Evidence from the 2010-11 Pakistani Floods∗

Home , 2010 Pakistan floods, Geography of Pakistan, Syed Qaim Ali Shah

Economic Shocks and Civic Engagement: Evidence from the 2010-11 Pakistani Floods∗

C. Christine Fair† Patrick M. Kuhn‡ Neil Malhotra§ Jacob N. Shapiro¶

This version January 23, 2014.

Abstract How countries transition from autocracy to democracy is a key question in both economics and political science. Using natural disasters to identify the political consequences of economic shocks, a number of recent studies have found a positive relationship between economic shocks and political liberalization. Scholars have interpreted these correlations as support for theoretical models that posit that lowered opportunity costs for rebellion due to negative economic shocks incentivizes the state to provide rights in equilibrium. This interpretation rests on two crucial assumptions: (1) the economic consequences of natural disasters are not strongly influ- enced by the governments’ potentially endogenous reactions to them; and (2) there are no strong non-economic channels through which natural disasters impact political attitudes and behavior. Using evidence from the 2010-11 floods in Pakistan we show that neither assumption is sustainable, at least in this one critical case. Leveraging diverse data sources (geo-spatial measures of flooding in 2010-11 and historical flood risk, constituency-level election returns, household economic survey data, nightlight illumination data, district-level emergency relief data, and an original survey of 13,282 Pakistani households), we show that the massive 2010-11 floods had no persistent economic impact, but find that Pakistanis living in flood-affected places became significantly more politically engaged than their unaffected peers; they turned out to vote at substantially higher rates and acquired greater political knowledge. We provide evidence that these results are consistent with an informational mechanism in which those affected by the floods learn about the role government plays in their lives and so become more civically engaged. These results call into question the interpretation of a broad set of papers and tie into a rich literature in political economy showing that disasters can have complicated political effects beyond their economic impacts.

∗We thank Saurabh Pant for research assistance on flood relief data. Major General Saeed Aleem’s team at the National Disaster Management Authority and Umar Javeed at the Center for Economic Research in Pakistan also provided useful information on relief efforts. Amarnath Giriraj and Mark Giordano at the International Water Management Institute developed historical flood data for us. Valeria Mueller at the International Food Policy Research Institute shared survey results on migration in flood-affected areas. Tahir Andrabi, Eli Berman, Graeme Blair, Mike Callen, Jishnu Das, Rubin Enikopolov, Asim Khwaja, Rebecca Littman, Rabia Malik, Hannes Mueller, Maria Petrova, Paul Staniland, Basit Zafar, and seminar participants at the 2013 AALIMS Conference at Rice, the 2013 Northeast Workshop in Empirical Political Science (NEWEPS) at New York University, Duke, University of Chicago, University of Michigan, and Yale all offered helpful comments and feedback. All errors are our own. This research was supported, in part, by the Air Force Office of Scientific Research, grant #FA9550-09-1-0314. †Assistant Professor, School of Foreign Service, Georgetown University; Email: c christine [email protected]. ‡Post-doctoral Research Associate, Empirical Studies of Conflict Project, Princeton University; E-mail: [email protected]. §Associate Professor, Stanford Graduate School of Business; E-mail: Malhotra [email protected]. ¶Assistant Professor, Woodrow Wilson School and Department of Politics, Princeton University, Princeton, NJ 08544; Email: [email protected]. Corresponding author.

1 Introduction

Considerable theoretical and empirical literatures study the political consequences of economic shocks (see e.g., Miguel, Satyanath and Sergenti, 2004; Burke and Leigh, 2010; Chaney, Forthcom- ing). In the theoretical literature the key mechanism by which economic shocks aﬀect governance is by decreasing the opportunity cost of rebellion for citizens, which in turn shifts the threat of rebellion conditional on government policies and thus the equilibrium behavior of the government

(Acemoglu and Robinson, 2006; Besley and Persson, 2010, 2011).

