Why they moved emigration from the Swedish countryside to the United States

Transcription

1 Why they moved emigration from the Swedish countryside to the United States Jan Bohlin and Anna Maria Eurenius Department of Economic History School of Business, Economics and Law University of Gothenburg Paper presented to the 8th Conference of the European Historical Economics Society, Geneva, 3 6 September 2009 Abstract The determinants of emigration from the Swedish countryside to the United States, , are explored by means of a panel data set with yearly observations for 24 counties. The emigration rate declined over the long term, which is explained by an improvement in the standard of living and employment opportunities. Persistent regional differences in the emigration rate are explained by regional differences in population density and emigration tradition.

2 1. Introduction Emigration from Scandinavia began to accelerate around Between 1870 and 1910 just over a million, corresponding to roughly one fifth of the average total population in this period, left Sweden. The overwhelming majority went to the United States. In the 1880 s Swedish emigration rates were only superseded by those of Norway and Ireland, and still in the 1890 s Swedish emigration rates were comparatively high (O Rourke and Williamson 1999, p. 122). The majority of the population lived in the countryside. Accordingly, the bulk of emigrants also originated from the countryside. At the height of the emigration in the 1880 s not much concern was voiced over the massive emigration. On the contrary, some voices, notably the Swedish economist Knut Wicksell, argued that emigration was a safety valve that diminished overpopulation and demographic pressures in the countryside. Twenty years later, when the last emigration wave before World War I blossomed, opinion had changed. Much concern was now raised over emigration. Landowners and other agrarian interest groups worried about scarcity of manpower and blamed emigration for it (Kälvemark 1972, chapter 4). Growing concern over emigration led to the appointment in 1907 of a large public investigation, Emigrationsutredningen, on the causes and consequences of the transatlantic emigration (Kälvemark 1972, ch. 5-6). It resulted in a main report, written by Swedish demographer Gustav Sundbärg comprising almost 900 pages (Sundbärg 1913), and 20 special investigations. As had been pointed out already by Emigrationsutredningen, a conspicuous feature of transatlantic emigration was that it fluctuated wildly over time. The most intensive period was between 1880 and 1893, when a total of 550,000 Swedes chose to leave the country. From its height in the 1880 s and early 1890 s the long-term trend in emigration was nevertheless pointing downwards. Emigration also varied characteristically between regions. Most of the emigrants originated from the western and southern parts, while the emigrants from the eastern and northern regions were considerably fewer. Our ambition is to explain the long-run downward trend in emigration rates as well as its regional variations. To do so we also need to control for the temporal fluctuations in emigration. We use a panel dataset of 24 Swedish counties over the period The dataset makes it possible to study the dynamics of emigration, which is easier to do with multiple time series instead of only one aggregated 1

3 time series as in time series studies, as well as the regional variations in emigration rates. To the best of our knowledge such a dataset is unique in studies of late nineteenth century transatlantic emigration. 1 Before we embark on our own empirical investigation we first, in the next two sections, give some typical characteristics of Swedish emigration and review earlier research on the subject. 2. An overview of Swedish emigration 2.1. Data Our data on emigration derive from official statistics, which are based on information from local parishes, which were put together by the parish priests and sent to the Central Bureau of Statistics. These emigrant data, contained in the so-called condensed population reports (summariska folkmängdsredogörelserna) served as the main source material for the official emigration statistics since 1861 (Tedebrand 1976, p. 85). The reports are known to contain some deficiencies. To some extent they are due to the fact that it was sometimes difficult for the parish priests to separate emigration from other kinds of migration. There were a significant amount of unregistered emigrants, which implies that the numbers reported by the local priests in fact were too low. It could be due for example to emigrants who moved without securing an address change certificate for the purpose of emigration or to factual errors in the residential registration. Another cause of errors in the statistics could be the fact that only the persons who had filed an address change certificate were actually registered as emigrants. People who indubitably had emigrated but without the proper kind of certificate where not registered as emigrants. However, these deficiencies in the recorded emigration refer primarily to the early years of Swedish mass emigration. The Emigration Ordinance of 1884 implied a prohibition against emigrant agents to convey emigrants abroad without presenting a proper address change certificate to the police authorities. This resulted in a decrease in the unregistered emigration (Bidrag till Sveriges officiella statistik, series A, 1885, p. XVII). Historical research has established that the condensed population reports, put together by parish priests, truthfully recorded the demographic information contained in the church books. There seems to be a consensus, however, that the official emigration statistics underestimated factual emigration, especially before the mid 1880 s, although the problems with the official statistics were not as large as for example Gustav Sundbärg believed. Its main deficiency seems to be that it does not account for the substantial short-term trans-atlantic work emigration that took place. In an international perspective Swedish emigration statistics is nevertheless considered exceptionally good (Tedebrand 1976, pp ). The official statistics is also the only available source that gives a regional break-down of the emigration streams. 1 Timothy Hatton and Jeffrey G. Williamson (1993) studied Irish emigration using a panel dataset of Irish counties for the period , but they only had data for four census years, not for the entire period. 2

4 2.2. Characteristics of Swedish emigration A first characteristic of Swedish emigration to the United States was its sharp fluctuations from year to year. This is clearly brought forward by Figure 1, showing the emigration rate in the period Apart from a peak in 1869 emigration was low until the 1880 s, when emigration peaked in 1882 and 1887/88. Another peak in the emigration rate occurred in the beginning of the After 1893 the emigration rate subsided and was low for the rest of the 1890 s until a new wave of emigration started around the turn of the century. The latter peaked in 1903, after which there was another short burst with a peak in Emigration fluctuated wildly, but as can be seen from Figure 1 there was also a downward trend in the emigration rate after the 1880s. Swedish emigration cycles was part of an international pattern with comparatively low emigration rates in the 1870 s and 1890 s and high emigration rates in the 1880 s and the first decade of the twentieth century. In the literature this has been connected to a pattern of inversely related long swings in the Atlantic economy. When the business cycle turned up in the US it turned down in the UK and other west- European countries, and vice versa (Thomas 1954; 1972). The 1870 s and 1890 s were decades of upswings in the business cycle in western Europe and downswings in the US, while in the 1880 s and 00 s the phases of the business cycles on the two continents were reversed. The transatlantic emigration rate not only varied in correspondence with long swings in the Atlantic economy, but also from year to year, mirroring short-run fluctuations in economic activity. Figure 2 shows yearly emigration rates along with a two year moving average of the difference in yearly GDP growth rates between the US and Swedish economy for the years As can be seen there is clear correlation between changes in emigration rates and changes in the business cycles as measured by differences in GDP growth rates. From 1881 we have data that separates emigration from the countryside and the cities. As shown by Figure 3 fluctuations in emigration were common for the countryside and the towns, but, particularly in the 1880 s emigration rates were higher for the towns. The cycles of emigration were common also for men and women, but, as shown by Figure 4, emigration rates were generally lower for women. This was particularly the case when emigration rates were high, as for example in the 1880 s. In other words, fluctuations in emigration rates were less pronounced for women than for men. Women did not react as swiftly as men to the allure of upswings in the business cycle in the US or on push effects resulting from downswings in the Swedish business cycle. This may be 3

