Commodity Market Integration (2024)

What Is Market Integration and Why Is It Relevant?

The existence of well-organized markets is an essential precondition for division of labor and ultimately for the orderly functioning of any nonsubsistence economy. Most economic historians of the preindustrial world, a period of slow and intermittent technical progress, regard the development of markets as an essential source of (“Smithian”) economic growth (Epstein, 2000; Johnson & Koyama, 2017; Persson & Sharp, 2015). However, Van Bavel (2016) strikes a note of caution, indicating that the benefits might be temporary because traders could collude and capture market institutions. The development of markets in history was a long-run, multifaceted, wide-ranging process, which affected markets for goods, land, capital, and labor, and featured a massive institutional change. It is plainly impossible to cover all aspects in this article, which will deal with the analytical works on the evolution of commodity markets. When did markets develop, why, and how much did their development increase welfare and the rate of growth of the economy? These questions can be addressed with two different sets of data—prices and quantity (Figure 1 relates them in the simplest demand and supply framework).

Producers in location (country) I can supply a good (or a composite of goods) at price p_i and send to location (country) J with a total cost t_ij so that consumers in J will pay a price p_j. In long-run equilibrium, the price differential is equal to trade costs and trade is inversely proportional to them: a reduction (increase) in trade costs will cause a decline (increase) in price gaps and a rise (fall) in trade. Thus, one can measure the development of markets with data on traded quantities (trade) and/or with data on prices (market integration). In an ideal world of perfect data, the analysis of trade would provide the big picture and the analysis of prices would address micro-issues such as the efficiency of markets or the effect of specific institutions with more detailed data. In the harsh reality of limited historical sources, the two literatures have been evolving separately, focusing on different research questions and on different periods.

This article discusses the literature on market integration, emphasizing whenever possible the results on the preindustrial period, where data on trade are very scarce.1 The underlying model of markets is very simple. There are two or more “locations” (usually cities), which may or may not trade. The price gap between them is equal to trade costs if they trade (cf. Figure 1) and smaller if they do not. In this case, prices move randomly or are determined by trading with a third market (Coleman, 2007; Federico, 2012). Prices are subject to random localized shocks (e.g., from harvests), which could cause them to exceed the price in some other location plus trade costs (known as commodity point in the literature). This creates opportunities for extra profits, which rational traders exploit by increasing flows or by starting to trade. Arbitrage brings the price gap back to commodity points: the faster it returns to its equilibrium level the more efficient (à la Fama [1970]) the market is.2 Similarly, profits could be made by arbitraging in time if expected future prices exceed current prices plus the cost of storage.3

This model clearly implies two different conditions that were stated almost 200 years ago by French mathematician Antoine Augustin Cournot. He characterizes an integrated market as “an entire territory of which the parts are so united by the relations of unrestricted commerce that prices take the same level throughout with ease and rapidity” (emphasis added; Cournot, 1971, pp. 51–52). In other words, (a) the equilibrium level of prices must be equal (the Law of One Price) and (b) prices must return easily and quickly to this level after any shock. These conditions are clearly separable because prices can slowly return to the same level or quickly change to a different level. They are both necessary but not sufficient for integration and, as argued in more detail elsewhere (Federico, 2012), each statistical technique can test only one condition. The state-of-the art time series econometrics, which many authors proudly boast about using, test Cournot’s second condition and thus, somewhat inadvertently, the issue of efficiency has gained prominence in the literature. Yet, arguably, efficiency is less relevant than price convergence, because only extreme levels of inefficiency, up to actual harassment of traders, would really hamper trade and seriously damage seriously. In contrast, price convergence affects the level of prices, which determines the daily decisions of consumers and producers.

This article deals with the two conditions separately. It first sketches out what we know about trends in price gaps (Cournot’s first condition) and on market efficiency (Cournot’s second condition). The section on The Causes of Integration surveys the research on the origins of integration and The Effects of Integration reviews the (very limited) estimates of the benefits of integration. The conclusion provides an agenda for further research.

Integration as Price Convergence

If taken literally, Cournot’s first condition would almost never be met: prices are extremely unlikely to be equal because trade costs should be nil. One could argue that condition can be met if price gaps are sufficiently small or, as suggested by Ejrnæs, Persson, and Rich (2008), equal to transaction costs. Neither solution is convincing. The former would introduce an arbitrary distinction between small and large gaps—that is, how big is big (McCloskey & Zecher, 1984)? The latter would yield a paradoxical conclusion: markets would be defined as equally integrated with hugely different price gaps (say 1% or 100% or even 1000%).4 Thus, all testing adopts a dynamic approach: A market is defined as integrating (disintegrating) if the gap between two locations narrows (widens) or, equivalently, prices converge (diverge). However, the interpretation of results can be cumbersome with a large number of locations, and thus of possible pairs. Thus, many large-scale studies estimate σ‎-convergence with changes in the coefficient of variation across all locations in each year. This is a dimensionless measure, is easily comparable in time and space, and the sequence of coefficients can be analyzed with time-series analysis looking for discontinuities and estimating long-run trends. The cumulated change as a percentage of the initial dispersion is a simple measure of the extent of integration (or lack of it), while detected discontinuities could be related to specific events, such as changes in trade policy or in transportation infrastructure. The results are likely to be imprecise if quality differs across locations, but they are not necessarily systematically biased (Federico, 2011).