Scholars have tested these theories by using natural disasters to instrument for economic shocks as disasters provide plausible random variation in economic damage unrelated to government activity. Br¨uckner and Ciccone (2011), for example, argue that the strong positive correlation between negative rainfall shocks and democratization provides evidence for the opportunity-cost mechanism

first articulated by Acemoglu and Robinson (2001). Negative income shocks arising from abnormal rainfall increase citizens’ willingness to participate in rebellion, which in turn creates incentives for politicians to democratize and provide public goods in order to avoid costly repression. Uncovering a similar result, Ramsay (2011) instruments for positive income shocks to oil-producing states using out-of-region damage that natural disasters inflict on oil-producing countries, and finds a robust negative relationship between such shocks and standard measures of democracy such as the Polity

IV and Freedom House scales.1 But while the observation that natural disasters increase the risks of rebellious behavior has broad empirical support (e.g., Nel and Righarts, 2008; Burke, Hsiang and Miguel, 2013),2 the claim that an observed positive relationship between natural disasters and democratic governance provides evidence for opportunity cost mechanisms is harder to sustain.

There are two major concerns here. First, identifying correlations between disasters and democratization (or conﬂict) provides little direct evidence on the underlying theoretical mechanism

(Burke et al., 2009). If the government (and international) response is suﬃciently eﬀective, there should be no persistent economic shock. Thus the very existence of the channel in any given case might be correlated with unobservables that make the government more likely to respond poorly

1Note this result is inconsistent with Brückner, Antonio and Tesei (2012), who find that positive GDP shocks driven by oil price shocks lead to increased democratization. Ramsay (2011) notes that price shocks internationally can be correlated with within-region events (e.g., inter- and intra-state wars) that also have a direct effect on incentives for democratization within the region. 2Though it is not uncontested (e.g., Slettebak, 2012).

2 (e.g., having little concern for citizens), which would itself be correlated with subsequent governance outcomes. Second, there appear to be many non-economic alternative channels by which disasters can impact political views per a large literature in political science. In developed countries, researchers have long noted that voters punish politicians for natural events beyond their control (Achen and Bartels, 2004; Healy and Malhotra, 2010), and new evidence suggests they do so in developing countries as well (Cole, Healy and Werker, 2011). Disasters also lead to reactions by governments and other organizations that can shift both citizens’ beliefs and the opportunities for would-be rebels. Andrabi and Das (2010), for example, find that the provision of aid by outsiders in the wake of the devastating 2005 earthquakes in Khyber Pakhtunkhwa (KPK) province and Azad Kashmir in Pakistan led to a substantial increase in trust towards outsiders. Berrebi and Ostwald (2011) show that terrorism increases after natural disasters and that the effect is concentrated among less-developed countries, which they attribute to the twin effects of turmoil after a catastrophe (which could allow militants to exploit the exacerbated vulnerabilities of the state) and the government allocating resources away from maintaining order and towards disaster relief.

Using evidence from the 2010-11 floods in Pakistan, we show that even when government action effectively ameliorates economic effects there can still be large political consequences from a disaster.

This means that the mechanism by which disasters affect political outcomes cannot be purely economic, suggesting the role of informational effects and citizen learning. The 2010 flood affected more than 20 million people, caused between 1,800 and 2,000 deaths, and damaged or destroyed approximately 1.7 million houses, making it the worst flood in Pakistan’s modern history.3 The 2010

ﬂoods were driven by an unusual monsoon storm that dropped historically unprecedented levels of moisture on the mountainous northwest regions of the country. According to government accounts,

KPK received 12 feet of rain from July 28 to August 3, four times the province’s average annual total

(Gronewold, 2010). Those exceptionally high rainfall rates in mountainous areas compounded what was already an unusually rapid snowmelt to trigger flash floods that vastly exceeded anything in historical memory. As the water drained from KPK during the first week of August, a more typical monsoon storm inundated the Indus flood plain, rendering it incapable of absorbing the dramatic inflows from the mountainous regions and overwhelming many water management structures. The

3The EM-DAT International Disaster Database records approximately 20.4 million people aﬀected and 1,985 killed from the 2010 ﬂoods.

3 following year Pakistan got hit by an unusually strong monsoon storm, causing another round of devastating floods in the southern plains. In both cases the surging waters hit some places more than others due in part to the random combination of human action, prior differences in soil moisture, micro-topographic differences, and complex fluid dynamics.

