Filling the Void: Evidence from Two Natural Disasters on the. Determinants of Taliban Support

Transcription

1 Filling the Void: Evidence from Two Natural Disasters on the Determinants of Taliban Support Federico Masera Hasin Yousaf This version: August 15th, 2015 Preliminary and incomplete Since the inception of war-on-terror there has been considerable increase in the Taliban insurgence and presence in Pakistan. In this paper we empirically analyze an important determinant of political support for Taliban: state capacity. We show that the state and the Taliban compete in the provision of some public goods. Shortfall in the state capacity may lead to crowding out of government support and the Taliban may gain popular support by filling this void. We study this mechanism by exploiting a flood in 2010 as a source of exogenous variation. We show that areas which were affected by the flood experienced higher increase in electoral support for Taliban. We provide evidence that this change is due to state capacity after the flood: using an instrumental variable approach we show how areas with higher gap between required and received rehabilitation funding observed higher increase in support for Taliban - a 10 percentage point increase in the funding gap leads to a 2 percentage points increase in the vote share of the Taliban parties. We finally show that alternate explanations such as increase in religiosity after natural disasters, anger against the incumbent for poor performance and political substitution among non-incumbent political groups cannot account for these results. Masera: Department of Economics, Universidad Carlos III de Madrid, Calle Madrid 126, 28903, Getafe, Spain ( fmasera@eco.uc3m.es); Yousaf: Department of Economics, Universidad Carlos III de Madrid, Calle Madrid 126, 28903, Getafe, Spain ( myousaf@eco.uc3m.es). We are indebted to Jesus Carro, Irma Clots-Figueras, Luis Corchon and Ignacio Ortuno-Ortin for providing advice and support at all the stages of the paper. We would like to thank Julio Caceres-Delpiano, Federico Curci, Natalia Fabra, Andres Garcia-Suaza, Angel Hernando, David Jaeger, Matilde Machado, Monica Martinez-Bravo, Christos Mavridis, Jaime Millan, Diego Moreno, Roger Myerson, Sebastian Panthoefer, Jan Stuhler and Ulrich Wagner, seminar participants at the Universidad Carlos III de Madrid, Third CEMFI-UC3M Micro Workshop, Jan Tinbergen Peace Science Conference, University of Warwick, International Conference on Terrorism and Organized Crime, and 16th Annual Conference of Public Economic Theory. 1

2 1 Introduction States in the developing countries, as in the developed countries, provide goods and services to address to the needs of its citizens. However, the state capacity to provide these goods is very limited in many developing countries 1. How does state capacity affect individual preferences? Do non-state actors in the developing world take advantage of the limited state capacity to gain popular support? These questions are difficult to answer because of several identification issues - without source of exogenous variation in the state capacity, causality is difficult to establish. In this paper we study the effects of state capacity in providing for its citizens on support for non-state actors using a natural disaster as source of exogenous variation in the state capacity. Specifically, we exploit the 2010 floods in Pakistan to study how the electoral support of extremist Islamic parties changes in the flood affected areas. The areas which were affected by the flood required post-disaster rehabilitation fund and support, while the unaffected areas did not require such aid. This leads to an exogenous increase in the demand for critical public goods for the individuals residing in the affected areas. The 2010 floods in Pakistan were the biggest floods in the history of Pakistan: more than one-fifth of the land was under water and more than 20 million people were affected. Many areas which were never ever flooded before were flooded. The government response to the floods was inadequate both due to mismanagement and limited humanitarian aid. The government initially underestimated the impact of the rainfall and flood, and thus delayed the rehabilitation efforts. Moreover, the humanitarian aid from the international community was slow to arrive which exacerbated the situation. Many affected areas had unmet demand for rehabilitation of more than 50 percent even three five months after the floods. We use data from the national and provincial electoral commissions and the flood extent 1 A recent policy evidence report by the IGC (Bandiera et al. (2014)) provides an excellent review of the literature on the state capacity. The report admits that the evidence on state capacity is very scarce and little is known about how to strengthen state capacity. 2

3 to assess the effect of flood on support for Taliban. We use a Difference-in-Difference (DiD) empirical strategy to identify the causal effect of exposure to the flood on support for the political parties sharing the Taliban ideology. We find that the areas exposed to the flood saw an increase in the support for political parties sharing Taliban ideology. In the baseline specification, the areas affected by the flood witnessed an increase of 3.6 percentage points increase in the vote share of such parties compared to the areas which were not affected by the flood. This is a big change: the increase corresponds to more than 50 percent increase in the electoral base of these parties. To highlight the importance of the proposed mechanism, we then use data on the percentage of funding which was unmet five months after the onset of the floods. We employ similar DiD approach to identify the effect of this funding gap on the support for political parties with Taliban ideology. In our context, given the extent of the flood and severe shortage of funds, the limited funds could not be supplied selectively. However, to alleviate endogeneity concerns, we instrument funding gap with the electoral district exposure to the flood, distance from the national and provincial capitals and the extent of damage from the flood. We show that the areas with higher funding gap had a higher increase in the vote share of Taliban parties. A 10 percentage point increase in the funding gap leads to a 2 percentage points increase in the vote share of the Taliban parties. We show that alternate explanations cannot account completely for the observed political results. One alternate explanation is of a normal model of political competition. If the incumbent party performs badly in some regions the party will be particularly punished in that region and non-incumbent parties may gain support specifically in those regions. We show that the results are not in line with this explanation. The incumbent party indeed lost political support after the floods but not significantly more in the flooded areas. The other non-incumbent parties, including the major political parties, saw an increase in their support but not particularly more in the flooded areas. The only nonincumbent parties that gained disproportionately more support in the flooded areas were the Taliban parties. The results seem to indicate that the only political party that gained significantly more support in the flood affected areas were the parties with connections 3

4 with the Taliban. This observed change may reflect increased popular support for such parties due to direct help from the Taliban during the floods, when the state failed to provide for their needs. However, an alternate explanation could be increase in religiosity following the floods. We provide evidence against this alternate explanation by exploiting the results from the 2002 and 2008 national elections and the earthquake of The government was widely praised for the good management of this disaster which also received a lot of international aid. We show is that the areas affected by the earthquake saw a significant decrease in the support for the Taliban. This result embodies the other side of our proposed mechanism. In this case the state capacity was adequate to meet the demands of the citizens and this crowed-out the support for Taliban in these regions. Additionally we show how our results are robust to controlling for several other potential confounding variables. First, the effect of flood on the individuals of that area can depend on the composition of the population and the level of development and infrastructure of that area. For instance, areas with less educated population and lower access to water, health and other facilities may be more likely to support Taliban. We control for the demographic and developmental indicators for the electoral areas. Second, there are time invariant electoral area specific factors, such as distance from the river, the suitability of the land for agriculture, ruggedness of the land, e.t.c., which may impact support for Taliban and the probability of flooding. We control for electoral region fixed effects to explicitly account for all such factors. Third, the flood may not be entirely random for the areas which are frequently flooded. The flood only hits as a shock to the areas which are rarely flooded. Thus, we control for the ex-ante risk of flooding using the flooding history of the electoral district. Moreover, in the most conservative specifications, we only focus on the regions which had low ex-ante risk of flooding, and replicate our results. The results are robust to battery of robustness checks and falsification tests. First, we demonstrate that the result hold for a finer geographic area. That is, we replicate our results for the support of Taliban ideology parties in the provincial assembly, which is a much smaller geographic area than the national assembly electoral district. Second, 4

5 we show that the results are not driven by inclusion of certain electoral districts. We drop the electoral regions which share the border with Afghanistan, and show that our results are virtually unchanged. We do similar exercise by excluding the regions which have very high margin of victory for the Taliban ideology party in 2008 elections, and find similar results. Third, if there are unobserved electoral districts specific time trends which foil our estimation to be causal, then assigning pseudo treatment to the electoral regions should generate an effect. We carry out falsification test by randomly assigning whether a district was affected by the flood and estimating our main specification several times. The estimated effect from this exercise is indistinguishable from zero. Finally, we provide evidence in favor of our identification assumption. The identification assumption employed for the causal estimation is the parallel trends assumption. That is, in the absence of the flood, the areas which were unaffected by the flood would have evolved in the same way as the areas which were exposed to the flood. Since, we have two pre-flood periods, we directly test this assumption and show that the parallel trends assumption is indeed satisfied. That is, the areas which were affected by the flood in 2010 had similar change in the support for the Taliban between 2002 and 2008, compared to the areas which were not flooded in This finding strengthens that what we find is indeed a causal effect of flood on the support for Taliban. This paper consists of 11 sections. In the next section, we provide brief overview of the literature on natural disasters, civil war and how individuals react post-disaster. In section 3, we present the detailed background of the context. Specifically, we capsulize the electoral system of Pakistan, the onset of the flood of 2010, and the failure of the government in relief efforts. Section 4 provides the data sources. In the fifth section, we sketch our theoretical framework and outline our empirical methodology. In section 6, we provide results from the baseline specification and specifications which control for numerous other variables. In section 7, we develop the empirical methodology for our mechanism, and also show the results. In section 8, we test for alternate mechanisms which can explain the rise in the support for Taliban. In section 9, we show heterogeneity 5

