The Impact of Access to Rail Transportation on Agricultural Improvement

The Impact of Access to Rail Transportation on Agricultural Improvement: The American Midwest as a Test Case, 1850-1860. During the 1850s, land in U.S. farms increased by more than a third—100 million acres—and almost 50 million acres, an area almost equal to that of the states of Indiana and Ohio combined, were converted from their raw, natural state into productive farmland. The time and expense of transforming this land into a productive agricultural resource represented a significant fraction of domestic capital formation at the time and was an important contributor to American economic growth. Even more impressive, however, was the fact that almost half of these total net additions to cropland occurred in just seven Midwestern states which comprised somewhat less than one-eighth of the land area of the country at that time. Using a new GISbased transportation database linked to county-level census data, we estimate that at least a quarter (and possibly two-thirds or more) of this increase in cultivable land can be linked directly to the coming of the railroad to the Midwest. Farmers responded to the shrinking transportation wedge which raised agricultural revenue productivity by rapidly expanding the area under cultivation and these changes, in turn, drove rising farm and land values. The Impact of Access to Rail Transportation on Agricultural Improvement: The American Midwest as a Test Case, 1850-1860 Introduction Growth accounting suggests that capital formation accounted for approximately one-third of the growth of per capita GNP in the United States between 1800 and 1840 and almost one-half between 1840 and 1900 (Gallman 2000). While some of this growth in the capital stock came from foreign sources, much was created through domestic savings and investment. According to Gallman (2000), for example, the domestic savings rate in the 1850s was between 17 and 20 percent depending on one’s assumptions about prices and home manufacturing. During the same period, a considerable portion of the domestic investment took the form of land clearing and associated improvements to make the land suitable for agriculture. While the federal censuses prior to 1850 did not collect information on the land in farms or improved acreage, we do know from other sources that, during the 1840s, the federal government sold off over 19 million acres—about 30,000 square miles—of public land despite the bursting of the 1830s speculative land bubble (Carter 2006: Table Cf79). Much of this land went (eventually) to farmers and, by 1850, the Census reported over 293 million acres of land in farms, of which about 38.5% had been improved (United States. Census Office and De Bow 1854: Table 183, p. 169). Public land sales during the 1850s increased more than two-and-a-half fold to more that 52 million acres. This represents an area 80 percent larger than the state of Ohio or 40 percent larger than Illinois (Carter, Gartner et al. 2006: Table Cf79). Although it took time for the new Page 1 land to be cleared and enter agricultural production, improved acreage grew by almost 50 million acres during the decade and land in farms expanded by well over 100 million acres. As a result, by 1860, a little over 40 percent of all land in farms had been improved (United States. Census Office. and Kennedy 1864: p. 184). Such additions to the stock of improved acreage may have accounted for more than half of all capital formation before 1820. Even as late as 1860, Gallman (2000) estimates that 22 percent of annual capital formation came from additions to the stock of improved acreage. However, while the quantitative significance of these improvements to farm land has been widely recognized (Primack 1962; Primack 1977; Atack, Bateman et al. 2000), the quantitative significance of the causal factors underlying such investments is less clear. Using a newly developed dataset, this paper sheds light on one such causal factor: the coming of the railroad which lowered the costs of internal transportation and thus expanded commerce. In our analysis, we focus on the decade of the 1850s in the American Midwest since it was there that almost half of the increase in improved farmland during the decade occurred and was also where much of the railroad construction took place, especially between 1853 and 1856. Indeed, in 1856, almost 40% of the nation’s new railroad construction occurred in a single Midwestern state, Illinois. Our new data consist of county-level information on the spread of the railroad network in seven Midwestern states (Illinois, Indiana, Iowa, Michigan, Missouri, Ohio and Wisconsin) during the 1850s which has been linked to existing county level data on farm outcomes at the benchmark census dates. The transportation data, collected as part of a larger project, derive from the geographic information systems (GIS) processing of digitized nineteenth century railroad maps and related sources as described below. Page 2 We measure the causal impact of the railroad on improved acreage using a difference-indifferences (DID) analysis.1 This compares the change in farm outcomes between 1850 and 1860 in a treatment group of counties – identified as those which gained access to the railroad during the period– versus a control group which did not. We have excluded from consideration those counties which already had rail links in 1850 so that our treatment group consists of counties that gained rail access during the 1850s versus those counties which did not have access until after 1860. Our analysis is complicated by the fact that some county boundaries were redrawn during the period while others did not even exist until the 1840s. Consequently, we have also restricted our analysis to the 278 counties in the Midwest that were in existence as of 1840 and did not change their boundaries during the 1850s.2 The location of these counties is shown in Figure 1 below. We have supplemented our analysis of the share of improved acreage with two robustness checks. First, because improved land was more valuable than unimproved land, we should also find increases in the value of farm acreage associated with the coming of the railroad. We measure this as the average value of an acre of land in farms in the county using census estimates of the “value of farms.” This value includes the value of the land as well as improvements and structures and our results confirm that the coming of the railroad did, indeed, have a treatment effect on farm values. Moreover, when we repeat this DID analysis of farm values but control for the percent of land that was improved, we are only able to explain some of the increase in farm values, suggesting that the railroad’s effects on land value were general – that is a pure location effect. 