%20-%20https://sketch.com%20--%3e%3ctitle%3e3ECBBF50-FD08-412A-943D-2CFC5E025A16%3c/title%3e%3cdesc%3eCreated%20with%20sketchtool.%3c/desc%3e%3cg%20id='Stylesheet'%20stroke='none'%20stroke-width='1'%20fill='none'%20fill-rule='evenodd'%3e%3cg%20transform='translate(-926.000000,%20-668.000000)'%20id='BUTTONS'%3e%3cg%20transform='translate(100.000000,%20597.000000)'%3e%3cg%20id='BTN-/-RIGHTS'%20transform='translate(816.000000,%2065.000000)'%3e%3cg%20id='ICON-/-Rights'%20transform='translate(10.000000,%206.000000)'%3e%3cpath%20d='M13.5752841,0%20L1.34943182,0%20C0.59375,0%200,0.586666667%200,1.33333333%20L0,22.6666667%20C0,23.4133333%200.620738636,24%201.34943182,24%20L17.6505682,24%20C18.40625,24%2019,23.3866667%2019,22.6666667%20L19,5.36%20L13.5752841,0%20Z%20M13.5752841,1.89333333%20L17.0838068,5.36%20L13.5752841,5.36%20L13.5752841,1.89333333%20Z%20M1.34943182,22.6666667%20L1.34943182,1.33333333%20L12.1988636,1.33333333%20L12.1988636,5.36%20C12.1988636,6.10666667%2012.7926136,6.69333333%2013.5482955,6.69333333%20L17.6235795,6.69333333%20L17.6505682,22.6666667%20L1.34943182,22.6666667%20Z'%20id='Shape'%20fill='%23595959'%20fill-rule='nonzero'%3e%3c/path%3e%3ccircle%20id='Oval'%20stroke='%23595959'%20stroke-width='1.5'%20fill='%23595959'%20cx='9.5'%20cy='13.5355339'%20r='5.5'%3e%3c/circle%3e%3cpolyline%20id='Path-2'%20stroke='%23FFFFFF'%20stroke-width='1.5'%20transform='translate(9.500000,%2013.035534)%20rotate(45.000000)%20translate(-9.500000,%20-13.035534)%20'%20points='8%2016.0355339%2011%2016.0355339%2011%2010.0355339'%3e%3c/polyline%3e%3c/g%3e%3c/g%3e%3c/g%3e%3c/g%3e%3c/g%3e%3c/svg%3e)
It has become all but a truism in Political Science that painful economic reform induces societal backlash. In practice, this has most commonly referred more specifically to the consequences of imposing market-oriented policies of austerity, deregulation, privatization, and trade and financial opening. These are widely seen as economically and politically costly in the short run, risking societal backlash – which induces pro-reform politicians to marginalize opponents, insulate bureaucracies, and/or move to render changes quick and irreversible (Aslund Reference Aslund2002; Przeworski Reference Przeworski1991; Williamson Reference Williamson1990; for empirical studies see, inter alia, Almeida and Walker Reference Almeida and Walker2006; Arce and Bellinger Reference Arce and Bellinger2007; Bellinger and Arce Reference Bellinger and Arce2011; Coxshall Reference Coxshall2010; Simmons Reference Simmons2016; Espinal Reference Espinal1995; North Reference North2000; Walton and Seddon Reference Walton and Seddon1994).
But what if this truism – that painful liberal reform causes protest – isn't so true? We contend that the existing consensus rests on a flawed methodological foundation, and that not only does liberal reform not necessarily induce backlash, but mobilization in response is actually less vigorous than would be expected given ex ante economic and social conditions that gave rise to liberalization in the first place. This is not a paradoxical testament to the popularity of such reforms, or of ‘hard times’ more generally. Rather, it suggests that the reforms themselves may induce dynamics that undermine collective action even in the face of extremely powerful grievances. If accurate, the implications are wide ranging, for reforms have often been implemented in ways that degrade democratic practice in order to mitigate potential backlash (for example, Hellman Reference Hellman1998; Stokes Reference Stokes2001).