6 of our results. We show how the results vary depending on the intensity of the flood, the distance from the Afghanistan, and the margin of victory or loss in the previous elections. Section 12 provides battery of robustness checks and falsification tests to demonstrate the results are well-founded. Section 11 provides a discussion of the results and concludes the paper. 2 Previous Literature This paper is closely related to the literature that studies the causes of civil conflict, war and terrorism (for a review of the literature refer to Blattman and Miguel (2010). These causes may be ethnic differences (Esteban and Ray, 2011; Esteban et al., 2012), extreme climate conditions (for a review Hsiang et al. (2013)), political instability (Fearon and Laitin, 2003), rise in international price shocks (Besley and Persson, 2008; Dube and Vargas, 2013) among many others. More closely related to our paper there have been some recent studies exploring the effect of foreign aid on conflict and violence. For example Berman et al. (2011b) study a model of competition between a government providing reconstruction program and violent rebels. They then test the model using panel data from Iraq and find that reconstruction spending reduces insurgent violence. Other papers have tried to identify, causally, how different forms of international aid affect conflict. Examples of this are Beath et al. (2012) that use a randomized controlled trial, Crost et al. (2014) using a regression discontinuity approach while Nunn and Qian (2014) uses an instrumental variable approach. They all have identified that development aid has either no effect or a detrimental effect on civil conflict. This reduced form approach to studying this problem has brought very interesting results but one existing gap in this literature comes from the fact that the connection between foreign aid and violence is still a black box and is not clear which is the mechanism that links both of them. What we do in this paper is testing one of these possible mechanisms by studying whether support to local terrorist groups is affected 6

7 by foreign aid. As suggested by Crost et al. (2014) then the increase of conflict could be a violent response to a decrease in support of insurgents caused by development aid. To our knowledge we are the first to empirically study this mechanism and we do so by exploiting the unique political situation of Pakistan where is possible to directly observe the support for the Taliban terrorist group. Additionally, we contribute to the literature that tries to understand the support for terrorist or rebel groups. A very detailed review is provide by de Mesquita (2008). Some more recent works that have dealt with this issue are Jaeger et al. (2012) that shows how radicalization of the Palestinian population is influenced in the short-run by Israeli violence while in the long run by major political events like the Oslo negotiations or the first Intifada. Berman et al. (2011a) find no evidence of the opportunity-cost theory that states that only individuals with a low opportunity-cost (unemployed, poor) will rebel more and use violence. Similarly Blair et al. (2013) find that the income has no link to personal support for militant and terrorist organizations. In a study in Iraq instead Iyengar et al. (2011) find a positive correlation at the district level between spending in labor-creating projects by the US military and violence reduction. In this paper we try to causally identify a mechanism similar in nature to Berman and Laitin (2008) that see the support terrorist groups as a way of receiving local public goods when neither the government or the markets can deliver it. Finally our paper also contributes to the literature that studies the effects of natural disasters and aid relief. A comprehensive discussion on the topic can be found in Stromberg (2007). For example it has been shown that international aid delivery (or the lack of) may have economic and political consequences (Alesina and Dollar, 2000). In particular, Drury et al. (2005) show that large disasters if not handled properly by the international community may destabilize local governments. More closely related to the effects of natural disasters on terrorism Berrebi and Ostwald (2011) show in a cross-country comparison that natural disasters are positively associated with terrorist attacks. Looking specifically at Pakistan, Fair et al. (2013) show how people more harshly affected by the 2010 flood 7

8 in Pakistan increased their turnout to elections, had more political knowledge and they were more demanding towards the government. Andrabi and Das (2010) show a positive effects of the 2005 Pakistan earthquake on trust towards foreigners cause by a prompt delivery of foreign aid. What we do here is study the potential use of international aid relief against terrorism and radicalization of the population. 3 Context In this section we briefly discuss the context of our setting. Specifically, we give an overview of the political system of Pakistan, giving summary about the elections, major political parties and the Islamic parties. Then, we present summary of the flood of 2010, followed by the overview of the relations between U.S. and Pakistan during the flood. 3.1 Political System of Pakistan The Figure 1 shows the electoral districts of Pakistan along with the geographic details of its location. Pakistan shares its borders with Afghanistan, China, India, and Iran. The provinces of Khyber Pakhtunkhwa and Balochistan share the border with Afghanistan, while the provinces of Punjab and Sindh share the border with India. Islamabad is the capital city of Pakistan. Administratively, it is also equivalent to a province known as the Federal Capital. Lahore and Karachi are the two main urban areas of Pakistan with population of 15 million and 25 million respectively. The governing structure of Pakistan is a parliamentary system. The parliament of Pakistan has bicameral structure composed of the Senate and the National Assembly (NA). The National Assembly has 342 members, of which 272 members are elected through voting 2. The tenure of the NA is 5 years, and the elections are held after every 5 years. 2 Out of the 70 non-elected members, 60 seats are reserved for the women and 10 seats for the minorities. These seats are elected through an indirect proportional representation list system, whereby political parties submit their lists of women candidates for reserved seats to the Election Commission 8

9 In our analysis, the elections were held in October 2002, February 2008 and May 2013 respectively. The 272 members of the NA are elected through the general elections held in 272 National Assembly constituencies. Out of the 272 electoral districts, 148 are in Punjab, 61 are in Sindh, 35 are in Khyber Pakhtunkhwa, 14 are in Balochistan, 12 are in the Federally Administered Tribal area (FATA), and 2 are in the Federal Capital, Islamabad 3. The electoral districts are very heterogeneous in terms of area, but have the same population in each electoral district 4. The voting structure is first-pass-the-post system. Each candidate can belong to at most a single political party or decide to run independently without any political affiliation. Historically, the two biggest parties in the political system of Pakistan are the Pakistan Muslim League Nawaz Sharif, PML (N) and the Pakistan s People Party (PPP). The PML (N) has strong base and support in the province of Punjab, while the PPP has its strength in the province of Sindh. Since, 1989, the PML (N) and PPP has had major representation in three governments. PPP secured the most number of votes in the 2008 elections, while the PML (N) secured the most number of votes in the 2013 elections. Apart from the two main parties, there are several other political parties. Muttahida Qaumi Movement (MQM) and Pakistan Tehreek-e-Insaf (PTI) are among the other political parties. Apart from these four major political parties, there are several political parties with rightist and Islamic ideology. The three major Islamic parties include: Jamiat-e-Ulemae-Islam (JUI-F), Jamiat Ulema-e-Pakistan (JUP) and Jamaat-e-Islami Pakistan (JI). prior to the election (Constitution, Article 51). 3 The administrative division of Pakistan has several layers. The most coarse layer is the provinces. There are 6 provinces: Punjab, Sindh, Khyber Pakhtunkhwa, Balochistan, FATA and the Federal Capital. The Gilgit Baltistan and Azad Jammu & Kashmir are also divisions within Pakistan. Thus, Gilgit Baltistan and Azad Jammu & Kashmir are not represented in the National Assembly and Senate of Pakistan. However, they have an autonomous government which is part of Pakistan. The second layer is of the districts. There are roughly 30 districts in each province. 4 For instance, the Karachi city has 20 electoral districts, where as the whole province of Balochistan has only 14 electoral districts. 9