1 The difference in differences approach was largely pioneered by David Card (see (Card and Krueger 1994)) and is now widely used (see, for example, the mini courses given by Jeffrey Wooldridge (2007) to various organizations including the NBER, and Angrist and Pischke (2009)). For criticisms of the methodology, see, for example, the special issue of the Journal of Business and Economic Statistics, especially Meyer (1995). 2 This restriction does not substantively change the results but it makes our analysis cleaner Page 3 A second robustness check addresses a key assumption in our DID analysis that, conditional on various observable determinants, the coming of the railroad to a county was a random event. If this assumption is false then the estimated treatment effect of the railroad on farm improvements would be biased. In particular, if county residents anticipated getting a railroad it would make economic sense to “build-in-advance” – that is, improve acreage in advance of the railroad. If this were the case, then our DID treatment effect would be biased downwards and our estimates of the effect of railroad would understate the true impact.3 To address this possible bias we compare our DID results with those using an instrument variable (IV). This instrument is derived from various federal government transportation surveys conducted in the 1820s and early 1830s—that is to say, right at the start of the railroad age and well in advance of when we observe the possible effects of the railroad on these Midwestern counties. These results are consistent with our initial DID findings of a relative large effect of the coming of the railroad such that it accounted for at least a quarter of the increased fraction of improved land and suggest that our DID estimates of the impact of the railroad may indeed be biased downwards. Nineteenth Century Land Clearing While the first step to becoming a farmer was to acquire title to the land either through private purchase or directly from the federal government under the terms of the prevailing land legislation, the purchase of raw land was generally a minor cost compared with the cost of clearing it and readying it for cultivation. For example, according to contemporary estimates by 3 This, for example, is a key insight in the work of Levinson (2008) and Xie and Levinson (2009) Page 4 Seaman (1852: 484 and 656), the cost of clearing and fencing land averaged between $11 (in 1850) and $12 (in 1840 when more of the land was wooded) per acre while Lebergott (1985, Table 3) estimates the median cost of improving an acre of land in the Midwest at between $14 and $25 in 1850. In contrast, the average price of raw land bought using military land warrants was less than a dollar per acre (Swierenga 1968, Table 6.1; Oberly 1990). These cost differentials reflect the fact that where the land was forested, trees had to be cut down and burned or hauled away. Even then, the more laborious and labor intensive task of stump removal remained. In areas where the land was naturally clear of trees, as on the Midwestern prairies, clearing expenses remained high as heavy, specialized equipment and crews were needed to break the prairie sod (Stellingwerff and Swierenga 2004). Even when the land had been cleared, it still had to be fenced to keep the livestock in and the predators out. Many of these activities were expensive, arduous, and time-consuming even when they could be done by the farmer and the family using only those tools and implements that came readily to hand on a nineteenth century farm. As a result, farmers rarely cleared all the land they needed or wanted at once. According to Primack (1962; 1977), farmers could clear 5-10 acres a year. Thus, clearing the approximately 1.3 million acres of land (much of it forested) that was cleared annually in the Midwest in the 1850s must have occupied the time of about one-sixth of the Midwestern labor force at the time. Needless to say, with this kind of investment in cash and/or time, cleared land was highly valuable compared with virgin stock and accounted for much of the resale price of farm land. The Impact of Transportation Improvements on Land Values The economic logic linking transportation costs to land improvements can be illustrated using von Thünen model of agricultural land rents first proposed in 1826 (Thünen and Hall Page 5 1966). To fix ideas, suppose that each farmer has a farm of size one acre (of improved land), located at distance t from a “central place” where trade with the hinterland (i.e., the farmer) takes place. The farm supplies labor z per acre at an opportunity cost of w. The costs of transporting output per unit are c and the price of output (at the central place) is p. In a competitive market these assumptions imply that the rental price of a farm at distance t, R(t), from the central place is R(t) = q (p – w·z – c·t) Note that, if t = 0 – the farm is located on the boundary of the central place and R(t) is at a maximum as determined by the yield of the land (q) given the labor input (z), output price, and the opportunity cost of the farmer’s time. Conversely, there will exist a t* that defines the distance from the central place beyond which it is not profitable to farm, t* = (p – w·z)/c. Economists sometimes refer to this as the extensive margin for market production. It is straightforward to show that dR(t)/dc = - q·c < 0 and dt*/dc = (p – w·z)/c2 > 0; that is to say, a reduction in transport costs “flattens” the relationship between land rents and distance to the central place and, therefore, increases the distance t* at which farming is still just profitable. A fall in transport costs therefore makes it economical to bring more land into production. However, before this land can be brought into production it has to be improved – that is, cleared and fenced– hence, the link between transport costs and land improvements. Of course, other factors will also determine t* and therefore affect the amount of land in production. A rise in p, for example, or a reduction in w, will also increase t*. Moreover, in an explicitly dynamic setup the permanent versus transitory nature of changes in these parameters will influence the likelihood of improving land and bringing it into production, as will Page 6 expectations about future values. In particular, railroad building was (in general) a permanent change in the environment. It could thus be expected to generate a greater response in land improvements than, say, a shift in output prices which might prove to be transitory. One reason for this was that much of the railroad investment was not only fixed but also sunk in that it was embedded in location-specific improvements. A railroad, once build, could not be easily relocated and the benefit of that investment could only be fully captured by another transportation medium—investments in grading, filling and cuttings, for example. According to the 1880 Census (United States. Census Office. 