The Global Impact of the Great Depression

Preparing to load PDF file. please wait...

0 of 0
The Global Impact of the Great Depression

Transcript Of The Global Impact of the Great Depression

Economic History Working Papers

No: 218/2015

The Global Impact of the Great Depression

Thilo Albers
London School of Economics
Martin Uebele
University of Groningen
Economic History Department, London School of Economics and Political Science, Houghton Street, London, WC2A 2AE, London, UK. T: +44 (0) 20 7955 7084. F: +44 (0) 20 7955 7730

The Global Impact of the Great Depression
Thilo Albers
Department of Economic History, London School of Economics and Political Science
Martin Uebele
Faculty of Arts, University of Groningen
This paper provides monthly economic activity indicators for 30 countries on six continents for the period 1925–36 based on more than 1200 historical time series. Aggregating these to a global economic activity indicator shows that the global recovery after 1931 was slower than much-cited contemporary evidence suggests. On a disaggregated level, we find that the majority of European countries experienced recessionary tendencies already in the mid-1920s, which puts the notion of a US-originated Great Depression into perspective. Our evidence cautions against employing industrial production to assess crises and recoveries across space as manufacturing catch-up growth occurs less developed countries. In this vein we find that in contrast to established historiography Spain, albeit floating her currency, was severely affected by the crisis, and Japan was hit harder than annual industrial production suggests. Finally, mapping the Depression suggests that economic improvements of major trading partners could have served as a catalyst for a country’s recovery. As a methodological contribution, we develop a framework to aggregate non-stationary series using principal component weights, and we scale the resulting indicators to an interpretable dimension using the standard deviation of annual industrial production indices.
Keywords: Great Depression, Economic Activity Indices, 20th Century, Business Cycles 2000 MSC: JEL codes:, C38, E32, E58, N14
1. Introduction
The worldwide economic crisis in the early 1930s serves as a yardstick for the depth, speed and international extent of all economic crises ever since. Understanding its root causes and how it spread is key to preventing similar events from happening again. However, not only are economists divided about why it happened, there is still a considerable discussion about what actually happened, i.e. where it originated and how the crisis spread. Given the importance and nature of the subject, more than 70 years should have been enough to bring about a data set with (i) reliable and comprehensive economic activity indicators at (ii) monthly frequency and for (iii) a large number of countries – yet reality is far from this. The state-of-the-art data sets fulfil at most two of these three requirements: GDP estimates such as Barro and Ursua (2010) typically come
Email addresses: [email protected] (Thilo Albers), [email protected] (Martin Uebele)

at annual frequency, while other authors focus on particular variables (see e.g. Accominotti, 2012, for the financial sector) or consider not more than about a dozen countries (see e.g. Mathy and Meissner, 2011, for industrial production). This paper introduces a new data set of economic activity indicators combining all three properties – a broad coverage of the economy, high frequency, and a more global geographical coverage.
The new economic activity indices suggest that global economic recovery after 1931 was slower than previous evidence suggested (Wagenführ, 1938; Almunia et al., 2010). In this light, the recovery after 2008 appears less gloomy than Almunia et al. (2010) argued. Moreover, we show that many European countries were in a weak state already in the mid-1920s, lending credibility to the idea of a “Long European Depression” (Kindleberger, 1986; Wolf, 2010) and raising doubts about US-centred explanations for the global Great Depression (Romer, 2004). Inter alia, these results stem from our third finding: Industrial production is not a representative indicator for economic activity in lesser developed countries. Since their industrial sector was small before the crisis, unconditional convergence in the industrial sector led to high growth rates, which masks the true impact of the crisis on the broader economy in terms of unemployment and agricultural depression.1 As a prime example, Spain that has so far been thought to having escaped the crisis (Choudhri and Kochin, 1980; Temin, 1993) did not do exceptionally well, especially when one takes into account the duration of the crisis. This is all the more important as Spain was not on the gold standard, apparently without too much of an effect. In the historiography of the interwar years, however, the gold standard is thought to be the “single best predictor of how severe the Depression was” (Temin, 1993, p. 92). Finally, mapping the Depression suggests that the recovery of important trading partners could have served as a major catalyst for a country’s recovery.
Why did previous research not improve the high frequency data coverage of the Great Depression? Our understanding of the historiography of the Great Depression is that methodological and more importantly computational constraints impeded previous research on its global transmission and that there was thus little need for such data. The literature on the Great Depression developed in four major waves with contemporary economists constituting the first wave. Especially the League of Nations (e.g. 1931, 1945) took an internationally comparative perspective but lacked the statistical tools and computational power to handle large, high-frequency data sets. This did not change when Friedman and Schwartz (1963) pioneered research on the US Great Depression, arguing that policy mistakes had caused it. In contrast to their predecessors, they used more monthly data to assess the timing of events. This second wave however focused on single countries and did not take into account the international dimension explicitly (Bernanke, 1995, p. 3). The third
1This mirrors recent findings about unconditional convergence in the manufacturing sector by Rodrik (2013) and Bénétrix et al. (2012).

