A number of aggregate-level studies have examined the effect of violence on migration and have generally found a strong connection between the two, though debate remains about whether the effect is direct or indirect. In his analysis of emigration from El Salvador to the United States, Jones (1989) concluded that the effect was indirect, with violence producing local economic dislocations that, in turn, led to emigration. Likewise, Morrison and May (1994) found that conflict-related economic turmoil was more important than violence in predicting out-migration between provinces in Guatemala. Schultz (1971) also found that rural violence was significant in predicting migration to Colombian cities, though its effect was small compared with other socioeconomic and demographic variables. Stanley (1987) found that military sweeps rather than killings per se were more strongly associated with variations in the flow of undocumented migrants from El Salvador to the United States.

Most other aggregate-level studies have discerned a direct connection between violence and migration, however. Shellman and Stewart (2007), for example, found that trends in Haitian emigration to the United States were strongly predicted by surges in political violence, even after the influence of economic conditions was held constant. Morrison (1993) likewise found that out-migration between provinces in Guatemala was strongly predicted by violence as well as economic conditions, and Morrison and Perez Lafaurie (1994) obtained similar results in Colombia. Morrison (1993) noted an apparent threshold effect in the relationship between violence and internal migration in Guatemala, such that the likelihood of movement increased only once violence reached a certain level, suggesting that at low to modest levels of violence, people may feel safer remaining in their own homes.

In addition to the foregoing country studies, cross-national comparative analyses also suggest a direct connection between violence and migration. A pooled time-series analysis by Schmeidl (1997) found that violence predicts international refugee flows more strongly than national economic conditions, and the fixed-effects model of Davenport et al. (2003) revealed that genocide, political violence, and civil war each strongly predicted refugee migration independent of economic circumstances. Another study by Moore and Shellman (2006) found that state violence toward civilians tended to produce international refugees, whereas high levels of dissident violence and civil warfare tended to produce internally displaced people (IDPs). Melander and Öberg (2007) have suggested that migration is more a product of the geographic scope of violence and the extent to which it touches urban areas rather than its intensity. In her review of the macro-level literature, Schmeidl (2001:85) concluded “that refugee flows are affected by state implosions and/or the formation of new states, genocidal violence, and internal struggles (particularly those fueled by foreign military interventions).”

In contrast to the relative abundance of aggregate studies, few analyses have examined the connection between violence and migration at the individual or household level. In their survey of displaced and nondisplaced persons in Colombia in 2000, Engel and Ibáñez (2007) found that the threat of violence and the presence of paramilitary and guerilla groups were strongly associated with out-migration, and Ibáñez and Vélez (2008) demonstrated that these associations hold up well under a variety of controls. Lundquist and Massey (2005) undertook an event history analysis of out-migration from households surveyed in Nicaragua and found that violence during the U.S.-sponsored Contra War strongly predicted out-migration to the United States, whereas conflict-related economic distress promoted migration to neighboring Costa Rica. Alvarado and Massey (2010) showed that Nicaraguan migration to the United States also rose in response to increases in lethal violence that accompanied the imposition of structural economic adjustment policies during the 1990s. In Mexico and Costa Rica, however, they found that the increases in lethal violence were more modest and were generally associated with a lower probability of emigration to the United States.

Research to this point thus suggests a clear connection between civil violence and migration, though the degree to which this association is mediated through the effect of violence on economic conditions is unclear. Also unclear is whether the effects of violence are similar for short- and long-distance moves and for internal versus international movement. Furthermore, only a few studies to date have examined the influence of violence on microlevel decision-making, and no study has yet considered the determinants of forced migration at multiple levels. Finally, it is unclear whether the influence of violence is characterized by threshold effects, such that rising violence induces movements only above a certain point, and below that violence might actually reduce the odds of movement.

In this article, we seek to rectify these shortcomings in the existing literature, although we are unable to test whether the effects of conflict are direct or indirect because this exercise would require time-varying measures of economic conditions, and our economic measures are static. Using fixed economic conditions as controls, we undertake a systematic event history analysis of how violence unleashed during Nepal’s Maoist insurgency of 1996–2006 affected the likelihood that people moved locally, internally, or internationally, holding constant other individual and household characteristics. Nepal offers a good test case for estimating the influence of violence on migration because its insurgency lasted a full decade, during which the scope and intensity of the conflict fluctuated considerably to produce substantial variation in the variable of interest.