10th census 1880., Walker et al. 1884), over 80 percent of railroad investment went into construction costs, of which only one or two percent represented the cost of the land itself; the rest was spent on surveying, grading, removing or bridging obstacles, and laying the track. While the railroad ties, ballast and the rails might be reused elsewhere and the land itself could be resold, the grading, cuttings, embankments, bridges, and drainage ditches had few alternative uses other than serving the needs of a transportation system—especially in the nineteenth century.4 Indeed, railroads (and interstate highways) today still follow many of the routes blazed by the nineteenth century railroads.5 If there were any adjustment costs involved in making improvements to their land—as would seem likely—it would be rational for farmers to improve some land in advance of construction if they anticipated receiving rail access in the near term—that is farmers should be 4 Nowadays, of course, they find recreational uses in the “rails to trails” movement And, of course, thanks to GIS, these can be quite precisely located. One might thus be tempted to begin with modern railroad maps and force the nineteenth century locations to conform to these. On average, this is probably a good strategy even though it would misrepresent earlier track locations that were changed through realignments. There are two downsides to this strategy—the potential for copyright infringement since these GIS SHP files are proprietary commercial ventures and the large number of individual segments which comprise the modern rail network. For example, the 177,000 mile rail network from Bureau of Transportation Statistics and distributed ESRI with ArcGIS contains over 131,000 segments. See (United States Department of Transportation. Bureau of Transportation Statistics. 1998) 5 Page 7 forward-looking rather than merely reactive.6 We will return to this point later in our instrumental variable analysis, as it relates to a possible downward bias in our difference-indifferences estimation. The Data We use a newly constructed geographic information system (GIS) database on the Midwestern transportation infrastructure in the mid-nineteenth century United States which we have linked to an updated version of the well-known ICPSR county level census database. This GIS-based methodology has several advantages over earlier approaches which have generally involved matching historical transportation maps to county boundary maps by hand (and eye).7 Our GIS database was created from geo-referenced digitized historical maps showing the transportation system—canals, navigable waterways and railroads—at different benchmark dates.8 We have not yet tried to render information on roads and trails. The canal data were taken from Poor (1860) and Goodrich (1961) as well as digitized maps from the Library of Congress “American Memory” web site such as the maps prepared by Williams (1851) and Disturnell (Burr 1850). Information on navigable rivers was taken from 6 For a similar point, see, for example, (Levinson 2008) See, for example, (Craig, Palmquist et al. 1998) who visually compared historical maps to county boundaries (which generally did not appear on the historical maps). For a discussion of the hand-matching procedure used by Craig and the pitfalls that can arise, see footnote #2 of (Atack, Bateman et al. forthcoming) 8 Geo-referencing refers to the process of fixing specific points with known geographic coordinates between the digitized image--which was invariably drawn and printed with error (and which may also be subject to parallax error as a result of the digitization process)--and the geographically accurately rendered base boundary files (a shapefile in ESRI’s parlance). Algorithms within the GIS software then distribute the error (the difference) between the historical images and the boundary file across the space between fixed points. In essence, the process treats the historical image as if it were printed on a sheet of rubber which is then stretched over the boundary file with pins holding it in place at fixed reference points between the two. Once done, it is then possible to “trace” features from the historical image onto the geographically coordinated boundary file. The resulting files can then be manipulated and used for computations using the GIS software. Accurate historical county boundary files are freely available from the National Historical Geographical Information System at the University of Minnesota (http://www.nhgis.org/) along with a wide variety of historical U.S. census data. 7 Page 8 U.S. Army Corp of Engineers GIS data (Vanderbilt University. Engineering Center for Transportation Operations and Research 1999) supplemented by information from Hunter (1949) and contemporary newspaper accounts regarding steamboat service on specific rivers. Since very little change occurred in the canal and river systems after the 1840s until the Corp of Engineers embarked on its ambitious navigation plans beginning with the establishment of an “Office of Western River Improvements” in 1866, we have measured access to water-borne transportation as of a single benchmark date – 1850 – rather than by change over time. The railroad portion of our GIS database is based on digitized state-level maps for 1911 from The Century Atlas (Whitney and Smith 1911) since these maps appear to be accurately drawn and the rail network was largely completed by that time but had not yet begun to shrink through closures.9 This 1911 railroad network was then traced back through time to earlier dates using digitized transportation maps from the Library of Congress and other sources such as the maps created by Taylor and Neu (1956). Our implicit assumption in this approach is that most railroad investment was literally sunk in location-specific grading and other immoveable features, a presumption that is strongly supported by the available data. For the Midwest which is our focus here, we have supplemented these more general sources with detailed information for the seven states in our Midwestern region. Especially detailed information about the spreading rail network in five of these states (Indiana, Illinois, Michigan, Ohio and Wisconsin) exists in the form of a series of crudely-drawn maps and a data appendix prepared over ninety years ago at