10 Jamiat-e-Ahle Hadith, and Pakistan Isami Tehrik (ITP) (formerly Tehriq-e-Jafaria (TeJ)) are among the other two Islamic political parties. In 2002, these five parties had a political alliance, Muttahida Majlis-e-Amal (MMA). The alliance was formed as a result of direct opposition to the policies led by President Pervez Musharraf to support for the United States war in Afghanistan (Adel et al., 2012) Islamic Parties and connections with Taliban Since the inception of war-on-terror, the Jamiat-e-Ulema-e-Islam (JUI-F), Jamiat Ulemae-Pakistan (JUP) and Jamaat-e-Islami Pakistan (JI) have voiced their disapproval of the Pakistan s support to the United State equivocally. Jamiat Ulema-e-Islam has its roots in the anti-colonial movement and pro independent Muslim state movement back in the 1920s. It has strong presence in the Khyber Pakhtunkhwa and Balochistan. Jamaat-e- Islami Pakistan also traces back its roots before the independence of Pakistan. All of these parties have presence in the provinces of Khyber Pakhtunkhwa and Balochistan, which share the border with Afghanistan. The three parties merged together in 2002 as a result of common opinion and to provide strong opposition to the President Musharraf s unconditional support to the United States for the war-on-terror (Norell, 2007). This pro-taliban stance was very popular among the province of Khyber Pakhtunkhwa, which shared the same ethnicity with the Afghanistan, in which the party gained majority. The MMA was able to form a coalition government in the province of Balochistan. All these parties emphasis on strict Islamic morals and principles in every day life. These parties have strong connections with the Afghan Taliban and share the same ideology: All the individual parties in the MMA have links to militant groups, hence 5 However, the political alliance was broken just before the 2008 elections. The JI wanted to boycott the elections, while the other two parties wanted to run for the elections. This difference of opinion led to the break-down of the alliance into the three political parties as they were before. 10

11 this coalition is of great interest when examining Pakistani links to the Taliban. (Norell, 2007, pg. 69) The Muttahida Majlis-e-Amal (MMA), which President Musharraf made the official opposition party in the Pakistan Parliament after the 2002 elections, provides political cover to the extremists, openly supports the Taliban as the legitimate government of Afghanistan, and maintains close ties to its leadership (Johnson and Mason, 2008, pg ) Norell (2007) presents detailed historical analysis of the relations between the Taliban and the MMA. Khyber Pakhtunkhwa, Balochistan and FATA are the areas sharing border with Afghanistan. The major political parties in these provinces are also the parties which are pro-taliban. Not only do the MMA and Taliban have connections, but they also trade benefits with each other. For instance, When the Taliban came to power in Afghanistan in 1996 it enjoyed support from the JUI-F, which in turn gained popular support from Pashtuns living in NWFP, Baluchistan and the Federally Administered Tribal Areas (FATA) (Norell, 2007, pg. 70) The MMA has vociferously objected to Pakistani and U.S. operations in the tribal areas to ferret out Al Qaeda and Taliban fugitivesdespite knowledge that such persons are lodged in these areas. Continued successful political evolution of the MMA could have deleterious consequences for the U.S.-led war on terrorism through encouraging militant recruitment, erecting operational barriers to Pakistani and U.S. forces seeking access to the tribal areas, and continuing to provide safe havens to Al Qaeda and Taliban fugitives. (Fair, 2004, pg. 498) Despite these connections, do the individuals associate MMA with the Taliban? More relevant question for our analysis is: do the individuals who vote for MMA are actually 11

12 voting for Taliban? There are almost no studies which try to understand the demand for the militant groups 6. The only paper we are aware of that tries to analyze the relation is Shapiro and Fair (2009). They found some evidence to the question above i.e. the support for the militants is correlated with: a desire for change-positive or negative- in the perceived role of Sharia in Pakistan (Shapiro and Fair, 2009, pg. 116), and political grievances. None of the other conventional factors: poverty, education, religiosity and desire for stronger democracy were correlated with the support for militant groups. They do not correlate the actual voting behavior of the individuals, but rather rely on selfreported measures. 3.3 The natural disaster: Flood 2010 The 2010 floods in Pakistan have been called the greatest humanitarian crisis in recent history by the United Nations, with more people affected than were affected by the South- East Asian tsunami and the 2010 earthquake in Haiti combined Ferris (2011). In terms of economic damages, the flood of 2010 caused an estimated damage of $ 19 billion. The flood was caused by abnormal monsoon rains in late July, 2010 which resulted in floods across all the provinces of Pakistan. Approximately, one-fifth of the Pakistan was flooded in terms of the area flooded and around 20 million people were affected due to the floods. It was the worst flood in the history of Pakistan, causing three times more people affected than the second worst flood which hit the country in In terms of the economic costs, it is the worst natural disaster to hit Pakistan ever. It caused more than three times the economic damage compared to the second worst natural disaster of Pakistan: the earthquake of 2005 (Pakistan, 2011). The fig. 2 shows the extent of the flood in September As it is apparent from the figure, the flood were widespread all over Pakistan. The areas from all the major provinces were affected. Areas around the river Indus were severely affected, while rest of usual. 6 Almost all of them, exclusively, focus on the supply of militant groups and how it affects business as 12

13 the areas were moderately affected. Most of the eastern part of Punjab and south-western part of Balochistan was unaffected by the flood. Khyber Pakhtunkhwa was the earliest to get affected. On 20th July, Peshawar, the capital city of Khyber Pakhtunkhwa received heavy rain-fall. Soon, the rest of the province, in series of rain-falls, was affected. The flood in the mid-augusts, reached the southern part of the country and affected southern Punjab, northern Sindh and Balochistan. The direct damages to the infrastructure was estimated to be around $11 billion, while the cost of re-constructing the infrastructure was estimated to be around $ 8 billion. The flood caused large-scale damage not only to the houses and infrastructure of the area, but also resulted in wide-scale agricultural damages. More than 700,000 acres (3,000 km2) of cotton, 200,000 acres (800 km2) acres each of rice and cane, 500,000 tonnes of wheat and 300,000 acres (1,000 km2) of animal fodder were destroyed by the flood (Pakistan, 2011). The government response to the flood was very poor. The NDMA was completely illequipped and unprepared to deal with a natural disaster of such an extent. The NDMA was in complete disarray (Ahmed, 2013). At first, the government mis-calculated the gravity of the situation. The first heavy rain of the monsoon which lead to flooding in the Khyber Pakhtunkhwa took place on 20th July However, the Flash Appeal for the relief and early recovery was not released until 20 days after the flood i.e. 9th August 2010 by that time most of the Pakistani land was under water. The poor rehabilitation and relief efforts was a result of both government incompetence and lack of international support. The government initially underestimated the effect of the floods and did not act promptly. The President of Pakistan at that time, Asif Ali Zardari, continued his trip to the United Kingdom despite the gravity of the situation. As of August 9, the international governments had committed less than $45 million. As a comparison, in the first ten days after the earthquake in 2005, the international governments had committed $247 million 7. The poor response by the government raised concerns about the stability 7 Other natural disasters around the world received intermediate amount of funding. For instance, in 13

14 of Pakistan and its strength to counter terrorism (Kronstadt, 2010). The poor rehabilitation was visible among the disaster affected regions. For instance, Doocy et al. (2013) surveyed households in the affected areas six months after the 2010 flood 8 and showed that the need for flood aid was uniformly present in all the affected areas. But, only half of the affected areas reported receiving food aid ever 9. 4 Data Sources We use data sets from several different sources. The political outcome data set is constructed using the official election outcomes from the Election Commission of Pakistan (ECP). The electoral outcome in each electoral district is collected from the ECP and recorded. For each district, we have information on the number of votes won by each candidate and his/her political affiliation. The candidates can choose to run as independent candidates i.e. without any political party affiliation. In the 2002 elections, the Muttahida Majlis-e-Amal (MMA) participated as a single political party composed of the coalition of the following five parties: Jamiat-e-Ulema-e-Islam (JUI), Jamiat Ulemae-Pakistan (JUP), Jamaat-e-Islami Pakistan (JI), Jamiat-e-Ahle Hadith, and Pakistan Islami Tehrik (ITP). The number of votes secured by the MMA over the total votes casted is used to determine the proportion of votes secured by the MMA in an electoral district. the first ten days of Cyclone Nargis, which hit Myanmar in May 2008, the international governments committed $110 million. Similarly, the earthquake in January 2010 in Haiti also saw a commitment of $742 million during the first ten days of the disaster. 8 They reported that 95% of the households reported damage to their house, and 82% reported permanent damage to the house. Moreover, 85% of the households were displaced for more than two weeks. 9 While half of the respondents cited as receiving aid in the first three months after the flood, only one-third of the respondents cited the same in the subsequent three months. More than 60% of the food needs were unmet across the flood affected areas six months after the flood. Farmers and daily-wage laborers were among the worst hit by the flood. They found that the common targeting practices were not kept in mind when distributing the food aid. That is, they showed that larger households, female headed households and internally displacement households were not significantly more likely to receive food aid compared to other households. The urban households were more likely to receive food aid despite the fact that rural households were the ones worst affected, both in terms of magnitude of people affected and magnitude of the damage. 14

