1 Introduction

Wars have far-reaching effects on various aspects of politics, the economy, and society. On a direct level, wars lead to morbidity, mortality, the destruction of infrastructure and healthcare systems, and the displacement of populations. However, their repercussions often extend beyond the immediate aftermath. According to Catani (2018), war-affected populations may experience long-term psychological stressors. Studies by Devakumar et al. (2014) and Phadera (2021) highlight intergenerational health consequences, indicating that the impact goes deeper and persists across generations.

With rising global instability, understanding the impact of war on demographic changes is crucial. Previous studies have found mixed evidence on the effect of conflict exposure on fertility. For instance, Lindstrom and Berhanu (1999), Agadjanian and Prata (2002), and Blanc (2004) provide evidence of a reduction in fertility due to conflict. On the other hand, Kulczycki and Prem (1999) and Ladier-Fouladi and Hourcade (1997) find no overall effect, while Kraehnert et al. (2018) identify evidence of increased fertility resulting from conflict. Despite a growing body of literature on the impact of war on fertility, the consequences of the Sri Lankan civil war remain relatively unexplored in this context.

This paper aims to explore the effect of exposure to armed conflict on female fertility in Sri Lanka and highlights the role of ethnicity in this context. The 26-year-long civil war in Sri Lanka (1983–2009) is considered one of the world’s most violent and prolonged secessionist movements since the end of the Second World War (DeVotta, 2004). Using the Uppsala Conflict Data Program Georeferenced Event Dataset (UCDP GED) on conflict events, I classify districts above the mean of the total number of conflict events during 1992–2009 as treatment districts (high-conflict zone) and the others as control (low-conflict zone). Using vital statistics data, I compute crude birth rates as a measure of fertility. Employing a difference-in-difference approach, estimates suggest that on average fertility declined by 5.42 births per 1000 people in the high-conflict districts of Sri Lanka. Moreover, the study suggests that an increase in female age at marriage in these areas may contribute to this observed fertility decline.

This paper emphasizes that, as one of the understudied aspects of heterogeneity in conflict literature, the effects of exposure to armed conflict differ across ethnic groups. The history of the Sri Lankan conflict suggests that ethnicity had a major role in shaping the dynamics of the conflict. The divide between the Sri Lankan Tamil minority and the Sinhalese majority percolated through the years and ultimately manifested into intolerance between the two communities, resulting in war in 1983. Consequently, exploring heterogeneity across ethnic groups is valuable in understanding their sensitivities and their ability to adjust their fertility preferences or responses to such shocks, which is further useful to guide policymakers.

The difference-in-difference estimates in the study provide evidence of differential fertility adjustments of the two ethnic groups in response to the conflict. Across various specifications, the estimates suggest that the Sinhalese reduced their fertility by 8–10 live births per 1000 people while Sri Lankan Tamils reduced theirs by 0–4 live births per 1000 people in response to exposure to armed conflict. This implies that the Sinhalese had a more pronounced fertility reduction. Additionally, using data on conflict-specific mortality from the UCDP database, this study finds evidence of a replacement effect among the Sri Lankan Tamils, contrasting with the absence of such an effect for the Sinhalese.

This paper contributes to a growing literature by evaluating the dynamics between conflict, ethnicity, and fertility outcomes. In line with recent studies (Schindler & Brück, 2011; Kraehnert et al. 2018; Thiede et al., 2020; Boehnke & Gay, 2020), this paper finds that changes in female age at marriage contribute to an overall reduction in fertility in Sri Lanka and evidence of ethnic group level replacement effects explaining a lower reduction in fertility for Sri Lankan Tamils. To the best of my knowledge, this is the first paper to focus on the effect of exposure to conflict on fertility in Sri Lanka.

The rest of the paper is organized as follows: Section 2 discusses related literature; Section 3 details the history of the ethnic divide and evolution of the conflict; Section 4 describes the data and the methodology; Section 5 discusses the overall effect in Sri Lanka and a related mechanism of increasing female age at marriage; Section 6 focuses on the ethnic group differential and the presence of replacement effect for Sri Lankan Tamils; Section 7 concludes.

2 Related literature

In general, fertility behaviors are governed by a plethora of factors such as income, education, age, and other socio-economic factors. Bongaarts’ model of the proximate determinants of fertility categorizes the set of factors affecting fertility levels into two broad categories: behavioral and biological (Bongaarts, 1987). Within this framework, three biological factors influence fertility decisions: permanent sterility, the probability of conception, and intrauterine mortality. Behavioral determinants, on the other hand, include factors like the prevalence of contraceptive use and effectiveness, duration of breastfeeding, practice of induced abortion, and patterns of union formation.

The literature on conflict identifies both biological and behavioral mechanisms through which armed conflict exposure can impact fertility. The four major pathways discussed in previous studies are income effect, marriage market effect, sexual violence against women, and replacement effect.

The income effect posits that factoring in the cost of raising children, coupled with a reduction in wealth during the war (Collier et al., 2003), can often result in reduced fertility. For instance, Kraehnert et al. (2018) found that the death of a woman’s sibling during the civil war in Rwanda lowered the probability of giving birth. The study suggests that if the deceased person played a role in monetarily supporting the upbringing of the child, the income effect would explain the reduction in fertility.