Our study improves on existing research in several ways. First, we consider competing geographic destinations in the same model. To date, most studies have either modeled the movement of refugees internationally (see Apodaca 1998; Iqbal 2007; Schmeidl 1997; and Shellman and Stewart 2007) or focused on internal displacements (see Ibáñez and Vélez 2008; Morrison 1993; Morrison and Perez Lafaurie 1994; Schultz 1971). Relatively few studies have considered both kinds of migration at the same time (see Davenport et al. 2003; Melander and Öberg 2007; Moore and Shellman 2004), and so far, none have distinguished between local and long-distance internal moves. Our analysis also controls for the effects of human capital, social capital, physical capital, and demographic characteristics because these factors are expected to continue to promote voluntary migration even during periods of civil conflict. Following Schmeidl (1997) and Clark (1989), we predict the odds of migration by considering the determinants of migration classified into one of three basic categories: root causes, proximate causes, and intervening factors.

We begin with a brief history of the Maoist insurgency in Nepal, followed by an introduction to the study site and its connection to the conflict. We then describe the multilevel longitudinal survey from which we take our data. After presenting the analytic model and discussing how we operationalize key variables, we offer new findings about how violence affects migration to local, internal, and international destinations and draw relevant theoretical, substantive, and policy conclusions.

After 30 years of autocracy, the people’s movement for democracy won a new constitution in 1990 that reinstated democratic rule and transformed the King into a constitutional monarch. The new constitution did not fully meet the demands of several long-suppressed groups, however, which continued to chafe at domination by higher castes. Even worse, the democratic government seemed in many ways to extend the Panchayat system and was characterized by corruption, bitter power struggles, and the continued exploitation of peasants and ethnic minorities. Nepal remained a multiethnic country run by higher-caste Bahuns and Chhetris who continued to subjugate other ethnicities, religions, and castes under a Hindu theocracy (Thapa and Sijapati 2004).

In response to this history of inequality and exploitation, the Communist Party of Nepal (Maoist) submitted demands for change to the government on February 4, 1996, and when these demands were ignored, it launched an armed insurgency known as the “people’s war” on February 13, 1996. Owing to the neglect and marginalization of minorities and the poor under a feudal system that actually worsened under democratic rule, communist ideology became rooted in the Mid-Western district of Rolpa (see Fig. 1), which constituted the stronghold from which Maoist guerillas launched their insurgency to establish a “People’s Republic of Nepal.” Soon, the surrounding districts of Rukum, Salyan, Pyuthan, and Jajarkot also became Maoist strongholds.

Although socioeconomic inequality and discrimination by caste and ethnicity were major causes of the conflict, one cannot ignore spatial aspects of the insurgency (Murshed and Gates 2005). In a country where more than 80% of the population remains rural, development efforts nonetheless historically focused on urban areas. Among Nepal’s five development regions (see Fig. 1), the Eastern, Central, and Western Regions contain the most important cities and display the highest levels of development. In terms of the Human Development Index (HDI, which combines normalized measures of life expectancy, literacy, educational attainment, and GDP per capita), the 1998 value for the Western Region stood at .35, and that for the Central and Eastern Regions was .34 (on a 0 to 1 scale). In contrast, the Mid-Western and Far-Western Regions had respective HDIs of just .26 and .27 (Nepal South Asia Centre 1999). At the district level (see Fig. 1), the degree of variation was even greater, with scores ranging from more than .60 in Kathmandu (in the Central Region) to under .20 in Mugu, Kalikot, and Bajura (in the Far-Western Region).

Once the insurgency began, low-development provinces generally experienced the greatest violence (see Do and Iyer 2006). Empirical analyses by Murshed and Gates (2005) reveal that variation in the intensity of the insurgency across Nepal’s75 districts was positively and significantly related to the degree of income inequality. According to these authors, a low-intensity conflict is one with fewer than 25 battle-related deaths per year and no more than 1,000 battle-related deaths in total; a medium-intensity conflict has between 25 and 1,000 battle-related deaths per year and accumulated deaths in excess of 1,000; and a high-intensity conflict has more than 1,000 battle-related deaths per year.

During the insurgency, violence began to escalate on September 25, 2000, when Maoists overran Dunai, the headquarters of Dolpa district, killing 14 police personnel (Hutt 2004). Still, until November 21, 2001, the conflict was only of medium intensity and involved mainly insurgents and the police. After peace talks failed, however, the government proclaimed a state of emergency and denounced the Maoist rebels as terrorists in November 2001. Thereafter, the insurgency grew into a high-intensity conflict that involved the Royal Nepalese Army (Murshed and Gates 2005), and by late 2002, violence between the Maoists and government forces was reported in 73 of Nepal’s 75 districts (Kok 2003).