Figure 2 reproduces the results of the most recent work by Federico, Schulze, and Volckart (forthcoming) on price convergence in Europe using this approach. The “unbalanced sample” line refers to the whole database, featuring almost 600 locations, with about 68,000 yearly observations, whereas the “balanced sample” includes 15 cities from a core area, stretching from England to Northern Italy and to Southern Germany.

The series fluctuates a lot, especially in the early centuries, with jumps in dispersions often related to wars. Thus, the 1528–1530 peak in the coefficient of variation for the balanced sample was caused by the very high price in Milan during the war of the League of Cognac, while the 1586–1588 and 1590–1592 peaks reflect the impact of the final stage of French religious wars on prices in Paris and Albi. The polynomial trend shows that the price dispersion in the core area (the balanced sample) declined substantially in the 16th century, rose in the early 17th century, and declined again somewhat from about 1650 to the French revolution.5 The liberalization of wheat trade in the 1830s caused a fast and massive process of integration that was partially reversed by the protectionist backlash of the 1880s and 1890s.

The dispersion of prices of the unbalanced sample is higher for most of the time, and its movements seem smoother. This hints at differences in the process of integration between the core and the rest of the continent, but the unbalanced sample series as such is not representative. The underlying sample changes over time; for example, the number of locations grows from about 30 in the 15th century to almost 300 in the early 19th century. However, the regional dimension of European integration has been explored in detail by Chilosi, Studer, and Tuncer (2013) and Federico et al. (forthcoming). They cover different periods with different datasets and group locations in regions with different methods, but the results are consistent.6 England was fully integrated as of the 16th century, and the whole area around the North Sea (most notably the Netherlands) in the 17th century. In contrast the landlocked regions were smaller and initially less integrated but in the long run they also joined the European market, Spain being the last one. In the 19th century the dispersion across boundaries declined and then increased, while the domestic one, unencumbered by trade barriers, decreased or remained constant (Chevet & Saint-Amour, 1991; Federico, 2007; Metzler, 1974).

There are no studies of domestic price convergence in the early modern period outside Europe, but there is a substantial body of research on transatlantic trends. There is no doubt that price dispersion declined across the Atlantic Ocean in the 19th century (Chilosi & Federico, 2015; Federico & Persson, 2007; O’Rourke & Williamson, 1994) and there is evidence of integration in the late 18th century (Sharp & Weisdorf, 2013). In contrast, trends across the Indian Ocean during the early modern period are more controversial. In a seminal paper, O’Rourke and Williamson (2002) argued that the increase in trade between Southeast Asia and Europe was not matched by price convergence because the monopoly power of the Dutch East India Company (VOC) kept price gaps for spices far above the simple transportation costs. This view was first questioned by Ronnback (2009), with a larger sample of goods and a wider geographical coverage. He stresses the differences by period, product, and route but on balance he finds that convergence was more frequent, especially between Europe and South America, than divergence. More recently, De Zwart (2016) returned to the issue of convergence between Southeast Asia and the Netherlands, with a systematic collection of data from the archive of the VOC. He finds differences in timing and trends. Convergence started for some products in the 17th century and continued in the 18th century for 10 products out of 16. Yet, in the early 19th century, the markup on Indian and Southeast Asian goods in Europe was still quite large, from 70% to 220%, and substantially greater than the price gaps between Europe and the United States (Chilosi & Federico, 2015). Convergence resumed in earnest in the 19th century and the last years of the so-called long nineteenth century marked the heyday of integration in the Indian Ocean.

In spite of these major scholarly achievements, there still are substantial gaps in our knowledge of long-run trends, which clearly depends on the available sources. Price data are simply not available for most periods, products, and areas, either because they were never collected or because the records have not survived (a typical case being the Roman Empire). Second, almost all series refer to commodities, with a strong majority of prices of cereals. This is hardly surprising: cereals were the staple food for the majority of the population and urban markets were tightly monitored, as rulers feared that high prices could trigger riots. However, the very presence of a strong market organization could mean that cereal prices are not representative. Indeed, there is substantial evidence of differences in trends between products, including cereals, for the same period and the same locations. Klovland (2005) compares the pattern of convergence (or lack thereof) between Germany and the United Kingdom from the 1850s to World War I and finds wheat to be an outlier. In about half the cases, the price differential declined, while the gap for wheat increased by 15 times (from 2 to almost 20%), the second largest increase out of 39 commodities. There are also sizable differences between patterns of convergence (or lack thereof) for cereals and wax and tallow candles in the first half of the 19th century (Federico, 2011). Some of these differences can be accounted for by the trade policy, but only up to a point. Barriers to trading coal in 19th century Europe were few, but domestic convergence was earlier and faster than inter-European convergence (Murray & Silvestre, 2020). Differences also appear in domestic markets. Prices of beef converged in Austria but not in Hungary, while table wine prices converged in both parts of the dual monarchy (Good, 1984). There was also no convergence in prices of wine and oil in Spain (Carnero Arbat & Sanchez Albornoz, 1981). In Italy, prices of agricultural products did converge (with the notable exception of olive oil), but the process was significantly slower than for wheat (Federico, 2007). This evidence is clearly not conclusive, but still suggests caution in extending the results for cereals to all commodities, let alone to industrial products.