the University of Wisconsin by Professor Frederick L. Paxson (1914) and his students. They were able to use extant contemporary travel guides published 9 These digitized images were purchased from Goldbug.com but the original source may be found in many libraries around the country including Vanderbilt’s. Page 9 between 1848 and 1860 as their principal sources of information.10 Since then, however, many of these travel guides which were cheaply printed ephemera seem to have been lost or disintegrated. This is unfortunate as there is little doubt that they are, or rather would have been, the single most valuable resource for our research. These guides served the needs of the traveling public, providing up-to-date route maps and timetables. Consequently, they were published at fairly frequent and regular intervals by a number of different companies and, presumably, competition among them should have guaranteed that only the best (that is, the most accurate, complete and useful) would have continued publication for an extended period of time.11 Our data for Iowa and Missouri are taken from digitized contemporary state maps from the on-line David Rumsey Map Collection (http://www.davidrumsey.com/) for 1856 (Colton 1856; Colton 1856), 1857 (Colton 1857) and 1859 (Mitchell 1859; Mitchell 1859) supplemented by national maps from the Library of Congress collection for 1858 (Sage 1858) and 1860 (Colton 1860). While GIS software is capable of generating a wide variety of different measures of transportation access, at the present time we use a simple county-level binary variable, ACCESS. This takes on a value of 1 if a railroad crossed or formed a county’s border. Our choice of this measure reflects the still-preliminary nature of our GIS database. While we are certain that it can create the binary “0-1” access variable accurately, we are not yet convinced that the original 10 These travel guides first appeared in the 1840s and include Disturnell's Guide (Disturnell 1846), Doggett's Gazetteer (Doggett 1847), Appletons' Guide (Appleton D. and Company. 1848), Dinsmore's Guide (Cobb 1850), Lloyd's Guide (Lloyd 1857), Travelers' Guide (National Railway Publication Company. 1868) and The RandMcNally Guide (Rand-McNally 1871). Some of these were published monthly; others, semi-annually or annually. Each typically went through many editions. All of the guides that we have physically handled are fragile, especially the multi-page fold-out maps, and not sturdy enough for scanning or copying, although a few guides have been digitized and are available on-line. See, for example, the June 1870 copy of the Travelers’ Official Railway Guide at http://cprr.org/Museum/Books/I_ACCEPT_the_User_Agreement/Travellers_Guide_6-1870.pdf from the Central Pacific Railroad Museum. There are also at least two different editions of Appleton’s Guide on Google Books such as http://books.google.com/books?vid=UOM39015016751375 as well as a number of other guides. See http://www.lib.utexas.edu/maps/map_sites/hist_guide_sites.html 11 However, as we note below, these sources do not seem to have been infallible—certainly at any moment of time. Page 10 maps and our “tracings” of them are sufficiently precise to generate more refined measures of access, such as the miles of track in the county. Another measure of “access”—proximity to a railroad stations or depots—can only be measured beginning in the mid-1850s when maps begin to record this important information. For example, the first digitized map in the Library of Congress collection to mention stations and depots explicitly is by Ensign, Bridgman and Fanning (1856). Paradoxically, it shows many more named places along the right of way than the map from Dinsmore’s travel guide for same year which one might have thought would show prospective travelers their station destinations. Nor does it appear that the Ensign, Bridgman and Fanning map was overly sanguine in its reporting of railroad stations. Another map published two years later (Sage 1858) shows most of the same stations as Ensign, Bridgman and Fanning plus additional ones along the new rights of way as well as a few others scattered between some of the 1856 stations and depots. Beginning in the 1870s, Rand McNally (Morgan 1969) along with other publishers began to produce “commercial” maps for the convenience of shippers, showing destinations to which freight could be consigned making the reporting of depots and stations more reliable and systematic. As a practical matter, however, trains could potentially stop and load almost anywhere (albeit inconveniently)—much as they do in parts of Alaska even today—and stations were built only when freight and passenger volume had passed some critical economic threshold. Despite our measure of access to transportation being crude and limited, it is still the best current measure of the historical spread of the U.S. railroad system at the county level for this period.12 These transportation access data have then been linked with the county-level Haines- 12 One obvious problem is that the access variable does not capture within-county variation in access and counties differed in size. In our econometric analysis we weight by county land area but we are unable at the present time to adjust for cases in which the railroad lay close to a county boundary but where the adjacent counties do not have rail Page 11 ICPSR census data using county FIPS codes which are common to both databases. In 1840, the seven states and territorial areas in our study had 391 “counties” (these include, for example, Clayton “county” in the Iowa Territory which then comprised most of what we now know as the state of Minnesota) (Thorndale and Dollarhide 1987). By 1860, however, all of the area under consideration had been organized into states and was divided into 623 counties. Because of boundary changes, we have restricted our analysis to those counties that were (1) present beginning in 1840, (2) had the same county boundaries (as determined by square mileage in 1850 and 1860), and (3) did not have rail access as of 1850. Taken together these restrictions produce a balanced panel to 278 Midwestern counties. They are distributed as shown in Table 1 while their locations are mapped in Figure 1.13 Qualitatively, our results are unchanged if we relax the requirement that boundaries remain unchanged. We have divided this panel into a treatment group and a control group. The treatment group did not have rail access in the 1850 but gained it at some point between 1850 and 1860. Our control group is the set of counties in the panel that did not have rail access in either 1850 or 1860. The 1840 data for these same counties then allows us to control for