In this short article we will thus treat what is often assumed – that painful reforms induce societal backlash – as an empirical question. It is not, however, an easy one to answer. The data we must use is observational, the treatment (reform) is often induced by the very economic pain it also tends to heighten in the short run, and the history of reform is likely to affect its subsequent consequences. Given these constraints, our task is to produce as clean an estimate of the effect of reform on protest outcomes as possible. Our approach utilizes a conservative design strategy to disentangle the causal effects of liberalization from its confounders. We begin by constraining the analysis to the Latin American region from the 1970–2003. This region was selected in part because economic liberalization there did not have the favorable valence it had, for example, in the post-communist world where it was directly associated with political democratization. That reform was painful and widely unpopular in Latin America is crucial, as it sets out a particularly conservative test for our demobilization claim.Footnote 1 We elaborate on this design strategy below.
We disagree with the conventional wisdom that liberalization induces protest, as the studies that demonstrate such findings suffer from fundamental methodological flaws. Why is this so? First, the available observational data on the political consequences of liberalization in Latin America are the result of a data-generating process in the real world that has produced simultaneous trends toward both economic liberalization and the initiation of democratic politics (and thus the reduction of repressive control over protest this implies). The latter trend may well be confounding the former (Solt et al. Reference Solt2014). More concerning, however, is that in the real world the assignment to treatment (that is, economic liberalization) was far from random: in almost all cases it was a consequence of disastrous ex ante economic conditions. But hard times and protest are unlikely to be unrelated. Further complicating the analysis, the same economic hardships that drove liberalization policies are also in part consequences of economic liberalization – which typically produces job losses attendant on privatization, increased current account deficits, severe fiscal retrenchment, devaluation and real wage declines (and thus falling living standards), surging poverty, and rising inequality. There is no easy divide between pre- and post-treatment here, which confounds the analysis. Finally, previous empirical work implicitly assumes that protest outcomes are independent of each other, at least up to the level of country fixed effects. But if what we have learned from the social movement literature is accurate, and collective action builds on prior histories with collective action and/or involves the creation of repertoires that emerge and are employed over time, this within-country ahistorical assumption is troubling. Moreover, it ignores the potential effects of the history of treatment itself as it applies to these and other dynamics in each country. We seek to address these challenges.
Observational Data and the Liberalization/Protest Nexus
The principal difficulties associated with properly assessing the relationship between economic liberalization and protest thus begin with two different types of selection problems. The first, is the tendency at least in the case-based literature to focus on ‘positive’ cases; ‘non-mobilization’ is rarely the focus of study. This, of course, is problematic if at least one of the inferential goals is to discover whether liberalization induces protest. Of course, shifting to a cross-national frame helps address this problem by scoring the universe of cases, and thus allowing us to at least calculate a frequency of protest given liberalization and compare it to cases where liberalization was not undertaken. This is what the cross-national econometric work has basically accomplished so far – relying on the assumption that proper model specification has resolved issues of omitted variable bias and confounds, and that further selection problems do not remain.
Sadly, even utilizing all available data in the relevant period (from before liberalization in the 1970s until after the end of the major liberalization episodes in the early 2000s) is not nearly enough to facilitate a proper assessment of the impact of liberalization. The problem begins with the distribution of the data. It is simply a historical fact that economic liberalization and continent-wide transitions to democracy in Latin America occurred at similar times: both experienced sharp gains in the 1980s and 1990s (see Figure 1). This poses a challenge for ordinary estimation, since political freedom is undoubtedly a key factor in explaining the observed level of political protest, the dependent variable. As a consequence, the causal effect we are interested in is the effect of economic reform conditional on the political regime. Indeed, especially in the first half of the dataset there is little information on essential counterfactuals for both the critical regime and liberalization variables: the cases cluster heavily in the autocratic/statist pairing early on and trend toward the liberalized/democratic pair as time passes. Thus, with both covariates trending together, and relatively few observations having solid common support, regression-based inferences can be unreliable at best, and seriously biased at worst.
