Evaluating Russian Economic Growth without the Revolution of 1917

Transcription

1 Evaluating Russian Economic Growth without the Revolution of 1917 Ivan Korolev July 5, 2017 Abstract This paper uses modern econometric techniques, such as the lasso and the synthetic control method, to construct the counterfactual GDP per capita series for Russia for The goal of this paper is twofold: first, to predict how the Russian economy might have developed without the Revolution; second, to evaluate and compare various econometric methods for computing the counterfactual GDP per capita series. The counterfactuals based on the preferred method, the synthetic control, suggest that without the Revolution Russia might have grown at about 1.6% a year in I am grateful to Ran Abramitzky and Frank Wolak for continuous support and encouragement. I also thank Simeon Djankov, Sergei Guriev, Caroline Hoxby, Andrei Markevich, and Hans-Joachim Voth for valuable comments and discussions. I gratefully acknowledge the financial support from the Stanford Graduate Fellowship Fund as a Koret Fellow and from the Stanford Institute for Economic Policy Research as a B.F. Haley and E.S. Shaw Fellow. All remaining errors are mine. Department of Economics, Stanford University, 579 Serra Mall, Stanford, CA, ikorolev@stanford.edu. Website: 1

2 1 Introduction A major question in Russian economic history and a subject of ongoing public debate in today s Russia 1 is how economic growth in the Soviet Union in s would have compared to counterfactual growth of Russia if the Russian Revolution had not happened in Some people believe that Soviet industrialization allowed the Soviet Union to grow faster than ever before and to achieve what otherwise would have been impossible, while others argue that fast economic growth of 1930s was just a return to the pre-revolution trend and that Tsarist Russia would have achieved the same level of economic development. There are several studies that try to understand the consequences of the various aspects of the Revolution on economic development of Russia. In particular, Hunter and Szyrmer (2014) evaluate Stalin s economic policies in using a multi-sector and multi-period linear model; Allen (2003) looks at Russian industrialization that took place in the beginning of 1930s and tries to understand what would have happened without centralized industrialization; Cheremukhin et al. (2017) look at the effects of Stalin s industrialization on economic growth in Russia using a two-sector macroeconomic model. These studies mostly relied on theory-based simulations, which has its advantages and disadvantages. On the one hand, this approach allows the authors to analyze particular policy changes and consider a wide range of scenarios; on the other hand, it might be difficult to model all factors that are important for economic growth (e.g. productivity, technology, international trade, and so on) together. This paper complements the existing literature by studying the consequences of the Russian Revolution of 1917 using a data-driven approach. The goal of this paper is twofold. First, I construct the counterfactual series of Russian GDP per capita from 1917 on, and hence I am able to predict how Russia might have developed without the Revolution of Second, I evaluate various econometric methods that could be used to construct the counterfactual series, and then I provide insights about their performance that can be useful in 1 For a brief summary of Russians opinion on Stalin see, e.g., russia-joseph-stalin-victory-day-opinions-contributors-cathy-young.html. 2

3 a variety of other settings, especially when the number of potential control variables is large relative to the sample size. My results can be viewed as a credible scenario of what might have happened without the Revolution, but not as an estimate of the causal effect of the Revolution. Essentially, I predict Russian GDP per capita after 1917 based on the pre-1917 data for Russia and both pre-1917 and post-1917 data for other countries. Even though I will routinely use the notion of counterfactual series for alternative scenarios and forecasts of Russian GDP per capita throughout the paper, it should be clear these counterfactuals do not answer any particular causal questions. Since the Revolution took place in Russia because of certain political, economic, and social issues, I cannot measure the causal effect of the Revolution by comparing Russia to other countries. Moreover, the Revolution consisted of several events and included various policy changes, and my methods do not allow me to analyze particular events separately. Instead, my counterfactuals can be viewed as an attempt to find a stable predictive relationship between Russia and other countries before 1917 and, assuming it would have remained unchanged after 1917, construct predictions according to this relationship. For instance, one might think that because of international trade, technology spillovers, or other causes the growth rates in different countries are interconnected. Then I can recover the relationship between the Russian economy and economies of other countries using econometric methods, without modeling this relationship explicitly, and then I can make forecasts based on this predictive relationship. In order to do so, I need to assume that the predictive relationship was stable over time and would not have changed after 1917 if the Revolution had not happened. I try to validate this assumption by conducting placebo tests, in which I construct the counterfactual GDP per capita series for the countries that did not experience a revolution and show that these counterfactual series are reasonably close to the actual series. I use two methods to estimate the predictive relationship between Russia and other coun- 3