5 because many of the jobs that women applied for, notably maids, were less sensitive to business cycle fluctuations. From the literature on late nineteenth century emigration it is well-known (Carlsson 1976, pp ) that emigration was predominantly concentrated to the age group years old. As can be seen form Table 1, 71 percent of the emigrants from Sweden to the US in the period belonged to this age group at the time of emigration. Of the emigrants in this period 16 percent were children (age group 0 14 years old). They travelled in company with their parents, so the substantial number of children that emigrated may also be seen as a result of the high emigration rate among men and women in their twenties. However, most of the emigrants were unmarried; otherwise the number of child emigrants would have been higher. In Table 1 we also show that emigration rates varied strongly among the age groups. In the age group years old, on average almost 2 percent emigrated each year in the period As can be seen from Table 1 emigration rates were also high in the adjacent age groups. In the age groups above 35 years of age the probability of emigrating quickly dwindled. Emigration rates among children (age group 0 14 years of age) were also low, which is another way to express the fact that most emigrants from the emigration sensitive age groups were unmarried at the time of emigration. The emigration varied not only over time but also among regions. Figure 5 shows how emigration rates in the Swedish countryside differed between counties during the period A clear concentration of high emigration rates is seen in the southern and western regions, with the exception of the counties Göteborg and Bohuslän in the west and Malmöhus in the south, where many people from the countryside migrated to the cities Göteborg and Malmö. 2 The highest emigration proportion is shown for the counties Halland, Värmland, Kronoberg, Älvsborg, Jönköping, and Kalmar. These six accounted for just over 28 per cent of the population on the countryside, but for ca 44% of the country s emigrants during the period The yearly emigration rate in these counties amounted to between 7 and 10 persons per thousand inhabitants. The counties in eastern and northern Sweden show a substantially lower proportion of emigration, with a yearly emigration rate between 1 and 4 emigrants per thousand inhabitants. The regional differences in emigration corresponded to other differences between the same regions, as was pointed out already in the vast public investigation on emigration, published in The main 2 It may also have been the case that trans-atlantic emigration is under-reported in that some emigrants may have emigrated via Denmark or Germany, in the case of Blekinge and Malmöhus county, and Norway, in the case of Värmland, Göteborg and Bohus county. This is indicated by the substantial difference between total emigration and trans-atlantic emigration for these counties, especially for Göteborg and Bohus county. For Sweden as a whole, emigration to the US accounted for percent of total emigration. For the counties mentioned above it accounted for percent, for Göteborg and Bohus county 65 percent. 4

6 report, written by Swedish demographer Gustaf Sundbärg, explained the regional differences in emigration by regional differences in the previous demographic development. According to Sundbärg there were three main demographic areas in Sweden based on characteristic differences in marital fertility: the east, the west and the north. Sundbärg s west region stretched from central Kopparberg county in the northeast to Kalmar county in the southeast (Figure 6). This region exhibited high marital fertility rates which in combination with low mortality rates led to rapid population growth. In his eastern region, stretching from Jämtland county in the north to Östergötland county in the south marital fertility rates were much lower. More rapid population growth in the west stimulated emigration which in its turn moderated the regional differences in population pressure (Sundbärg 1910, p. 8). Nils Wohlin, one of Sundbärg s colleagues in the public investigation committee on emigration, pointed out that the regional differences in demographic behavior corresponded to differences in economic conditions and in the structure of landownership in the countryside. Land reclamation and parceling out of land into smaller farm units was much more common in western Sweden where land to a large extent was cultivated by peasants owning their land. In the core eastern region on the other hand, where large estates and ownership by the nobility was much more common, parceling out of land into smaller farm units was much less prevalent. Estates and other large farm units relied on hired labour. According to Wohlin there was a causal connection between the much higher parceling out of land into small units in the west and higher birth rates which in its turn spurred emigration (Sundbärg 1910, pp. 4,9; Wohlin 1909, pp ). The causal link between regional differences in economic structure and the structure of landownership have been developed by later historians, foremost among them Christer Winberg. The agrarian revolution in the first half of the nineteenth century signified increased grain prices and increased labour demand. In the core eastern counties, where large farms and large states were much more prevalent than in the west, it resulted in increased demand for wage labour. In the typical western counties, on the other hand, the result was further sub-division of landownership and land reclamation. In small farms cultivated by the owner and other family members children were often seen as economic assets adding to the labour force at the farm, while among wage labourers and other landless people the incentives for having many children were less persistent. The higher birth rates in Sundbärg s eastern Sweden may thus, according to Winberg, be explained by differences in ownership structure which led to different responses to the demand forces unleashed by the agricultural revolution (Winberg 1975, pp ; Winberg 1988, p. 49 ff). In northern Sweden, the third of Sundbärg s main demographic region encompassing Västernorrland county, Västerbotten county and Norrbotten county, economic conditions were quite different compared to the other regions, even though there were similarities with the west in demographic behaviour. High 5

7 fertility rates and low mortality led to rapid population growth, but in the northern region it was also stimulated by an inflow of migrants from other parts of Sweden. northern Sweden, Norrland, with its vast natural resources was the site of the typical resource based Swedish export industries: the sawmill industry, the pulp and paper industry and the iron ore mines. Many small farm units were set up in the countryside by settlers who often combined farming with seasonal wage labour in the forestry sector and sawmills. However, agricultural yields were generally lower in the north region, and many farmers were much dependent on the vicissitudes of the forestry sector and the sawmill industry. This may explain why the time profile of emigration from the northern counties is quite dissimilar to the west and east regions. In the east and west region there is a downward trend in transatlantic emigration, which is not seen in the north (Figure 7). On the contrary, there is even a tendency for emigration to go up in the northern counties towards the end of our period, in connection with a slackening of growth and profitability in the forestry and sawmill sector (Sundbärg 1910, pp ). 3. Earlier studies on the causes of late nineteenth century Swedish-US emigration Earlier research on the causes of late nineteenth century emigration have typically used two types of data: time series data in order to explore temporal variations in emigration and cross-sectional data on regions, where the dependent variables often is the average rate of emigration for a given period, to explore the regional variations in emigration rates Time series studies The international literature on emigration is dominated by studies on time series data. Since Swedish data are exceptionally good, emigration from Sweden has featured prominently in this literature. The first generation of studies was dominated by the question whether emigration was caused by push or pull factor. In a seminal study Jerome (1926) argued that upswings in the US business cycles, resulting in increased labour demand, were the most decisive causal factor, which also explains the fluctuations in emigration, whereas push forces in the sending countries where of lesser importance. Simon Kuznets agreed that pull factors were decisive, otherwise it was difficult to explain the international character of the emigration waves, since it was highly unlikely that push factors would operate at the same time in different countries (Gould 1979, pp ). However, Brinley Thomas (1954; 1972) argued that the US and British economy were causally connected in the framework of an Atlantic economy. This should also be the case for other European countries, not the least Sweden, since their business cycles were tied to the British by means of foreign trade, in which case it explains why push forces operated simultaneously in many countries. For the case of Sweden Dorothy Swain Thomas argued for the importance of push 6