Integration as Market Efficiency

Cournot quoted the speedy return to commodity points (or in his view to the equal price) as the second essential feature of an integrated market; and this can be interpreted a measure of efficiency. The measurement of speed of return was made possible only by advances in time series analysis in the 1980s, but since the 1960s scholars have explored the efficiency of markets (or “integration”) by computing pairwise coefficient of correlation or by running OLS (ordinary least square) regressions for pairs of markets. The intuition is straightforward: the faster the adjustment to (equally sized) shocks, the closer prices in different locations would move together. Running an OLS regression adds the further hypothesis that prices were set in a central market and then transmitted to other cities via arbitrage (Ravaillon, 1987). This naive approach would return an upwardly biased coefficient if prices shared a common trend (e.g., from currency devaluation) or weather shocks. Other studies tackle the first problem by computing correlations of first differences (e.g., Studer, 2008) or detrending price series (e.g., Chilosi et al., 2013). The most recent versions of this co-movement approach use the dynamic factor analysis (DFA), which is an evolution of the principal component analysis (Andersson & Ljungberg, 2015; Federico et al., forthcoming; Uebele, 2011). This analysis measures the co-movement of prices, and thus the efficiency of the market, with the average of the shares of the variance of the common factor on the variance of individual series over a specific period. It is straightforward to get a series of coefficients with a rolling approach—that is by moving the reference period one year at a time.

A parallel strain in the literature has measured integration with the variance of prices for a single market or, as suggested by Engel and Rogers (1996), with the ratio of prices in two locations (Dobado, Garcia-Hiernaux, & Guerrero, 2012, 2015). This approach assumes that in a closed economy prices are inversely related to domestic production and that in an open economy arbitrage would reduce their volatility. Neither assumption is necessarily true. The volatility also depends on the covariance with prices of other goods, given the elasticity of substitution, and on the extent of storage or intertemporal arbitrage (Foldvari & van Leeuwen, 2011). Furthermore, any interpretation of decline in variance as evidence of increasing efficiency of the market assumes that the size of underlying shocks is constant, and this may not be the case. For instance, irrigation or other modern technology may reduce the effect of weather on crop yield and thus the size of weather-related price fluctuations. Last but not least, integration may increase the variance of prices: the wider the area, the more likely shocks are to be uncorrelated.

In a more general vein, the interpretation of the results of variance and co-movement analysis, as of price gaps (see “Integration as Price Convergence”), is meaningful only in a dynamic setting. Any threshold to define efficient a market is bound to be arbitrary. Is a coefficient of correlation of 0.6 enough? How big is big? The co-integration revolution of the 1980s seemed to offer an unambiguous answer to this question. The basic auto-regressive (AR) model explains the change in price gap between two markets from time t to time t+1 with the size of the price gap at time t. In an efficient market, the coefficient of the explicative variable is expected to be negative and significant—an initial disequilibrium causes the price gap to move in the opposite direction. Furthermore, with the coefficient it is possible to compute the half-lives of price shocks as a standard measure of the speed of return to equilibrium. This approach is still used, with more complex specifications including, for instance, dummies for specific shocks (e.g., wars) or the changes in the average price across locations as a measure of common shocks (Bernhofen, Eberhard, Li, & Morgan, 2016). The first-generation models have a serious flaw. They implicitly assume that prices would converge even if the price gap were smaller than the transaction costs, when arbitrage is unprofitable. This logical shortcoming can be addressed by running a so-called threshold auto-regressive (TAR) model. It assumes that prices converge only down to the commodity points and that they move randomly within them. As a bonus, the model estimates, with a grid-search procedure, the most likely commodity points, which can be used in further analysis. Alas, neither TAR model is flawless. First, as for all co-integration tests, it would return unbiased results if and only if the frequency of available data matches the real time of adjustment. The speed would be biased upward if the frequency of data is lower than the actual speed of adjustment (e.g., if data are monthly and adjustment takes one week; Brunt & Cannon, 2014; Taylor, 2001). Second, the grid-search procedure assumes that most observations are around the commodity points. This would not be the case if the market was very inefficient (price gaps being very frequently outside the commodity points) or if the two locations did not trade (price gaps being permanently lower than the trade costs). Last but not least, the TAR model needs a minimum number of observations and yields a single set of transaction costs. It thus assumes them to be constant during the period of estimation, in contrast with the whole research agenda in market integration.7 This assumption may be harmless if the period is short (i.e., for high frequency data; 240 observations correspond to less than one year), but can be wrong for monthly ones and is surely wrong for yearly series.