4 tries. The first method is based on running a usual OLS regression of Russian GDP per capita on GDP per capita of other countries using the pre-1917 data and then constructing the forecast of Russian GDP per capita after 1917 as fitted values. However, the problem with this approach is that I do not have enough years before the Revolution to run an unrestricted regression: I have 32 countries in the control group and only 32 observations before the Revolution. Hence, I use the lasso to select the best regressors and then run an OLS regression with only the selected regressors to construct the counterfactual series as the fitted values from this OLS regression. From now on, I will call this method simply the lasso. Another method is a recently developed synthetic control method. This method has been used in Abadie and Gardeazabal (2003), Abadie et al. (2010), and Pinotti (2012). For more a more theoretical discussion of the synthetic control method and its relation to other approaches, see Abadie et al. (2015) and Doudchenko and Imbens (2016). The idea of this approach is to use the pre-treatment data to find a so-called synthetic control, a counterfactual observation that had been most similar to the treated observation before the treatment took place, and then to look at how this counterfactual observation would have behaved after the treatment. Here I am interested in counterfactual development of Tsarist Russia without the Revolution, so the treatment is essentially the Russian Revolution that took place in I compare these two methods using placebo tests that involve other countries, where revolutions did not happen, so that I know how the true growth path without revolutions looks like, or Russia before I discuss these placebo tests in detail in Subsection 4.1. The remainder of the paper is organized as follows. Section 2 gives historical background of the Russian Revolution. Section 3 describes various approaches I use in this paper. Section 4 describes the way to evaluate and compare various methods and then presents the main findings. Section 5 concludes. 4

5 2 Background In this section I give a brief overview of the historical context of the Russian Revolution of In the second half of the 19th century and beginning of the 20th century Russia was an agrarian country, poorer than other European countries and approximately as rich as Japan. Russia was facing economic and political challenges that, in particular, led to the Revolution of 1905, which, however, did not result in dramatic political changes. To be precise, the Revolution of 1905 led to a constitutional reform, but Russia remained a monarchy and Tsar Nicholas II stayed in power. The government started to implement large-scale agrarian reforms in 1906, but these efforts arguably were cut short by World War I in Despite political and economic challenges, the economy was steadily growing before World War I. In 1914 Russia entered World War I, which led to more serious economic, social, and political problems, and in 1917 the Revolution took place. In fact, the Russian Revolution of 1917 consisted of two revolutions. First, in March 1917, Tsar Nicholas II was forced to abdicate and the provisional government was formed. Then, in November 1917 there was another revolution, and as a result, the Bolsheviks came to power, replacing the provisional government. The Civil War (loosely speaking, the war between the Bolsheviks Red Army and the Anti-Bolsheviks White Army) started as a result of the Revolution. There is no universal agreement among historians on the exact dates of the Civil War: historians believe it started in November 1917 or in the summer of 1918 and ended between 1920 and 1922 (see Bradley (1975), Mawdsley (2007), Bullock (2008)). The Bolsheviks won the war, and the Soviet Union was created in December During 1920s, the economic and political systems in the Soviet Union became more and more centralized and controlled by the government, and by the beginning of 1930s the Soviet Union became a country with planned economy. Starting in 1928, the government attempted a centralized industrialization in order to quickly turn an agrarian country into an industrialized one. The cost of industrialization was high: according to Wheatcroft and Davies (2004), between 5.5 and 6.5 million people died during the famine of ; then, according 5

6 to Pipes (2001), about 680,000 people were executed during political repressions known as the Great Purge in The benefits of the industrialization are still unclear: some historians (e.g. Allen (2003)) claim that it led to faster economic growth in the country, while others (e.g. Cheremukhin et al. (2017)) argue that the Russian economy would have grown at comparable rates even without the forced industrialization. Figure 1 plots the evolution of Russian GDP per capita over time in order to illustrate the main patterns of economic growth in Russia and the Soviet Union in As a benchmark, it also plots the evolution of American GDP per capita. I would like to stress that the GDP per capita data for Russia and the Soviet Union was constructed by Markevich and Harrison (2011) retrospectively. Hence, I am not concerned about possible attempts from the Soviet government to manipulate the data. Figure 1: Russian and American GDP per Capita, The natural logarithm of actual Russian GDP per capita is plotted as a solid line. The natural logarithm of American GDP per capita is plotted as a dashed line. The vertical line marks 1917, the year when the Revolution took place. In this paper I try to estimate the counterfactual GDP per capita in Russia had the Russian Revolution not taken place in 1917 and compare it with the actual GDP per capita. It is worth noting that I do not discuss whether it was possible for the Russian government 6