pre-1850s trends in certain variables in our analysis (see below). access. County boundaries, in other words, are arbitrary ways of delineating local economies; the average farmer in a geographically small county in our control group that happened to be adjacent to a large county in the treatment group might have had better access to the railroad than the average person in the large county. If this were common it would produce a downward bias in our estimated treatment effects. Although we do not believe this bias to be important empirically we admit that the question is open. 13 By “balanced” we mean that the same counties appear in 1850 and 1860; no new counties enter the sample during the 1850s. Balancing ensures that county fixed effects are “differenced away” when we compute the change in economic outcomes from 1850 to 1860; this would not be the case if new counties entered the sample in the 1850s. We restrict our basic analysis to counties with fixed land area because the ICPSR census data are not adjusted for changes in land area over time. Results are qualitatively the same if we do not impose the restriction that the county not have rail access by 1850 and if we do not impose the requirement that county boundaries be the same in 1850 and 1860. Page 12 Table 1: Distribution of Balanced Panel of Sample Counties State Number of counties Percent of Total (%) Illinois 71 25.5 Indiana 66 23.7 Iowa 17 6.1 Michigan 12 4.3 Missouri 41 14.7 Ohio 57 20.5 Wisconsin 14 5.0 Total 278 Notes: To be included in the sample, counties must be (1) present in all three census years: 1840, 1850 and 1860; (2) have fixed county boundaries; and (3) either not have rail access in 1850 but gained rail access by 1860 (=treatment group) or did not have rail access in either 1850 or 1860 (=control group). By construction, none of the counties in our sample had rail access in 1850. By 1860, however, 195 of the 278 counties, representing almost 72 percent of the land area in the panel, had gained direct access to at least one railroad (See Table 2 and Figure 2). The share of farm acreage that was improved rose by 13 percentage points between 1850 and 1860 across all counties in the panel whereas the logarithm of per acre value of farms, which we use in our robustness checks (see below), increased by 0.796 log points, or 123 percent (= (exp[0.796] – 1) x 100)). Table 2 also shows the sample averages for percent improved and the logarithm of per acre farm values conditional on their treatment status. In the treatment counties, the percent of land that was improved rose by 15.6 percentage points from 1850 to 1860 or almost double the 8.4 percentage point increase in the control counties. The difference between these—7.2 percentage points—is the difference-in-differences estimate of the treatment effect of gaining rail access Page 13 Figure 1: Percentage Change in Improved Farm Land, 1850-1860 in Sample Panel of Counties with Stable Boundaries Page 14 with no other control variables in the regression (see below). This is equal to 18 percent of the mean proportion of land that was improved in 1850. As such, the effect is economically large. The equivalent difference-in-differences estimate of the treatment effect of rail access on per acre farm values is 0.072, or about 7.5 percent. Both estimates are statistically significant at the 5 percent level. Table 2: Sample Statistics of Key Variables in Aggregate and in Control and Treatment Counties 1850 1860 ∆ (1860 – 1850) Rail = 1 0 0.717 0.717 % Improved Acres in 0.390 0.526 0.136 Agriculture Treatment 0.411 0.567 0.156 Counties Control Counties 0.337 0.421 0.084 Treatment 0.074 0.146 0.072 Control Ln (Per Acre Farm 2.119 2.915 0.796 Value) Treatment 2.211 3.028 0.817 Counties Control Counties 1.887 2.632 0.745 Treatment 0.324 0.396 0.072 Control Source: See Table 1. There are 278 counties in the sample. Treatment = 1 if county gained rail access between 1850 and 1860. Control = 1 (Treatment = 0) if the county does not have rail access before the Civil War. There are 195 treatment counties and 83 control counties. Observations are weighted by square miles prior to calculating sample means. Ideally, in a difference-in-differences analysis, the treatment and control counties will be sufficiently “similar” or “matched” such that treatment can be (plausibly) argued to be randomly assigned. However, it is immediately evident from Table 2 that this was not the case with rail access. Counties that gained rail access in the 1850s already had a greater fraction of their farm acreage improved and more valuable land by 1850 – that is, before gaining rail access. Page 15 Figure 2: Midwestern Railroad Network as of 1880 (in green) Shown Relative to Panel of Counties with Stable Boundaries (outlined in black) (Railroads shown in blue cross-hatching were built before 1850. Those shown in red were built between 1850 and 1859) Page 16 The observations that the treatment counties already had more of their farm acreage improved and more valuable farms (in part, because the land was more improved, see below) are manifestations of the same general issue –railroads did not arrive randomly. In particular, railroad promoters and investors in the Midwest paid close attention to agricultural development because the profitability of the railroad depended upon it (Gates 1934; Fishlow 1965). This is illustrated in Table 3, which shows the correlates, agricultural and otherwise, of gaining rail access in the 1850s using a simple linear probability model (the dependent variable equal 1 if the county gained rail access in the 1850s – that is, the county is in the treatment group – and zero if not – the county is in the control group). In column 1 of Table 3, we have included just three covariates: the logarithm of the value of agricultural output per acre (a measure of the land’s “yield,” in this case its revenue productivity); the percentage of total agricultural output in 1840 represented by wheat production; and the change in the percent wheat between 1840 and 1850.14 Wheat is a crop of particular importance at this time as it was the primary “cash crop” of the period and, therefore, likely a key determinant of whether a rail line would be profitable. Each of these variables had a positive and significant effect on the likelihood of gaining rail access; for example, a 10 percentage point increase in the wheat share boosts the probability of gaining rail access by 5.3 percentage points. 