8 factors: the most consistent increase in emigration occurred when pull and push coincided prosperity in America was highly important as a stimulus to emigration from Sweden, but cyclical upswings in Sweden were a far more powerful counter-stimulant than is generally recognized (Thomas 1941, p. 169). The first generation of time series studies on emigration used simple techniques, such as bivariate correlations and trend calculations, to explore the causes behind emigration. Later economists and economic historians have used multiple regression analysis. They have typically used three types of explanatory variables: data on economic activity indicating employment opportunities in the sending and receiving countries, data on wages or income levels, and demographic data in the emigrant countries indicating the presence of Malthusian pressures. Lagged emigration and in some cases also stocks of previous emigrants have also been employed as explanatory variables in time series regressions. In a time series model of Swedish emigration Wilkinson (1967, pp ) regressed the log of emigration from Sweden to the US against itself lagged one year and the log of manufacturing output in the two countries indicating employment opportunities. He concluded from the model that push and pull factors were equally important. So did also Quigley (1972) in a study where he regressed the log of Swedish-US emigration against itself lagged one year and against the log of wage rates in the two countries. He also included births lagged 26 years in the regression to control for Malthusian pressures. Quigley estimated separate equations for emigration originating in the industrial and agricultural sector and also included a variable on harvests in the regressions. Similar time series models have been estimated for other countries. In a survey of the literature up until the 1980 s Gould (1979, pp ) summarized some of the results from time series models of emigration. By and large variables serving as proxies for employment opportunities give clear and consistent results. Income variables give less clear results, especially when combined in the same regression as variables indicating employment opportunities. When wage/income variables are included in regressions that do not include variables indicating employment opportunities they tend to pick up the influences from the latter. However, while there was a trend in the Swedish-US wage ratio in that Swedish wages caught up on US unskilled wages (Williamson 1995; Prado 2010), there was no trend in employment opportunities. In time series models short-run fluctuations in activity levels dominate estimation results rendering wage variables insignificant. 3 3 Hatton (1995a) combines the variables relative wage rate (the US wage rate divided by the Swedish) and the ratio of employment opportunities in the US and Sweden (defined as the percentage deviation of GDP per capita from its trend) into one variable. He does so by multiplying the relative wage rate by the ratio of employment opportunities raised to 1.5. The log of the combined variable enters the regression as one of the explanatory variables. In other words, he restrains the coefficient for the log of the ratio of GDP per capita deviations to be 1.5 times the size of the 7

9 A variable that always is highly significant in time series regressions and with a large impact is lagged emigration. In other words, emigration the previous year(s) has a large impact on current emigration. It has been interpreted as a lagged response, for example emigration in the first few months in a year reflected conditions in the autumn the previous year, but also as mirroring a friends and relatives effect. Several scholars have also tried to incorporate the latter by including the emigrant stock as an explanatory variable (Hatton 1995a; 1995b; Hatton and Williamson 1993). In many cross-sectional studies the stock of previous emigrants have turned out to be an important explanatory variable, but it is more cumbersome to include in time series regressions since the stock of previous emigrants is more or less equal to cumulative previous net emigration. It is not obvious why such a cumulative sum, which monotonously increases, should be able to contribute to explaining in a time series context the year to year variations in emigration or even its trend. The statistical results of incorporating it in regressions for various countries have also been very mixed (Hatton 1995a, p. 561; Hatton 1995b, p. 412; Hatton and Williamson 1993, p. 584). Demographic variables have been incorporated in many forms in time series studies on emigration. The most important types of variables seem to be lagged births or the percentage share of emigration sensitive age groups, e.g years of age, in the population at risk of emigrating. Although it is clear from micro data that age is an important variable for explaining emigration it has been difficult to show any sizable impact from demographic variables in a time series context. The reason seems to be that although there is a variation over time in the share of emigration sensitive age groups in the total population, it seems to be of a more long-swing character, and is therefore not pertinent to the explanation of the shortrun year to year variations in emigration rates (Gould 1979, p. 642) Cross-sectional studies on regional variations in emigration There is a voluminous historical literature on Swedish emigration in the late nineteenth century. As we have already mentioned, a comprehensive study was carried out in Sweden at the beginning of the 1900s through the state survey Emigrationsutredningen ( ) under the leadership of Gustaf Sundbärg. The survey s ca 900-page main report was written by Sundbärg (1913) himself and contained, among other things, a statistical overview of each county, a review of the emigration legislation and a part that dealt with the causes of emigration and the measures that could be taken for its restriction. The survey also contained twenty supplements by various authors and several statistical studies and accounts of coefficient for log wage-ratio. This is motivated by a utility function derived in Hatton (1995b). We have been able to by and large reproduce Hatton s estimation results (not exactly because there obviously are some differences in the data) but when we let the wage rate enter the regression separately from the GDP/capita deviation ratio, the wage rated completely looses all statistical significance. 8

10 widely differing subjects, such as economic statistics, demographic information, the applicable legislation and ordinances concerning, for example, emigration agents and emigration ships, reports from emigrating Swedes and so on. Sundbärg claimed that the principal reason for the emigration was the inability of the Swedish countryside to maintain an increasingly larger population. He did not approve of the ongoing rationalisation in the agricultural sector, but recommended a more extensive development of agriculture. He also opined that industrial development had a significant role to play in reducing emigration, but that it would not suffice. Hence, it was absolutely necessary to develop the agricultural sector (Sundbärg 1913, pp ). The report was published in 1913 but never became the basis for Swedish emigration policy that it was intended to be. Soon after it was published the First World War brought a stop to Swedish mass emigration. Another large research project dealing with late nineteenth century Swedish emigration was carried out at the Department of History, Uppsala University, between 1963 and 1976 and was summarised in the book From Sweden to America (Runblom and Norman 1976). The Uppsala project mainly resulted in numerous descriptive regional studies 4 that examined various aspects of the emigration process. From these local studies scholars have concluded that countryside parishes with a high degree of industrial production tended to have higher emigration rates than purely agricultural parishes, although there were exceptions to this pattern. Moreover, emigration rates tended to be lower in rural parishes with a geographical proximity to urban industrial centers, presumably because migration to the cities partly substituted for overseas emigration for inhabitants of these parishes. Another conclusion from local studies was that emigration tradition, that is, a prior history of high emigration, induced further high emigration (Carlsson 1976, pp ; Norman 1976, pp ). Only in a few instances did researchers in the Uppsala project use statistical methods, more advanced than simple descriptive statistics. An exception is Hans Norman s study of a cross-section of 214 parishes from Örebro county, Västmanland county and Värmland county in central Sweden (Norman 1974; Norman 1976, pp ). Norman s data for these parishes consisted of average emigration rates for the period which he matched against data on internal emigration rates, emigration tradition (measured as average previous emigration rates), geographical distance to cities, and various socioeconomic variables emanating mainly from around the turn of the century He did not use regression analysis but instead a method called AID (Automatic Interaction Detector) analysis which purportedly assesses which variable contributes most to explain the cleavage between high and low emigration parishes. 5 The problem with this type of method seems to be that it is impossible to assess the 4 See for example Nilsson (1970), Norman (1974), Rondahl (1972), Tedebrand (1972). 5 See for example Sonquist and Morgan (1970), Åkerman (1979; 1971) 9