The co-movement, variance-based, and co-integration models have been used to study the “integration” (or more precisely “efficiency”) of the European market in the long run, with different samples of cities for different periods and, unsurprisingly, with somewhat contrasting results. We will quote here three representative examples. Jacks (2005) deals with the integration of the wheat market in Europe and in the United States in the 19th century, from 1800 to World War I, measuring efficiency with speed of return to equilibrium and proxying convergence with changes in transaction costs and from a TAR model.8 Both measures suggest that integration started in the early 19th century, but levels and trends differed. Domestic integration was common to all countries, while the initial level of efficiency and its subsequent increase was broadly proportional to the level of development of each country. Two other works this analysis backward in time with yearly prices for small-time invariant samples of cities. Chilosi et al. (2013) covers the period to 1620, computing coefficient of correlations over detrended series and volatility for a sample of 13 cities, and Federico et al. (forthcoming) use DFA for 15 cities. Both studies find some relevant differences with trends in price dispersion for the same sample of cities (see “Integration as Price Convergence”). Federico et al. (forthcoming) show that the two series moved together in the very long run but not in the medium term, especially in the 18th century. Efficiency grew significantly in the early decades, collapsed around midcentury, and rose fast from the 1770s to the 1840s, with only a minor pause during the French wars.

Interesting results in this line of research speak to a major debate on the history of China. Pomeranz (2000) argued that 18th-century China was as advanced as Western Europe and quoted the development of markets as evidence for this claim. His optimistic view was buttressed by Shiue and Keller (2007), who compared China with Europe in the 18th century with three different measures of efficiency. They argued that China as a whole was not less “integrated” (i.e., efficient) than Western Europe, although its most advanced area, the Lower Yangtze valley, lagged behind England. As posited by Pomeranz (2000), the difference between China and Europe widened in the 19th century when the far-reaching integration in Europe was not matched by a parallel process in the Heavenly Empire. Bernhofen et al. (2016) offers a much less optimistic view with an enhanced AR model and data on prices of wheat for 80 cities in North China and of rice in 131 cities in South China. The length of return to equilibrium was already growing (i.e., the market was becoming less efficient) toward the end of the 18th century, well before the mid-19th-century crisis of the empire, and, in stark contrast to Shiue and Keller (2007), it was on average about six times longer than in the (admittedly smaller) European “national” markets for series of comparable frequency. This more pessimistic view tallies well with the results by Li (2000), who shows a massive divergence in prices for wheat and millet in the northern province of Hebei, which includes Beijing, from 1738 to 1911.

The Causes of Integration

The basic model of integration implies that integration depends on trade costs, which Anderson and van Wincoop (2004, p. 691) define as

all costs incurred in getting a good to a final user other than the marginal cost of producing the good itself: transportation costs (both freight costs and time costs), policy barriers (tariffs and non-tariff barriers), information costs, contract enforcement costs, costs associated with the use of different currencies, legal and regulatory costs, and local distribution costs (wholesale and retail).

This yields a model similar to the gravity models of trade theory (Head & Mayer, 2014):

$$\text{MI}=\text{f}(\text{Tc},\text{B},\text{I},E,X),$$

where $\text{MI}$ is a measure of convergence, usually the price gap between markets or, less frequently, of efficiency (the speed of return to equilibrium), $\text{Tc}$ measures transport costs, $\text{B}$ barriers to trade, $\text{I}$ information costs, $X$ all other trade costs, and $E$ is a set of dummies to capture specific large shocks (e.g., wars). In this setting, the expected sign of explicative variables varies according to the dependent variable. A positive sign corresponds to less integration if the dependent variable is a measure of dispersion and to more integration if the dependent variable is a measure of efficiency. This framework has been used in quite a few studies for different periods and areas, with different set of explicative variables. It is impossible to report in detail all results, so we will only give a flavor of the main ones.

The availability of transport infrastructures usually fostered integration but the effect varied according to the competition of other means of transportation. Paved (free) roads reduced the gaps in rye prices in Westphalia in the middle decades of the 19th century, but the effect of waterways was mixed. Railroads had little impact, possibly because the network was not yet fully developed (Uebele & Gallardo-Albarran, 2015). Keller and Shiue (2008) estimate that building railways accounted for about four-fifths of price convergence in Germany in the 19th century. The case of Italy was somewhat different (Federico, 2007). The building of railways substantially reduced the price gaps between landlocked markets in the north, but had little impact for the rest of the country, where rail transportation was outcompeted by coastal trade. For the whole continent, Jacks (2006) confirms that the mere existence of a railway connection between two locations reduced (his TAR-based estimate of) trade costs by 5%, while the effects on efficiency are mixed; the speed of adjustment is positively related to the existence of a connection but negatively related to its length.