Conflicts can also exert an impact on the marriage market and thus impact fertility. This occurs when the conflict results in a demographic deprivation of young men/husbands, either by reduced spousal cohabitation with men recruited and away at war (Blanc, 2004; Urdal & Che, 2013; Vandenbroucke, 2014) or by altering the sex ratio (Boehnke & Gay, 2020; Kraehnert et al., 2018). Exposure to armed conflict can also result in a shift in marriage patterns due to a change in bride prices (Corno et al., 2017) or a change in the age at marriage. The increased uncertainty and absence of men could encourage postponement in marriage and thus delay marital fertility (Thiede et al., 2020). On the contrary, Hacker et al. (2010) also emphasize that in some situations, war can be a catalyst for marriages. Faust (1996) underscored that war resulted in a dread of “spinsterhood” in southern women during the American Civil War. Barber (2000) reported that the American Civil War had led to a “marriage frenzy” in Richmond.

Additionally, conflict may positively affect fertility rates due to increased sexual and gender-based violence against women by armed combatants or perpetrated by intimate partners (McGinn, 2000; Østby, 2016; Leatherman, 2011; United Nations, 2014, Wood, 2018). Kottegoda et al. (2008) provide accounts of increased vulnerabilities of women and higher intimate partner violence in Sri Lanka during the civil war. The repercussions of conflict extend to the accessibility of reproductive health care for women (Kottegoda et al. 2008; Urdal & Che, 2013; Lindskog, 2016). On the contrary, conflict can also result in psychological stressors as well as poor nutritional status of the women and thus negatively impact fertility (Nowrojee, 1996; McKay, 1998; Abu-Musa et al. 2008; Dossa et al. 2015; Catani, 2018).

Finally, the effect of higher mortality on fertility is ambiguous (Schultz, 1997; Ben-Porath, 1976). The presence of a positive correlation suggests a replacement mechanism. Replacement effects can arise at an individual level if children are considered old-age insurance for the parent and an increase in mortality leads individuals to have more children in the expectation that only some of their kids will survive (Nugent, 1985; Rossi & Godard, 2019) or if people are trying to replace the loss of their loved ones (Schindler & Brück, 2011; Kraehnert et al., 2018). At the group level replacement effects can arise if pronatalist sentiments emerge during conflicts among the minority community to maintain the size of their group in the population (Chi et al., 2015; Palmer & Storeng, 2016; Alburez-Gutierrez, 2019). In the Sri Lankan context, there may be a replacement effect in place for the Sri Lankan Tamils based on the qualitative findings of Jordan and Denov (2007) who report evidence of marriage between members of opposite genders being strongly encouraged amongst LTTE militants to inspire a greater sense of commitment and to ensure traditional Tamil customs are practiced in the larger society.

With many counteracting mechanisms, the net effect of conflict on fertility remains to be tested empirically for Sri Lanka.

3 The Sri Lankan civil war

In 1981, Sri Lanka had a total population of approximately 14.8 million. The Sinhalese comprised 74% of the population, followed by Sri Lankan Tamils who were 12.7% of the population, and Indian Tamils who were 5.5% among other groups (Census of Population and Housing Report, 1981). Historically, the Sinhalese have inhabited the western, southern, and central districts, whereas the Sri Lankan Tamils were a majority in the northern and eastern parts of the country (Census of Population and Housing Report, 1971; Keethaponcalan, 2002). In line with that, the Census of Population and Housing Report (1981) finds that 65% of the Sri Lankan Tamils resided in the northern and eastern districts in 1981.

Keethaponcalan (2002) describes that the districts in the Northern and Eastern Provinces (Jaffna, Kilinochchi, Mullaitivu, Mannar, Vavuniya, Trincomalee, Batticaloa, and Ampara) were referred to as the “traditional homeland” by Sri Lankan Tamils. These areas were under the direct control of the Liberation Tigers of Tamil Eelam (LTTE) during the war and had to therefore witness a large amount of conflict between the LTTE and the Government of Sri Lanka (GoSL)Footnote 1. Figure 1 identifies the geographic variation in conflict intensity in the districts along with the dominant ethnic group in the district. The figure points to an overlap between the conflict zone and the dominant ethnic group in the district being the Sri Lankan Tamils. While there were some instances of conflict reported in almost all districts of Sri Lanka, they were largely concentrated in the Sri Lankan Tamil dominant districts (Appendix Table 4).

Fig. 1
figure 1

Map of Sri Lanka - areas of high-conflict and largest ethnic group in each district. Note: The classification of districts as high and low conflict is based on Appendix Table 4 using the UCDP conflict dataset. Data for the size of ethnic groups by district is obtained from the Census of Population and Housing Report (1981) and is provided in Appendix Table 5. This allows a graphical depiction of the largest ethnic group in each district, mapped with the conflict status of the district

Full size image

Though the war began in 1983, the roots of the ethnic conflict in Sri Lanka go back to the former as Ceylon under British Rule. With the ‘divide and rule’ policy underway, there were tensions between the communities even before 1983. It is claimed that there was an over-representation of Tamils in the colonial administrative positions, owing to British favoritism. The British often constructed English medium schools in Tamil areas that allowed greater access for Tamils to it and government jobs at the time. Tamils comprised about 30 percent of upper-level civil service jobs while comprising 22.7 percent of the population as per the census in 1946. The Sinhalese comprised 69.4 percent of the population and 57.6 percent of the upper levels of civil service (Keethaponcalan, 2002). This over-representation was considered unjust by the Sinhalese.