15 The candidate from the MMA appeared in 171 out of 272 electoral districts in Their representation is spread over all the provinces. The MMA is widely represented in the provinces of Balochistan and Khyber Pakhtunkhwa. They appeared in 13 out of the 14 electoral districts in Balochistan, and 32 out of the 35 electoral districts in Khyber Pakhtunkhwa. In the 150 electoral seats of Punjab and 61 electoral seats of Sindh, the MMA appeared in 90 and 36 electoral districts respectively. However, in the FATA the MMA did not contest from any electoral district 10. Before the 2008 elections, the alliance of the five parties was broken due to internal power struggle among the parties. This clash also resulted in split of Jamiat-e-Ulema-e- Islam (JUI) into Jamiat-e-Ulema-e-Islam, Fazl-ur-Rehman (JUI-F) and Jamiat-e-Ulemae-Islam, Samiul Haq (JUI-S). In order to associate the votes secured by the MMA in 2008 (and subsequent election, i.e. 2013), the votes are aggregated for these six political parties. There are electoral districts in which some of these six parties do not have a candidate representation. There are electoral districts in which there is no candidate from any of the six parties, while there are districts with candidate from all the six parties. We compute the share of the MMA as the share of the political parties that present in that electoral district. We compute this share if two of the three major parties of the previous alliance MMA participate from that particular district. That is, if in any district less than the two major parties participate, we label it as if the MMA did not participate in that particular electoral district 11. The data on the flood of 2010 is gathered from two different sources: United Nation Office for the Coordination of Humanitarian Affairs (OCHA) and the National Disaster Management Authority (NDMA). We take the details which areas were affected by the 10 This is not that surprising. The politics in FATA are more localized than other provinces. Usually, the representatives from FATA run without any political affiliation i.e. as independent candidates. 11 We tried with other definitions as well. For instance, labeling MMA as not participate only if all three major parties not participate from an electoral district; labeling MMA as not participate if two of the major parties do not participate in the electoral districts. The results with these different definitions of the share of MMA are presented in the Robustness section. All the different definitions of the vote share of MMA lead to the same qualitative result, and our consistent with our mechanism; the magnitudes are adjusted accordingly. 15

16 flood from the OCHA. OCHA reports the areas which were affected by the flood at subdistrict level. The data on funding gap is taken from the NDMA. NDMA reported the funding gap present in the flood affected and other areas which were indirectly affected by the flood (for instance, due to rehabilitation in the neighboring electoral district). They stored this information in form of a map representing the funding gap faced by an area in form of categorical variable. That is, the funding gap in a categorical variable with following categories: 0 20%, 20 40%, 40 60%, 60 80% and %. 5 Empirical Methodology In this section, we provide brief overview of the theoretical framework which forms the basis for our mechanism. Then, we provide overview of the empirical methodology employed to test our framework and provide causal effect of the flood on support for Taliban. The paper proposes lack of state mechanism to explain the differences in the level of support observed for the pro-islamic and pro-taliban party in the wake of the response to the natural disaster. Specifically, the paper argues that the government and nonstate actors compete in the developing countries for the political support. When the government is able to provide for the needs of the citizens it crowds out the non-state actors. On the other hand, government being unable to provide for these needs results in higher support for the non-state actors and crowding out of the government support. 5.1 Empirical Methodology To test our mechanism, we first explore the effect of natural disasters on the support for the MMA. We use the geographic heterogeneity of the effect of the flood along with a difference in difference methodology to analyze if the flood had any effect on the political outcomes. Specifically, we estimate, MMA it = α i + λ t + β(affected i d2013 t ) + X itγ + u it, (1) 16

17 where MMA it is the proportion of votes secured by the Muttahida Majlis-e-Amal in electoral district, i, at elections, t 12. The variable affected it is the variable denoting whether electoral district, i, was affected or not. X i are district level indicators of the average level of literacy in the district, percentage of people involved in agriculture, average household size, access to clean water and electricity, percentage of children immunized under 5 years of age, and percentage of urban area in the district. In addition to these demographic and economic variables, the control variables also include the distance from the Afghanistan, and the number of political parties in the election in that year. In order to avoid effect of flood on these control variables, we use pre-flood value for these controls and interact them with the binary indicator for the year The time trend common to all the states is captured by the term λ t. α i are the electoral districts fixed effects 13. The methodology has several advantages. First, it controls for the pre-existing differences among the electoral districts through rich set of controls or electoral district fixed effects. Thus, any difference between the affected area and unaffected area only stems from the fact that the area is affected, rather than pre-existing differences. Second, the specification allows for difference in the result of election result for the MMA between the 2008 and 2013 elections through the term, λ t. In order to obtain consistent average treatment effect of the flood on the proportion of votes secured by the MMA in the 2013 elections, the treatment should be independent of the unobserved error term 14. That is, the areas which were flooded compared to the unflooded areas should not be systematically different across the unobservables. Formally, E(u it affected i = 1, X it, α i ) = E(u it affected i = 0, X it, α i ). This expression might not hold in our case. For instance, due to poor inundation system, heavy monsoon rains and melting of snow in the northern mountains in the summer the areas around the river 12 Note that t = 2002, 2008, or Instead of using time invariant control variables, the specification also uses fixed effect specification to capture the effect of unobserved time invariant factors too. 14 For simplicity of the argument, consider that the flood has similar effect on all the flood affected areas. The argument presented below is valid for the heterogeneous effects of flood with minor changes. 17

18 Indus are flooded very frequently. In fact, some of the areas are flooded almost every year. Knowing that the areas around the flood are frequently flooded, people choosing to live around the river may be systematically different from the people choosing to live further away from the river. This would violate the above expression and generate inconsistent average treatment effects. In order to control for this potential factor that can lead to inconsistent estimates, we augment our specification above with the frequency of flood risk for each given electoral district. Specifically, we estimate the following specification: MMA it = α i + λ t + β 1 affected i d2013 t + β 2 frequency i d2013 t + β 3 affected i frequency i d2013 t + X itγ + u it, (2) where frequency i denotes a binary indicator for the frequently affected areas. The estimate of β 1 from the equation above would yield estimates which control for the propensity of flood in a given electoral district. That is, the above specification allows for the flood to have differential impact on the areas which are more likely (and are frequently) affected and areas which are usually not flooded by the flood, but were flooded in the 2010 flood. For these infrequently affected areas, the no selection bias expression, E(u it affected i = 1, frequency i = 0, X it, α i ) = E(u it affected i = 0, frequency i = 0, X it, α i ), is more likely to be satisfied. This is because the informal flood prevention mechanisms present in the frequently affected areas are most likely to be absent in the infrequently affected areas. Moreover, the infrequently affected areas are more likely to be surprised by the flood, and lack flood rehabilitation experience, which might make them more vulnerable to the flood. Since, the flood of 2010 took place on a massive scale which flooded almost one-third of Pakistan, there is enough variation in the term, affected it frequency i. Another thing to keep in mind is that we use fixed effects rather than lagged dependent variable in our regressions. That is, we assume that once controlling for fixed effects, the conditional independence assumption is satisfied. Following Angrist and Pischke 18

19 (BOOK), we know that if the true assumption required for conditional independence is the lagged dependent variable, the fixed effect will be too large. However, we notice that the estimates from the lagged dependent variable (not reported here) are larger than the estimates obtained from the fixed effects. Using lagged dependent variable, MMA increased their votes, on average, by 6 percentage points from 2008 to 2013 in all the electoral districts. In the areas affected by the flood, they increased the vote share by an additional 4.5 percentage points. 6 Baseline Results In this section, we discuss the results of flood on two outcomes: the vote share of the MMA in the electoral district. The results from the estimation of Equation 1 are shown in Table 1 below. Column 1 shows the unconditional effect of flood on the proportion of votes secured by MMA. The estimates allow for correlation of residuals between different observations from the same electoral seat, but not across electoral seats. That is, we cluster the standard errors at the electoral seat level 15. MMA increased their votes, on average, by 3.4% from 2008 to 2013 in all the electoral districts. In the areas affected by the flood, they increased the vote share by an additional 3.6%. The magnitude of the effect is very large as the MMA, on average, had 2% of the votes in an electoral district in The remaining estimates control for rich set of factors that might confound our estimates. In Column 2, since MMA are more likely to gain higher proportion of votes in areas closer to the Afghanistan border, we control for the distance from Afghanistan (in 100 km). The dynamics of voting might be slightly different in an electoral seat located in the provincial or national capital city, which is accounted for in the Column 3. Column 4 15 In all the estimates we cluster the standard errors at the electoral seat level. However, there might be correlation of residuals across the seats located in the same city or in the same district. We clustered the standard errors at city level and in another specification at the district level. The standard errors virtually remain unchanged, which hints that this correlation might be low. 19