14 The 1840 census reports the value of wheat production and the total value of agricultural output. For 1850 we use an estimate of the value of total agricultural output based on national prices multiplied by quantities; the wheat share is therefore the wheat output (in bushels) multiplied by wheat price divided by the estimated value of agricultural output. It is possible that our procedure for estimating agricultural values may overstate the growth in the percent wheat over the 1840s (because 1840 output was probably valued at local prices) but any such bias should be mitigated once we control for state fixed effects (since state-level variation arguably captures the most salient price variation). Our agricultural yield variable also uses the estimated value of agricultural output in its construction (the numerator). Page 17 Table 3: County-Level Agricultural Correlates of Gaining Rail Access in the 1850s: Linear Probability Regression Constant Log (Agricultural Yield in 1850) Percent Wheat in 1840 ∆ in Percent Wheat, 185040 Urban and Population Density Variables Included? Water Transportation Dummies Included? State Dummies Included? Adjusted R2 Coefficient Coefficient Coefficient 0.222* (0.089) 0.325* (0.077) 0.533* (0.244) 0.684* (0.205) No 0.365* (0.123) 0.290* (0.078) 0.349 (0.253) 0.652* (0.272) Yes 0.456 (0.150) 0.280* (0.092) 0.356 (0.283) 0.615* (0.275) Yes No Yes Yes No 0.111 No 0.190 Yes 0.211 Source: see text and Table 1. Notes: Unit of observation is the county (N = 278). Dependent variable = 1 if county gains rail access by 1860 (treatment group), 0 otherwise (control group). Agricultural Yield: value of agricultural output/(improved + unimproved acres in farms). Percent wheat: value of wheat output/value of total agricultural output. Urban and Population Density variables: percent urban in 1840, ln (population/square miles) in 1840, change in percent urban between 1840 and 1850, change in ln (population/square miles) between 1840 and 1850. Water transportation: canal = 1 if canal existed within county boundary (or part of boundary), river = 1 if navigable river existed within county boundary or part of county boundary, Great Lakes = 1 if county abutted one of the Great Lakes. Prior to estimation, observations are weighted by the number of square miles in the county. Standard errors are shown in parentheses. *: significant at 5 percent level. However, the significant positive coefficients on the agricultural variables may also be due to other factors that are omitted from the regression. For example, in an earlier paper (Atack, Bateman, Haines, and Margo, forthcoming) examined the relationship between gaining rail access and measures of urbanization, population density and our water transportation dummies Page 18 using the same panel. Accordingly, in column 2 of Table 3 we have included measures of urbanization and population density in 1840 as well as changes in these variables between 1840 and 1850. We have also included dummy variables for access to different modes of water transportation (river, canals, and abutting the Great Lakes) while in column 3 we have also included dummy variables for each state. Adding additional controls reduces the magnitudes of the coefficients of the agricultural variables but the coefficients on the wheat “yield” variable and the “pre-trend” (1840-50) in the wheat share remain statistically significant. Moreover (not shown) several of these control variables are themselves statistically significant – for example, the presence of a canal was positively associated with gaining rail access whereas having a navigable river reduced the statistical likelihood of gaining rail access. Because gaining rail access clearly was not a random event, it is important, at the very least, that we control for the factors shown in Table 3 in our difference-in-differences analysis as they may have had independent effects on the change in percentage of improved acreage between 1850 and 1860. Column 1 of Table 4 reports our base DID estimate (that is, absent any controls for the other correlates of rail access) for the percentage of total farm acres that were improved. This is the same as the figure reported in Table 2. In column 2, we report the DID estimate of the treatment effect, taken from a regression in which we have included interactions between the full set of control variables in Table 3 and the dummy variable for the 1860 census year. Including these control variables reduces the DID treatment effect from 0.072 (base estimate) to 0.048, or 4.8 percentage points. However, although the addition of controls reduces the impact of getting rail access, the treatment effect remains significantly positive: counties that gained rail access in the 1850s increased improved acreage to a significantly greater extent than counties that remained rail-free before the Civil War even after controlling for other factors. Page 19 Table 4: Difference-in-Differences Estimates of the Impact of Gaining Rail Access on the Percent of Farm Acres that were Improved: 278 Midwestern Counties, 1850-60 Rail Access = 1 Controls? Coefficient Coefficient 0.072* 0.048* (0.015) (0.015) No Yes The unit of observation is the county. The dependent variable is the percent of farm acres in the county that were improved as reported in the census of agriculture. The coefficient of Rail Access = 1 is the coefficient of a dummy variable for the treatment counties (Treat = 1) interacted with a dummy variable for year = 1860. In column 1 there are no control variables other than fixed effects for counties and year = 1860. In column 2 the regression also includes interaction terms between the control variables in column 3, Table 3 (see the notes to Table 3) and year = 1860. Prior to estimation, observations are weighted by square miles in the county. Standard errors shown in parentheses are clustered at the county level. *: significant at the 5 percent level. We have made two robustness checks on this finding. First, we estimate the treatment effect of gaining rail access on the per acre value of farms. If the logic underlying the von Thünen model is correct, then the impact of gaining rail access should also be reflected in higher per acre farm values. We have already shown that this was the case with the simple DID calculations in Table 2 but those estimates do not control for additional factors associated with gaining rail access. Moreover, the size of the treatment effect should be smaller still if we also control for the change in the percent improved (since improved land itself was more valuable) – again, something that is not reflected in DID estimate in Table 2. Table 5 shows the difference-in-differences estimates for the per acre value of farm land. In contrast to the results for improved acreage, the effect of gaining rail access is larger once we Page 20 add the additional controls. Note that, when we also add the change in the percent improved acres to the regression, the treatment effect is reduced in size, as the von Thünen framework would predict but controlling for the percent improved does not eliminate the effect of gaining rail access on land values. This suggests that transportation improvements raised the value of unimproved land as well. This, of course, generally should be the case since most unimproved land has an option value – the farmer could always choose to improve it in the future. Table 5: Difference-in-Differences Estimates of Rail Access on Ln (Value of Farms/Farm Acres): 278 Midwestern Counties, 1850-1860 Rail Access = 1 Coefficient Coefficient Coefficient 0.072* 0.139* 0.107* (0.033) (0.034) (0.032) % Acres Improved in Farms Controls? 