Post-independence in 1948, the bulk of the power in government landed in the hands of the Sinhalese elite who resented Sri Lankan Tamils. They passed several laws that were detrimental to Tamil representation and interests. They disenfranchised Tamil migrant plantation workers from India, also known as the Estate Tamils or Indian Tamils, or Upcountry Tamils. This meant the Tamils lost 40% of their parliamentarians because Estate Tamils could not be represented under this law. The Sinhala Only Act of 1956 declared Sinhala as the only official language of Sri Lanka. This made it mandatory for Tamils in the public sector to qualify in Sinhala within a designated short period, a lack thereof meant losing their jobs. The Sinhalese considered it as a move to reclaim the true Sri Lankan identity that was lost during British Rule. However, to the Tamils, this was an attempt to limit their economic opportunities, as from 1956–1976 of the 189,000 people recruited in the public sector, 95% were Sinhalese. From the 1970s there was virtually no employment of Tamils in the armed forces and police which are 99.9 and 95% respectively Sinhalese. In 1972, the government changed the country’s name from Ceylon to Sri Lanka and made Buddhism the nation’s primary religion, emphasizing a sense of culture and national identity. In an attempt to bridge a gap between educational disparities in the country, the ‘standardization policy’ of 1972 required Tamil students to score higher marks than their Sinhalese counterparts to get admission into Sri Lankan Universities. This act was aimed at providing support to the previously disadvantaged Sinhalese youth. Hoping to undo the injustice to the Sinhalese under the British Era, standardization ended up creating a further rift between Sinhalese and Tamils. Demands for equal status by Tamils radicalized over the years and led to the formation of a militant organization called LTTE in 1976 under the leadership of Velupillai Prabhakaran. Their demands for self-government and an independent state called the Tamil Eelam waged war in 1983.

3.1 Eelam War-I (1983–1987)

The first of four Eelam Wars began in 1983 and lasted 4 years. While it was the northern and eastern provinces that were direct conflict zones under direct control of LTTE, there were some instances of violence reported across Sri Lanka. In July 1983- known as “black July”- riots broke out in Colombo and other cities. LTTE had killed 13 army soldiers. In retaliation by the Government of Sri Lanka (GoSL), about 2500–3000 Tamils likely died. Violence escalated over the years as Tamil insurgents used car bombs, human bombs, and landmines against Sinhalese civilians and the military. India tried to intervene by sending peacekeepers in 1987, but its troops had to withdraw by 1990 as Tamil guerillas refused to step back.

3.2 Eelam War-II (1990–1995)

After the withdrawal of the Indian Peace Keeping Forces, the war got more intense. LTTE killed about 600–700 Sinhalese police officers in June 1990. A female LTTE suicide bomber assassinated India’s Prime Minister Rajiv Gandhi at an election rally in 1991. The government cut off shipments of food and medicines to the Jaffna peninsula and started intensive aerial bombardment. The LTTE responded by massacres of Sinhalese and Muslim villagers which led to Muslim self-defense units doing the same in Tamil villages.

3.3 Eelam War-III (1995–2002)

In January 1995, a peace agreement was signed with the new government. This however was violated after 3 months when LTTE destroyed ships and naval gunboats. This was followed by massacres in refugee camps in the Jaffna Peninsula and other civilian sites. By December, Jaffna was under government control for the first time. However, in July 1996, the LTTE carried an 8-day strike in Mullaitivu against 1400 government troops. More than 1200 government soldiers were killed and the LTTE also lost about 332 troops. Suicide bombers attacked Colombo, the Central Bank in Colombo, the Sri Lankan World Trade Center, and other southern cities repeatedly during the late 1990s.1999 also saw an attempt on President Kumaratunga’s life by a bomber. Norway tried to negotiate a peace settlement in December 2000. After various rounds of negotiations, LTTE and GoSL signed a peace treaty with LTTE foregoing the demand for a separate Tamil state. However, in October 2003, GoSL declared a state of emergency when Tigers announced their full control over Northern and Eastern Sri Lanka. Norway reports about 300 violations of the ceasefire by the GoSL and 3000 by LTTE during that period.

3.4 Eelam War-IV (2006–2009)

Eelam War IV from 2006–2009 was the final war that culminated in peace, and GoSL declared victory. After the October 2006 peace talks in Geneva failed, the government launched a major attack against the LTTE. The next few years were called a bloodbath as thousands died on both sides. Ultimately in May 2009, the GoSL declared victory. The estimated death toll by the UN over the years is 75,000–100,000 people with many other victims of kidnappings, sexual abuse, and disabilities due to war.

4 Data

The dataset for this analysis was created by digitizing the data from the Annual Reports of Vital Statistics from 1967–1980 (Vital Statistics, 1967–1980) and 1992–2009 (Vital Statistics, 1992–2009) and Census Reports over the years 1963–2012 (Census of Population and Housing Report, 1963, 1971, 1981, 2001, and 2012). Registration data exists for 1981–1991 as well, but I am unable to obtain the vital statistics for 1981–1991. The data on vital statistics has information on births and deaths at the district level disaggregated by ethnicity. Appendix Table 6 provides the descriptive statistics of the data for all ethnic groups in Sri Lanka. The registration system for births and deaths in Sri Lanka has been in place since 1867 and is of good quality (Levine, 2007; Jayachandran & Lleras-Muney, 2009) with completeness of at least 90% in most of the years of this analysis (United Nations Statistics Division, 1970–2010; United Nations Statistics Division, 2006). Assuming exponential population growth, I use the census year population by ethnic group for each district to interpolate the population for intercensal years.Footnote 2 Using the data on births at the ethnicity-district level, I create measures of crude birth rates. While crude rates have the drawback of not accounting for the age composition of the population, which may be consequential for the size of the next birth cohort, data availability restricts the analysis to using the said measure.