Economic reform and democracy, averages
We are left with a problem that is difficult to resolve using the available data: whether economic reform induces protest because of the social and economic dislocations it induces, or whether protest occurs simply because the authoritarian constraints that perhaps formerly held it in check have begun to lessen. For instance, what would the absence of protest tell us in a sharply authoritarian context? Consider that between 1973 and 1979 Chile implemented a profound wave of economic liberalization, which brought with it hyperinflation, mass unemployment and catastrophic declines in living standards. Yet there was comparatively little public dissent during this era. Is this evidence of the demobilizing effects of liberalization (as per the depoliticization literature), or the simple effect of savage repression, or both?
A traditional approach to dealing with the lack of common support in the data would be to utilize some form of matching. But this would still be problematic in this context, for both the data we utilize and the causal processes that generate political protest are dynamic. The first problem that this portends is that while some critical covariates – notably the economic consequences of liberalization in terms of growth rates, the balance of payments, GDP per capita – are in one sense post-treatment consequences of economic reform, but in a dynamic sense, however, they also affect the selection into treatment. Indeed, politicians rarely, if ever, impose costly strategies in good times when growth is strong, inflation is low, and workers can find employment. Neither the political will nor the political support to launch a painful adjustment strategy could be found in that context. Instead, sharp turns toward economic liberalization almost always take place in hard times. For instance, they have been launched after a hyper-inflationary episode (as in Peru, Argentina, Bolivia and Brazil), after multiple efforts to combat economic downturns in less recessionary ways (as in Mexico, Argentina, and Peru in the late 1980s), or even more strikingly after an almost complete collapse of market mechanisms, as in Chile after 1973. If such a selection into treatment is taking place, it would tend to artificially induce a correlation between protest and liberalization. For if hard times breed protest and predict moves toward economic liberalization (which they do in ex ante statist contexts like Latin America in the 1970s and early 1980s), it becomes difficult to separate the causal effect of economic reform on protest that operates through the additional pain it imposes, over and above the discontent and dissent generated by antecedent economic circumstances that are already nearing – or at – levels of economic catastrophe.
There is a further sense in which dynamic effects undermine efforts to estimate the causal effect of liberalization on protest. For it is a commonplace in the social movements literature (for example, McAdam et al. Reference McAdam, McCarthy and Zald1996; Tilly Reference Tilly1978; Tilly Reference Tilly, Castañeda and Schneider2017) that legacies of antecedent protest have a powerful effect on subsequent mobilization. This can happen because protest repertoires are created and learned, new community organizations or practices that are the locus of organization and mobilization develop, and/or because the human capital and organizational skills required to initiate and sustain mobilization are learned through participation in antecedent activity. And these are very much related to the history of treatment with economic reform.
Research Design and Methods
Our strategy for overcoming these methodological challenges involves using marginal structural models (Blackwell Reference Blackwell2013; Robins, Hernán and Brumback Reference Robins, Hernán and Brumback2000) to estimate the dynamic effects of liberalization on protest activities. Such models are an extension of propensity score matching that weight on the probability of treatment at time t, and the probability of treatment in the past, at t − 1, which allows us to capture the dynamic decision making associated with both the choice to liberalize, and then the probability of protest, given the history of liberalization.
Selection of Cases and Period
Our dataset covers 1970–2003, and includes seventeen Central and South American countries.Footnote 2 The outcome of interest is the number of protest events, and the treatment of interest is economic liberalization. These cases were carefully selected based on three criteria. First, we utilize the same cases, time period, and core data sources that were employed in prior studies that produced diametrically opposed findings to our own.Footnote 3 Our aim is to provide a directly comparable contrast to highlight the role that improved estimation has on the results and exclude the possibility that these are driven by changes to data sources, time periods, or case coverage.