wave, the gold standard literature, emerged in the 1980s and has shaped today’s understanding of the interwar crisis as well as the field of international macroeconomics. According to this research, the gold standard was not only instrumental to the global transmission of the crisis, but also causal to the downturn itself (Eichengreen, 1992). The methodological innovation of this literature was to expand the data to many countries and thus improve the econometric identifyability of the causes for the crisis (Bernanke, 1995, p. 1). However, this type of research has undermined another important dimension of the data – the time frequency dimension. With computational power still limited, the datasets of the gold standard literature would rest on annual data, which makes it hard to analyse how the crisis spread, given the speed of the recession that would eventually evolve into the Great Depression. The fourth wave consists of recently published literature and shows the merits of using higher frequency data in an international setting (e.g. Wolf, 2008; Mathy and Meissner, 2011; Accominotti, 2012). To the best of our knowledge, Ritschl and Sarferaz (2014) are the first to explicitly assess crisis transmission channels for the 1930s from a time series perspective. Contrary to conventional wisdom, they identify a financial transmission channel from Germany to the United States. With our data, this class of models could be applied to a wider range of countries, thus facilitating a more global identification of transmission channels.
Given methodological and more importantly computational constraints in the past, there was little scope for such identification of transmission channels until recently. The relaxation of these constraints and the promising findings of the fourth wave of research motivate us to introduce a novel dataset for the interwar years, which facilitates to analyse the Great Depression from a global perspective with high-quality, high-frequency observations. To produce indices comparable across time and space, we address two problems related to diffusion indices. Firstly, the “trend” in the short period of 1925–1936 (which makes the series non-stationary) is the actual object of interest and needs to be preserved. The solution we propose is to estimate the index weights with principal components on the stationary time series but apply the estimated weights to the standardised levels. Secondly, diffusion indices are not as readily interpretable as industrial production, which is by its nature linked to the physical world. Hence, we scale each country’s monthly activity index to the second moment of annual industrial production in order to make it interpretable and comparable across time and space.
In the remainder of the article, we describe the data sources, discuss the methodological problems and offer solutions, and finally present our results on the prelude, onset and recovery from the Great Depression across space.

2. Sources and Aggregation
2.1. Data
This section presents some key facts about the disaggregated data we employ. We provide more than 1200 monthly time series for 30 countries during the interval from 1925 to 1936,2 leaving us with more than 150,000 data points. Our main source is the Statistisches Handbuch der Weltwirtschaft (Statistisches Reichsamt, 1936, 1937) complemented with data from the International Abstract of Economic Statistics (Tinbergen, 1934; Methorst, 1938) and country-specific sources.
The Statistisches Reichsamt gathered the data from national statistical offices, periodicals such as The Economist, publications such as Lloyds Register of Shipping, and reports by private banks. Contemporaries praised the Handbuch for its coverage and accuracy (see e.g. Mitic, 1936). Data from the International Abstract, which was compiled by Jan Tinbergen (1934), has been used in numerous publications (see e.g. Eichengreen, 1982). We made several crosschecks of both publications with the original sources, and found no evidence for inaccuracies, reassuring us about the quality of the data source. In some cases, the indicator series from the two editions of the Handbuch were not entirely comparable, so we linked them via re-basing. In some exceptional cases, small extrapolations were unavoidable. Appendix B documents the source, the unit, and, if applicable, conversions for each single series.
Figure 1 maps the global coverage of our dataset. Different shades indicate the number of economic indicators. In comparison to one of the most extensive monthly cross-country data sets on industrial production or proxies thereof (Mathy and Meissner, 2011), we improve the coverage for the period from 1927 onwards by 16 countries, most of which were developing countries in the interwar period.
Naturally, number and type of indicator series vary by country. The arithmetic mean of indicators available is 40, the median 32, and the standard deviation about 24. Doubtlessly, one could argue that this induces a selection problem into our dataset. However, there is an alternative interpretation. For example, coal production might be a good indicator for Great Britain, whereas it is only of secondary importance for a less industrialised country such as Denmark. Therefore we favour the interpretation of a “qualitative filter:” the series were recorded for the simple reason that they were relevant for the economy, whereas omitted ones were of less relevance.
A more serious concern is the balance between nominal and real variables. Real indicators include count data (such as number of individuals unemployed) and those given in physical measures (such as tons of goods transported on railways), whereas all others are defined as nominal. Of the
2For a few countries the period is shorter, but always covers the immediate run-up to the Great Depression.