7 to avoid the Revolution. It might be the case that in order to avoid the Revolution, it would have been necessary to implement political or economic reforms that would have changed the Russian growth path significantly. 3 Data and Empirical Strategy The main source of data for my analysis is the updated version of the Maddison project, i.e. Bolt and van Zanden (2013), which has data on Russian GDP per capita from 1885 on as well as information about other countries. Table 1 lists all countries in the dataset. The construction of country characteristics and data sources are discussed in the table footnotes. The series of Russian GDP for , as reported in Bolt and van Zanden (2013), was constructed by Markevich and Harrison (2011). They built upon numerous previous studies, most notably Gregory (2004), to fill in the last gap in the Russian GDP series. An important concern about this data is that the territory and population of the country changed dramatically after the Revolution. However, Markevich and Harrison (2011) explicitly took these changes into account in order to construct a consistent time series of GDP per capita for the Russian Empire and then the Soviet Union. As usually done in time series econometrics, I mainly use a natural logarithm of GDP per capita in my analysis, because economists often believe that growth rates are stationary, and hence the first differences of the logarithm of GDP per capita are stationary as well. As for the forecast horizon, I focus on , because World War II started in 1939 and Germany invaded the Soviet Union in 1941, so that any forecasts or comparisons that go beyond that point do not make much sense. To summarize, I have data on , with 1917 being the year when the treatment took place, and my goal is to construct a counterfactual GDP per capita series for Russia from 1917 on. As I mentioned above, I can use several methods to do it. 7

8 Table 1: Countries Country Europe Independent Population, 1913 WW1 Losses WW1 Losses (thousands) (thousands) (% of population) Russia ,192 1, Austria 1 1 6, # Belgium 1 1 7, Denmark 1 1 2, Finland 1 0 3, # France ,463 1, Germany ,058 1, Italy , Netherlands 1 1 6, Norway 1 1 2, Sweden 1 1 5, Switzerland 1 1 3, UK , * Greece 1 1 5, Portugal 1 1 5, Spain , Australia 0 1 4, * New Zealand 0 1 1, * Canada 0 1 7, * USA , Argentina 0 1 7, Brazil , Chile 0 1 3, Colombia 0 1 5, Peru 0 1 4, Uruguay 0 1 1, Venezuela 0 1 2, India ,700 54* Indonesia , Japan , Sri Lanka 0 0 4, Europe equals 1 if a country is located in Europe and 0 otherwise. A majority of Russian population lived in Europe, so I treat Russia as an European country. Independent equals 1 if a country was independent throughout the entire period and 0 otherwise. I treat Australia, Canada, and New Zealand as independent because they had so-called responsible governments and arguably enjoyed significant amount of independence from the UK. Sweden and Norway formed a union under the Swedish monarch until 1905, so formally Norway was not an independent country until then. However, because these two countries still had separate laws, legislatures, armed forces, etc., I treat them as independent as well. Data on 1913 population is from Maddison ( WW1 Losses takes into account military losses only and comes from Encyclopedia Britannica ( /World-War-I/53172/Killed-wounded-and-missing). For the countries marked by * (that were parts of the British Empire), the data is from Ellis and Cox (2001), except India, for which the data is from Urlanis (1971). Austria and Finland, marked by #, did not exist in modern borders in For Austria, I use data from Erlikman (2004); for Finland, from International Labor Office ( ). 8

9 As a baseline, I use an ARIMA(1,1,0) model, i.e. estimate an AR(1) model in first differences: log(gdp pc ) t = α + log(gdp pc ) t 1 + ε t, where where t indexes time and then construct the predicted values from 1917 on. This approach would not allow me to account for possible structural breaks or for global macroeconomic factors that might have affected Russian economy, because essentially it consists of just continuing the trend that existed before the Revolution. However, if other methods cannot beat this method in placebo tests, it might indicate that other methods do not perform well. Because I have data on other countries, I can use it to predict Russian GDP as well. The first option is to run a cross-sectional regression of Russian GDP per capita on GDP per capita of other countries before 1917 and then construct the predicted GDP per capita after 1917 simply as fitted values from that regression using actual data for other countries: k log(gdp pc ) Russia,t = α + β j log(gdp pc ) j,t + ε t, where j indexes countries in the control group. It is also possible to combine these two methods and to estimate an ARMA model like p k q log(gdp pc ) Russia,t = α + ρ j log(gdp pc ) Russia,t j + β j log(gdp pc ) j,t + θ j ε t j + ε t j=1 j=1 j=1 j=1 The main challenge with this approach is that the number of the countries in the sample is almost as large as the number of the pre-revolution observations, so estimating such model does not make much sense: it is possible to fit the observed GDP almost perfectly with so many regressors. But such overfitting may lead to extremely inaccurate forecasts. However, I can overcome this challenge by constraining the regression coefficients. The first option to solve the overfitting problem is to use a lasso regression, first developed 9