Secondly, we focus on the Latin American region specifically, for it is here that economic liberalization and democratization are not conceptually intertwined. Were we to include, for instance, post-communist cases, where economic reform was clearly seen as part and parcel of a democratization project, we might well observe less protest than the attendant economic pain would suggest if citizens saw it as part of the cost of obtaining political freedom. But this would be a bias in favor of the demobilization finding, and we choose a more challenging test. If anything, in Latin America, where economic liberalization has long been associated with unpopular pressures emanating from the United States and the international financial institutions, protest is to be expected. Finally, we focus on the 1970–2003 time period in that region for two reasons: (1) to remain consistent with the existing literature and (2) because this is the period in which liberal reform was enacted and largely became consolidated throughout the region. To continue the data forward in time would be to assume that protest outcomes in the recent past are a consequence of liberalization decisions taken primarily in the 1980s and 1990s.Footnote 4 When the process of liberalization has exited the political agenda, it would be a stretch to imply that protest outcomes are driven by it, and given the fact that basic liberalization was implemented nearly everywhere in the region by the early 2000s, we end the analysis in 2003.
We binarize the [0,1] scale used for economic liberalization in prior studies, first computing the average liberalization score across all country years, and then, treating those countries greater than this mean as liberalized, and those that are less than this mean as not economically liberalized. We therefore look at the effect of liberalization in expectation, the expected value of the treatment, A, on the outcome, Y. We chose a binary treatment partly out of necessity, as marginal structural models that can reliably estimate the effects of continuous treatments are not yet available. It is also, however, a more conservative test of our theory. By binarizing the treatment, we average the results over weaker and stronger treatments in the data (in the sense of greater or lesser liberalization); the more cases that are found in the data that are close to the dividing point for this binarization, the weaker the estimated effect would be. As a consequence, the bias would be toward finding reduced treatment effects, and a greater likelihood of a null result. The fact that we show results despite binarization increases our confidence that there is in fact a demobilization effect, as the data reduction in the former is biased toward no result.Footnote 5
We then use a set of standard covariates to estimate the propensity for economic liberalization: the balance of payments, the current account balance, GDP per capita, economic growth rates, whether a country is a democracy, and country-level fixed effects.Footnote 6 We selected these economic covariates because they embody the crucial dilemma that we seek to overcome. They are macroeconomic characteristics that are known to be profoundly affected by economic liberalization, and that simultaneously are widely thought to provoke mobilized responses by citizens when they indicate economic downturn. Unfortunately, they are also important predictors of selection into a process of economic reform in the first place. Our marginal structural models approach is designed to correctly estimate the treatment effect of liberalization in observational data like this, where traditional approaches would founder since they are unable to correctly estimate the effect of liberalization given critical covariates that are both pre- and post-treatment.
Causal Inference with Longitudinal Data
Understanding the effect of economic liberalization on political protest requires asking a causal question: in the absence of liberalization, would the number of protests be significantly different? In an ‘ideal’ world, we could randomly assign polities to undergo economic liberalization, observe the counts of protests in treated and untreated groups, and subtract. Because economic liberalization is not random, we must model each country's propensity to liberalize in order to make causal claims based on this sort of observational data (see Morgan and Winship Reference Morgan and Winship2014).
One additional difficulty for the inferences is the role of time. Liberalization is not only the result of economic conditions in the current year, but also, of the previous year. This requires estimating E[Y|A t, X t, A t−1, X t−1], where A is the treatment, liberalization, and X are observed covariates, which contrasts with the standard E[Y|A t, X]. We can represent A t, X t, A t−1, X t−1 as 
$\bar{A}$ for convenience. Simply put, this means that estimating the propensity for treatment requires including the entire relevant treatment history and covariate history; ignoring values at t – 1 introduces omitted variable bias, and simply including all the treatments and covariates in a regression introduces post-treatment bias (Blackwell Reference Blackwell2013).