To create a panel dataset of female age at marriage, district-level data on pre-war years female singulate mean age at marriage (SMAM) has been collected for years 1953, 1971, and 1981 from the Census of Population and Housing Report (1981). The vital statistics dataset by GoSL provides information on female age at marriage for the post-war years of this analysis from 1992–2006 (Vital Statistics, 1992–2009). The data classifies females getting married into 3 categories based on their prior marital status: females who were never married, females who were divorcees, and females who were widows. For consistency in age at marriage information before and during the war, since SMAM reports mean age at first marriage, I use the information of female age at marriage for women who were previously never married so that this can mimic SMAM data in war years. This dataset is not available at disaggregated levels by ethnic groups.

I obtain conflict data from the UCDP GED database. UCDP is the world’s major data source for organized violence with the oldest ongoing data collection project for civil wars. It defines an armed conflict as a contested incompatibility that concerns government and/or territory over which the use of armed force between the military forces of two parties, of which at least one is the government of a state, has resulted in at least 25 battle-related deaths each year. UCDP has global coverage and provides data on three types of organized violence: state-based conflict, non-state conflict, and one-sided violence, both spatially and temporally. I aggregate conflict events at the district level. Appendix Table 4 provides the number of conflict events aggregated at the district level over the years 1992–2009. The district with the greatest number of conflict events is Mullaitivu with 806 total conflict events reported. Jaffna experienced about 559 of all conflict events reported in the dataset. The districts with the lowest conflict events were Ratnapura and Kalutara which reported 0 conflict events as per the dataset. The dataset classifies each conflict event as organized violence with at least one fatality in either the best, low, or high estimate categories and provides its date and location (Sundberg and Melander, 2013). Using the best estimates of conflict-specific fatalities from the dataset, I obtain an estimate of conflict-specific death rates at the district-year level.

Since the boundary of districts changed over time, I adjust the data based on the area of the district in 2012 to allow for comparability over time and thus have 25 districts in this analysis. Wherever districts were split to create new ones, I adjust its population to be based on its share of the parent district’s area. For instance, the Gampaha district with a 1387 km sq area was carved from the Colombo district which had an area of 2051 km sq area before the split. So, the population assigned to Gampaha before the split has been adjusted in proportion to the areas of the two districts (i.e., approximately 65% of the Colombo population before the split).

5 Overview: impact in Sri Lanka

Few studies on Sri Lanka have examined trends in fertility (Abeykoon, 1987; Puvanarajan & DeSilva, 2001), but none through the lens of the effect of armed conflict. To identify the causal effect of exposure to conflict on fertility, I estimate a difference-in-difference model with crude birth rate as the dependent variable. Using the information on conflict events, I classify districts above the mean of the total number of conflict events during 1992–2009 as treatment districts (high-conflict zone) and the others as control (low-conflict zone) in Appendix Table 4. Alternatively, given that conflict per capita may be a more appropriate measure of conflict exposure, for each district I calculate the average of the lagged number of conflict events per capita per year. Using this measure of conflict per capita, I assign districts as treatment group if the conflict intensity was greater than the mean across districts and as control group if they are below the mean intensity. Both these classifications result in the same districts being assigned to treatment and control groups. Appendix Fig. 4 demonstrates parallel trends in the average crude birth rate for the treatment and control group in the pre-treatment years. Since the timing and duration of conflict are unpredictable, the treatment can be considered exogenous.

I estimate the following equation to identify the effect of armed conflict exposure on the crude birth rate in Sri Lanka:

$${{br}}_{{gdt}}={{{\beta }}}_{0}+{{{\beta }}}_{1}{{treated}}_{d}\times {{post}}_{t}+{\theta }_{d}+{{{\delta }}}_{t}+{\epsilon }_{{gdt}}$$
(1)

where br is the crude birth rate for ethnic group g in district d at time t, \({\theta }_{d}\) and \({{\delta }}_{t}\) represent district and time fixed effects respectively. treated is a dummy variable that takes a value of 1 for high-conflict districts and 0 for others, post is a dummy that takes a value of 1 for years after 1983 that mark the beginning of the civil war (i.e., for 1992–2009, due to data constraints I do not have data from 1983–1991). Thus, treated × post is 1 for treatment districts that are the high-conflict districts for years after 1983 (i.e., for 1992–2009). β1 in this setup is the relevant coefficient, measuring the impact of exposure to conflict on the crude birth rate in Sri Lanka. I estimate the model using robust standard errors (Castelló-Climent et al., 2018).Footnote 3

Estimates from Table 1 column 1 and column 2 report the overall average effect of conflict on fertility in Sri Lanka. While Column 1 provides the unweighted estimates, column 2 provides district population weighted estimates.Footnote 4 On average, the effect of exposure to high conflict was a reduction of 3.96 to 5.42 births per 1000 people. The pre-conflict average crude birth rate in Sri Lanka was 27.95 and a reduction of approximately 3.96 to 5.42 5.42 in the crude birth rate is a moderate response of 14.16–19.39% fertility reduction compared to some other estimates in the literature. A Population Reference Bureau report (Madsen and Finlay, 2019) documents that fertility in conflict usually drops by up to one-third. For instance, it dropped by one-third due to the Khmer Rouge conflict in Cambodia in the 1970s. Curlin et al. (1976) find a similar response of the age-specific fertility rate of women aged 10–14 in Bangladesh which fell by 33% due to conflict. A much larger reduction is documented by Vandenbroucke (2014) who finds that WWI led to a 50% reduction in France’s birthrate.