11 quantitative impact of the different variables. Norman nevertheless confirmed results from local studies of a more descriptive character, namely that industrial parishes and parishes with an emigration tradition tended to have higher, and parishes with a geographical proximity to urban centers tended to have lower emigration rates. Another result of his study was that parishes with a high percentage of arable land of the total land area tended to have lower emigration rates. Since the studies of the Uppsala project were all of a regional or local character it may be argued that their conclusions cannot be generalized to other regions. The American scholar Briant Lindsay Lowell (1987) studied the causal factors behind the emigration by assembling a cross-sectional dataset for all 301 härads (an administrative unit intermediate between parishes and counties) of Sweden for the period , which he explored by means of regression analysis. Lowell s dataset consisted mainly of demographic and economic variables, which he obtained from the aforementioned Emigrationsutredningen, but also some other sources. Lowell confirmed some of the results of the Uppsala project, for example that geographical proximity to cities led to lower emigration rates and that emigration tradition was an important factor behind emigration. Based on the contribution of the various variables to the overall coefficient of determination, partial correlations and standardized regression coefficients he came to the conclusion that emigration tradition was the most important explanatory variable. It is, however, difficult to evaluate the effects of the different variables on the emigration rate merely by looking at correlations and regression coefficients, without considering descriptive statistics of the variables in the dataset, their means and standard deviations. In other words, Lowell did not provide any standard by which the reader can size up the quantitative importance of the various regression coefficients. Another drawback of the Lowell study is that although the dependent variable is the mean emigration per thousand inhabitants for the years , most of the explanatory variables come from one particular year around Hence, all variables do not pertain to the same date and no consideration is given to the fact that explanatory variables varied over time in a dissimilar way from region to region. Like the aforementioned study by Norman in the Uppsala project, Lowell s study endeavored to explain regional variations in late nineteenth century trans-atlantic emigration by exploring a cross-sectional dataset. However, as shown in the discussion of time series models of emigration, a conspicuous feature was its characteristic variation over time. As we argue below different types of variables explain the time series variation and the regional variation in emigration. Ideally, we should therefore have a dataset that combines the crosssectional and time series dimension. 10

12 4. Description of the dataset To explore the causes behind emigration we have constructed a panel dataset, containing yearly data for 24 of Sweden s 25 counties (län) over the period 1881 to 1910; Stockholm city is excluded since we confine ourselves to emigration from the countryside. The majority of the population still lived in the countryside and hence the majority of emigrants came from rural areas. Another obvious reason why we confine ourselves to a study of emigration from the countryside is that is possible to collect information on variables of interest from these areas to a much larger extent than is possible for urban areas. From 1881 it is possible for the first time to obtain county data emigration divided on the countryside and the towns. For this reason our study commences in We end our study in This thirty years period covers the main phases of Swedish-US emigration Variables used in the regressions Our dependent variable in the regressions reported below is the number of emigrants to the US per thousand inhabitants in the countryside. Several explanatory variables have been constructed by using our hypotheses of what determined emigration as a starting point. They broadly can be grouped under three headings: demographic variables indicating the size of emigration sensitive cohorts in the population; variables indicating the availability of land and other means to earn a livelihood in the countryside; variables indicating the evolution of the standard of living. It is well-known that most of the emigrants were unmarried men and women in their early twenties or slightly below. We do not have data on the age distribution of the various counties, except for the census years. However we have yearly data on births from the 1860 s. Births lagged 20 years may serve to indicate the size of the emigration sensitive cohort in the population. Accordingly one of our variables is the number of births lagged twenty year per thousands of current population. We expect emigration to vary positively with this variable. The marriage rate varied over time and between the counties. In time series regressions this variable has been shown to have some influence (Hatton 1995a). Thus we include the number of marriages per thousand inhabitants as one of our explanatory variables. We expect that an increase in the marriage rate would lead to less emigration since married people were less inclined to emigrate. Moreover, a higher marriage rate may also indicate better or more stable employment opportunities and access to land. High population pressure affected the possibility of earning a living in agriculture, which should have affected the propensity to emigrate. To test this proposition we have constructed a variable that we call 11

13 population density, defined as the quotient population/(arable+0.33 x meadowland). 6 All else equal we expect emigration to be higher the higher population density is. If a large proportion of the farming units were small, it could indicate that the possibility of obtaining land was greater, which should have led to a lower propensity to emigrate. In their panel data study of Irish emigration Hatton and Williamson (Hatton and Williamson 1993) included a variable showing the proportion of small farms. Swedish agricultural statistics give information on how the farming of cultivated land was distributed between owners and leaseholders and how many of the farming units were in the size classes -2 hectares, =>2 hectares 20 hectares, >20 hectares 100 hectares, >100 hectares. It is accordingly fully possible to include size variables in the regressions. However, counties with high population pressure generally also had a much higher share of small farm units. A high share of small farm units may also have had a contradictory effect on the emigration propensity since it may also indicate fewer opportunities for landless people to gain employment at larger farms. Moreover, the variation within counties of the size distribution of farm units was small, which makes it difficult to obtain reliable estimates. Experimentation with various panel regressions did not lead to any economically and statistically significant effects from size variables. We have therefore excluded such type of variables from the regressions presented below. Another variable, whose evolution might indicate population pressure, is the percentage share of crofters and cottagers of the rural population. To be a crofter (torpare) meant to dispose a tiny peace of land for personal use and a place to live in return for performing labour services at the main farm under which the crofter was subsumed. There was actually quite a big differentiation in social and economic conditions among the crofters. The well-situated among them were not much below the position of small copyholders, while the poorest among them were not much better situated than cottagers. The cottagers (backstugusittare) lived in small houses without any adjacent piece of land. They were the poorest social strata in the countryside and served as casual labour during harvest seasons but were often dependent on poor relief. The number of crofters and cottagers expanded rapidly and increased its share of the countryside population between 1750 and In the last decades of the nineteenth century, when population pressure declined, the proportion of crofters and cottagers dwindled. A large fraction of crofters and cottagers may indicate that many of the rural inhabitants were landless or could not get 6 Meadow land has been weighted by 0.33 based on its value compared to arable land. See for example Höijer (1921, p. 219) 7 It is possible to obtain separate numbers for crofters and cottagers in the census years (Wohlin 1908; 1909), but not for other years when we have to rely on the official agricultural statistics, where the number of the two groups are combined under the heading jordtorp och andra jordlägenheter. 12

14 employment under better conditions than that of crofters and cottagers, which should have led to increased emigration. The following two variables concern the impact of increased output of animal products and grain. We assume that increased output affected labour demand in the countryside and, all else equal, it also improved the rural population s economic well-being and should have led to lower emigration. A characteristic feature of Swedish agricultural development in the late nineteenth and early twentieth century was the increasing share of animal output. Expanding animal production improved the profitability of farm owners and should have led to increased and also more stable job opportunities, since labour demand in this line of production was less seasonally dependent than grain production. We do not have data on the value of animal output per county, but we do have data on the number of animals. Our variable for animal output is based on a measure in which the whole animal stock is reduced to cattle units (nötkreatursenheter), a term used in Swedish agricultural statistics. 8 To control for the effect of increased animal output on the emigration rate we use the variable number of cattle units per hectare arable and meadow land. We expect the emigration rate to decline with increased animal output. Grain output did not at all increase to the same extent as animal output. Increasing grain output per unit of arable land, had contradictory effects on emigration rates. On the one hand it increased living standards, which should have lowered emigration rates. On the other hand, increased labour productivity also led to decreased labour demand if it was not sufficiently compensated by an increase in the cultivated acreage. For these reasons we would expect the effects of grain output on the migration rate to be less clear-cut than animal output. In order to control for the effects of increased grain production, we have estimated the value of the production in fixed prices for the different counties and years. It is well known that Swedish agricultural statistics, at least before the 1890s, underestimated the size of arable acreage. The harvest per acreage unit, i.e. the reported production of crop products divided by the arable acreage size, should be a more reliable measure than total output. In the regression we use the variable crop product output per hectare arable land acreage. Both variables indicating the evolution of agricultural production enter the regressions in logarithmic form, since we assume that it is the rate of increase in agricultural production which is important in this context. Another factor affecting employment opportunities in the countryside was the availability of job opportunities outside the agricultural sector. By the term countryside population (landsbygdbefolkning) in Swedish population statistics is meant people living outside the towns. Most of them, but not all, earned their living in the agricultural sector. It would be useful to have data on how large a share of the gainfully 8 A cattle unit is equal to one cow, 2/3 horse, 10 sheep, 12 goats and four pigs. 13