The effect of railways on integration has been extensively studied for India as part of the debate about the economic benefits of British rule. Price dispersion was very high at the beginning of the 19th century and fell sharply in the 1830s and from the 1860s onward (Studer, 2008). Hurd (1975) argued that railways were the main driver of integration by a visual inspection of changes in coefficients of variation for districts with and without railroads. Andrabi and Kuelhwein (2010) estimate that railways accounted for only for a fifth of the convergence in wheat and rice prices with a time-variant dummy for the existence of a connection in a panel regression and suggest that most of the convergence can be explained by other integrating forces, such as a common currency, language, and administration. However, they use retail prices of cereals, and thus the (likely) stable markup on wholesale prices may have dampened the effect of a decline of trade costs on wholesale prices. This hypothesis is confirmed by estimates by Donaldson (2018) on the costs of transporting salt from a common production site to the whole of India by different means. He estimates that the elasticity of trade costs to distance was around 0.25, and that both road and, somewhat surprisingly, rivers were more expensive than rail transport, by eight and slightly less than four times, respectively.

As expected, there is strong evidence of a negative effect of trade barriers on integration. The recent research has substantially refined the interpretation by O’Rourke and Williamson (2002) about the convergence of prices (or of the lack of it) between Asia and the West. De Zwart (2016) relates the pattern of convergence in the 17th and 18th centuries (see “Integration as Price Convergence”) to the level of competition. Prices converged earlier and faster in competitive goods, including pepper, while gaps remained wide in products such as cloves and nutmeg, where the trading companies succeeded in enforcing their monopoly. Accordingly, Chilosi and Federico (2015) find a strong positive effect on price convergence of the abolition of monopolies of the British East India Company (EIC) on trade with India in 1815 and of the Dutch Nederlandsche Handel-Maatschappij (NHM), which had replaced the failed VOC around 1850. They also find a strong effect of the final repeal of the British Corn Laws in 1842 on price gaps between the United States and the United Kingdom. Many other works confirm the effect of trade liberalization on convergence in 19th century. The abolition of Piedmontese duty on wheat in 1849 accounted for about a fifth of the price convergence among Italian states before the unification of the country in 1860 (Federico, 2007). Keller and Shiue (2014) estimate that the entry of a state into the Zollverein reduced the wheat price gaps with other Zollverein cities by 28% on average, although the estimate may include the effect of other factors. In a more general vein, Jacks (2006) finds that outright prohibition or protection against imports increased transaction costs in wheat prices in Europe and slowed down the adjustment after a shock.

Very few studies include a specific variable for information costs, probably because of the dearth of suitable data. Brunt and Cannon (2014) use as proxy the diffusion of local newspapers in late 18th- to early 19th-century England. It had no effect on the speed of adjustment, probably because newspapers had to rely on slow physical transportation, but seems to have reduced the volatility of relative prices and increased the correlation of shocks (i.e., the opportunity for arbitrage). The real game changer in the 19th century was the telegraph. The opening of the first line between England and the United States in July 1866 slashed time of transmission of news from about 10 days to less than 1 and strongly reduced the mean and variance of gaps in information-adjusted daily cotton prices; that is, the gap between the New York price and the latest known Liverpool price (Steinwender, 2018). The telegraph contributed significantly to price convergence between Europe, North America, and South Asia (Chilosi & Federico, 2015) and within Europe (Murray & Silvestre, 2020).

Integration may have been affected by many factors other than transportation costs or barriers to trade. Arbitrage would be easier if the risks were lower, the institutions were better (e.g., stronger protection of property rights and a more efficient and fair judicial system with less privilege) and the level of trust between potential partners higher. Some authors have tried to take these factors into account rather than relying on fixed effects or time dummies in panel regression. Much depends on the available data.

We know the precise timing of changes in political borders and in currency regimes. Several authors have tested the effect of fixed exchange rates on integration, finding, as expected, that efficiency was higher under the gold standard and that was negatively related to the variance of exchange rates (Jacks, 2006). The existence of a political border implies at the very least different institutions and legal procedures, if not barriers to trade. Engel Rogers (1996) estimate that in the 1980s the Canadian–American border increased the volatility of relative prices (and thus reduced efficiency) by a fifth relative to the volatility within the same country. With the same method, Jacks (2009) finds that borders reduced the efficiency of the European wheat market, but the effect declined. Andrabi and Kuelhwein (2010) find a significant effect of borders between British India and the princely states on price gaps in India. The interpretation of the border effect depends on the specification of the model. In a bare bones model, a border dummy can pick up the effect of omitted or mis-measured variables. Federico (2007), in his analysis of the integration of the Italian market in the 19th century, includes among explicative variables a border dummy alongside transport costs and duties, and finds it to be negative (i.e., wrongly signed) but not significant. It is thus somewhat puzzling that border dummies remain significant in Jacks (2006), which features a rich set of proxies for other trade costs. The effect is, as expected, positive on commodity points and negative on speed of adjustment—that is, borders widened price dispersion and reduced efficiency.