0.657* (0.164) No Yes Yes The dependent variable is the natural logarithm of the value of farms as reported by the census (this includes the value of land plus structures, fencing, ditching etc.) divided by total farm acres (improved + unimproved). See notes to Table 4 for controls in columns 2 and 3. All regressions include fixed effects for counties and year = 1860. Standard errors are clustered at the county level. *: significant at the 5 percent level. Our second check on the robustness of our results is made using an instrumental variable. Although we can control for the various correlates in Table 3 in our DID analysis, it is still possible that the analysis is invalid because the treatment – gaining rail access – is correlated with the error term even after controlling for observable determinants of gaining rail access. This problem is of particular concern here because the 1840 census reported no acreage statistics Page 21 making it impossible for us to control for the pre-1850 trend in the percent of land that was improved. This will likely bias the treatment effect towards zero. To understand why this is so, consider the following “thought experiment.” Suppose that farmers in a county anticipated gaining rail access in the near future – a reasonable supposition for Midwestern farmers at the start of the 1850s.15 If the costs of improving acreage were increasing at the margin which was almost certainly the case since the most readily improvable land would be improved first, it would make economic sense to incur some of the costs prior to gaining rail access. If this were the case, counties that gained rail access in the 1850s could have experienced above average growth in improved acreage prior to 1850 and this would be reflected in a higher average percent improved in 1850 – as we observed was the case (recall Table 2). 16 Conversely, some of the control counties might have anticipated (correctly, as it happens) that they would only gain rail access after 1860. By this same logic, they would have experienced above average growth in improved acreage in the 1850s. To the extent that our control variables capture this process, the bias can be “corrected” but we cannot be sure we have captured all of it. Consequently, the percent improved will grow too slowly in the treatment counties in the 1850s and too fast in the control counties, causing the DID estimate of the treatment effect to be biased downwards. 15 For example, based on new rail construction in 1848 and 1849, farmers in five counties in the balanced panel (Boone, Dearborn and Washington counties in Indiana, Genesee county in Michigan and Henry county in Ohio) might reasonably have expected that they too would have a railroad in their county within a year or so. In fact, if a dummy variable (ANTICIPATE = 1) is included in the rail access regression in Table 3, the coefficient is positive (β = 0.314), indicating that such counties did gain rail access in the 1850s. However, the standard error of the coefficient is relatively large (s.e. = 0.195); more importantly, including this additional dummy variable has virtually no effect on the other coefficients in Table 3. Further, if include ANTICIPATE in the DID specifications (Table 4, or Table 5, full set of controls) the estimated treatment effects of rail access are unchanged. Presumably this is because the number of counties -- five -- is so small that the effect cannot be discerned;in the 1840s, most of the rail construction was taking place much further east. 16 If investment in improved acreage in the treatment counties in anticipation of the arrival of the railroad occurred after 1850 these would be reflected in the 1860 figure and would not bias the estimated treatment effect. Page 22 The appropriate correction for this sort of bias is to estimate the relationship using an instrumental variable—a variable that predicts gaining rail access in the 1850s when we control for other factors but which is otherwise uncorrelated with the outcomes we are examining. Using a variable which isolates plausibly exogenous variation in rail access such as we propose here is similar to what would have been the case if rail access had been randomly assigned. We then predict rail access in 1860 using the instrumental variable and examine the effect of that predicted access on the percent of land that was improved in 1860. The historical narrative of internal improvements in America, particularly that for canal construction, assigns an important role to government in promoting these advances (Goodrich 1961). One such important source of government assistance was the assignment of the Army Corp of Engineers to conduct surveys for potential transportation routes. Beginning in 1824, the President was granted authority to survey routes for “such roads and canals as he may deem of national importance, in a commercial or military point of view, or necessary for the transportation of the public mail” (U. S. Congress. 1823-4). Although railroads were not mentioned in the original act (hardly surprising since it predates even the Stockton to Darlington Railway in England, the first steam railroad in the world), it was not long before surveys conducted under this legislation also considered them. For example, in 1825 a survey to “ascertain the practicability of uniting the headwaters of the Kenawha [sic] with the James river and Roanoke river” expressly mentions railroads. Railroad routes soon came to dominate the surveys with perhaps as many as 61 such surveys being made before the law was repealed by Andrew Jackson’s administration effective in 1838 (Haney 1908, p. 277). Our instrumental variable is derived from the government surveys reported in American State Papers and compiled by Haney (1908, p. 283). For each survey, we have identified the counties Page 23 at the start and endpoint of the proposed line. For example, a 1831 railway survey plotted a route from Portage Summit on the Ohio Canal (near Akron) to the Hudson River (we used Albany as