15 employed were engaged in agriculture since it might be argued that a large share of employment outside agriculture indicate more job opportunities, which inhibited emigration. For the census years we have data on the occupational distribution of the population. For other years the share of the value of agricultural taxable properties of the value of all taxable properties may serve as a proxy for the proportions of agricultural and non-agricultural job opportunities in the countryside. We have experimented with this variable in various regressions but did not obtain statistically or economically significant estimates. 9 We have therefore dropped it from the regressions reported below. On further thoughts it is not clear-cut how an increase in the non-agricultural share of the labour force should affect emigration. On the one hand it increased employment opportunities outside of the agricultural sector, but on the other hand nonagricultural output shared the same cyclical variations as the agricultural sector and as we have shown in section 2 above emigration rates tended to be higher in towns than in the countryside. Two variables refer to the evolution of incomes and living standards in the countryside. They are formulated on the assumption that an important incentive for emigration was the increase in economic well-being that an emigrant could hope to achieve by moving to the US. The most important variable in this respect was the real wage differential between the USA and Sweden. The assumption is that the higher the real wage in the USA compared to Sweden, the higher the emigration rate, and vice versa. In defining the variable, Swedish agricultural labour wages for each county are first deflated with a Swedish cost of living index (Myrdal 1933, p. 128). The Swedish real wages thus obtained are then divided by a real wage index for unskilled labour in the USA (Williamson 1995). This variable is also expressed in logarithmic form since we assume that the prospective emigrant reacted on the percentage wage increase he or she could hope to obtain by moving across the Atlantic. The next variable with a connection to the living standard is the proportion of the population that were entitled to poor relief. A large proportion of the population on poor relief indicates difficult living conditions that should have spurred emigration. The variable is defined as the percentage share of poor relief recipients of the total countryside population. In the regressions reported below we also use the lagged dependent variable as one of our explanatory variables. Time series regressions have unequivocally shown that this is an important variable in modeling the dynamics of migration. Sometimes this variable has also been interpreted as showing the impact of the friends and relatives effect. There are, however, better ways to model the impact of the emigration tradition. In time series studies the emigrant stock has been used for this purpose, with mixed 9 We have also tried the variable employment share in agriculture in regressions for the census years, without achieving any significant results. 14

16 results, as we argued in section 3.1 above. The fact that we have a panel dataset enables us to form a variable that better serves our purpose of indicating the effect of the emigration tradition. In constructing the variable we have first for each county and year calculated the cumulative sum of previous emigrants to the US. For each county we have than calculated its share of total cumulated countryside emigration in a year and divided it by the county s share of the total countryside population in the given year. Our constructed variable for emigration tradition is a ratio between two quotients and its cross-sectional variations is much larger than its variations over time within the various cross-sectional units (see Table 2 below). We believe that it had its most important impact on regional differences in emigration, and, since the variable lacks a natural scale, we have in fact scaled it so that for each year its arithmetic mean equals unity. We expect that a high value for emigration tradition will have a positive impact on emigration rates Descriptive statistics for the variables In Table 2 descriptive statistics of the variables that we use in our study are displayed. Along with the means and overall standard deviations of the variables, the standard deviations within the cross-sectional unit and the standard deviations between the cross-sectional units, i.e. the standard deviations of the crosssectional means, are also given. As shown by the table, for some of the variables the most notable variation in the dataset is over time within the cross-sectional units. This is most particularly the case for the Swedish/US real wage ratio. For other variables, on the other hand, most particularly population density and emigration tradition, variation across the cross sectional units, was the most important source of variation in the dataset. In table 3 we show how the mean of the various variables for the census years , 1890, 1900 and As we discussed above, the emigration rate varied not only over time, but also in a characteristic way between the various counties. Earlier research on Swedish emigration, demographic behavior, and Swedish nineteenth century historiography more generally, has made much of the distinction that Swedish demographer Sundbärg made between eastern and western Sweden. Our ambition is to explain not only the evolution over time of the emigration rate, but also differences in emigration between the various parts of the country. In table 3 we therefore also show how the variables we use in our regressions evolved for Sundbärg s demographic regions Since we do not have data for most of the variables before 1881, we use 1881 and not the census year 1880 as our first year of comparison. 11 Sundbärg s regions were constructed from data at a lower administrative level than the counties. Some counties contained parishes from both the east and the west or the east and the north. To operationalize Sundbärg s 15

17 Between 1881 and 1910 the overall emigration rate (un-weighted county average) declined by about 5 emigrants per thousand both in the north and in the west, but it was always higher, on average by slightly more than 3 per thousands, in Sundbärg s western counties than in the eastern counties. Differences in demographic behavior were the basis for Sundbärg s division of Sweden into different regions. Those differences are also clearly brought forward in table 4. Birth rate lagged twenty years per thousands of current population declined over time, but it was always higher in the western than in the eastern counties, even though the regional differences diminished between 1881 and The same regional differences can be seen in the marriage rate. This variable also declined over time. It was always higher in the east than in the west, but in this case also the regional differences were less pronounced in 1910 than in Average population density for all counties did not show much of a trend between 1881 and However there were always large regional differences in this variable. The population density was consistently higher in the western counties than in the eastern counties, although the difference tended to diminish over time. The percentage share of crofters and cottagers in the countryside population showed a downward trend, reflecting a decline in the landless population. There were regional differences also in this variable. In typical eastern counties the share of crofters was larger. This was no doubt connected to the generally larger farm units in this area and the more wide-spread use of hired labour, which made it possible for landless people to earn a living in wage labour on larger farms and estates. The number of cattle units per hectare arable and meadow increased by about 20 percent in the period The regional differences in this increase were not pronounced, although the growth rate was somewhat higher in the western counties, reflecting the higher population density in these counties and also the prevalence of small farm units who tended to specialize more in animal husbandry than larger farms. Crop output per hectare arable tended to increase by about 35 percent between 1881 and 1910 in all 24 counties and by more than 45 percent in the eastern counties. Superficially it would appear from these figures that the increase in crop output was of more importance than the increase in animal output. This would be a false conclusion, however. The lion share of the increase in crop output was the increase in oats production, which was predominantly used to feed the animals. A growing stock of animals regions at the county level we have defined the following counties as belonging to the east : Stockholm, Uppsala, Södermanland, Östergötland, Gotland, Västmanland, Gävleborg. The following counties are defined as belonging to the west : Jönköping, Kronoberg, Kalmar, Blekinge, Kristianstad, Malmöhus, Halland, Göteborg and Bohus, Älfsborg, Skaraborg, Värmland. The following counties are defined as belonging to the north : Västernorrland, Västerbotten, Norrbotten. The remaining counties, i.e. Örebro, Kopparberg and Jämtland are referred as belonging to a mixed category, since they were difficult to categorize as belonging to any of the other regions. Actually Sundbärg categorized Jämtland as belonging to the east, based on demographic considerations. Due to its geographic position and economic characteristics that distinguish it from the typical eastern counties we prefer instead to refer it to the mixed group. 16