Table 1 Impact of conflict on fertility
Full size table

As a robustness check to this overall estimate, I explore alternative specifications in Appendix Table 7. There, column 1 assigns treatment based on Li et al. (2019), with an alternative measure assigning 8 districts as high-conflict (with more than 95% of all conflict events occuring in these districts and a larger presence of LTTE) and the rest as low-conflict zone. Column 2 assigns treatment and control districts to top and bottom quartile districts (6 in each group) ranked by the mean level of conflict per capita per year. Columns 3 and 4 remove the top 5% and top 10% of conflict districts respectively. The result remains qualitatively similar.

As highlighted by Bongaarts’ model (Bongaarts, 1987), one key determinant of fertility is age at marriage (operating via the marriage market effect). Early marriage or cohabitation tends to lead to longer reproductive periods and, therefore, higher fertility rates. In line with this hypothesis, prior studies have found that age at marriage has been an important determinant of fertility in Sri Lanka with a negative relationship for all ethnic groups (Abeykoon, 1987).

To understand if conflict resulted in an increased age at marriage that exerted a negative effect on overall fertility noted in Table 1, I estimate the following equation:

$${{aam}}_{{dt}}={{{\beta }}}_{0}+{{{\beta }}}_{1}{{treated}}_{d}\times {{post}}_{t}+{\theta }_{d}+{{{\delta }}}_{t}+{\epsilon }_{{dt}}$$
(2)

where aam is the female age at marriage in district d and year t. As before, treated is a dummy variable that takes a value of 1 for high-conflict districts and 0 for others, post is a dummy that takes a value of 1 for years after 1983 that mark the beginning of the civil war (i.e., for 1992–2006) and 0 for pre-war years (1953, 1971, and 1981). The coefficient β1 measures the effect of conflict on age at marriage. \({\theta }_{d}\) and \({{\delta }}_{t}\) represent district and time fixed effects respectively. I estimate the model using robust standard errors (Castelló-Climent et al., 2018). Footnote 5

The estimate in Table 2 suggests that on average, exposure to high-conflict led to an increase in female age at marriage by approximately 2 years in high-conflict zones in Sri Lanka. A higher age at marriage would delay fertility and thus result in a lower birth rate. This increased age at marriage could be driven by a marriage market squeeze if war impacted the local sex ratio and resulted in demographic deprivation of young men/husbands either due to men being increasingly recruited in combat or due to spousal deaths.

Table 2 Impact of conflict on age at marriage
Full size table

6 Digging Deeper: Heterogeneity & Replacement Effect

6.1 Heterogeneity by Ethnicity

Next, I evaluate the heterogeneity by ethnic groups. Appendix Figs. 5–9 present preliminary evidence of the differential in fertility trends using the size of the birth cohort and the crude birth rate.Footnote 6

To formally estimate the heterogeneity in fertility adjustments to conflict, I estimate the following equation:

$${{br}}_{{gdt}}={{{\beta }}}_{0}+{{{\beta }}}_{1}{{treated}}_{d}\times {{post}}_{t}+\mathop{\sum }\limits_{g=2}^{n}{{\beta }_{g}{ethnicity}}_{g}\times {{treated}}_{d}\times {{post}}_{t}+{\gamma }_{g}+\,{\theta }_{d}+{{{\delta }}}_{t}+{{{\alpha }}}_{{gt}}+{{{\mu }}}_{{dg}}+{\epsilon }_{{gdt}}$$
(3)

where βg captures the heterogeneity in the effect of conflict by ethnic group. Since my interests lie in looking at the two ethnic groups that were directly involved in the conflict, I simplify my analysis as follows: in the main specification, I classify ethnic groups as Sinhalese, Sri Lankan Tamils, and Others (non-Sri Lankan Tamil and non-Sinhalese category).Footnote 7 As a robustness check, in an alternate specification in Appendix Table 9, I also disaggregate ethnic groups as Sinhalese, Indian Tamils and Sri Lankan Tamils, Moors and Others (aggregating any remaining ethnic groups, <1% of the total population).Footnote 8 Regardless of specification, I included the dummy of n-1 ethnic groups interacted with treated × post, omitting it for Sinhalese and making Sinhalese the reference or base category. \({\beta }_{1}\) in this setup is the average effect of being in a high-conflict zone on Sinhalese fertility. \({\beta }_{1}+{\beta }_{g}\) would give the net effect of conflict for community g. I estimate the model using robust standard errors (Castelló-Climent et al., 2018).Footnote 9γg, \({\theta }_{d}\) and \({{\delta }}_{t}\) represent ethnic group, district, and time fixed effects respectively. I further control for district-ethnic group fixed effects, \({\mu }_{{dg}}\) and year-ethnic group fixed effects, αgt to account for unobservables and prevent omitted variable bias such as accounting for some geographic differences within ethnic groups (for instance accounting for some behavioral differences between Sri Lankan Tamils in the northern part relative to those Sri Lankans who lived elsewhere) or for accounting for shocks/policy in some years across Sri Lanka that could have affected ethnic groups differently respectively (for instance the introduction of a nationwide family planning program that could differentially affect ethnic groups based on their cultural acceptance of contraceptives).