Figure 1: Geographical Coverage
1211 disaggregated series, about 44% are real and 56% nominal indicators. While economists generally believe in market prices as indicators of scarcity, one might always argue that the price level might be driven by monetary policy rather than real factors. In the interwar period this would particularly be connected to the international monetary system (Bernanke, 1995, p. 15). Although our robustness checks show that real and nominal variables carry very similar bits of information we deflate nominal indicators with wholesale prices in order to meet these concerns.3
Another characterisation of the variables is a sectoral one (see Figure 2). We assign each variable into one of the following seven groups: (i) Official Production Indices, which were calculated by the respective contemporary statistical departments and are only available for twelve countries, (ii) Trade, (iii) Production, Sales, Employment, Transport and alike (iv) Prices, including wholesale and consumer prices, (v) Private Banking, including variables such as clearings and market interest rates (vi) Central Banking, and (vii) Stock Market, including variables such as stock issues. We are agnostic about the relevance of the individual sectors for the business cycle, but expect nearly all of them to carry some information. For example, we include trade, because exports are a very important indicator for the well-being of primary product producers in this period. Moreover, in several countries such as the Netherlands and France, statistical offices included export quantities as proxies, when they calculated industrial production indices (Wagenführ, 1938, p. 96).4 Moreover, the import of capital goods teaches us about the investment mood. Banking
3However, deflating the indicators might underplay the severeness of the crisis for primary good exporters. 4Wagenführ was rather sceptical about trade as a proxy for output.

Number of Indicators: 1211
Official Production Indices Trade Production, Sales, Transport, Employment Prices Private Banking Central Banking Stock Market

< 1% 5%


11% 14%

Figure 2: Sectoral Distribution of the Data
variables such as clearings can contain valuable information as one can infer from the contemporary Index of Business Activity for Britain calculated by The Economist (1933). While there is some variation across countries concerning the availability of real indicators, we have a reasonable number of these for each country (see Appendix).
In sum, the nominal versus real categorisation and the sectoral distribution of our indicator variables suggest a good balance of the sample. It allows us to show that the inclusion of deflated variables does not substantially affect the results. For country-specific information, we refer the reader to our extensive Appendix, where we document every individual series. The remainder of this section discusses how we treat and aggregate the underlying data to country indices. Before doing that, however, we illustrate why industrial production might be an insufficient proxy for “macroeconomic health.”
2.2. Economic Activity versus Industrial Production The most popular indicator of economic well-being is per capita GDP, but in its absence and
if monthly observations are needed, industrial production often serves as a proxy for short run fluctuations of economic well-being instead (see e.g. in Wolf, 2010; Mathy and Meissner, 2011). Industrial production is defined as the sum of the physical output in different industries weighted by employment shares for example.