Leveraging that plausibly exogenous variation along with a diverse set of data sources—multiple measures of ex ante flood risk, geospatial measures of flooding in 2010-11, election returns, and an original survey of 13,282 households—we show first that the economic impact of the floods was very small one to two years after they occured; with the exception of a depressed asset index among farmers there is no evidence of adverse economic impacts in flood-affected areas. We then show that

Pakistanis living in flood-affected places had substantially different levels of civic participation than their unaffected peers. They turned out to vote at higher rates in the 2013 elections and exhibited increased electoral participation relative to the last election (2008) by an even larger margin. These effects are substantively large and quite robust. Turnout in the Pakistani national and provincial elections rose roughly 11 percentage points between 2008 and 2013 (from 44% to 55%), a massive increase. Our results suggest that approximately 19% of this change can be attributed to the impact of the 2010-11 floods.4 Put differently, the increased civic engagement due to the floods led to a 2 percentage point change in the absolute level of turnout.5

Turning to potential mechanisms we find little evidence that these results are driven by rewarding aid spending or what was generally considered to be an effective response (particularly compared to previous floods in Pakistan).6 The incumbent Pakistan Peoples’ Party (PPP) lost massively in the 2013 election, dropping 16.5 percentage points of vote share from 2008 in the average constituency. While their loses were roughly 25% smaller in flood-affected constituencies on average, the conditional correlation between their performance and the intensity of the floods varies; it is positive in the Provincial Assembly elections but close to zero in the National Assembly one. More broadly, while the two main enduring national-level parties which were in government at the national or provincial level when the floods hit—the Pakistan People’s Party (PPP) and

Pakistan Muslim League-Nawaz (PML-N)—gained vote share disproportionally in ﬂood-aﬀected

4The 95% confidence interval on this effect ranges from 7% to 31% of the change. 5This is almost half the 5.4 percentage point increase between the 2000 and 2004 U.S. presidential elections and is larger than the 1.4 percentage point increase in turnout in the United States between 2004 and 2008, typically attributed to an unusually motivated electorate turning out in support of a historic candidate. 6The government was especially effective at coordinating the large flows of foreign aid.

4 areas relative to 2008, the effects are quite noisy and rarely reach traditional levels of statistical significance. We also show that the increase in turnout is extremely unlikely to be driven by compositional effects as a result of post-flood mobility.

Instead, we provide two pieces of suggestive evidence that the results are driven by a change in the importance individuals attach to influencing government action. First, we demonstrate that respondents affected by the flood have greater political knowledge than their unaffected counterparts. A one standard deviation increase in the proportion of the population exposed to the floods predicts a .077 increase in our political knowledge index, a shift of .1 standard deviation. Sec- ond, using a vignette experiment, we find that citizens exposed to floods are significantly more supportive of aggressive approaches towards demanding public services and believe them to be more effective than their non-exposed counterparts. A one standard deviation movement in the proportion of the population exposed to floods at the tehsil level (the third-level administrative unit) corresponds to a .16 standard deviation increase in perceived effectiveness of the aggressive action and a .18 standard deviation increase in approval for that approach. Overall, flood exposure appears to increase civic engagement, leading people to invest in political knowledge and become more politically active, even in the absence of a persistent economic shock.

The remainder of this paper proceeds as follows. Section 1 provides the relevant background on the 2010-11 Pakistan floods. Section 2 describes how we measure key variables. Section 3 outlines our research design and identification strategy. Section 4 analyzes the 2010-11 floods’ economic effects. Section 5 presents our core results on the floods’ impact on civic engagement. Section 6 provides robustness checks of those core results. Section 7 provides evidence on the mechanisms underlying the main results. Section 8 concludes by discussing policy implications and laying out directions for future research.