10 in Tibshirani (1996). A lasso estimator is defined as follows: min β s.t. ( log(gdp pc ) Russia,t α t β j t It is equivalent to solving j ) 2 k β j log(gdp pc ) j,t j=1 min β ( log(gdp pc ) Russia,t α t k j=1 β j log(gdp pc ) j,t)2 + λ j β j, where λ is a penalty parameter. I choose the penalty parameter using leave-one-out crossvalidation for each country separately. A detailed discussion of this procedure is presented in Appendix A.1. The idea behind the lasso is to penalize for having too many regressors by zeroing some of the coefficients: because the augmented objective function is not smooth, the problem typically has a corner solution in which some of the parameters are set to zero. Hence, using the lasso helps to solve the problem of having insufficient number of observations in the data by zeroing out the coefficients on the variables with low explanatory power. In this paper I apply the lasso as follows: first, I normalize the control variables so that they all have zero mean and unit variance, then choose the penalty parameter as discussed in Appendix A.1, then run the lasso with this selected penalty parameter using the observations for After that, I keep only the selected countries and run OLS with these countries and a constant (again using ) in order to avoid biasing the coefficients towards zero (hence, I use the lasso for the selection but not for the estimation). Finally, I construct the fitted values for using the resulting estimates. In addition to running the lasso with only the current levels of GDP per capita in other countries as controls, I run it with other sets of controls as well. Table 2 lists all specification that I use. 10

11 Table 2: Control Variables for Lasso Specification Name Control Variables ( ) 31 Lasso 1 Baseline log(gdp pc ) j,t j=2 ( ) 31 log(gdp Lasso 2 Lags & Russia" pc ) j,t, log(gdp pc ) j,t 1, log(gdp pc ) j,t 2, j=2 log(gdp pc ) Russia,t 1, log(gdp pc ) ( Russia,t 2 ) 31 Lasso 3 Lags log(gdp pc ) j,t, log(gdp pc ) j,t 1, log(gdp pc ) j,t 2 ( j=2 log(gdp pc ) j,t, log(gdp pc ) j,t 1, log(gdp pc ) j,t 2, Lasso 4 Lags & Leads & Russia ) 31 log(gdp pc ) j,t+1, log(gdp pc ) j,t+2, Lasso 5 Lags & Leads j=2 ( log(gdp pc ) Russia,t 1, log(gdp pc ) Russia,t 2 log(gdp pc ) j,t, log(gdp pc ) j,t 1, log(gdp pc ) j,t 2, log(gdp pc ) j,t+1, log(gdp pc ) j,t+2 ) 31 j=2 The second option to solve the overfitting problem is to use the synthetic control method. It is somewhat similar to the lasso, but uses different restrictions on the coefficients. It finds a combination of control units (i.e other countries in the sample) with non-negative weights that sum up to one in a way that minimizes the weighted sum of squared differences between the treated and synthetic units during the pre-treatment period: min β s.t. w t (log(gdp pc ) Russia,t t β j = 1, β j 0 j j ) 2 k β j log(gdp pc ) j,t The synthetic control method is similar to the lasso in the sense that it also imposes constraints on the ordinary or weighted least squares coefficients, but the constraints are j=1 more demanding. Since the synthetic control method requires all coefficients to be nonnegative and sum up to one, the resulting synthetic control unit can be interpreted as a weighted average of the members of the control group. The synthetic control method has its advantages and disadvantages. The main advantage 11

12 is that it is very intuitive and it clearly shows how the counterfactual would have looked like. The cost of that, however, is that the method is completely atheoretical: it is not based on any economic model, so it cannot take into account possible policy changes which could have changed the predictive ralationship. For example, before World War I began, Russian government started some important reforms, but it did not have time to implement them fully, and it is not clear what the effect of those reforms would have been if they had been completed. As for now, I abstract from the potential effect of the reforms, because this question constitutes a separate research topic. Given that the data for the end of the 19th and the beginning of the 20th century is quite limited, I focus only on GDP per capita as the main explanatory variable, even though I would like to use some variables that describe the sectoral composition of the economy as well. There is still a big question: how to choose a pre-treatment period over which to minimize the difference between the treated and synthetic control units: the data is available starting from 1885, and Russian Revolution took place in 1917, so I have only about 30 years for matching. Ideally, I would want to do matching without using some observations just before 1917 in order to be able to use them as a placebo: to check that the synthetic control approximates the behavior of the treated unit right before treatment well enough. However, there are two issues with this approach: first, a relatively small number of observations before treatment; second, the fact that Russia took part in World War I, and it caused its GDP per capita to fall substantially in The problem is that if I use pre-treatment years as a placebo, I cannot account for this decline in GPD per capita when constructing the synthetic control unit. In order to solve these problems, I separate the tasks of selecting the best method and constructing the counterfactual. In order to evaluate different methods and select the best one, I start with placebo tests for the cases when I know how the true growth path without a revolution looks like. 12