Coincident economic indicators or “diffusion indices” constitute an alternative approach, which was pioneered by Burns and Mitchell (1946) and revived in the late 1980s by Stock and Watson (1989).5 In contrast to industrial production, a “diffusion index” (a technical term for what we call economic activity index) is defined as an unobserved (latent) variable capturing the current state of economic activity, and measured by a statistical model that uses a wide variety of economic time series as inputs – just as in this article. In industrialised economies such as the US, the two indicators can be assumed to move closely together, but there are good reasons to suspect that this does not apply to industrialising economies. Given the popularity of industrial production for measuring the crisis of the interwar years this question is highly relevant.
The Nordic production-employment paradox illustrates why industrial production and economic activity might not move together.6 For example, the Danish industrial production index grew by 53% from 1925 to 1936 (League of Nations, 1945, p. 142) and Finnish production even doubled during the same period (Mitchell, 2014).7 However, qualitative accounts in the International Abstract of Economic make us somewhat sceptical of these miraculous growth records. For example, while the official annual industrial production index records a 15% increase in 1933 for Denmark, the experts of the International Statistical Institute observe only a “slight improvement” (Methorst, 1938, p. 61). These qualitative records are easily backed up by employment numbers. According to official statistics Danish unemployment grew by five percentage points between 1925 and 1936 (Methorst, 1938, p. 68) and employment in Finland regained its 1926 level only in 1935 (Methorst, 1938, p. 98). In these cases, qualitative accounts and unemployment records are extremely hard to reconcile with industrial production exhibiting such strong growth rates.
Finland and Denmark were hardly exceptions. Such “growth miracles” occurred elsewhere as well and seem to be associated with small industrial employment shares (Figure 3).8 This type of unconditional convergence limited to the industrial sector has recently been discussed by Rodrik (2013) for more recent decades and by Bénétrix et al. (2012) going back as far as 1870.9 In thrust with industrial production’s popularity as an indicator of broad economic activity, this cautions
5See also Spiethoff (1955) and Spree (1977, 1978) for earlier applications to Germany. 6Grytten (2008) argued that it can be explained by positive labour supply shocks. 7However, if we take annual gross production value from Methorst (1938, p. 98), and convert it to 1925 dollars using Board of Governors of the Federal Reserve System (1943, p. 670) and Williamson (2014), the growth miracle is reduced to 50%. This alone may suffice to illustrate the questionable scale of interwar industrial production estimates. 8We took the employment shares from Statistisches Reichsamt (1936). They do not refer to the exact same years for all countries but this should not substantially affect the results as they typically behave smoothly in time. We calculated compound growth rates using the respective first and last annual values in our sample, mostly 1925-1936. 9That there is no substantial evidence for unconditional GDP convergence in the interwar years (Roses and Wolf, 2010) might surprise the reader in this context, but unconditional GDP convergence has not been found for any broader sample (Rodrik, 2013, p. 1).

Figure 3: Annual growth of Industrial Production between 1925 and 1936, and initial industrial employment share
against the use of cross-sectional data (such as in Eichengreen and Sachs, 1985) for assessing the recovery without controlling for initial industrial development. Hence, we favour economic activity indices including the industrial sector over indices on industrial production only, and use information from as many economic sectors as possible to arrive at a more representative account of interwar business cycles.
Having said that, we will still use one particular piece of information from industrial production indices. While their levels might be subject to convergence issues, there is no a priori reason to assume that their second moments are. The next chapter illustrates how industrial output fluctuations can be used to rescale diffusion indices, making them interpretable in terms of amplitude and comparable across the sample and with indicators in the literature.
2.3. Aggregation - Index Problems Now and Then Estimating latent business cycle indices has seen important developments in the 1990s and early
2000s (see e.g. Stock and Watson, 1991; Otrok and Whiteman, 1998; Kose et al., 2003). These linear models are formulated for stationary data, which for business cycle research is usually the correct specification. But what if a particular cycle is the object of interest? In the case of interwar economic time series, removing the “trend” means to remove the actual crisis, and thus renders the procedure pointless.
One way to tackle this problem is to use annual GDP for the levels and a latent monthly indicator for interpolation to get to a monthly indicator as Mitchell et al. (2012) did for interwar Britain.

While their mixed frequency approach is useful when annual national accounts are reliable, we are skeptical of its applicability to our sample of countries, in which reliable GDP data is sometimes lacking and the commonly used industrial production indices (League of Nations, 1945) are subject to the criticism raised above.10 A second way is to realise that Mitchell et al. (2012)’s mixed-frequency model is actually based on a non-linear aggregation model suitable to our needs (Proietti and Moauro, 2006). However, instead of using this non-standard model requiring highly specialised skills we apply Occam’s razor and solve the problem with textbook econometrics.11 The resulting algorithm is illustrated in Figure 4. It combines all necessary steps: seasonal adjustment, deflating, detrending and principal component analysis of the indicator series, index aggregation, and finally re-scaling.
Figure 4: Aggregation Procedure
Seasonal adjustment is a problem contemporaries had to deal with as well, especially with the fact that it differs across the indicator series. First, the intensity of seasonality differs. For instance, the seasonality of interest rates, if any, is much weaker than the one of unemployment. Second, it
10We also created a set of interpolated indices following the method proposed by Fernandez (1981) that was implemented in Matlab by Enrique M. Quilis (2013). We refrain from showing the results as we find the approach not suitable for our sample. The results are available from the authors upon request.
11Stock and Watson (1991, p. 75) suggest another alternative: They re-trend the index after aggregation and calibrate its scale to the US Department of Commerce’s business activity index but this requires first differencing, which does not always render our series stationary.