18 required a larger output of fodder grain, but the animals were also given more and better fodder, which led to a much larger increase in animal output than is shown by the number of cattle units. Unfortunately this increase in animal output is not shown by our data since we do not have data at the county level of animal output. The Swedish/US real wage differential increased by almost 30 percent between 1881 and There were also regional differences in the wage rate. Wages were generally higher in the eastern counties than in the western counties. This difference was not large but persistent. It even increased slightly. In other words agricultural wages increased somewhat more in the eastern than in the western counties. The share of poor relief recipients of the countryside population declined in our thirty years period, reflecting the betterment in economic conditions. The decline was particularly prevalent in the western counties. In 1881 the share of poor relief recipients was almost one percentage unit higher in western than in eastern counties. By 1910 this regional difference had virtually disappeared. The way we have scaled the variable emigrant tradition its overall mean is set to unity for all time periods. As can be seen from Table 2 the most important variation is the between variation across the counties. As seen in Table 3 there was a persistent difference between the west and the east regions, which increased slightly over time, reflecting the higher emigration rates in the western counties. 5. The determinants of emigration 5.1. Econometric methodology Our dataset is a panel, which is important since different variables determine change over time and crosssectional variation in the emigration rate. A panel dataset makes it possible to control for cross section specific effects, observed or unobserved, and at the same to study change over time. We study the following type of model: Y = α + λ + β X + e (1) it i t j j, it it j where the subscript i stand for cross sections and t for time periods. The dependent variable Y, in our case emigration per thousand inhabitants in the countryside, is conditioned on three sets of variables. We are primarily interested in vector of regression coefficients for X j, a set of variables which vary over time and between cross sections. X j that are assumed to be constant over time periods and cross β j is a 17

19 sections. Cross-section specific constants, α i, reflect the effects of observable and unobservable variables that take a specific value for each cross section but are constant over time. To control for effects that influence each cross-section alike in a given time period, such as the state of the business cycle in the US and in Sweden and the industrialization level in Sweden, we also include a set of time fixed effects, λ t. The eit s are observation specific idiosyncratic errors. If the cross section specific constants are correlated with variables X j,an ordinary least squares estimate with a common intercept for all cross sections yields biased estimates of the regression coefficients of interest, β j, since they would pick up the influence of cross specific but time constant variables that are not controlled for. In our case, common reasoning as well as statistical tests tells us that this is indeed the case. We have therefore used the fixed effect estimator. A drawback of the fixed effect estimator is that it is impossible to obtain estimates for variables that vary between cross-sections but are constant over time within each cross-section. From this it also follows that it may be difficult to obtain reliable estimates for variables where the variations within the cross-sectional units are small. In most econometric theory for panel data it is typically assumed that the dataset consists of many crosssections observed over only a few time periods. However, in our case we have about as many crosssections as time periods, 24 and 30. This raises a host of other problems, but also possibilities, which have lately come to be studied by econometricians (Pesaran et al. 1999). On the positive side, panels consisting of multiple time series of reasonable length make it possible to model dynamics. In other words, we may include lags of the independent variables and also lags of the dependent variable among the regressors. It has been shown that in such models the fixed effects-estimator yields biased (too low) estimates for the lagged dependent variable, since it is inevitably correlated with the error term. Instrumental variable estimators have been proposed instead (Baltagi 2001, ch. 8). But these are also more appropriate for datasets consisting of many cross-sections. Fortunately, in our case the bias of the fixed effects estimator deriving from the inclusion of lagged dependent variables among the regressors will be small. 12 Another concern is that the variables included in the dataset are non-stationary, as suggested by a battery of panel unit root tests, not reported here. It is well known that a regression with time series data that are not stationary may lead to spurious results unless the variables are cointegrated. With panel data that include many time periods, about as many as the number of cross sections, as in the present case, the situation is less clear (Baltagi 2001, ch. 12). According to Phillips and Moon (1999; 2000), if we have 12 With infinitely many time periods the bias disappears. When we have as many time periods as in our case it will also be quite small (Nickel 1981). 18

20 independent cross sections then asymptotically the regression coefficients obtained with panel data using the fixed effect estimator are correct even with non-stationary dependent and independent variables that are not cointegrated. The idea is that independent cross section data in the panel adds information and this leads to a stronger overall signal than the pure time series case (Phillips and Moon 2000, p. 271). The problem with this result is that in practice time series are not infinitely long and cross sections are not independent. Entorf (1997) concludes that in panels with about the same size in the cross section and time series dimensions as the one used in the present study there is a spurious regression problem in fixed effect regressions when the individual time series are non-stationary. Wooldridge also warns that when there are many time periods (for example 30) and the number of cross sections is not very large (for example 20) we must exercise caution in using fixed regression estimates with non-stationary data, since the spurious regression problem may arise (Wooldridge 2003, pp ). It is therefore of some importance of testing for cointegration between the variables used in the regression. Several panel cointegration tests have been proposed in the literature. Using Westerlund s error correction based test (Westerlund 2007) we can reject the null hypothesis of no cointegration between the dependent variable and each of the independent variables. 13 We therefore proceed with estimating our model. To model the determinants of emigration we use an autoregressive distributed lag model, which allows us to model dynamic relationships. After experimentation with the dataset we have decided on two autoregressive terms, which removes autocorrelations in the residuals. We have included one lag on the other independent variables, i.e. they are assumed to influence the emigration rate in the same 14 and in the next period. In other words: (2) Y = ρy + ρ Y + β X + β X + λ + α + e it 1 it 1 2 it 2 j,0 j, it j,1 j, it 1 t i it j j In estimating the model it is useful to rewrite in the following form: ( ) (3) Y = ρy + ρ Y + β Δ X + β + β X + λ + α + e it 1 it 1 2 it 2 j,0 j, it j,0 j,1 j, it 1 t i it j j For interpretation it is convenient to rewrite (3) in error correction form: 13 We have used the xtwest procedure in Stata (Persyn and Westerlund 2008) by running bivariate tests with the dependent variable and each of the independent variables. 14 The variable crop product per hectare is lagged one period since the harvest took place in the autumn. It is therefore reasonable that the harvest in a year primarily affected behavior in the nest year. 19

21 Δ Y = ( ρ + ρ 1) Y β + β X ρ Δ Y + β Δ X + λ + α + e (4) j,0 j,1 it 1 2 it 1 j, it 1 2 it 1 j,0 j, it t i it j 1 ρ1 ρ 2 j We may then interpret β + β j,0 j,1 1 ρ ρ 1 2 as the long-run multiplier of variable X j, while j,0 β is the short-run multiplier Estimation results In Table 4 our estimated regression model is presented. We present estimates for all emigrants as well as estimates of regressions for men and women separately. In the following discussion we concentrate on the interpretation of the long-run multipliers. Increased population density stimulated emigration both in the short-run and long-run. The short-run impact multiplier of population density is not so large as the long-run multiplier and is also not statistically significant at conventional levels. To interpret the long-run coefficient of 8.3 we have to look back at the descriptive statistics in Table 2, which shows that the average population density was Accordingly, the estimated coefficient means that if population density had doubled emigration would have increased by 9.21 per thousand inhabitants, which is a powerful effect. However, since average population changed very little over time it cannot account much for changes in the average long-term evolution of the emigration rate. There were important differences in population density between counties, though, so this variable is important for explaining regional differences in emigration rates. In counties with lower population density it was easier to get access to land or other means of earning a living, which inhibited emigration. The estimated long-run coefficient of 8.3 for population density pertains to the non-northern counties. Sundbärg s non-northern counties were much more sparsely populated with an agricultural sector that was much more extensive with less importance for arable production. We have measured population density by the ratio population over arable and meadow land. Due to the relatively small share of arable acreage of the total land area in the northern counties our measure for population density is less suitable for them. It seems reasonable to expect that an increase in population density did not have the same effect there as in other counties. We have therefore interacted a dummy variable north, standing for Sundbärg s northern counties, with population density. An estimated coefficient of 4.9 for this variable signifies that the effect on emigration rates of an increase in population density in the northern counties was only half as powerful as in the other counties. 20