20 controls for rich set of pre-existing socio-economic and demographic factors, measured at the district level adopted from the 1998 census, such as male and female literacy, percentage of households working in agricultural sector, percentage of households having access to electricity and piped water, percentage of children immunized, population density of the area and the percentage of urban area. Higher political competition may make matter in order to get more votes from a particular constituency, which is accounted for in Column 5 by controlling for the number of political parties contesting from a particular electoral seat. Finally, there can be other factors that might correlate with the support for MMA and being affected by the flood such as ruggedness of the land, suitability of the land for farming, e.t.c. Column 6 employes electoral seat fixed effects to control for all such time invariant factors. As shown in the table 1, the results are virtually unchanged in all of the above specifications. The results do not change the magnitude a lot and remain highly statistically significant. As discussed above, in order to find causal effect of being affected by the flood, the flood affected areas should not be systematically different from the unaffected areas and the flood should be completely random. Equation 2 takes this into account and controls for the ex-ante propensity of flooding for an electoral area. The ex-ante propensity of flooding for an electoral area is calculated using US Satellite Observatory Data. The data records the incidence of flood for a 10 km by 10 km area in the last 50 years. This information is used to construct the propensity of flood. The propensity of flood for an electoral seat is then a continuous variable. In the table 2, average number of times an electoral area is affected by the flood in the last 50 years is used 16 Table 2 shows the results. Column 1 shows the unconditional estimates, while Column 2 shows the estimates with all the controls. Column 3 controls for time-invariant electoral seat specific factors. The parameter of interest is the coefficient on the interaction between 2013 time dummy and af f ected dummy. This parameter captures the change in vote 16 The results with minimum and maximum number of times an electoral area is affected by the flood are similar. 20

21 share of MMA from 2008 to 2013 in the areas that had very low ex-ante probability of being flooded but were affected by the flood compared to the areas that also had very low ex-ante probability (no flood in the last fifty years) of being and were unaffected by the floods. The results show that the vote share of MMA increased by 4.4% in areas with low ex-ante probability of flooding but which were affected by the flood compared to the areas with low ex-ante probability of flooding and which were unaffected by the flood. The effect of being affected by the flood is decreasing in the ex-ante probability of flooding. That is, with one more year of flooding in the past fifty years, the effect decreases by 0.026%. The effect is significant for the areas for which the number of flooding in the last fifty years is less than Around 20% of electoral regions have more than 32 floods in the last fifty years. This finding further strengthens are argument; the effect stems from the areas which had very low ex-ante probability of flooding. It is highly plausible that for these areas the flooding was random and unanticipated. The areas with low ex-ante probability of flooding may experience increase support for Taliban, while the areas with high ex-ante probability of flooding did not experience a change in support, due to several reasons. First, it could be that the areas with high exante probability of flooding may already have in place informal mechanism and disaster management funds allocated in case the area is hit by a flood. This would substantially limit the help that Taliban can provide to gain more support. Second, the areas with high ex-ante probability of flooding are more likely to have pre-allotted governmental funds in their favor due to high probability of risk of flooding. 7 Mechanism: Funding Gap In order to shed more light on the mechanism underpinning these changes, we utilize the funding gap data. The National Disaster Management Authority (NDMA) reported the 17 The t-test for the hypothesis: β β 3 = 0 is rejected at 10% significance level, whereas, we fail to reject the null hypothesis: β β 3 = 0 at 10% significance level. 21

22 funding gap present in the flood affected and other areas which were indirectly affected by the flood (for instance, due to rehabilitation in the neighboring electoral district). Funding gap is the difference between the reported funding required after the flood and the actual funding received in the five months following the flood. The UNOCHA stored this information in form of a map representing the funding gap faced by an area in form of categorical variable. That is, the funding gap in a categorical variable with following categories: 0 20%, 20 40%, 40 60%, 60 80% and %. We use this information to generate funding gap corresponding to an electoral district Vote Share Following the previous sections, we estimate similar difference-in-difference specification: MMA it = α i + λ t + β(f undinggap i d2013 t ) + u it, (3) where F undinggap i represents the funding gap in district i. Funding gap is re-scaled between zero and one. One potential problem with the specification above is that funding gap may not be exogenous. The funding gap as defined by the difference between funding required in the electoral district and the funding supplied. Being affected by the flood provides an exogenous increase in the funding required for the flood. The intensity of the flood, measured by moderately v.s. severely affected, effects the funding required by the district. These two variables are exogenous 19. The funding efficiency and capacity of the government can be proxied by the distance of that electoral district from the national and provincial capitals. Thus, we instrument funding gap with whether the area was affected, whether the area was severely affected and the distance from the national and provincial capital. 18 The funding gap information on the map is finer than the electoral district level. This helps us construct an estimate of the funding gap at the electoral district level. 19 This is specially true if we concentrate only on areas with low propensity of flooding. 22

23 (TALK ABOUT VALIDITY OF THE INSTRUMENT) Table 3 shows the results from both OLS and IV estimates. Column 1 reports unconditional estimates, Column 2 reports estimates with controls and Column 3 reports estimates with fixed effect. Column 4 to 6 report the unconditional, specification with controls and fixed effects for IV estimates. The results show that the funding gap is highly positively significant: an 10% increase in the funding gap is associated with 0.71% increase in the vote share of the MMA. The IV estimates are larger than the OLS estimates. IV estimates reveal that an 10% increase in the funding gap is associated with 1% increase in the vote share of the MMA. (TALK ABOUT ECONOMIC SIGNIFICANCE OF THE RESULTS) In order to identify the areas which are driving the results, we estimate the equation above for different sub-samples. Table 4 shows the results from the estimation for different areas. The results indicate that the effect is not driven from a particular sub-sample, but rather is present in all the different regions. We interpret the results for the areas with low ex-ante propensity of flooding. Column 2 shows that an 10% increase in the funding gap is associated with 1.01% increase in the vote share of the MMA in the moderately affected areas, while the 10% increase in the funding gap is associated with 1.07% increase in the vote share of the MMA in the severely affected areas. (TEST IF THE RESULTS ARE STATISTICALLY DIFFERENT FROM EACH OTHER) We then estimate the non-parametric relation between the change in MMA vote share and the funding gap. This is very similar to the DiD estimation. Similar to the DiD, we are comparing high funding gap areas to the low funding gap areas and analyzing how much do the MMA gain in the high funding gap areas relative to the low funding gap areas. However, we are not imposing a linear relation between funding gap and change in MMA vote share. Instead, we are calculating the effect on MMA vote share for each local value of the funding gap. Non-parametric estimation has several advantages over the parametric one. The estimation does not impose any functional form on the relation as in parametric estimation. Instead, it fits the best polynomial which explains the relation. Moreover, it finds relation at every point in the distribution of the dependent variable i.e. 23

24 local regressions throughout the distribution. This is more informative than the average effects. The Figure 6 plots the non-parametric relation between change in MMA vote share and the funding gap along with the 95%confidence interval. The figure shows that the relation between change in MMA vote share and funding gap is concave with very close to being linear. The relation is stronger in the interval where the funding gap is between 20 and 60%. Areas with funding gap of 40% experienced a three-fold increase in MMA vote share, while areas with funding gap 80% experienced a four-fold increase in the MMA vote share compared to the areas with no or very little funding gap. The results clearly indicate that funding gap is an important determinant of increase in the vote share of MMA. These results highlight the importance of funding gap as one of the mechanisms through which the change in support for MMA is working. The areas which had higher funding gap witnessed greater increase in the vote share of the MMA. 7.2 Participation Another consequence of the funding gap could be change in representation of MMA. The MMA is not present in all the districts in the 2008 elections which were held before the flood. As a result, the MMA can decide to file candidates in more districts. The filing can increase disproportionately in the areas with high and low funding gap. We test precisely these implications. That is, we test whether the MMA reacted along the intensive margin to file more candidates in the areas with differential funding gap. Table 5 shows the results. We see that there was increase in the probability of MMA candidate running from an electoral district from 2008 to However, there was no significant difference between the increase in MMA candidate running between areas with high or low funding gap. The results highlight that the MMA did not respond along the extensive margin, rather responded more on the intensive margin. 24