13 First, I use other countries in my sample for placebo tests: I use different methods to construct the counterfactual series for these countries starting from 1917, and then I compare the counterfactual and the actual series. Since other countries did not experience a revolution, I evaluate the methods based on the difference between the actual and counterfactual series after Second, I use Russia before 1917 as a placebo test. I divide the pre-revolutionary period in two parts, the matching part and the testing parts, using the Russian data to see how well the methods I use predict Russian GDP per capita out-of-sample, but only before the Revolution, when no treatment took place. Finally, based on the performance in placebo tests, I choose the best methods and use them in the main analysis, now using the entire pre-1917 period for matching. I also consider using different subsamples of countries as controls. One may want to match the counterfactual and actual units along other dimensions in addition to matching on GDP per capita. I have data on whether a country was independent at that time, whether it participated in World War I, and whether it is located in Europe. Restricting the sample only to European countries seems problematic. First, even though Russia is usually considered a European country, it does not entirely lie in Europe. Second, since Russia was probably the poorest European country at that time, it might be hard to find an appropriate synthetic unit. Asian or South American countries are likely to be a better match. As for World War I, it would be great to match on it, but there are several problems with it as well. First, it is hard to come up with a reasonable and objective definition of being a participant. Second, the countries that were mostly affected by World War I were European countries, but, as discussed above, these countries did not match Russia well in terms of economic development. Hence, using matching on World War I may be problematic, and I discuss the challenges and possible solutions in the next section. As for being independent, arguably, it is an important factor for economic development of 13

14 a country, and then restricting the sample to independent countries can improve the quality of counterfactual series. At the same time it is relatively easy to determine when each country became independent. There are two ambiguities here. First, it is somewhat difficult to define exactly when Canada, Australia, and New Zealand became independent from Great Britain; however, these three countries had so-called responsible governments by the end of the 19th century and arguably enjoyed a significant amount of independence. Hence, I keep them in the sample for the purposes of this exercise. Second, Norway and Sweden formed a union under the Swedish monarch until 1905, so formally Norway was not an independent country until then. However, since these two countries still had separate laws, legislatures, armed forces, etc., I treat them as independent as well. The countries that I drop from the full sample are Finland, India, Indonesia, and Sri Lanka. Finally, I should note that one might want to match Russia with other countries based on population to control for the overall size of the economy instead of just using GDP per capita. The difficulty with using population in my analysis is that Russia was one of the most populous countries in the world at that time and second most populous in my sample, which makes it almost impossible to find a good match in terms of total population. Consequently, I focus only on GDP per capita in my analysis, assuming that all economies were scalable (or, in economic terms, that there were constant returns to scale), so that Russia can be represented as a small country replicated many times, and that the population density does not matter for economic growth. 4 Results This section presents the findings of my paper. It consists of several subsections: Subsection 4.1 discusses which method to use, Subsection 4.2 presents the main results of my paper, then Subsection 4.3 discusses some robustness checks and presents additional findings, and finally Subsection 4.4 evaluates the performance of the preferred methods using selected 14

15 countries from the control group as placebos. 4.1 Choice of Preferred Specification Before I present the main findings of the paper, I discuss how to pick the preferred method from the methods considered above using placebo tests. I use two types of placebo tests. The first one evaluates performance of various methods using the post-1917 period for the other countries in my sample, while the second one uses the pre-1917 period for Russia by breaking it into the matching and testing parts Other Countries as Placebo Tests In this subsection, I drop Russia from the sample, construct counterfactual series for all other countries using all methods under consideration, and then pick the method with the best performance. I use the quadratic loss function and the absolute deviation loss function in my analysis, so I concentrate on the sum of squared residuals and the sum of absolute deviations of residuals as the measures of performance. The main underlying assumption of this exercise is that other countries were similar to Russia, so that if a method performs well for other countries it also performs well for Russia. I have seven methods to choose from: the time series (ARIMA(1,1,0)), five different kinds of the lasso (described in Table 2), and the synthetic control. Moreover, for every method, I can use various pools of countries as controls: all countries in the dataset, only independent countries, or only World War I participants. In this subsection I do not match based on participation in World War I for several reasons. First, it is hard to come up with a clear definition of being a participant. Second, even if I could come up with such definition, not only the sample of potential controls, but also the sample of countries for which I could run placebo tests would be limited to the war participants. But such countries were mostly rich European countries which were significantly different from Russia and do not necessarily form a good comparison group for Russia. 15