The insurgency finally ended on November 21, 2006, with the signing of a Comprehensive Peace Agreement, making the “people’s war” in Nepal one of the longest and most intense conflicts in contemporary times. Between February 1996 and December 2006, 8,377 people were killed by government forces and another 4,970 were killed by Maoist insurgents, yielding a total of 13,347 deaths (Informal Sector Service Center 2008). The civil war gave rise to widespread human rights abuses, such as abduction, forced conscription, torture, rape, extortion, use of civilians as human shields, forced billeting, arbitrary taxation, and local wildcat strikes. Together, the ongoing violence and these abuses created a state of terror, resulting in the displacement of thousands of people.

Given its relatively high level of economic development and flat terrain, which was less suited to guerrilla tactics compared with the hills and mountains, Chitwan was exposed to relatively moderate levels of violence during the decade-long civil war, particularly compared with districts in the Mid- and Far-Western Regions of Nepal. The insurgency resulted in a total of 200 deaths in Chitwan district over a decade, while the number of deaths in some of the hardest hit districts reached over 900 (Informal Sector Service Center 2008). Chitwan thus offers a unique opportunity to study real-time migratory responses to relatively moderate levels of violence while controlling for a number of variables known to influence migration.

Our analysis is made possible by the existence of a representative survey of households in Western Chitwan that compiled detailed, multilevel data over the entire period of civil conflict, which we were able to combine with monthly data on civil violence not only in Chitwan but also in surrounding districts. The Chitwan Valley Family Study (CVFS) used a combination of ethnographic and survey methods to create a detailed database of social, economic, and demographic characteristics for individuals, households, and communities. For sampling purposes, the western portion of the valley was divided into a set of mutually exclusive neighborhoods of 5–15 households, and 171 neighborhoods were selected on an equal probability basis using multistage cluster sampling (Barber et al. 1997). The migration data come from a prospective monthly panel survey that began in February 1997 and ended in January 2006. Instead of the 171 original neighbor-hoods, however, only 151 were followed in the panel owing to budget constraints. Household members from these 151 neighborhoods were reinterviewed each month even if they left the sample neighborhood, except when the entire household moved outside of Nepal. The panel survey thus yields accurate monthly data on place of residence, age, ethnicity, gender, and marital status.

We pair this longitudinal data set with corresponding cross-sectional data sets compiled at the individual and household levels for 1996, thereby creating a comprehensive file that links individual migratory behavior both to fixed and time-varying independent variables defined at multiple levels. In our sample, we include only those 3,802 respondents between the ages of 15 and 69 who resided in the 151 neighborhoods at the outset of the panel survey in 1996. After we merged the different data sets, fewer than 2% of the person-months contained missing values for any variable in the analysis, leading us to discount missing data as a significant source of bias in our analyses. In order to measure violence from Nepal’s civil conflict, we compiled monthly data on violence from the Maoist insurgency, relying on special data sets compiled by the South Asia Terrorism Portal (SATP) and the Informal Sector Service Center (INSEC). The data compiled by the SATP were downloaded directly from the group’s website, whereas those from INSEC were entered manually from its yearly Human Rights Yearbooks. The former source kept monthly tallies of bomb blasts, landmine explosions, and major clashes between government and insurgent forces, along with the casualties from these various incidents, whereas the latter reported dates of Maoist violent events by district.

Using the aforementioned person-month file, we then estimated multinomial logit models to predict the effect of independent variables defined in month t or at the baseline on migration outcomes in month t+1. Before choosing the multinomial logit model for our analysis, we tested for one of the key assumptions of the multinomial logit model, which is the independence from irrelevant alternatives (IIA). The IIA requires that the odds of choosing alternative 1 over alternative 2 is not influenced by the presence of other alternatives. We used the Hausman test in Stata to confirm that the assumption holds. Among the four tests reported, three were negative. Based on Long and Freese (2006:244–45), we conclude that these negative results offer sufficient evidence that IIA was not violated, which reinforces our decision to use the multinomial logit model. The model can thus be written as where i denotes individuals; j represents the three possible values for migration destinations 1, 2, and 3 versus 0, which is the reference category representing no migration; and t represents time period, which is month in our study. The multinomial regression produces separate regression equations for the three different migration destinations represented by j. The migration outcome is a nonlinear transformation of the natural logarithm of the predicted value of the odds or relative probability, p /1 − p, of migrating to three competing destinations (j =1,2,or3) versus not migrating. β_0j is the constant, β_1j to β_nj represent the coefficients; ε_ij is the error term for individual i with migration outcome j; and finally, X_1i to X_ni represent the explanatory variables, which are grouped under four categories: proximate causes, root causes, intervening factors, and demographic variables.