To estimate the effect of treatment history on the outcome, we thus also rely on marginal structural models, which provide a way to estimate the effect of treatment history on the outcome at an observed endpoint, without introducing omitted-variable or post-treatment bias (Robins, Hernán and Brumback Reference Robins, Hernán and Brumback2000). Central to the idea of the marginal structural model is the inverse propensity score weight, which allows us to weight observations according to their propensity for treatment. Rather than simply estimating pr(liberalization), we estimate 
$IPW\left( A \right) = E\left[ A \right]/pr\left( {{\rm liberalization}} \right)_t$ for the treated observations, and IPW(A) = 1 − E[A]/1 − pr(liberalization)t for the untreated observations. We estimate these weights for each year, and then multiply these weights together, to calculate the per-country inverse propensity score during the period under study.
To estimate the treatment model, we use Bayesian Additive Regression Trees (BART), a weakly additive sum-of-trees model, with a regularizing prior (Chipman, George, and McCulloch Reference Chipman, George and McCulloch2010). We use this model because the tree structure allows for nonlinearities in the data, and the model does not make strong parametric assumptions. The regularization from the Bayesian priors helps estimation here, as there are relatively few observations, but many covariates.
Modeling the effect of liberalization on protest, we focus on the effect of recent liberalization for each country-year, and sum over the rest of the treatment history (following Blackwell and Glynn Reference Blackwell and Glynn2018), so the outcome model becomes 
$Y = \left({\beta_0 + \beta_1A_t + \beta_2A_{t-1} + \beta_3\left({\mathop \sum \nolimits_{t-2}^T } \right)} \right)\times \bar{A}$, where we regress the number of protests on the treatment at time t, the treatment at t – 1 and the summed previous treatment history, weighted by the inverse probability of treatment weights 
$\bar{A}$. The dependent variable here is a count of protest events in each country-year, and we thus use two models to estimate this outcome equation: a standard poisson model, and a zero-inflated poisson regression, as there are a large number of zeros in the outcome data. We observe four counterfactual treatment outcomes from this model: Y 1,1, Y 1,0, Y 0,1, Y 0,0; treatment during both periods, treatment only at time t, treatment only at t – 1, and no treatment. To return valid standard errors, we then bootstrap the estimates, sampling by country.
Results and Discussion
Table 1 presents the results, and Figure 2 plots the coefficients. The results from the Poisson model and the Poisson component of the zero-inflated Poisson are shown, and are substantively similar. Across both models, the treatment effects are statistically significant when considering the liberalization in the current year and previous year, and when considering only the contemporaneous treatment. The average treatment effect is negative, and for the effect at A T + A T−1, it is consistent across both models.Footnote 7
Average treatmeant effects from Poisson and zero-inflated Poisson marginal structural models
Note: black lines denote 95 per cent confidence intervals, and grey lines 90 per cent confidence intervals.
Average treatment effects
Note: table includes bootstrapped 95 per cent confidence intervals.
What do these finding tell us? In practical terms, they suggest that mass political resistance to economic liberalization is less than expected given the generally harsh conditions in which it is typically imposed, and the economic pain that typically follows it. It does not tell us that liberal reform is popular; indeed, there is substantial evidence that at least in Latin America it generally is not. Nor does it tell us there is only a trivial protest response to harsh economic reform. For, of course, scholars are quite right to have pointed out the important instances of civil society mobilization against liberalization. Our argument is different: on the margins, liberal reforms are consistent with notably lower levels of societal resistance than one would expect given the level of economic dislocation and societal disruption they induce.
Of course, such a finding is compelling only insofar as the effort to estimate it is derived from a plausible causal mechanism. The demobilization claim we evaluate here has been tied by its proponents to two different forms of collective action problems: organizational destruction and the loss of a focal point for grievance (see, for example, Garretón Reference Garretón, O'Donnell, Schmitter and Whitehead1986; Kurtz Reference Kurtz2004a Kurtz Reference Kurtz2004b; Wolff Reference Wolff2005). With respect to the former, protest against economic grievances in Latin America had, in the era before liberalization, long been driven by institutionalized actors – most typically, labor unions and professional associations. Leadership skills, organizational talent, logistical and financial support, etc. were concentrated in these groups. But these groups were also among the principal victims of economic crisis and economic liberalization. The massive waves of privatization, bankruptcy, and de-industrialization produced as liberalization took hold struck at the membership core of the trade unions. At the same time, a massive rise of informality in the labor force made it difficult to reconstruct organizations, for now individuals largely operated separately from a large-scale employer, with no easy locus of organization or any sort of legal infrastructure available to support collective action.