1 Background on the Floods

The scale of the 2010-11 ﬂoods dwarfed any Pakistani natural disaster in recent memory. The 2010

floods affected more than 20 million people (i.e., about 11% of the total population), temporarily displaced more than 10 million people, and killed at least 1,879, with the 2011 floods affecting another 5 million, displacing another 660,000 people, and killing at least 5,050 more (Dartmouth

5 Flood Observatory (DFO), 2013; Center for Research on the Epidemiology of Disasters (CRED),

2013). A survey of 1,769 households in 29 severely flood-affected districts found that 54.8% of households reported damage to their homes, 77% reported at least one household member with health problems, and 88% reported a significant household income drop (Kirsch et al., 2012).

Figure 1 shows the combined maximal extent of the 2010 and 2011 ﬂoods.

INSERT FIGURE 1 HERE

The 2010 disaster was a significant outlier in Pakistan’s flood history. Figure 2 shows standardized values for the number affected, displaced, and killed for each flood between 1975 and 2012.

The upper graph uses data from the International Disaster Database (EM-DAT) hosted by the

Center for Research on the Epidemiology of Disasters (2013) (data range 1975-2012) and the lower graph draws on data from the Global Active Archive of Large Flood Events of the Dartmouth

Flood Observatory (DFO) (2013) (data range 1988-2012). In terms of the number aﬀected and the number displaced, the 2010 ﬂoods were the largest in the modern history of Pakistan by several orders of magnitude and according to the EM-DAT almost twice as devastating as the next largest

ﬂood.7

INSERT FIGURE 2 HERE

Commensurate with the devastation, the 2010-11 floods also led to an unprecedented reaction by Pakistani civil society (the central, provincial, and district governments) as well as by the international community (Ahmed, 2010). Pakistan’s Economic Affairs Division took the overall lead on donor coordination, while Pakistan’s National Disaster Management Agency (NDMA) directed and coordinated the various relief efforts. The NDMA maintained close working relationships with relevant federal ministries and departments, Pakistan’s armed forces, and donor organizations supporting the relief efforts to ensure that resources were mobilized consistent with local needs.

At the provincial level, the chief minister of each province was responsible for making sure that various line ministries and the Provincial Disaster Management Authorities acted in concert with each other and with the international and domestic relief eﬀorts. At the district level, district

7The next largest was the 1992 ﬂood which aﬀected 12.8 million, displaced 4.3 million, and killed at least 1,446 people.

6 coordination officers were responsible for those activities of local governance that are devolved to that administrative level (Office for the Coordination of Humanitarian Affairs (OCHA), 2010;

National Disaster Management Agency, 2011).

This complex system often made it diﬃcult to discern who was responsible for success or failure.

In December 2010, one of the authors visited two relief camps, one each in Nowshera (KPK) and one in Dadu (Sindh). The Pakistan military was most visible through their air, road, and water missions by uniformed personnel. Tents and other supplies were usually branded by the donor (e.g., the United Nations, USAID, etc.), but the distribution of these goods was done through Pakistani personnel. This was deliberate as the intention of many donor and bilateral organizations was to foster the impression that the Pakistani government was eﬀectively managing relief eﬀorts.8

In addition, there were spontaneous localized self-help efforts that emerged during the initial phase of the crisis and continued throughout. These included victims and their kin’s own efforts to save their belongings but also included survivor-led repair of local access roads and bridges after the floods receded. This was in addition to an enormous civil society response that tended to spon- taneously coalesce at very local levels (mohallas, union councils, villages, etc.). Such local groups collected and distributed truckloads of relief items. Countless local as well as national organizations set up collection sites for donations of goods and cash and then undertook the distribution of the same. Individual philanthropists, professional bodies, and even chambers of commerce donated money and supplies to the victims. Scholars associated with Pakistan’s Sustainable Development

Policy Institute note the importance of these local forms of assistance but contend that they are virtually unknown (and thus poorly documented) beyond the local level (Shahbaz et al., 2012).