16 First, Table 3 reports the mean and median SSE across countries for each of the three samples of countries. The detailed discussion of the construction of these measures is presented in Appendix A.2.1. As we can see from the table, when all countries are used as controls (Panel A), Lasso 2 yields the lowest mean SSE across countries, while Lasso 4 yields the lowest median. The synthetic control method yields both second-lowest mean and median in this case, and also it has the best performance in terms of both mean and median when I restrict the controls to independent (Panel B) or seriously affected by World War I (Panel C) countries. Hence, I find that the results in Table 3 might suggest that the synthetic control is the best method to use. I should note that the results in Table 3 are based on the sample of countries including Germany and Italy, which experienced dramatic social, political, and economic changes in s. Thus, using Germany and Italy for placebo tests can be problematic. However, the ranking of the methods is not affected by dropping them from placebo tests. Table 3: Comparison of Methods Panel A: All Countries Method SSE Time Series Lasso 1 Lasso 2 Lasso 3 Lasso 4 Lasso 5 Synthetic Mean Median Panel B: Independent Countries Method SSE Time Series Lasso 1 Lasso 2 Lasso 3 Lasso 4 Lasso 5 Synthetic Mean Median In Panel A I use all 30 countries as controls, compute the SSE for each of these countries, and report the mean and median across these countries. In Panel B I use only 26 independent countries as controls, compute the SSE for each of these countries, and report the mean and median across these countries. I mark the lowest mean and median in each panel in bold. I discuss how to compute the SSE in more detail in Appendix A.2.1. The penalty parameters for the lasso are chosen via cross-validation, as described in Appendix A.1. Note, however, that because the samples of countries for which I compute the SSE differ across two panels of the table, I cannot directly compare different panels of the table: it is possible that it is easier to predict GDP per capita for some countries and harder for others. 16

17 Hence, if I want to compare methods across panels, I need to make sure that I compute the SSE for the same countries even if I use different samples of countries as controls. In order to solve this problem and compare the methods across various controls samples, I next compare the performance of various methods with different samples of countries as controls, but keeping the sample of countries for which I compute the SSE fixed. In Table 4 I use all countries or only independent countries as controls, while computing the SSE only for independent countries. As we can see from the table, now the synthetic control method is the best for both samples of controls, and restricting the sample of control countries to independent countries only improves both the mean and median SSE. Table 4: Comparison of Methods Panel A: All Countries Method SSE Time Series Lasso 1 Lasso 2 Lasso 3 Lasso 4 Lasso 5 Synthetic Mean Median Panel B: Independent Countries Method SSE Time Series Lasso 1 Lasso 2 Lasso 3 Lasso 4 Lasso 5 Synthetic Mean Median In Panel A I use all 30 countries as controls, compute the SSE only for 26 independent countries, and report the mean and median across these countries. In Panel B I use only 26 independent countries as controls, compute the SSE for each of these countries, and report the mean and median across these countries. I mark the lowest mean and median in each panel in bold, and the lowest mean and median across all samples in bold italic. I discuss how to compute the SSE in more detail in Appendix A.2.1. The penalty parameters for the lasso are chosen via cross-validation, as described in Appendix A.1. To summarize this section, overall I find the synthetic control to be probably the most preferable method, but in some cases Lasso 2 and Lasso 4 also perform well. To shed more light on the performance of the various methods, I next do the following: eliminate all strictly dominated methods (i.e. the methods which are beaten by at least one other method in all panels of Tables 3 and 4), which leaves me with Lasso 2, Lasso 4, and the synthetic control method, and evaluate them using the pre-1917 data for Russia. 17

18 4.1.2 Pre-1917 Russia as Placebo Test In this subsection I divide the pre-revolutionary period into two parts and use it for placebo tests, as described in detail in Appendix A.2.2. I restrict my attention to the three methods that performed well in the previous group of placebo tests, Lasso 2, Lasso 4, and the synthetic control method, but I enrich the set of specifications by using matching on World War I. I have three methods and four choices of control groups. As before, the first pool of controls includes all countries; the second one is restricted to independent countries. The third and fourth use only World War I participants with different definitions of participation. The first one includes countries that lost at least 0.05% of their population or 10,000 people in World War I; the second one includes only countries that lost at least 0.05% of their population. The only difference is India that satisfies the former criterion but not the latter, but it might play an important role in matching because it was a poor country which might enter the synthetic control unit with a relatively high weight. Table 5 presents the results of these placebo tests. As we can see from the table, if we compare methods for the same choice of control groups, the synthetic control method yields the lowest SSE for almost every choice of the control sample. As for the comparison across control groups but within a given matching period, the synthetic control yields the best results in two cases out of three, but it is not clear whether restricting the control group helps. As for the various lasso-based methods, it is difficult to rank them: they perform very differently depending on the matching period and the controls choice. What is clear, however, especially with Lasso 2, is that restricting the sample typically helps. It is quite intuitive: the lasso is very flexible, and having too many potential controls may lead to overfitting. Hence, even though the lasso tries to select appropriate controls in a data-driven way, restricting the pool of controls beforehand may be helpful. To conclude, I find that while the synthetic control method typically compares favorably to the other methods under consideration, the placebo tests are inconclusive regarding the 18