22 We may interpret a decline in the percentage share of of crofters and cottagers as indicating a decline in the share of landless people. Between 1881 and 1910 the percentage share of crofters declined by one percentage unit, from 4.47 to 3.56 (Table 3). According to our long-run point estimate of 0.3 this should have contributed to a decline in the emigration rate by roughly 0.3 per thousands. There is a statistically discernable influence in the short run of an increased number of people in emigration sensitive ages, as indicated by our variable births lagged twenty years per thousands of current population. An increase in this variable by one unit in a year would have increased emigration by 1.2 per ten thousand inhabitants, which is a fairly small effect. The long-run multiplier of 0.04 is even smaller and it is also not statistically significant at conventional levels. If we nevertheless accept ourt point estimate, its influence on the long-term evolution of the migration rate is meagre. Between 1881 and 1910 this variable declined from 28.3 to 26 meaning that it contributed to a decline in the emigration rate by 0.1. Its importance for regional differences is only slightly larger. Another demographic variable, the marriage rate, has more impact on the emigration rate, both in the short and the long run. This variable can possibly also be taken to reflect economic conditions since the marriage rate tended to vary with the business cycle, but it also show a trend of long-term decline and it varied characteristically between counties. Between 1881 and 1910 an un-weighted average of county marriage rates declined from 6.1 to 5.6. Since our estimated long-run coefficient is -0.65, this decline in the marriage rate should have increased emigration by roughly 0.3 per thousand when other causal influences are accounted for. As we will discuss below, it had some impact on regional differences in the emigration rate. Going over to the economic variables, an increase in crop products per hectare arable has a statistically discernable short-run effect on the emigration rate. The effect is however practically unimportant; an increase in crop output per arable by ten percent in a given year would have reduced emigration by slightly less than one per ten thousand inhabitants. Consequently it was practically unimportant as crop products per arable seldom increased by more than one percent per year. The long-run impact is even smaller, of unexpected sign, and is also not statistically significant. Of considerable more importance for the emigration rate is the evolution of the number of cattle units per hectare arable and meadows. 15 According to our estimated long-run multiplier a ten percent increase in the number of cattle unit per hectare would have reduced emigration by 0.7 per thousands. The number of cattle units per hectare increased by more than twenty percent between 1881 and 1910 (Table 3), so this variable is of considerable importance for explaining the declining trend in emigration form the Swedish countryside. 15 For essentially the same reason as with the variable population density we have interacted a dummy variable north with the log of cattle units per hectare arable and meadows. In the northern counties animals to a large extent grazed on mountain pasture (fäbodvall). Thus the denominator in the ratio defining this variable did not have the same importance in the northern counties. 21

23 The impact of increased animal output on emigration should probably primarily be interpreted as reflecting an increase in and a more stable level of labour demand in the agricultural sector. The evolution of wages for agricultural workers is more directly indicative of the evolution in the standard of living. Increased wages led to less emigration both in the short and in the long run. The estimated long-run coefficient for the Swedish/US real wage rate of 2.7 shows that a ten percent increase in agricultural wages led to a decrease in emigration by almost 0.3 per thousand inhabitants. 16 This is a powerful effect since the Swedish/US real wage ratio increased by roughly thirty percent in the period Another variable which illustrate an increase in the living standard is the decline in the percentage share of poor relief recipients that took place between 1881 and 1910; the county average declined from 4.28 to Our estimated long-run multiplier tells us that an increase by one percentage unit in the share of poor relief recipients would have increased emigration by 0.5 per thousands. Accordingly the decline in share of poor relief recipients in the countryside led to a decrease in emigration by 0.4 per thousands. It appears that the poverty rate had its greatest impact on emigration in the short run. Our estimated shortrun multiplier suggests that an increase in the share of poor relief recipients by one percentage unit in a year would have led to a rise in emigration by 2.3 per thousands, which is a substantial effect. Many historical studies on emigration have pointed out the importance of emigration tradition for explaining regional differences in emigration. Our variable emigration tradition has a statistically significant and substantially important effect on the emigration rate both in the short and long run. The way we have constructed the variable it cannot explain changes over time in average county emigration rates, but as will be shown below it is an important variable for explaining regional differences in emigration rates. In Table 4 we also show estimates for men and women separately. If we look at the long-run multipliers the results are similar to the ones presented in the first column of the table. One important difference, though, is that those variables reflecting economic motives behind emigration and conditions in the rural labour market yield higher and also more significant estimates for men than for women. For example, changes in population density, number of cattle units and especially the Swedish/US wage rate had larger effects on male emigration rates. We believe that this reflected first of all that the male emigration rate was generally higher and also that women s emigration appeared to have been less sensitive to the state of 16 It should be noted that this result is not dependent on the chosen Swedish cost of living index to deflate nominal wages or the US real wage index used to deflate Swedish real wages. The variable enters the regression in logarithmic form and the same cost of living index and US real wage index is used to deflate all county wages. Since we include time dummies for each year the estimated slope coefficients are not affected by our choice of deflators, only the time dummies are affected. We could equally well have used nominal wages in the regression and still obtained the same slope coefficients. 22

24 the business cycle, perhaps because employment in the typical jobs that women obtained in the US, such as maids, were less sensitive to fluctuations in the business cycle. 6. Simulations with the model 6.1. Accounting for long-run changes in emigration The number of emigrant from the Swedish countryside (un-weighted average of 24 counties) to the US declined between 1881 and 1900 by 5.16 per thousands, from 8.35 to Between 1881 and 1910 the corresponding change was In Table 5 we summarize what our regressions tell us about the determinants of this decline in the emigration rate. We calculate the predicted change in the number of emigrants per thousand by multiplying the point estimates for our long-run multipliers of the various variables with the actual change that took place in these variables in and in We do this calculation for an overall average of all counties as well as for averages of the eastern and western counties separately. 17 For Sweden as a whole, as well as for Sundbärg s eastern and western regions, variables indicating increased economic wellbeing and increased and more stable labour demand (number of cattle units per hectare, the Swedish/US real wage rate, percentage share of poor relief recipients) explain the lion share of the long-term decline in emigration rates. Demographic variables are fairly unimportant for explaining the long-term evolution of emigration rates. Only in the case of the western region was a long-term decline in population density of substantial importance for the long-term decline in emigration rates. In these counties, where populations pressure was the highest in the 1880 s, emigration and migration to the cities, and also a decline in the natural population increase, resulted in a long-term decline in population density. As can be seen the model generally predicts the evolution of emigration better for the western than for the eastern counties. This is to be expected since there are more counties in the western region and therefore more observations influencing the estimation results. For the entire period our model explains 72 percent of the decline in emigration rates. The corresponding figure for the western counties is 83 percent and for the eastern counties 52 percent Accounting for persistent regional differences in emigration We have seen that there were substantial and persistent regional differences in emigration rates. Since the same type of economic variables by and large explained the long-term evolution of emigration rates in 17 Our model generally predicts emigration from the northern counties badly, due to the radically different nature of the agricultural sector in these counties, so we do not present any separate simulations for the northern counties. 23