25 8 Alternate Explanations In this section, we discuss other mechanisms that can yield the same result. In particular, we discuss three alternate mechanisms. The observed increase in the MMA support can be due to anger against the incumbent, political substitution or increase in religiosity. We show below that the evidence is inconsistent with each of these mechanisms. Note, however, that the mechanisms tested below are not necessarily substitutes to our explanation, but can also act as complements. In other words, the explanation that anger against incumbent resulted in increased support for MMA is not entirely substitute for the explanation that funding gap lead to the increase in support. However, we show that other mechanisms are not in play to provide additional evidence to cement that funding gap is the main mechanism in play. 8.1 Anger against incumbent The individuals can blame the incumbent political party, PPP, for not providing proper rehabilitation after the flood. This can be true especially when the individuals hold the political party directly responsible for poor recovery. Thus, the accountability or anger against the incumbent can be thought of as a function of the funding gap. It is documented in the literature that the individuals punish the incumbents for natural disasters and other shocks which the politicians have little control over (Cole et al., 2012). According to this argument, the vote share of the incumbent should decrease more in the areas with high funding gap compared to the areas with low funding gap. We test whether this was actually true. The Table 6 shows the results. Column 1 and 2 show the result for the whole sample, while Column 3 and 4 show the results for the sub-sample which have low ex-ante propensity of flooding. The estimates show that PPP decreased the vote share by about 20 percentage points from 2008 to However, the vote share was unresponsive to the funding gap. That is, the vote share for PPP did not decrease more in the high funding 25

26 gap areas compared to the low funding gap areas. If anything, the PPP lost less votes in the high funding gap areas. Furthermore, on average, PPP had similar vote share in the areas with low funding gap and high funding gap. The difference is economically small and statistically insignificant. This can be taken as suggestive evidence that the funding was not strategically assigned by the government along maximizing the chance of re-election lines. Since, according to the literature on political incentives for the politicians, maximizing the probability of re-election is the single most important incentive for the politicians (Besley, 2005), this can be taken as suggestive evidence that the funding was not strategically distributed by the government 20. This helps us rule out that the individuals voted more for MMA due to their anger against the incumbent political party. 8.2 Political Substitution Another natural explanation for the observed change in MMA vote share can be political substitution. This explanation is closely related with the previous one. The overall decreased vote share for the incumbent has to be replaced by the other political parties. If a simple political substitution mechanism was under play, we should expect other major political parties to experience the same increase in vote share. That is, we should see a similar increase in the vote share of the other major political party, PML-N, in the high funding gap areas compared to the other areas. We test this alternate explanation. The Table 7 shows the results. Column 1 and 2 show the result for the whole sample, while Column 3 and 4 show the results for the sub-sample which have low ex-ante propensity of flooding. The results show that the areas with high funding gap were areas with lower vote share for the PML-N, the major opponent of PPP which is also the incumbent political party in 2013), in the 2008 elections. The PML-N had 30 percentage points less 20 The un-strategic distribution of funding by the government can be due to the magnitude of the disaster and the limited resources availability. Government did not had enough resources, given the magnitude of the disaster, to make strategic decisions. Rather, it was constrained by the resources to distribute it as evenly as possible along the affected areas. This is also evident from the Figure on funding gap. Instead of having clusters of high funding and low funding areas, the funding gap is almost uniformly distributed all across the country. 26

27 votes in the areas with high funding gap compared to the low funding gap areas. The PML-N gained, on average, around 10 percentage points votes more in the 2013 elections compared to the 2008 elections in all the areas. However, they did not gain more vote share in the high funding gap areas. If anything, they gained less in the high funding gap areas compared to the low funding gap areas. 8.3 Increase in Religiosity Another possible explanation could be that the areas that were affected by the flood became more religious. Most common reason cited for such a change is that the people in the affected areas saw the flood as a wrath of god and turned more religious (Steinberg, 2000). We show that the increase in religiosity alone through being affected by the flood can not explain the observed changes in support for MMA. If natural disasters lead to increased religiosity and hence higher support for MMA, we should observe similar increase in the support for MMA in another natural disaster. We exploit the earthquake of 2005 to test these claims. The Appendix provides details about the earthquake. In order to test these claims, we use similar approach employed in the previous section. Specifically, we test whether the MMA won more proportion of the votes in the 2008 elections in the areas affected by the earthquake compared to the areas which were unaffected by the earthquake. The following equation summarizes the equation of interest: MMA it = α i + λ t + βeq it + u it, (4) where EQ it is a measure for whether the electoral district, i, at election year, t, was affected by the earthquake or not. The estimation controls for electoral-district fixed effects and time effects. The Table 8 shows the results from estimation of the above equation. The robust standard errors clustered at the electoral-district level are reported in the parenthesis below the estimates. Three different measures of whether an electoral district was affected 27

28 by the earthquake or not are used. The first two measures use a binary indicator for whether an area was affected or not. Specifically, in column 1 and 4, an area is coded as being affected by the earthquake if it is within the 350 km radius of the epicenter of the earthquake 21. Column 2 and 5 use a stricter definition of affected by only coding areas within the 250 km radius of the epicenter of the earthquake as affected, and areas outside the 150 km radius as unaffected by the earthquake. Finally, column 3 and 6 use a continuous definition of affected. That is, using the log(distance) of the center of an electoral district to the epicenter of the earthquake as a measure of how much an electoral district was affected 22. In the column 1, of Table 8 shows the result from considering the 350 km radius around the epicenter as affected areas by the earthquake. The areas in the 350 km radius of the epicenter of the earthquake experienced an 8.9 percentage points decrease in the vote share compared to the areas outside of 350 km radius of the epicenter. Similarly, the areas inside the 250 km radius of the earthquake epicenter experienced a higher increase (14.7 percentage points) in the vote share of MMA compared to the areas outside 250 km. The vote share increases by 9 percentage points for every doubling of the distance from the epicenter. The results clearly rule out that after every natural disaster the religiosity increases which can explain the observed results. However, the results provide further evidence for our mechanism. This is essentially the other side of the coin. That is, an instance in which the government was able to meet the needs of the individuals after the disaster (See Appendix). Our mechanism predicts that the support for these non-state actors should decrease as a response of efficient government rehabilitation. This is essentially what we showed above. The results further strengthen the mechanism under operation 21 The maximum distance from epicenter at which the after-shocks of the earthquake were felt was around km away from the epicenter (Peiris et al., 2006) 22 We used several other measures of distance. For instance, the minimum distance of the earthquake epicenter to an electoral district, the maximum distance of the earthquake epicenter to an electoral district, and the distance by connecting roads from the epicenter to the electoral district. All these different measures of the distance yield similar results. The results are available upon request. 28

29 is lack of state. 9 Heterogeneity of results In this section we present heterogeneity of the results along two dimensions: the distance from Afghanistan and the margin of victory/loss for MMA in the 2008 elections. These two dimensions capture the important implications of the mechanism. The MMA and the Taliban have higher accessibility in the areas closer to the Afghanistan. For two areas which are affected by the flood with the same funding gap, the area closer to the Afghanistan has higher probability of receiving rehabilitation support from the MMA and Taliban, ceteris paribus. Similarly, for two affected areas with the same funding gap, the MMA and Taliban will exert more effort in the areas which are closely contested compared to the areas in which they are clear winners or clear losers, ceteris paribus. We check the heterogeneity of the results along these two dimensions. 9.1 Intensity of the floods The flood in 2010 affected different electoral areas differently; while some regions were completely destroyed, the other were partially affected by the flood. In the previous sections, we did not distinguish between different intensities with which the flood can affect a region. In this section, we explore whether the results are driven entirely from the areas that were severely affected by the flood, or from areas that were moderately affected by the flood or the results are driven from both the regions with varying magnitude. According to our mechanism, the expected political outcomes in the areas severely affected by the flood should be different from the areas moderately affected by the flood. Specifically, the severely affected areas should see a higher increase in the political support for MMA compared to the moderately affected areas. Consider the case in which government distributes the humanitarian aid equitably across 29