19 preferable choice of the control group for this method. Hence, I will use the synthetic control method with different control groups as the baseline, and I will present the results of the lasso when I do robustness checks. Table 5: Comparison of Methods Panel A: Matching Period Method Lasso 2 Lasso 4 Synthetic All Independent WW WW Panel B: Matching Period Method Lasso 2 Lasso 4 Synthetic All Independent WW * WW Panel C: Matching Period Method Lasso 2 Lasso 4 Synthetic All Independent WW * 0.333* WW * I discuss how to compute the SSE in detail in Appendix A.2.2. I mark the lowest SSE in each row in bold, and the lowest SSE in each panel in bold italic. The penalty parameters for the lasso are chosen via cross-validation when the entire period is used for matching as described in Appendix A.1, and I use these parameters for all matching sub-periods. The only exceptions, marked by *, are when the crossvalidated penalty leads to no controls being selected. All refers to the specification that uses all countries in the pool of controls. Independent refers to the specification that uses only independent countries in the pool of controls. WW1 1 refers to the specification that uses countries that lost at least 0.05% of their population or 10,000 people in World War I in the pool of controls. WW1 2 refers to the specification that uses countries that lost at least 0.05% of their population in World War I in the pool of controls. 19

20 4.2 Main Findings In this section I present the results of the synthetic control method with different choices of control groups, since the synthetic control method performs well in the placebo tests but there is no clear ranking of choices of the control group. I consider four variations of the synthetic control method. I have already discussed two specifications in detail: the one that uses all countries as potential controls and the one that uses only the independent countries as potential controls. The third specification uses all countries but Finland as potential controls. The reason is that Finland, which is given a high weight in the specification with all countries, was a part of Russian Empire before the Revolution, so it likely was affected by the Resolution as well. Hence, to rule out possible spillovers, I run this specification as a robustness check. The fourth specification uses only World War I participants as controls, and it uses the first definition of participation: military deaths equal to at least 0.05% of country s population or at least 10,000 people. Before discussing the counterfactual GDP per capita series, I first look at the properties of the synthetic control units. Table 6 presents the composition of the synthetic control unit in various specifications, while Table 7 compares the actual Russian country characteristics with the characteristics of different synthetic control units. Several things regarding the composition of the synthetic control units and their characteristics are worth noting. 20

21 Table 6: Weights for Synthetic Control Units Specification Country All Countries No Finland Independent WW1 Finland Sweden Portugal Argentina Peru Venezuela India Japan Sri Lanka The table presents the composition of the synthetic control units in various specifications. The first one uses all countries as potential controls, the second one uses all countries but Finland as potential controls, and the third one uses only the independent countries as potential controls. Entries with 0 mean that the country was in the pool of potential controls but was assigned zero weight; entries with mean that the country was dropped from the pool of potential controls. Table 7: Balancing of Actual and Synthetic Control Units Characteristic Unit Europe Independent Population, 1913 WW1 Deaths WW1 Deaths (thousands) (thousands) (% of population) Actual ,192 1, All , No Finland , Independent , WW , The table compares the characteristics of Russia with the characteristics of the synthetic control units in various specifications. The first one uses all countries as potential controls, the second one uses all countries but Finland as potential controls, and the third one uses only the independent countries as potential controls. All characteristics of the synthetic control units are computed simply as weighted averages of individual characteristics of the countries that compose the synthetic control unit. 21

22 First, in most specifications a relatively low weight is given to European countries. Second, the first two specifications assign relatively low weight to independent countries. Third, all specifications yield the synthetic control units with significantly lower World War I deaths as a percentage of population than Russia actually experienced. This is consistent with Russia being poorer than other European countries at that time, so that most European countries do not serve as good controls for Russia. In fact, the only European country that is assigned high weight in these specifications is Finland (with more than 45% weight in the first specification), which was a part of the Russian Empire before the revolution and also was relatively poor as compared to other European countries. At the same time it were European countries that actively participated in World War I, and European countries constituted a large share of independent countries as well. In other words, there is a trade-off between making the synthetic control unit similar to Russia in terms of economic development, i.e. GDP per capita, and making it similar to Russia across the characteristics such as being European, being independent, or participating in World War I. Russia was quite unique in that it was an independent and European country and it played a significant role in World War I, but at the same time it was far poorer than other European countries. Now I move on to the main results of my paper: the performance of the counterfactual unit in terms of GDP per capita. Figure 2 presents the results from four specifications of the synthetic control method. Even though there are differences between these specifications, the overall pattern of the results is similar. They all match the pre-1917 behavior well, with the quality of matching being higher when I use less restrictive pools of potential controls. As for the post-1917 behavior, the second scenario is pessimistic, the third one is optimistic, and the first and fourth ones are moderate, but in general the behavior of the synthetic control unit is similar: all synthetic control units steadily grow in 1920s, then experience a crisis associated with the Great Depression, and then start recovering in the second half of 1930s. Depending 22

