rigorous quantitative manner. For the sake of data limitations, many variables capturing possible mechanisms such as public trust, fiscal spending, and pressure of interest groups cannot be constructed at the local and regional level. For this reason, a more nuanced and plausible qualitative assessment can be buttressed.

The deterioration of institutional quality can be seen through the erosion of the freedom of expression and civil liberties (De Haan and Siermann 1996), more distortionary regulation (Daude and Stein 2007), and more widespread power abuses (Damania et al. 2004) alongside a steady and more gradual erosion of the rule of law (Helmke and Rosenbluth 2009). Several possible channels of transmission can be established. First, the political instability has eroded the trust in public institutions. The assessment of trust in public institutions by OECD in 2015 has shown that Slovenia has had the lowest level of trust in both political and legal system in the OECD together with Portugal, Greece, and Spain.^[10] By eroding trust in public institutions, political instability has decimated the confidence of citizens and firms in the institutional framework, curtailed government integrity, further weakening the institutional fabric, which has first-order implications for both quality and economic outcomes such as per capita income, total factor productivity, innovation, and unemployment, among several others.

Second, the onset of political instability has exacerbated the institutional backsliding that is perhaps best exemplified by the deterioration of economic freedom.^[11] Country-level assessment of economic freedom by Kim (2023) suggests that the level of economic freedom in Slovenia after the beginning of political instability has consistently declined. The estimated drop in the Heritage Foundation’s economic freedom index between 2009 and 2017 is around 5 basis point, which represents the largest deterioration among OECD member states. The largest drop has been was observed with respect to the integrity of government administration and tax burden. Similar evidence of deterioration of the quality of economic institutions is reported in Fraser’s index of economic freedom (Gwartney et al. 2022). In response to political instability, prolonged worsening of economic institutions has led to stalled economic and structural reforms associated with austerity measures after the economic and financial crisis, which have been repeatedly resisted and blocked by powerful interest groups, positing a source of coordinated disruption and other less institutional forms of policy influence (Scartascini and Tommasi 2012) that appear costly to economic growth, R&D, innovation and creative destruction (Aghion and Howitt 1992). The decline in government integrity further worsened both fiscal health and stability. Prior to the political instability, Slovenia has maintained both relatively balanced budget without continued deficit spending and one of the public debt-to-GDP ratios in the OECD. In particular, the share of public debt in GDP increased from 21 % in 2008 to 82 % in 2015, respectively, while government net borrowing spiraled from a balanced budget in 2007 into the fiscal deficit of 14 % of GDP in 2015.

A combination of soaring public debt and prolonged deficit-based spending proved costly to economic growth and macroeconomic stability (Rodrik 2018), especially in the presence of weak and unstable political environment. It reflected the sizeable losses of economic growth and institutional quality emanating from the institutional environment based on clientelistic networks and dominated by small elites and widespread cronyism in various branches of government and public sector, especially public procurement, health care, as well as across the private sector where the interplay between widespread state ownership and interference has been particularly pervasive. Political instability also eroded the incentives for institutional reforms and modernization. The gridlock imposed by the weak and unstable political environment has been particularly noteworthy in the judiciary. A series of reports by World Bank’s Doing Business projects emphasized long and cumbersome judicial delays as one of the major stumbling blocks against private sector development. In particular, together Italy and Cyprus, Slovenia’s judiciary has been characterized by long delays that have consistently exceeded 1000 days. Alternative estimates by Council of Europe’s European Commission for the Efficiency of Justice (CEPEJ) also indicate and confirm long judicial delays in Slovenia’s judiciary as one of the highest ones in Europe. The erosion of judicial efficiency further reinforced a vicious cycle of low public trust in the judiciary, increased transaction costs, and ameliorated the resistance to institutional reform by powerful interest groups in the judiciary. In turn, the interplay between deteriorating economic conditions and worsening judicial effectiveness in the presence of political instability and low social trust increased the demand for regulation, taxes, and redistribution (Alesina and Giavazzi 2008).

The increased demand for over-regulation and state interference (Gratton et al. 2021) posited a series of major disincentives to innovate and engage in various other forms of productive behavior. By increasing the demand for regulation, taxes, and redistribution, political instability further amplified the power of vested interest groups by incentivizing powerful actors, particularly labor unions and public sector employees to resist regulatory and institutional reforms, austerity measures, and privatization of state-owned enterprises through coordinated disruption, lobbying, influence peddling, and other forms of entry barriers (Parente and Prescott 2002). It should be noted that at the height of political crisis and gridlock, state-owned enterprises’ share of employment represented almost 10 % of overall employment, one of the highest ratios among OECD member states.

Furthermore, the increased demand for regulation and redistribution encouraged a deep co-optation between the economic and political elites where the hostility and aversion toward the reforms became a mode of political survival and ignited social conflict (Ippolito and Cicatiello 2019). The hostility toward reforms raised the overall cost of doing business, leading to a disproportionate drop of foreign direct investment, which prolonged both the length and severity of the economic downturn (Kim 2010). By inducing a permanent institutional sclerosis, political instability further eroded both medium-term and long-term fiscal sustainability, worsening adverse demographic pressures arising from rapidly aging population and high implicit liabilities of the pension system (Röth et al. 2018). At the same time, the political consequences of instability should not be neglected (Guriev 2018). A vicious cycle between political instability and deteriorating economic conditions fueled the demand for both right-wing and left-wing populism (Boeri et al. 2021), leading to the runoff victory of euroskeptic and right-wing Slovenian Democratic Party in 2020 (Taggart and Pirro 2021). Hence, instability helped to erode institutional quality particularly through the erosion of postmodern cosmopolitan liberal values (Beugelsdijk et al. 2022), increased political polarization, and undermined social cohesion coupled with distrust in science and less effective combat of COVID19 pandemic (Algan et al. 2021) that proved costly not only to institutional quality but also to human capital accumulation, talent acquisition, research and development, public health, and economic growth.