The regions marked by intense conflict were predominantly populated by Sri Lankan Tamils.Footnote 10 The high conflict districts were the districts that that represented the traditional Tamil homeland and were fought over to create the independent state of Tamil Eelam. Given that ethnicity played a crucial role in the evolution of the conflict, I hypothesize that the Sri Lankan Tamils and the Sinhalese would not have a similar demographic response since one of them is a majority and the other is a minority. The socio-political vulnerability of the group could shape its demographic adjustments to shocks. As per Table 1 column 3, the effect of high-conflict exposure leads to a decline in Sinhalese fertility by 8 live births per 1000 people. Column 4 which accounts for population weights suggests a similar effect.Footnote 11 For the effect on Sri Lankan Tamils, the treated × post × Sri Lankan Tamil coefficient is positive and significant. While the net effect on Sri Lankan Tamils as per column 3 is of reduction in fertility by 4 live births per 1000 people, column 4 suggests that the net effect on Sri Lankan Tamils is equivalent to a null effect (−10.66 + 11.41 = 0.75 ~ 0). These estimates in Table 1 underscore that considering only the overall effect on the population masks the heterogeneity by the ethnic groups. Sri Lankan Tamils experienced a significantly lesser reduction in their fertility in response to the conflict compared to the Sinhalese. For the Sri Lankan Tamils, their behavioral responses could be driven by not just an effort to maintain the size of their community and thus their political representation (De Silva et al., 2020), but to also preserve their culture and tradition (Jordan and Denov, 2007). But to the contrary, a majority community would not necessarily be under such pressure and can thus reduce their fertility much more in response to the war.

As a robustness check for these heterogeneous effects by ethnicity, Appendix Table 8 explores alternative specifications. There, to explore an alternative treatment classification, column 1 assigns treatment based on Li et al. (2019), assigning 8 districts as high-conflict (with more than 95% of all conflict events being in these districts and a larger presence of LTTE) and the rest as the low-conflict zone. Column 2 assigns treatment and control districts to top and bottom quartile districts (6 in each group) ranked by the mean level of conflict per capita per year. Lastly, columns 3 and 4 remove the top 5% and top 10% of conflict districts respectively to ensure that these are not driving the main results. Across these columns, the result remains qualitatively similar to the main estimates in Table 1 suggesting that there was a lesser fertility reduction for the Sri Lankan Tamils relative to the Sinhalese.

Appendix Table 9 describes additional results for a finer disaggregated effect across all 5 ethnic groups as a further robustness check. Columns 1-4 are similar to those in Appendix Table 8 described above. The estimates for Sinhalese and Sri Lankan Tamils are qualitatively and quantitatively similar to those in Table 1 and Appendix Table 8, suggesting that aggregating remaining ethnic groups into the ‘Other’ category does not bias the main estimates.

For completeness, Appendix Table 10 explores estimates with a continuous treatment intensity which allows variation in treatment intensity and timing across districts in column 1. A placebo check based on this continuous specification in column 2 confirms that changing the timing of treatment intensity to pre-war years produces a null effect of war on fertility. The estimates here are obtained by dropping observations during the war and as a placebo, imposing the continuous conflict intensity from the years 1992 to 2005 on the years pre-war (1967 to 1980). This null estimate using a placebo conflict intensity suggests that the main estimates are not caused by some spurious pre-conflict trends. However, there are several reasons why it is not favorable to consider the continuous treatment specification as the main specification.Footnote 12

As an extension of the difference-in-differences model, I examine the dynamic effects of war on fertility trends using the event-study design. I estimate the following specification to look at ethnicity-specific effectsFootnote 13:

$$\begin{array}{ll}{{br}}_{{gdt}}={{{\beta }}}_{0}+\mathop{\sum }\limits_{i=1967}^{1979}{{{{\beta }}}_{i}{treated}}_{d,i}+\mathop{\sum }\limits_{i=1992}^{2009}{{{{\beta }}}_{i}{treated}}_{d,i}\\\qquad\qquad+\,\left\{\mathop{\sum }\limits_{j=1967}^{1979}{{{{\beta }}}_{j}{treated}}_{d,j}+\mathop{\sum }\limits_{j=1992}^{2009}{{{{\beta }}}_{j}{treated}}_{d,j}\right\}x\,{{sri}\,{lankan}\,{tamil}}_{g}\\\qquad\qquad+\,\left\{\mathop{\sum }\limits_{k=1967}^{1979}{{{{\beta }}}_{k}{treated}}_{d,k}+\mathop{\sum }\limits_{k=1992}^{2009}{{{{\beta }}}_{k}{treated}}_{d,k}\right\}x\,{{other}}_{g}+{\theta }_{d}+{{{\delta }}}_{t}+{\gamma }_{g}+{\epsilon }_{{gdt}}\end{array}$$
(4)

Here, \({{treated}}_{d,i}\) is a dummy variable that takes a value of 1 high-conflict districts for year i and 0 otherwise. The years 1967-1979 form the lag years and the year 1980 is the base category and is therefore dropped from the specification. The years 1992 to 2009 comprise the lead years of this analysis. Sinhalese are the omitted base category, as before. \({{sri\; lankan\; tamil}}_{g}\) is a binary variable that takes the value 1 for Sri Lankan Tamil ethnic group and 0 otherwise. \({{other}}_{g}\) takes a value of 1 for the other ethnic groups. \({\gamma }_{g}\), \({\theta }_{d}\), and \({{\rm{\delta }}}_{t}\) represent the ethnicity, district and time fixed effects respectively.