the terminus is not otherwise specified) (Haney 1908, p. 286). In some cases, we inferred both endpoints such as for a 1832 survey for a route between “the Mad River and Lake Erie” in Ohio (Haney 1908, p. 286). We used Springfield and Sandusky as the termini of this projected railroad. Having identified the starting and ending counties, we then drew a straight line between the center of the “start” and “end” counties. Counties that lay along this straight line received a value of one, while those that did not were coded as zero. That is, if a railroad were built, our instrument presumes that it would be built in a straight line as the shortest distance between the two points.17 These U.S. Army Corp of Engineers’ surveys provided valuable information to the general public and prospective railroad promoters alike regarding topography and other factors that would affect potential construction costs. Since the costs of these surveys were borne by the public purse, their existence should have raised the likelihood that a railroad would eventually be built by lowering its private costs. Indeed, George Rogers Taylor (1951, p. 95) even argued “[a]s trained engineers were still very scarce … the government rendered a uniquely valuable service by making its experts available for such surveys.” Moreover, as Haney (1908, p. 284) observed “it is of some significance that in most cases the routes of these government surveys were early taken by railways … in the great majority of cases these early surveys have been closely followed” [emphasis added]. Indeed, many of the interstate highways today follow these same routes. Presumably, therefore, the surveys were found to be very useful and so it would seem 17 Our use of a “straight-line” instrument is inspired in part by Bannerjee, Duflo, and Qian (2006) who construct a similar instrument for their study of the impact of rail access in modern China on wages. Or course, many features of topography other than the shortest-distance criterion – grade, hills or mountains, and so on – influenced railroad building but these features of topography likely affected density and urbanization directly, and thus are not candidates for instrumental variables (they fail the exclusion restriction). Page 24 that our “Congressional Survey” instrument should be well suited to predicting whether or not a county gained rail access. Table 6: Instrumental Variable Estimates of the Treatment Effect of Gaining Rail Access on the Percent of Farm Acres that were Improved: 278 Midwestern Counties, 1850-60 First Stage, IV Rail Access = 1 Second Stage (2SLS) 0.192* (0.073) Congressional Survey =1 0.316* (0.068) IV – DID = 0, Significance Level Sample Controls? 0.060 1860 1860 Yes Yes The first stage shows the coefficient of the Congressional Survey instrumental variable from a linear probability regression of rail access. Controls are the same as in column 3 of Table 3. The second stage shows the coefficient of the predicted value of rail access from a regression of percent improved using the 1860 observations from the panel; control variables are the same as in column 1, above. IV – DID = 0: significance level of a test of the difference between the second stage coefficient (0.192) and the DID coefficient from Table 4, column 2 (0.048). Table 6 reports our estimates using this instrumental variable. The Congressional Survey instrument does quite well in predicting treatment (gaining rail access in the 1850s) even when we control for all of the other variables included in the DID analysis (this is the “first stage” coefficient shown). We have then used this first stage regression to predict the probability of Page 25 gaining rail access, and used the predicted values of treatment in the second stage of the twostage least squares (2SLS) regression of 1860 outcomes. The 2SLS coefficient is positive and much larger than our DID estimate (0.192 versus 0.048). The difference between the two is also statistically significant at the 6 percent level. The much larger IV coefficient strongly suggests that the anticipatory logic described earlier was present, leading to a downward bias in the estimated treatment effect of gaining rail access on improved acreage. Concluding Remarks We have used a novel data set on the antebellum transportation networks in the Midwest derived from applying GIS software to digitized historical maps to estimate the impact of gaining rail access on investments in improving land for agricultural activity. Such investment during the nineteenth century was a major component of U.S. capital formation at the time. Counties that gained early access to the railroad – in our case, prior to 1860 rather than later – experienced significantly greater increases in the percentage of farm acres that were improved. Such counties also experienced above average increases in the value of their farms. A part of this was because of the rise in the share of improved acreage but it also reflected the fact farm land in general was more valuable because of its higher revenue product and locational advantage with the advent of the railroad. We can use the estimated treatment effects to gauge the overall explanatory power of the railroad as a factor behind these agricultural improvements. Using the difference-in-differences estimate (0.048, Table 3, column 4), then predicted change in percent improved acreage between 1850 and 1860 due to gains in rail access is 0.034, or 3.4 percentage points (0.048 x 0.717 = Page 26 0.034). This predicted increase is 25 percent (0.034/0.136 = 0.25) of the increase in percentage of improved acreage over the 1850s. If, however, we were to use the much larger IV coefficient (0.192), then predicted increase in percent improved acreage is correspondingly larger: 0.138. Taken at face value, this would imply that the coming of the railroad accounts for all of the increase in percent improved acreage over the decade—indeed, that expansion would have been even more rapid but for some (unidentified) retarding factors. Although the IV coefficient is almost significantly different from the DID coefficient at conventional significance levels, it is not significantly different from, say, the average of the DID and IV coefficients. If we use this average ([0.048 + 0.192]/2), 0.12, then we can attribute slightly more than two-thirds (68 percent = [0.12 x 0.771]/0.136) of the growth in improved acreage to the spread of the railroad. In short, whatever else might have led Midwestern farmers to undertake the backbreaking labor of clearing their land for crops and livestock, it is likely that no other single factor was as important as the potential gains from trade deriving from the coming of the railroad. Page 27