In terms of physical capital, 80% of the households to which respondents belonged owned farmland. As for household amenities, 49% of respondent households had a roof made of slate, tin, or concrete; only 25% had a floor made of concrete. These materials represent better quality than materials such as straw to thatch the roof or mud, brick, or wood for the floor. Higher-quality materials are not generally affordable to poor households and thus indicate a higher socioeconomic status. On average, 46% of respondent households did not have their own source of drinking water, while 36% had no toilet, and 66% had no electricity. Among the goods owned by households, 52% had a radio and 62% had a bicycle, but only 12% had a TV and just 3% had a motorcycle; 7% reported owning a cart, 5% owned a biogas plant, and just 1% and 3% had a tractor or an irrigation pump, respectively. The average household owned 20.2 chickens and ducks, 1.15 pigeons, 0.57 bullocks, 0.36 cows, 0.16 male buffaloes, 1.25 female buffaloes, 1.46 sheep and goats, and 0.06 pigs.

As already mentioned, these variables were used to create the factor scales of violence (under proximate causes) and physical capital (under root causes), with each index standardized for ease of interpretation with a mean of 0 and variance of 1. Table 4 presents means, standard deviations, and ranges for predictors of migration and counts for migration to different destinations. Education, which is a measure of human capital, is generally quite low. Although years of schooling ranged from 0 to 16 in the typical person-month under observation, it averaged only 3.1 years. Another measure of human capital is whether the respondent held a salaried job in1996, but in only 5.3% of person-months did respondents hold such a job as of 1996. Age is another measure of human capital, and in the average month when respondents were exposed to the risk of migrating, the typical person was 39.5 years old, though people ranged in age from 15 to 69 years. From the statistics on social capital, we see that in the average person-month, around 6.3% of all respondents had at least one household member who had migrated within Chitwan by 1996, another 8.8% had someone who had migrated to other districts within Nepal, and 9.7% had someone with international experience.

Prior work has generally hypothesized and found a strong positive connection between violence and migration, though at least one study has uncovered threshold effects of violence (Morrison 1993) and another study found that low levels of violence may negatively influence the odds of movement (Alvarado and Massey 2010). In order to test the threshold hypothesis systematically, we introduced a squared term for violence to capture the nonlinearity. As can be seen in Table 5, both the violence scale and its square are highly significant in predicting migration, indicating a curvilinear relationship. Whereas the main effect of the violence scale is strongly negative, its square is strongly positive, which means that, controlling for other relevant determinants, violence lowers the odds of out-migration to all three destinations up to a certain point, after which it increases the odds of out-migration. The effect of violence on migration also seems to rise as the distance of the move increases, such that violence has the most profound effect on migration to international destinations, followed by migration to other districts, and then local migration within Chitwan.

In order to see directly how violence influences the likelihood of local, internal, and international migration, we used the data shown in Table 5 to generate predicted probabilities of movement to each destination. Specifically, we varied the violence scale from the observed minimum to the observed maximum and applied the associated coefficients for violence and its square while applying other coefficients to mean values of the associated variables. As can be seen in Fig. 2, violence is most influential in suppressing international migration and less influential in suppressing internal migration to other districts in Nepal. The probability of international migration drops to near zero at even low levels of violence, whereas the likelihood of internal migration drops to near zero at moderate levels of violence, and the odds of moving to either destination remain near zero over the entire range of violence observed in Chitwan and surrounding districts. The probability of migration within Chitwan is greater no matter what the level of violence, but rising violence also has the effect of suppressing local mobility as it goes from low to moderate levels. Unlike the situation with internal and international moves, however, as violence rises into the high range, it eventually begins to increase the odds of making a local move.

Thus, at the levels observed in and around Chitwan, violence served to lower the odds of internal and international migration but had a curvilinear effect on local mobility, reducing the likelihood of moving within Chitwan at low to moderate levels but increasing the probability of mobility within Chitwan as it approached high levels. Because we were forced to assume that the level of violence stood at zero before 2002—owing to the lack of published data—we reestimated the model using person-months after this date to make sure that our conclusion was not an artifact of this assumption. The reanalysis yielded the same curvilinear effects, confirming the validity of our findings.