The estimates from Eq. (4) are reported in Appendix Table 11. Column 2 provides the coefficients for the Sinhalese ethnic group over the years (\({{\beta }}_{i})\). Column 3 provides the heterogeneity estimates from the interaction term in Eq. (4) for the Sri Lankan Tamils (\({{\beta }}_{j}\)). This represents the differential between the fertility of the two ethnic groups. Adding column 2 and column 3 gives the total effect on Sri Lankan Tamils, which is listed in column 4 (\({{\beta }}_{i}+{{\beta }}_{j}\)). The estimates from columns 2 and 4 have been graphed in Figs. 2 and 3 to show the dynamic effect for the two groups.

Fig. 2
figure 2

Dynamic effect of conflict on crude birth rate of Sinhalese. Note: This figure plots the coefficient and 95% CI of the effect of conflict on the crude birth rate for the Sinhalese ethnic group

Full size image
Fig. 3
figure 3

Dynamic effect of conflict on crude birth rate of Sri Lankan Tamils. Note: This figure plots the coefficient and 95% CI of the effect of conflict on the crude birth rate for the Sri Lankan Tamil ethnic group

Full size image

For the Sinhalese, Fig. 2 depicts that the effect of conflict on the Sinhalese birth rate was negative. The coefficient remains negative through all war years, whereas pre-war years saw a fluctuation around 0. A close look at Fig. 3 suggests that for the Sri Lankan Tamils, fertility did not drop much in the war years. Therefore, comparing Figs. 2 and 3, it is evident that the reduction in fertility was of a larger magnitude for the Sinhalese. The estimates from column 3 of Appendix Table 11 further indicate that the difference in fertility behavior between the two groups during war years is statistically significant. These results are in line with the heterogeneity evidence from the difference-in-difference estimates of the two groups presented in Table 1. Therefore, the dynamic effects are largely consistent with the difference-in-difference estimates.

Overall, estimates are qualitatively robust across the definition of treatment assignment. Estimates of the interaction term coefficient for Sri Lankan Tamils across the specifications are significant and positive which implies Sri Lankan Tamils did have a significantly different behavioral response to the war, and most specifications suggest a lesser reduction in fertility compared to Sinhalese. These results are in line with the overall findings of Puvanarajan and De Silva (2001). While their study does not estimate the impact of civil war on fertility, they broadly trace trends in fertility by ethnicity in Sri Lanka and find that the Sinhalese witnessed a stronger reduction in fertility than Sri Lankan Tamils during the 1990s. They also find evidence of an increasing age at marriage in Sri Lanka.

To the extent that I am unable to control for migration, these results should be interpreted with some caution. Several sources highlight that the civil war had resulted in large emigration from Sri Lanka (Fazil, 2019; Ibrahim et al., 2021; Sen, 2020). Sriskandarajah (2002) reports that about 25% of the entire Sri Lankan Tamil population left the country. It is possible that those who left were the ones who had access to wealth and higher education which increased their possibility of finding refuge elsewhere. One might argue if the more educated/wealthy had higher fertility rates because they could afford to have kids, then their exit could artificially cause the fertility estimates of my analysis. For instance, Zimmerman and Carter (2003), Thiede et al. (2020) and Dehejia and Lleras-Muney (2004) find evidence that with access to greater resources, the educated and wealthy could be more able to afford to raise children in the presence of exogenous shocks like wars and famines. However, if 25% of the Sri Lankan Tamils left, because the number leaving is so large, there were likely people from across the entire income distribution who left and not just those on the top (which typically comprises 1% of the population). Further, there is also evidence in the literature that the more educated have a lower fertility rate (Schultz, 1994; Duflo et al., 2015; Jayachandran, 2017). So apriori, it is unclear as to what was the composition of the people who left during war and what it would mean in terms of bias in the estimates of this study.

6.2 Replacement effect

I test for the presence of a replacement effect at the ethnic group level as a mechanism for the heterogeneity findings. Previous studies often utilize survey data to measure this effect at the individual level where they trace the woman’s family history to see if she experienced any wartime deaths and altered her fertility in response to it. In this study, since I don’t have access to individual-level survey data, I measure the response of fertility at the ethnic group level to lagged conflict-specific mortality rate at the district level.Footnote 14,Footnote 15

I estimate the following equation:

$$\begin{array}{ll}{{br}}_{{gdt}}={\beta }_{0}+{\beta }_{1}{{treated}}_{d}\times {{post}}_{t}{+\beta }_{2}{{treated}}_{d}\times {{post}}_{t}\times {{dr}}_{d,t-1}\\ \qquad\quad\,\,+\,{\beta}^{{\prime}}_{g}\mathop{\sum }\limits_{g=2}^{n}{{ethnicity}}_{g}\times {{treated}}_{d}\times {{post}}_{t}\\\qquad\quad\,\,+\,{{\beta }^{{\prime\prime} }}_{g}\mathop{\sum }\limits_{g=2}^{n}{{ethnicity}}_{g}\times {{treated}}_{d}\times {{post}}_{t}\times {{dr}}_{d,t-1}\\ \qquad\quad\,\,+\,{Pairwise}\,{Interaction}\,{terms}+{\gamma }_{g}+{\theta }_{d}\\\qquad\quad\,\,+\,{{{\delta }}}_{t}+\,{\nu}{{dr}}_{d,t-1}+{{{\upalpha}}}_{{gt}}+{{{\upmu}}}_{{dg}}+{\epsilon }_{{gdt}}\end{array}$$
(5)

where br is the crude birth rate for ethnic group g in district d at time t. Likewise, dr is the conflict-specific death rate for ethnic group g in district d at time t-1. treated is a dummy variable that takes a value of 1 for high-conflict districts and 0 for others, post is a dummy that takes a value of 1 for years after 1983 that mark the beginning of the civil war. As before, I include the dummy variables of n-1 ethnic groups interacted with treated × post, omitting it for Sinhalese and making Sinhalese the reference or base category. Thus, as in the main specification, I include the dummy for Sri Lankan Tamils and for others category (n = 3). Pairwise interactions include the interaction of treated with ethnicity dummies, post with ethnicity dummies, treated with death rate and its lag, post with a lag of death rate, as well as ethnicity dummies with a lag of death rates. \({\gamma }_{g}\), \({\theta }_{d}\) and \({{\delta }}_{t}\) represent ethnic group, district, and time fixed effects respectively. I further control for district-ethnic group fixed effects, \({\mu }_{{dg}}\) and year-ethnic group fixed effects, \({{\alpha }}_{{gt}}\)Footnote 16.