Measuring the quality of domestic institutions is much more difficult than adding dummies for borders and results of the few attempts to capture it have been somewhat disappointing. Keller and Shiue’s (2020) analysis of the causes of differences in average price gaps between pairs of German cities over the period 1820–1880 indicates a proxy institution with the abolition of guilds, equality in the law, and the redemption of feudal land. Jointly these variables explain about a quarter of the differences, but the relevance of these specific institutions for wheat trade is not immediately clear. Bateman (2012) proxies quality of institutions in early modern Europe with indexes of fiscal centralization or the activity of (traditional) parliaments, but neither variable is significant. It is possible that the result reflects the poor quality of these variables. Malinowski (2019) succeeds in precisely measuring the activity of the Sejm, the Polish parliament before the 18th century partition of the country, and finds that its inaction hampered market integration and reduced the state revenue, ultimately causing the demise of the kingdom. However, the Sejm was a particularly dysfunctional institution because its voting procedure with individual veto rights made it extremely difficult to reach any decision.

There is no simple way to measure cultural differences, which are routinely proxied by data on ethnic composition or dummies for common languages. Yet the results suggest that the impact might have been relevant. Jacks (2006) finds that the common language reduced price gaps between pairs of European cities and increased the speed of adjustment, even if the latter effect is small. A measure of differences in ethnic composition of different cities of the Habsburg Empire is positively related to price gaps, and its inclusion as an explicative variable causes the dummies for post-1918 states to become not significant (Schulze & Wolf, 2009).

Market integration is by its nature a long-term process, but, as is clear from the analysis of movements in price dispersion (see the section on “Integration as Price Convergence”), it can be heavily affected by short-term shocks, most notably wars. Thus it is important to control for their effect in the statistical analysis of causes of integration. Chilosi and Federico (2015) systematically test the effect of specific market interventions and of a number of political events on intercontinental price gaps, with mixed results. Domestic shocks, including the Indian mutiny, do not seem to have affected the price gaps, while World War I increased price differentials across the Indian Ocean but not across the Atlantic. In the latter case, the effect of war was fully captured by the series of freights. Jacks (2006) tests the impact of six different combinations of civil and foreign wars relative to the no-war default case. In all these combinations except “neutral” (defined as the combination of a market in a neutral country and another in a country at war), the war dummy is significant in that wars increased dispersion and slowed down the adjustment.

This short review highlights two points. First, the same variable may yield different results in explaining convergence or efficiency. This is not really surprising because convergence and efficiency are affected in different ways by changes in trade costs. In an efficient market, a shrinking in commodity points causes convergence but does not necessarily increase the speed of adjustment after a shock.9 However, an improvement in information flows (a decrease in costs or an increase in speed of transmission) or an institutional change (e.g., the abolition of restrictions to arbitrage) may increase the speed of adjustment but does not necessarily affect the equilibrium price gaps. The measured effect on differentials may be spurious if changes in timing of adjustment are not matched by parallel changes in the frequency of data.10 Actually, an improvement in information flows may reduce the size of shocks and thus the opportunity for arbitrage. Upon receiving a relevant piece of news (e.g., a poor harvest), rational traders would anticipate the movement in prices in the affected market and adjust local prices accordingly before exceeding the commodity points. Likewise, wars had an unambiguously negative effect on price dispersion, while there is a margin of uncertainty on their effect on efficiency. They disrupted the orderly working of markets, but in some cases they could increase circulation of information with a positive effect on efficiency.

Second, a list of significant variables is not sufficient. It is necessary to know how much they matter or, as Jacks (2006) asks, “what drove market integration?” The answer would be the more precise the wider the range of relevant explicative variables is. As is clear from the discussion so far, Jacks (2006) is the most comprehensive work, as he tests the effect of no fewer than 19 variables in his TAR-based estimates of trade costs and efficiency. Jacks (2006, Table 5) sums up his results by ranking the impact of continuous variables separately (as measured by the impact of standard deviations on the dependent variable) and dummies (as measured by their coefficients). This division makes the comparison between the two categories of variables difficult. The author nevertheless concludes that ”trade costs seem to be more responsive to changes in the choice of monetary regimes than changes in the underlying technology of transport,” while ”speeds of price adjustment present a more balanced account as transport, monetary, and commercial variables all seem to play a part” (Jacks, 2006, p. 405). The gravity-type models in a log-log specification yield a straightforward measurement of contribution of each variable to convergence as the coefficients from a log-log specification times the cumulate change of the variables. With this method, Chilosi and Federico (2015) show that not all significant coefficients are equally relevant and that the causes of globalization changed during the long 19th century. The convergence was mostly determined by the reduction in barriers to trade during the “early globalization” of first half of the 19th century and by better information (the telegraph lines) and technical progress in sea transportation during the “heyday of globalization” after 1870. The earlier period accounted for most of the overall convergence, and thus ultimately the abolition of barriers to trade mattered more than any other cause.