25 both regions they cannot explain much of those persistent regional differences. We now turn to an exploration of what determined them. In the left panel of Table 6 we give the arithmetic means for our variables calculated over the entire period , for all counties and for the eastern and western regions separately. From the differences between the regional means and the overall means of the variables we then calculate what change in comparison to the overall mean in the emigration rate our estimated long-run multipliers predict for Sundbärg s regions. For example, compared to the overall mean population density in the east deviated by 0.24 (= ). This leads to a lower predicted emigration rate in the east compared to the overall mean by 2.0 (=8.32 x 0.24). If we sum all differences between the columns in Table 6 representing the eastern and western regions, we see that the model on average predicts an higher emigration rate by 3.66 per thousand in the west compared to the east. The average factual difference in the emigration rate between the two regions was 3.34, so the model performs well in explaining regional differences in emigration. Most of the difference between the east and the west is explained by two variables: population density and emigration tradition. Already Sundbärg and Emigrationsutredningen pointed out that population pressure were considerably higher in the typically western counties than in the typically eastern counties. This led to a higher rate of emigration from these counties which in its turn stimulated further emigration through the friends and relatives effect. To a lesser extent the demographic variables, marriage rate and twenty years lagged births per current population, also point in the same direction. Together these two variables predict a yearly average of 0.6 more migrants per thousand inhabitants in the west than in the east. However, they are compensated for by economic variables, particularly cattle units per hectare, which predict slightly lower emigration in the west than in the east. 7. Conclusions We have set ourselves two main tasks in this paper: first to explain the long-run decline in emigration from the Swedish countryside to the United States, from its height in the 1880s to its last wave in the first decade of the twentieth century; secondly to explain persistent regional differences in the emigration rate. We have done that by building a panel dataset for 24 Swedish counties observed over the period , which we have explored with panel regression methods. Our data makes it possible to examine in a common framework the time series and regional dimensions of the emigration. It turns out that the long-run decline in emigration rates is explained by economic variables, which we interpret as indicating increased standard of living and employment opportunities. Increased well-being in the Swedish countryside, a catching up on US unskilled wages and more employment opportunities at home made overseas emigration less attractive in 1910 than in the early 1880 s. However, these persistent 24

26 forces affected regions with high and low emigration rates alike and therefore cannot be used to explain persistent regional differences. These regional differences are rather primarily explained by previous differences in demographic behavior, with historical roots in different land-tenure patterns, which had resulted in higher population density in counties which experienced high emigration rates. More people on available land made it harder to gain access to land or employment, which stimulated emigration. This in its turn led to further emigration through the friends and relatives effect. 25

27 Table 1: Age distribution of emigrants and emigration rates in different age groups, averages for the period Percentage shares of all emigrants Emigration per thousands in age group Table 2: Descriptive statistics for variables mean standard deviation, overall standard deviation, between standard deviation, within Emigration to the US per thousands Births lagged 20 years per thousands of current population Marriage rate Population density Percent crofters Cattle units per hectare arable and meadow land Crop per hectare arable Swedish/US real wage 1) Percentage share, poor relief recipients Emigration tradition 2) ) Since this variable is a ratio between index numbers we have scaled the overall mean to unity 2) This variable is a ratio of two quotients. It does not have a natural scale and we have scaled the average value for each year to unity. 26

28 Table 3: Annual averages of variables for 1881, 1890, 1900 and 1910 for all counties and for different regions Change Change All East West All East West All East West All East West All East West All East West Emigrants to the US rate, per s Population density Crop per hectare arable b) Cattle units per hectare b) Swedish/US real wage a) b) Percent poor relief recipients Percent crofters Marriage rate Births t 20 per 1000s of current pop. Emigration tradition Notes: a) The overall mean of the Swedish/US real wage ratio is set to unity in 1910 b) Changes and are in percentages 27

29 Table 4. Fixed effects regressions for all emigrants, male and female emigrants (dependent variable: number of emigrants per thousand inhabitants, mid year population). All Male emigration Female emigration coeffic Stand. p- coeffic Stand. p- coeffic Stand. p- ient err value ient err value ient err value Dependent variable t Dependent variable t Δ Population density Δ log(crop products per hectare) t Δ log(cattle units per hectare) Δ log(north x Cattle units per hectare) Δ log(swedish/us real wage) Δ Poor relief recipients (percent) Δ Crofters and cottagers (percent) Δ Marriage rate Δ(Births lagged 20 years/current population) Population density t North x Population density t log(crop products per hectare) t log(cattle units per hectare) t North xlog( Cattle units per hectare) t log(swedish/us real wage) t Poor relief recipients (percent) t Crofters and cottagers (percent) t Marriage rate t (Births lagged 20 years/current population) t Emigration tradition t N R 2 within R 2 between R 2 overall Long-run multipliers Population density North x Population density log(crop products per hectare) log(cattle units per hectare) North x log(cattle units per hectare) t log(swedish/us real wage) Poor relief recipients Crofters and cottagers Marriage rate Births lagged 20 years/current population Emigration tradition Note: Robust standard errors that allow for correlations inside the cross sectional units. 28

30 Table 5: Predicted change from estimated model and Predicted change Predicted change Long run overall east west overall east west multiplier Population density Crop product per hectare Cattle units per hectare Sweden/US real wage Poor relief recipients Crofters and cottagers Marriage rate Births t 20 per 1000s of current pop Emigration tradition t Predicted emigration rate Factual emigration rate Table 6: Predicted regional differences (Sundbärg s regions) from estimated model Arithmetic mean of variables Predicted difference in emigration rate from overall mean Long run multieast west east west Predicted difference plier overall east west Population density Crop product per hectare Cattle units per hectare Swedish/US real wage Poor relief recipients Crofters and cottagers Marriage rate Births t 20 per 1000s of current pop. Emigration tradition Sum predicted deviation from overall mean in emigration rate Factual emigration rate

31 Figure 1.Number of emigrants per thousand inhabitants, Source: Bidrag till Sveriges officiella statistik, series A 30

32 Figure 2.Difference in yearly GDP growth rates (2 year moving averages) between USA and Sweden and total numbers of Swedish emigrants to the US, , , , , , Difference in US-Swedish GDP growth rates, percentages, 2-year moving average (right scale) Number of Swedish emigrants to the US (left scale) Sources: US GDP (Maddison 1991, p. Table A6), Swedish GDP (Krantz and Schön 2007) 31

33 Figure 3. Number of emigrants to the US per thousand inhabitants, by gender, Emigration per thousands, women Emigration per thousands, men Sources: Bidrag till Sveriges officiella statistik, series A 32

34 Figure 4. Number of emigrants per thousand inhabitants, rural and urban areas, Emigration per thousands, countryside Emigration per thousands, towns Sources: Bidrag till Sveriges officiella statistik, series A 33

35 Figure 5: Emigration by county from the Swedish countryside to the USA, emigrants per thousand inhabitants (average ) Sources: Bidrag till Sveriges officiella statistik, series A Note: The number of counties in each size class in parenthesis after the legends 34