30 the electoral districts 23. On the other hand, some electoral regions were more affected than the others. This implies that there is a deeper void to fill in the areas which were severely affected by the flood compared to the areas moderately affected by the flood. This makes it more likely for the areas that are severely affected by the flood to be more vulnerable to take help from the non-mainstream political actors. We test the implication of the mechanism formally. We test whether the areas that were severely affected by the flood saw a higher increase in the political support for the MMA compared to the moderately affected areas. The affected by the flood binary variable is broken down into an indicator whether the area was severely affected or if the area was moderately affected. The definitions of the severely and moderately affected are adopted from the National Disaster Management Authority (NDMA) of Pakistan 24. Equation 5 illustrates the specification of interest. MMA it = α i + λ t + β s severe it + β m moderate it + u it, (5) where severe it is a binary indicator equal to one if the area was severely affected by the flood, and zero otherwise. moderate it indicates whether an area was moderately affected. The estimation controls for electoral-district fixed effects and time effects. Out of the 112 electoral districts that were affected by the flood, 55 were severely affected, while 57 were moderately affected.?? shows the results. The severely affected areas saw a higher increase in the vote share 23 This is plausible due to many reasons. First, the government did not have sufficient aid to fulfill the needs of any electoral district. Under such a situation, the government can opt to satisfy the need of all the electoral districts partially, rather than satisfying the needs of some districts, while leaving out the other districts. Second, (CITE SOMETHING) 24 The NDMA coded the severely affected and moderately affected areas as follows: If more than 50% of the land area of an electoral region was flooded, then the area is severely affected. On the other hand, if less than 50% of the land area of an electoral region was flooded then the area is moderately affected. (CONFIRM THIS). We employ another parallel definition of severely and moderately affected area in terms of the population of the region that was affected. That is, the area is moderately affected if less than 50% of the electoral region s population was affected by the flood, and an area is severely affected if more than 50% of the electoral region s population was affected by the flood. The two different measures of severely and moderately affected areas have many overlap. The results from this other definition are omitted due to brevity. The results are available upon request. 30

31 of MMA compared to the moderately affected areas. The moderately affected areas, on average, saw a 3.3% increase in the vote share for MMA compared to the areas which were unaffected by the flood. On the other hand, severely affected areas saw an increase of 4% compared to the unaffected areas. Column 4 and 5 show the estimates for areas with low ex-ante propensity of flooding, while column 6 and 7 show the results for high ex-ante propensity of flooding. The areas are characterized as low and high ex-ante propensity of flooding based on the mean 25. The results in Column 4 and 5 indicate that the effect is mainly driven by the areas with low ex-ante propensity of flooding. The areas with low ex-ante propensity of flooding experienced an 3.6% and 7.2% increase in the vote share for the MMA in the moderately and severely affected areas respectively. On the other hand, Column 6 and 7 reveal that there was no statistically significant change in the vote share of the MMA in the moderately or severely affected areas. The change in the severely affected areas is also economically insignificant. However, in the moderately affected areas, the point estimate is close to the point estimate yielded from low ex-ante propensity of flooding. The results may be insignificant due to low power due to small number of observations. 9.2 Distance from Afghanistan The mechanism we propose depends on the level of support provided by the Taliban in the areas compared to that of the government. The capacity of the Taliban to provide support in a region greatly depends on the distance of the region from the Afghanistan border. This is mainly due to two reasons. First, the Talibans originate from the area close to the Afghanistan border. The Pakistan version of Taliban have their headquarters in the North Waziristan and Mohamand agency in FATA, next to the Afghanistan border. Second, the Taliban are of Pashtun-origin. The inhabitants of the Khyber Pakhtunkhwa province, and to a smaller extent the people in the Balochistan province are of the same 25 The areas with ex-ante probability of flooding greater than the mean are characterized as high ex-ante propensity of flooding areas. The results with the cut-off at median are similar. 31

32 ethnic origin. On the other hand, the Pakistan government is known for its political preference for the provinces of Punjab and Sindh. The government has been often alleged to overlook the needs of the people of Khyber Pakhtunkhwa and Balochistan, and favor the Punjab and Sindh for development 26. This implies that the areas closer to the Afghanistan border are more likely to receive greater Taliban support, everything else equal. This generates testable implication according to our mechanism. Specifically, since the areas closer to the Afghanistan border are more likely to receive support from Taliban relative to the government, the areas closer to the border of Afghanistan should yield a higher increase in the political support for the Taliban in the 2013 elections compared to the areas further away from the border. The following equation captures the mechanism formally: MMA it = α i +λ t +β 1 affected it +β 2f f(dist afgh i )+β 3f affected it f(dist afgh i )+u it, (6) where affected it is the dummy for affected areas interacted with the dummy for the year 2013, and dist afgh i is the distance of the electoral region, i, from the Afghanistan border. The Figure 4 plots the marginal effects for the above equation. Specifically, the plot shows how β 2f +β 3f varies with the distance from the Afghanistan. A smooth polynomial is fitted to account for non-linear effect heterogeneous effects. The graph shows nice smooth decreasing marginal effects as a function of distance from Afghanistan. The areas closer to the border saw an increase of more than 15 percentage points, compared to the areas which were unaffected by the flood. The effects last until around 300 km distance from the border. After the 300 km, the marginal effect is statistically indistinguishable from zero. 26 In fact, the two major political parties, Pakistan Muslim League-N and Pakistan People s Party originate from the provinces of Punjab and Sindh respectively. Both the parties share strong political support based on their ethnicity. 32

33 9.3 Margin in previous election The margin could be used as an proxy for the strategic value of the area to the MMA 27. Areas where MMA had very high (very low) vote share relative to the runners-up (winner), MMA has little political incentives to help. The areas with margin close to zero are the areas which have the highest strategic value to the MMA. The MMA can strengthen their vote share for areas barely won and tip the scales in there favor for areas where they barely lost. This implies that we should observe greatest change in the areas with margin close to zero. We test this prediction using similar estimation as in the previous section. That is, we estimate: MMA it = α i +λ t +β 1 affected it +β 2f f(margin i )+β 3f affected it f(margin i )+u it, (7) where affected it is the dummy for affected areas interacted with the dummy for the year 2013, and margin i is the margin of victory or loss of MMA in the electoral region, i, in the 2008 elections. The Figure 5 plots the marginal effects for the margin. The f(margin i ) is a fourth order polynomial to trace out non-linear marginal effects. The graph shows a quadratic relation between the marginal effect and the margin. The marginal effect is greatest at the areas where MMA had lost by less than 20 percentage points. The marginal effect is statistically significant for the areas where MMA lost by less than 50 percentage points and won by less than 10 percentage points. The effect of flood on areas where MMA lost by more than 50 percentage points or won by more than 10 percentage points is statistically insignificant. 27 We construct the margin as the difference between the MMA vote share and the runner-up in areas where MMA won and difference between the vote share of the winner and MMA in the areas where MMA did not win. So, the variable is positive if MMA won and negative if MMA lost. 33

34 10 Robustness In this section we check for the robustness of our results. First, we replicate the main results using data from the provincial assembly. Second, we carry out falsification tests to confirm that the results are not driven from pre-existing trends in the data. Last, we partially test the parallel trends assumption required for consistent estimates of the difference-in-difference results Results from Provincial Assembly In the general elections, the individuals vote simultaneously for candidates from national and provincial assembly. The provincial assembly is finer than national assembly and overlooks the provincial matters. The provincial assembly has its own budget and it has discretion on how to allocate the budget across the province. We test if the results hold for provincial assembly, a more finer level 28. Table 9 shows the results of estimating equations 1 to 3 on the Provincial Assembly data. Column 1 shows the estimates for Equation 1, Column 2 shows the estimates from Equation 2, while results from estimation of Equation 5 are shown in the Column 3 for low ex-ante propensity of flooding areas 29, while Column 4 shows results for high ex-ante propensity of flooding areas. All the specifications have provincial electoral seat fixed effect and the standard errors are clustered at the provincial assembly seat level. The results are similar to the results obtained from the national assembly data. Column 1 indicates that the affected areas experienced an 1.6% increase in the vote share of the MMA compared to the areas which were unaffected by the flood. The results are smaller 28 Note, however, that the dynamics might be different in the provincial assembly. The provincial assembly might be better able to divert the blame of lack of rehabilitation to the national assembly and the government. Moreover, the provincial assembly politics are more localized in nature: MMA is not strongly represented in the provinces of Punjab and Sindh. Thus, pacifying the effect. On the contrary, the individuals might hold the provincial representatives more accountable. (CITE SOME- THING/ARGUE MORE). 29 As before, the low and high ex-ante propensity of flooding is decided by whether the risk of flooding is lower or higher than the mean value of risk. 34