Other independent variables in the model generally display effects that are consistent with prior theory and research. Like Massey and Espinosa (1997), we find that ownership of physical capital generally reduces the odds of migration. Ownership of farmland, for example, significantly reduces the odds of moving within Chitwan and to other districts in Nepal, although it increases the odds of international migration (see Table 5). Similarly, access to household amenities significantly lowers the odds of moving within Chitwan. The effects of human capital are generally positive, with education and occupational skill increasing the likelihood of movement to all three destinations (though the effect of occupational skill is not significant in predicting international migration). As expected, social capital evinces strong place-specific effects on mobility, with ties to local movers predicting migration within Chitwan, ties to internal migrants most strongly predicting migration to other districts in Nepal, and ties to international migrants strongly predicting emigration outside the country. Unlike the results of many studies of migration, however, age generally had a decelerating negative effect on the likelihood of migration irrespective of destination: as respondents aged, they slowly grew less likely to migrate. Although this unusual pattern could reflect the absence of households that left Nepal without returning, the same effect has been observed in other studies of migration that used retrospective data from Chitwan and that were not subject to this potential bias (Bohra and Massey 2009;Massey etal. 2010).

Schmeidl (1997) identified seasonality as an important intervening factor that can play a significant role in increasing or decreasing refugee migration, and the dummy variables for month do indeed reveal distinct patterns of seasonal migration for different kinds of moves. Compared with December, within-Chitwan migration is elevated during all months except January, September, and October; migration to other districts in Nepal is less likely to occur in January but more likely to occur in February; and departures for international destinations are most likely to occur in February, March, and November.

In terms of personal demography, the effects of gender and marriage are complicated by a significant interaction that stems from the Nepalese practice of patrilocal residence following marriage. Among males, the effect of marriage on migration is simple: it raises the odds of moving within Chitwan by 68% (e^(0.516) = 1.675) and increases the odds of moving internationally by 82% (e^(0.597) = 1.817). In contrast, whereas unmarried women are always less likely to migrate than unmarried men, because of the interaction between gender and marriage, married women are more likely than unmarried men to move locally and internally, though not internationally. Compared with a single man, for example, a single woman is 24% less likely to move within Chitwan (1 – e(−0.271) = 0.237), 34% less likely to move to another district in Nepal(1 – e^(−0.408) = 0.335), and 95% less likely to move outside of Nepal(1 – e^(−2.949) = 0.948); but a married woman is 67% more likely to move within Chitwan than a single man (e^{(−0.271+0.516+0.269)} = 1.672) and 29% more likely to move elsewhere in Nepal (e^{(−0.408+0.067+0.597)} = 1.292). A married woman is still much less likely than a single man to move internationally, with the odds being 73% lower (1 – e^{(−2.949+0.597+1.028)} = 0.734).

Another demographic variable, rising household size, decreases the odds of leaving Chitwan for other districts and other countries. Each additional person in the household lowers the odds of migrating to other districts by around 4% (1 – e^−.045 = 0.044) and to other countries by 12% (1 – e^−.13 = 0.122). Finally, in terms of caste, Hindus and Hill Tibeto-Burmese are far more likely to move internationally than other groups, and the Hill Tibeto-Burmese seem to be the most mobile, with an elevated risk of migration to all three destinations. Thus, consistent with the earlier results of Bohra and Massey (2009), those who belong to ethnic groups with prior migratory experience are more likely to migrate than Terai Tibeto-Burmese, the indigenous people of the Chitwan Valley, whatever the level of violence.

Third, we were able to measure the independent effect of violence on the decision to migrate locally, internally, and internationally. So far, the literature on forced migration has tended either to study movement to one destination at a time or to lump all destinations together. We found that the effect of violence on the likelihood of migration was different for different destinations, increasing as the distance of the move increased. the effect of violence was weak in the case of local moves, stronger for moves outside of Chitwan, and the strongest for international moves. Fourth, we specified a more complex model of individual decision-making that estimated the effects of violence while controlling for leading predictors of voluntary migration, which we found influenced migration whatever the level of civil violence. Interactive models found that effects of any of the predictors of migration did not significantly vary by the level of violence.

Finally, our results are novel in showing that civil violence affects migration differently at different levels of intensity. Whereas most prior work has found a positive effect of violence on out-migration, our results are consistent with those few studies that have either found violence to have no influence on migration below a certain threshold or to reduce the likelihood of movement at low to moderate levels. We found a curvilinear effect of violence on decisions to migrate to all three destinations such that the odds of out-migration were reduced at low to moderate levels of violence and only increased as violence approached high levels. Inspection of predicted probabilities indicates that at the modest levels of violence observed in and around Chitwan, approaching the maximum observed there only increased the odds of making local moves but had no practical influence in raising the odds of internal or international moves. Of course, the higher levels of violence observed in other districts of Nepal could well be great enough to raise the probability of moving internally and internationally as well.