There would be evidence of a replacement effect if an ethnic group’s fertility response was positive in high-conflict zones in response to the lagged conflict-specific death rate. Table 3 presents these estimates. The sign of the coefficient treated × post × death rate lag (i.e. β2) would reflect replacement effects or lack thereof for Sinhalese, which was the reference category ethnic group. Lagged death rates seemed to negatively impact fertility for Sinhalese in conflict zones which implied that with an increase in conflict-specific mortality, Sinhalese in high-conflict districts reduced fertility and thus points to the absence of any replacement mechanism. This could be possible due to various reasons such as the majority status of the ethnic group which allows it to reduce fertility without any immediate threat to its group’s political representation in the country. Other reasons for increased mortality resulting in lower fertility could include fear and uncertainty that manifest into psychological and physiological stress and trauma.

Table 3 Replacement effect mechanism: impact of conflict on fertility
Full size table

On the other hand, treated × post × Sri Lankan Tamil × death rate lag, which is the relevant coefficient for Sri Lankan Tamils, is positive (i.e. β′′2). It seems that there is a high responsiveness of Sri Lankan Tamils to the conflict-specific mortality around them. The replacement effect coefficients suggest that due to 1 additional conflict-related death per 1000 people in the district, Sri Lankan Tamils responded by increasing their fertility by 32.5 births per 1000 people while the Sinhalese responded by reducing their fertility by 6.496 births per 1000 people. This could reflect some pro-natal sentiments in the minority group to try and maintain their group size during the war, an effort to replace the loss of a loved one or it may even reflect the insurance effect wherein parents increase fertility as a means of securing insurance for their old age if they fear only some of their kids would survive. Thus, there is evidence to suggest that there was an ethnic-group level replacement effect in place for the Sri Lankan Tamils, but not for the Sinhalese.

While this mechanism is not exhaustive, it is still indicative of a replacement effect partially contributing to the differential in fertility trends. Apart from this mechanism, other potential explanations include an increased risk of intimate partner violence and assault in conflict areas coupled with lower usage of contraceptives by Sri Lankan Tamils relative to Sinhalese. GoSL estimates suggest that 2% of all births in 1991–1998 were illegitimate and could potentially account for sexual violence channel during the war, however, that information is not available by ethnic groups at the district level for further analysis.Footnote 17 For the most part provision of healthcare was taken up by the LTTE in war zones (Stokke, 2006; Mampilly, 2011). If they were biased in the provision, access to health care in conflict zones might be selectively more available to Sri Lankan Tamils than to Sinhalese since social welfare provision was one aspect of LTTE state-building. This could have further contributed to a lesser negative effect on Sri Lankan Tamil fertility.

7 Conclusion

This paper provides evidence of a negative relationship between fertility and war in Sri Lanka on average. The increased female age at marriage in high-conflict zones contributes to this lower fertility. An exploration of the ethnic heterogeneity unveils that in response to high-conflict exposure, the Sinhalese had a larger reduction in fertility than their Sri Lankan Tamil counterparts. This study further contributes to the literature by finding evidence of the presence of a replacement effect for the Sri Lankan Tamils. This could reflect pronatalist sentiments in the minority group to maintain their group size during the civil war, an effort to replace the loss of a loved one or it may even reflect the insurance effect wherein parents increase fertility as a means of insurance for their old age if they fear only some of their kids would survive. Future research should explore these in more detail with survey micro-data which often ask questions on desired fertility, and record female fertility history along with information on the loss of loved ones. Future research may also study the marriage market channel disaggregated by the ethnic group if data becomes available.

From a policy perspective, there is enough evidence in the literature (Li et al., 2019) to suggest that children born during such a crisis can often suffer from lower birth weight and lower height-for-age (HAZ) or weight-for-age (WAZ) scores under age 5. If births are not well spaced apart, it can negatively impact the health of the mother and the child. Lower health outcomes could potentially transform into lower human capital in adulthood. So reproductive health policies and accessibility of neonatal care need to be targeted towards high-conflict areas.

It is also key for future research to analyze fertility trends post-war. Jayasundara-Smits (2018) highlights the sense of vulnerability in the minority post-war. She finds that war areas report a dominance of bodily grave crimes relative to the non-war areas where material grave crimes were dominant. With a distinct pattern in the nature of post-war crime in the two regions, the minority may continue to feel threatened. If the replacement effect in Sri Lankan Tamils was in part due to a fear of diminished old-age insurance due to war deaths or due to a pronatalist sentiment to maintain their identity in Sri Lanka, then the differential nature of the post-war crimes in conflict-affected areas could sustain this fertility differential. This has far-reaching consequences in the long term. It can change the demographic composition of the country and thus has the potential to impact the possibility of a future war in an already volatile political environment.