35 in magnitude than the results observed in the national assembly: the results obtained here are less than half the size of results obtained in the national assembly. Column 2 controls for the ex-ante propensity of flooding in an electoral area. The results show that the vote share of MMA increased by 1.7% in areas with low ex-ante probability of flooding (no floods in the previous fifty years) but which were affected by the flood compared to the areas with low ex-ante probability of flooding and which were unaffected by the flood. As the results in the previous column, the magnitude is less than half of the results from the national assembly. Unlike national assembly, the results in the provincial assembly are driven from the severely affected areas with low ex-ante propensity of flooding. The results show that the vote share of MMA increased by 5.6% in areas with low ex-ante probability of flooding but which were affected by the flood compared to the areas with low ex-ante probability of flooding and which were unaffected by the flood. There was no statistically significant difference between moderately affected areas with low ex-ante propensity of flooding and unaffected areas. Moreover, the areas with high ex-ante propensity of flooding did not witness statistically significant difference in the votes for MMA compared to the unaffected areas. In order to shed light on the mechanism, Equation 3 is estimated using both OLS and IV. Table 10 shows the results for different areas using Provincial Assembly data 30. Column 1 to 4 show results for the low ex-ante propensity of flooding areas, while Column 5 to 8 show the results for the high ex-ante propensity of flooding areas. Column 1, 2, 5 and 6 show the estimates for the moderately affected areas, while Column 3, 4, 7 and 8 show the estimates for the severely affected areas. The odd columns show the estimates for OLS, while IV estimates are shown in the even column. All the estimates take fixed effects and cluster standard errors at the provincial assembly level. The results show that all the effect of funding gap is stemming from the areas which 30 The Table 4 used National Assembly data. 35

36 had low ex-ante propensity of flooding. Moderately affected areas, and severely affected areas with low ex-ante propensity of flooding increased MMA vote share by 7.3% and 9.9%respectively, compared to the areas which were unaffected by the flood. There is no statistically significant change in the areas which were moderately or severely affected and had high ex-ante propensity of flooding. (DISUCSS MORE) 10.2 Testing Parallel Trends Assumption In the sections above, we employ difference-in-difference estimation to establish causal effect of being affected by the flood on the MMA vote share. Our identification strategy relies on the assumption that the trends in the affected and unaffected areas should be the same. That is, in absence of the flood, the average vote share for the MMA would have changed in the same way in the areas which were affected by the flood compared to the areas which were unaffected by the flood. Since we have electoral data for the 2002 elections, we can test this parallel trends assumption. We test whether the vote share of MMA changed differently between the 2002 and 2008 elections in the affected areas relative to the unaffected areas. Figure 7 shows the results. The graph shows that the change in MMA vote share in the affected areas between 2002 and 2008 elections is statistically the same as the change in MMA vote share in the unaffected areas 31. Having parallel trends consolidates that the result we find is a causal effect. There can still be some variables changing at the same time as the flood, i.e. between 2008 and 2013, that can confound the results (DISCUSS MORE). 31 In the graph, we included all the electoral regions. However, in between the 2002 and 2008 elections, earthquake took place which affected some electoral districts. We partially account for that by using only those areas that were 300 km outside the epicenter of the earthquake. The results are similar to the results using all the electoral districts. 36

37 10.3 Falsifications Tests In order to test if there are any unobserved trends in the data driving our results, we carry out falsification tests. In particular, we carry out two different falsification exercises. The first falsification exercise randomly assigns the status of affected and unaffected to half of the districts and estimates the results from Equation 1. The exercise is repeated 1000 times. The second falsification exercise is same as the first one, but randomly assigns affected status to the same number of electoral regions as those which were actually affected. The same exercises are carried out for the funding gap. In the first exercise, one-fifth are randomly assigned an integer value from zero to four. The Equation 3 is estimated and the exercise is repeated 999 times. In the second exercise, the funding gap values are assigned randomly to the actual number of areas (using the distribution of funding gap). Moreover, the equation with both affected and funding gap as the independent variables is also estimated in both cases. All the equations take electoral seat fixed effects and cluster the standard errors at the national assembly seat level. The Table 11 shows the mean, the standard deviation of the estimates obtained from each exercise in the first two rows. The third row reports the p-value from hypothesis test that the reported estimate is equal to zero. All the estimates are less than 0.001, and all of the p-values are greater than The estimates obtained strongly re-assure the fact that the results we got in the previous sections are not driven from pre-existing trends in the data Dropping Border areas The border areas can be systematically different from the areas which do not share the border with Afghanistan. This can be due to lower transportation cost and higher accessibility due to sharing the border. Moreover, the government might have lower control over these areas due to their proximity to the border (CITE SOMETHING ABOUT 37

38 POROUS BORDERS HERE). Thus, we drop the areas sharing border Afghanistan and re-estimate Equation 3. Table 12 shows the results. Column 1 and 3 show the OLS and IV results respectively for the whole sample, while Column 2 and 4 show the results for the sub-sample with exante low propensity of flooding. The results change very little compared to the baseline specification. The areas with high funding gap saw an increase of 5.2 percentage points compared to areas with low funding gap. The effect is similar in magnitude to the baseline specification Dropping Areas with high MMA vote share In this section we drop the areas with high MMA vote share in the 2008 elections. The areas with already high vote share in the 2008 elections may be systematically different from the areas with low MMA vote share. We drop the areas which lie in the top decile of MMA vote share in the 2008 elections and re-estimate the Equation 3. Table 12 shows the results. Column 1 and 3 show the OLS and IV results respectively for the whole sample, while Column 2 and 4 show the results for the sub-sample with ex-ante low propensity of flooding. The results are similar in magnitude to the baseline specification: areas with high funding gap experienced 7.4 percentage points increase in the MMA vote share compared to the areas with low funding gap. 11 Discussion & Conclusion The results shown above highlight important determinant of extremist ideology and support for extremist groups. The individuals respond to the way non-state actors and government provides for them. We find causal effect of flood on the support for extremist groups. Using flood of 2010 as an exogenous, heterogeneous increase in the needs of individuals, the paper shows that the government recovery effort post-flood have first-order 38

39 effects on the support for the groups with extreme ideology. Specifically, we show that the areas that were affected by the flood saw a 3.6 percentage points increase in the support for the parties with extreme ideology compared to the areas which were unaffected by the flood. We show that the percentage of funding gap reported by the electoral district is strongly positively correlated with the increase in MMA support. In particular, a 10 percentage point increase in the funding gap leads to 7.4 percentage points increase in the MMA votes. We show evidence that the results obtained are inconsistent with alternate explanations such as punishing the incumbent, political substitution among the political parties and increase in religiosity. Similarly, falsification tests and robustness checks strengthen the fact that what we find is causal effect of flood on the increase in support for the extremist groups. We propose the mechanism through which this causal effect operates. We propose that the efficiency of the government in the post-flood period is an important determinant through which these results occur. Future public policy and research should take into account the complementarity between government relief efforts and rise of extremist groups in areas with weak institutions and extremism. Our results show how government s capacity directly determines the support for NSOs. Our results shed light to how the same amount of funding can have multiple-fold higher impact on the individuals if given at the right time. In the other words, the results highlight the importance of giving aid at the right time. The estimated direct damage from the flood was $2.35 billion. Contrast this to the $1.6 trillion spent by the U.S. on the war on terror since its inception. The complete cost of damage would have been only 7 days of war on terror spending. The broader implications of our analysis could be when the government fails to provide services in relatively developed countries. Our results shed light, in a completely different setting, on the rise and success of left leaning political parties in the aftermath of global financial crisis in Europe, especially the Southern Europe. The global financial crisis led to huge increases in the needs of the citizens, the unemployment sky-rocketed and the there 39

40 was wide-spread recession, budget cuts and austerity measures. These circumstances fueled increase in the rise for support for the left parties like Syriza in Greece, and Podemos in Spain. Figure 1: The National Assembly Electoral Districts of Pakistan 40

41 Figure 2: Flood Affected Districts 41

42 Figure 3: Earthquake Epicenter 42