The analysis of causes has made impressive strides, but the results have to be taken with an abundant pinch of salt. Some authors worry about reverse causation and endogeneity. Existing trade flows or political decisions, such as the access to the Zollverein (Keller & Shiue, 2008, 2014) may have determined the choice of railways or roads to be built. In addition, transportation costs could be endogenously determined by the amount of trade (Jacks & Pendakur, 2010). In these cases, a simple OLS regression might overstate the contribution of new infrastructures or liberalization of trade to integration. These concerns seem excessive. It is surely plausible that total trade affected railway planning, but less likely that trade in a single commodity, even one as important as cereals, could determine it. It is even less likely that any specific commodity or route flow was large enough to affect the level of freights in the competitive world market for shipping.

In contrast with the concern for endogeneity, most authors do not seem to bother about the existence of trade flows between each pair of locations, even if this is an essential condition for price gaps to be a meaningful measure of integration. Sensibly, Brunt and Cannon (2014) explicitly state that they prefer to run their Vector Error Correction Model (or VECM an advanced version of the AR one) only with pairs of neighboring counties, which were more likely to trade than distant ones. Brunt and Cannon (2014) can afford to be restrictive because their database includes a very large number of nearby counties to work with. To be sure, data on trade flows between specific cities are often unavailable, but the anecdotal information is abundant, and thus authors should defend their choices of market pairs beyond the obvious ones (e.g., nobody would doubt that Amsterdam and Konigsberg traded in the 17th century).

The data problems are arguably even worse. First and foremost, there is a major problem of omitted variables for lack of data. A good example is the network of state granaries in China, set up by the Qing dynasty in the second half of the 17th century to reduce the effect of local harvests on local supply, and thus on price fluctuations (Will & Bin Wong, 1991). Its existence might explain the high level of “integration” (or efficiency) of Chinese domestic markets in the 18th century (Shiue, 2002, Figure 3) and its decadence in the early 19th century contributed to the “disintegration” of the market. Second, a lot of explicative variables are measured imperfectly. For instance, using the aggregate nominal protection (the ratio of custom revenue to total imports) as proxy for barriers to trade is bound to return biased coefficients if duties differ substantially between products or change in time, as was the norm in the 19th century. This should not be a big problem for the analysis of market integration, as finding data on duties for primary products is fairly easy. The issue is more serious for transportation costs. By definition, distance does not take into account the different geographical features that affect transportation costs, such as rivers and mountain ranges, and cannot capture changes in them. Most authors add geographical variables, such as dummies for access to rivers or the sea, and Keller and Shiue (2016) claim that ruggedness is a proxy for railway costs in 19th-century Germany. Changes in transportation costs have been modeled by interacting distances with time trends (Jacks, 2006, 2009) or segmenting the period of observation to get different coefficients (Uebele & Gallardo-Albarran, 2015). It is unclear whether these solutions can really substitute for precise measures of transport costs. Building fairly accurate series of costs of water transport is comparatively easy. The distances are fixed and data on seaborne freights, although not necessarily route and commodity specific, are fairly abundant, at least for the 19th and 20th centuries (Federico & Tena-Junguito, 2016). In contrast, the overland distances could vary according to the shape of the network (the opening of a new road or of a new rail line could shorten it) and the data on unit costs are either scarce, as for road transportation, or very abundant, but difficult to collect and unwieldy to manage, as for railways. There were no major changes in productivity of road transportation before the diffusion of the internal combustion engine in the 20th century, but costs depended on the quality of roads. For example, the advent of turnpikes (paved toll roads) in England in the second half of the 18th century slashed transportation costs by 40% (Bogart, 2005). In contrast, the productivity of railways was growing fairly steadily and thus one would expect rates to decline if the market were competitive. In most countries, this was not the case. Railways were state owned or state regulated and gains accrued to railway owners rather than being transferred to customers. In most countries, fares differed by product, route, size of shipment, and so on, and costs of rail transportation varied considerably across products and in time. In these cases, the coefficients for simple dummies for the connection would be biased.11

One might expect that the combined effect of poorly measured dependent variable and the prevailing use of imperfect series and dummies as explicative variables would yield poor results. But this is not the case. Most studies report significant coefficients, which means that either authors are really lucky or the underlying effects were quite strong.