Secondly, economic liberalization also marked a decisive shift in the potential target of any mobilization. In the era of protected industrialization and extensive industrial policy, it was obvious that the state was a crucial arbiter of economic outcomes, and as such a ripe target for mobilization. By contrast, after authority over prices and investment shifted to decentralized private markets, and competition came from both outside and inside the economy (as tariff barriers came down), malign outcomes experienced by individuals were no longer so obviously a product of the state's choices.
We believe it is very important to take the potential demobilizing consequences of liberalization seriously. A large body of literature has assumed the inevitability of powerful resistance to liberalization, and has focused on the institutional and other mechanisms that can be used to ‘ride it out.’ These range from ‘shock therapy’ approaches of rapid, hard-to-reverse reform, to institutional structures designed to insulate decision makers from popular (or legislative) resistance, to practices that directly undermine democratic accountability. But if we are right, these assumptions are miscast, and these often counter-democratic practices are unnecessary even if one believes liberal reform to be imperative. We believe that these results suggest that in the future, much greater empirical attention should be paid to the ways in which policy changes may induce or reduce collective action problems that change the ability of societal actors to resist them.
Data availability statement
Data replication files are available in Harvard Dataverse at: https://doi.org/10.7910/DVN/WU2WSE.
Acknowledgements
The authors would like to thank Andrew Schrank and Kenneth Shadlen, as well as the Comparative Politics Research Workshop at Ohio State University for insightful comments on earlier drafts.
'%3e%3cpath%20fill='%23595959'%20fill-rule='nonzero'%20d='M14.165%2014.217c-.955%200-1.881.376-2.547%201.071L9.07%2017.835C8.376%2018.53%208%2019.427%208%2020.382c0%20.955.376%201.881%201.07%202.547.696.695%201.593%201.071%202.548%201.071.955%200%201.881-.376%202.547-1.07l2.547-2.548c.695-.695%201.07-1.592%201.07-2.547%200-.724-.202-1.418-.607-1.997l2.663-2.663c.579.376%201.273.608%201.997.608.955%200%201.881-.376%202.547-1.071l2.547-2.547C27.624%209.47%2028%208.573%2028%207.618c0-.955-.376-1.881-1.07-2.547C26.233%204.376%2025.336%204%2024.381%204c-.955%200-1.881.376-2.547%201.07l-2.547%202.548c-.695.695-1.07%201.592-1.07%202.547%200%20.724.202%201.418.607%201.997l-2.663%202.663c-.579-.405-1.273-.608-1.997-.608zm6.165-5.586l2.547-2.547c.405-.405.955-.637%201.534-.637.579%200%201.129.232%201.534.637s.637.955.637%201.534c0%20.579-.232%201.129-.637%201.534l-2.547%202.547c-.405.405-.955.637-1.534.637-.318%200-.637-.087-.926-.203l.926-.926c.29-.29.29-.753%200-1.013-.29-.29-.753-.29-1.013%200l-.926.926c-.145-.29-.203-.608-.203-.926-.029-.608.174-1.158.608-1.563zm-4.631%2010.738l-2.547%202.547c-.405.405-.955.637-1.534.637-.579%200-1.129-.232-1.534-.637s-.637-.955-.637-1.534c0-.579.232-1.129.637-1.534l2.547-2.547c.405-.405.955-.637%201.534-.637.318%200%20.637.087.926.203l-.926.955c-.29.29-.29.753%200%201.013.145.145.318.203.521.203s.376-.058.52-.203l.927-.926c.145.29.203.608.203.926%200%20.579-.203%201.129-.637%201.534z'/%3e%3crect%20width='33.5'%20height='26.5'%20x='.75'%20y='.75'%20stroke='%23595959'%20stroke-width='1.5'%20rx='3'/%3e%3c/g%3e%3c/svg%3e)