23 on the specification, actual Russian GDP per capita catches up with the counterfactual one somewhere between 1933 and Figure 2: Actual and Synthetic Russian GDP per Capita The natural logarithm of the actual GDP per capita is plotted as a dashed line, the synthetic series is plotted as a solid line. The vertical line marks 1917, the year when the Revolution took place. The upper-left graph shows the results of the synthetic control method when all countries in the sample are included as potential controls. The upper-right graph shows the results of the synthetic control method when all countries except Finland are included as potential controls. The lower-left graph shows the results of the synthetic control method when only the independent countries are included as potential controls. The lower-right graph shows the results of the synthetic control method when only countries that lost at least 0.05% of their population or 10,000 people in World War I are included as potential controls. If we believe that these counterfactuals are plausible, without the Revolution Russia 23

24 would have grown at a rate comparable with, if not higher than, the developed countries. For instance, in (i.e. before the Great Depression) the GDP per capita annual growth rates for the optimistic and moderate scenarios are 2.6 3%, which is higher than the American (2.3%) or Japanese (1.6%) growth rates over the same period. Even for the pessimistic scenario the annual growth rate is 1.5%, which is pretty close to the Japanese one. In 1930s, during the Great Depression, there is an economic crisis in all scenarios, with the average annual growth rates being virtually 0 for the pessimistic scenario, about % for the moderate one, and almost 1.5% for the optimistic. As a comparison, the American average growth rate over the same period was 0.15%, and the Japanese was impressive 3.2%, since Japan was virtually unaffected by the Great Depression. The average annual growth rates over the entire period based on the synthetic control method is about 0.8% for the pessimistic scenario, 1.6 2% for the moderate one, and about 2% for the optimistic one. Again, as a comparison, the American average annual growth rate over the same period was 1.3%, and the Japanese was 2.4%. Table 8 concisely summarizes the results of the comparison between the synthetic control unit, USA, and Japan. It takes the lower of two moderate scenarios as a baseline. Table 8: Growth Rates in Comparison Period Counterfactual Russia USA Japan % 2.31% 1.65% % 0.15% 3.23% % 1.27% 2.40% The table compares growth rates of counterfactual Russia with the growth rates of Japan and USA. The results for Russia are roughly based on the synthetic control method when all countries are included as controls. It leads to higher growth rates than the specification when Finland is omitted from the pool of controls, but lower growth rates than when I restrict the sample to independent countries or to World War I participants. The growth rates are computed as geometric means over the corresponding periods. It is probably true that the Soviet industrialization of 1930s promoted faster growth and helped to achieve higher levels of GDP per capita than would have been possible without the Revolution. On top of that, the Soviet Union was virtually unaffected by the Great 24

25 Depression: even though the actual growth rates slowed a bit, the recovery was very fast, and growth rates in 1930s were quite high. It is worth noting, however, that the industrialization was also associated with the Soviet Famine of and the Great Purge, during which several million people died. It is worth noting that this decline in population in itself would have increased GDP per capita even if GDP had stayed constant. Overall, the Revolution probably allowed Russia to reach higher levels of economic development by the end of 1930s, but it also was associated with a huge decrease in the GDP per capita in 1920s, while if the Revolution had not happened, Russia would probably have grown more consistently. 4.3 Additional Results In this section, I present some robustness checks. I start by looking at the sensitivity of the synthetic control method to the choice of the matching period. I use three alternative choices: , , and That is, I match Russia with the control unit only using a particular subperiod of , and then construct the synthetic series for the entire period Table 9 presents the weights that are obtained when I use different subperiods for matching and compares them to the weights that I get when I use for matching. Table 10 describes the characteristics of the various synthetic control units. 25

26 Table 9: Weights for Synthetic Control Units Specification Country Finland Netherlands Portugal Australia are Argentina Peru Uruguay are Venezuela India Japan Sri Lanka The table presents the composition of the synthetic control units in various robustness checks specifications. The first one uses the entire period for matching. The second one uses only for matching; the third one uses only for matching; and the fourth one uses only for matching. Entries with 0 mean that the country was in the pool of potential controls but was assigned zero weight. Table 10: Balancing of Actual and Synthetic Control Units Characteristic Unit Europe Independent Population, 1913 WW1 Deaths WW1 Deaths (thousands) (thousands) (% of population) Actual ,192 1, , , , , The table compares the characteristics of Russia with the characteristics of the synthetic control units in various robustness checks specifications. The first one uses the entire period for matching. The second one uses only for matching; the third one uses only for matching; and the fourth one uses only for matching. All characteristics of the synthetic control units are computed simply as weighted averages of individual characteristics of the countries that compose the synthetic control unit. 26