The Effects of Integration

The conventional wisdom assumes that market integration brought substantial welfare gains and was a major cause of economic growth, in all likelihood the major one before the industrial revolution. However, transforming this claim into hard estimates has proven extremely difficult, even more so than in the parallel literature on trade (Federico, 2019). Bateman (2012) ran a series of regressions to explain levels and changes in real urban wages and urbanization in early modern Europe with three different measures of integration: domestic and international price gaps and price volatility. The results of this ambitious attempt are on the whole quite disappointing, as the integration variables are hardly significant, but this is not really conclusive. Integration is imperfectly measured, the proxies for development are questionable, and the model omits many potential variables.

On paper, estimating static gains from integration should be a more manageable task, but very few authors have tried and, as far as we know, none have attempted it for the preindustrial period. Ejrnæs and Persson (2010) and Steinwender (2018) estimate the efficiency gains from the layout of telegraph lines between the United States and the United Kingdom in 1866. They deal with markets for different products with different specifications, but come out with fairly similar figures: the telegraph increased American wheat exports by 2% and cotton exports by 8%, and gross domestic product (GDP) respectively by 0.0073% and 0.0015%. These low figures are not really surprising and one would expect the effects of change in price dispersion to be larger. Federico and Sharp (2013) use a very simple partial equilibrium model to analyze a case of regulatory failure in the U.S. interwar years. The productivity of railways was growing fast, and thus fares for wheat and other commodities could be cut, but the Interstate Commerce Commission, the federal monitoring agency decided otherwise, depriving American consumers of potential gains equivalent to at least 0.5% of GDP. With a more sophisticated multimarket model Chilosi and Federico (forthcoming) estimate that price convergence accounted for about two-thirds of the growth of world trade in cotton and for one-third of the increase in wheat trade from 1815 to 1913. The welfare gains were modest but not negligible for large countries; about a half a percentage point of GDP for United States, the United Kingdom, Western Europe, and India, even less for Eastern Europe, but massive for small specialized exporters such as Egypt.

These partial equilibrium estimates for one or two products are bound to underestimate the total gains from integration. Two papers on integration of domestic markets for all agricultural products find much larger gains. Costinot and Donaldson (2016) use linear programming to estimate the effect of the convergence of local (county) prices on New York prices in American agriculture. The results are impressive; price convergence caused production to increase by 62% from 1887 to 1920 and by 55% from 1954 to 1997, roughly as much as technical progress (respectively 30% and 70%). These figures imply that market integration accounted for about a tenth of total growth in American GDP before 1920.12 Donaldson (2018) explores the effects of railways on agricultural output of Indian districts from 1870 to 1939 with a three-stage strategy. He first shows that trade costs determined flows, then that the existence of a railway connection increased agricultural output by 16% (i.e., GDP by a tenth) and finally that about 85% of this increase can be explained by the additional trade from railways rather than by the substitution of rail for other means of transportation.

Conclusion: A Research Agenda

This survey, although unavoidably selective, shows how far the study of market integration has progressed since the 1990s. Federico (2012) lists 61 works on Europe published before December 2009, and since then several others have been added. Almost all the early studies dealt with measurement of integration, with a heavy emphasis on testing efficiency with time-series econometrics. Attention has since shifted to the causes of convergence, which scholars have explored with a variety of gravity models from trade.

This work has yielded some solid results on trends in Europe and in transoceanic trade. Integration progressed slowly, with setbacks and accelerations, throughout the early modern period and very fast in the “long 19th century”—especially its first half. The few works on the interwar years (Federico & Persson, 2007; Hynes, Jacks, & O’Rourke, 2012) show that the early 1920s were a period of low price dispersion while the Great Depression caused a sudden disintegration of the wheat market. We know much less about trends in domestic than international integration and about trends outside Europe, with few forays on integration in China and India. The 19th century was the golden age of integration because of the coincidence of trade liberalization and a spurt of technical progress in transportation. But throughout history, from the Middle Ages to World War II, waves of integration/disintegration were largely determined by changes in barriers to trade, while political events such as wars were a major source of short-term shocks.

In contrast, the literature on the effects, which arguably should be the key issue in the whole research agenda on integration, is still in its infancy. So far, there are very few estimates of static gains, with a clear trade-off between simple partial equilibrium models, which may underestimate the effects, and more ambitious general equilibrium ones, which are often too data-intensive for the historical research. As far as we know, there is not a theoretically sound and empirically implementable strategy for estimating dynamic effects of market integration (and trade). Finding such a strategy is a key task for the future, but it is not the only one. The analysis of trends first and foremost requires more price data, with research extending to other periods (including post 1950), to products other than cereals, to integration of domestic markets, to non-European countries, and so on. These data would also be useful for the analysis of causes, which needs a more precise measurement of trade costs and some imagination in finding proxies for other determinants of integration. The task is surely easier for market integration, which focuses on a single product, than for trade. Much has been done in this area, but much remains to be done.