Such volunteerism was not unique to the 2010 ﬂood; rather, it is a common feature in Pakistan’s domestic response to such calamities. Halvorson and Hamilton (2010), for example, document extremely high levels of volunteerism following the 2005 Kashmir earthquake.9

By the end of July 2010 the government had appealed to international donors for help in responding to the disaster, after having deployed military troops in all aﬀected areas together with

21 helicopters and 150 boats to assist aﬀected people (Khan and Mughal, 2010).10 In response,

8Author ﬁeldwork in Nowshera and Dadu in December 2010 and January 2011. 9Note that this is not unique to Pakistan either. Scholars have also noted this elsewhere in South Asia (see e.g., Haque, 2004; Rahman, 2006; Ghosh, 2009). 10The United States provided an additional seven helicopters as part of their relief eﬀorts.

7 the United Nations launched its relief eﬀorts calling for $460 million to provide immediate help, such as food, shelter, and clean water. Countries and international organizations from around the world donated money and supplies, sent specialists, and provided equipment to supplement the

Pakistani government’s relief efforts. According to the United Nations Office for the Coordination of Humanitarian Affairs (UNOCHA) (2010), by November 2010, a total of close to $1.792 billion had been committed in humanitarian support, the largest amount by the United States (30.7%), followed by private individuals and organizations (17.5%) and Saudi Arabia (13.5%).11

The government’s effort and the massive influx of foreign aid made the response quite effective.

The ratio of killed to 1,000 people aﬀected from the EM-DAT data, and the ratio of deaths to

1,000 people displaced for each flood between 1975 and 2012 from the DFO data, provide a proxy for the effectiveness of the government’s response. For the 2010 flood, the ratios are 0.10 and 0.19, respectively, which is the smallest ratio in the DFO series (1988-2012) and the seventh smallest ratio in the EM-DAT series (1975-2012).12 Strikingly, the 2010 ratio is only 21% of the median ratio of killed to 1,000 displaced in the DFO data, so roughly one fifth as many people died as would have been expected given the median response in the last 37 years. Compared to modeled mortality risk using worldwide data, the number killed was only one-quarter what would have been expected given Pakistan’s level of development, population distribution, and where the flood hit

(Maskrey, 2011, ch. 2, p. 30). Overall, the government’s performance in handling the immediate challenge from the 2010 ﬂoods appears to have been quite good.13

2 Data Sources

We leverage multiple data sources to measure our variables of interest, including: (1) economic outcomes; (2) ﬂood exposure; (3) constituency-level voting behavior in the National and Provincial

Assembly elections; and (4) attitudinal and behavioral outcomes at the individual level. Appendix

Table A.1 provides summary statistics for all variables used in the subsequent analyses.

11By April 2013, this total had increased to more than $2.653 billion with the three largest donor groups being the United States (25.8%), private individuals and organizations (13.4%), and Japan (11.3%) (UNOCHA, 2013). 12Compared to the 1992 flood, the only flood of comparable magnitude in the last 30 years, the 2010 ratio is 72% smaller when using the DFO data and 8% smaller in the EM-DAT data. 13For comparison, Hurricane Katrina killed 1,833 people in the Gulf Coast in 2005 even though many fewer people were directly affected, approximately 500,000 according to the EM-DAT database.

8 2.1 Economic Outcomes

We use two pre-existing sources of data to assess economic outcomes. First, we use household- level data from the 2007 and 2011 Punjab Multiple Indicator Cluster Surveys collected by the

Punjab Bureau of Statistics. These data were collected using a tehsil-representative sample for the province of Punjab (population approximately 91.4 million in 2012) and include standard asset, expenditure, and income modules for 91,075 households in 2007 and 95,238 in 2011. The 2011 ﬂoods were concentrated in Sindh and had little impact in Punjab, meaning the 2011 Punjab MICS was eﬀectively administered 10-15 months post-treatment for the vast majority of the sample.

Second, we use data on nighttime illumination from human settlements detected by the Defense

Meteorological Satellite Program’s Operational Linescan System (DMSP-OLS). The DMSP-OLS

“nighttime lights of the world” images are processed specifically for the detection of change, and are made available by the National Geophysical Data Center. We use the cloud-free composites images produced using all the available archived satellite images of DMSP-OLS during a calendar year. These composites are scaled onto a geo-referenced 30 arc-second grid (approximately 1 km2) where each grid cell takes on a 6-bit scale digital number (DN), from 0 to 63. For each year, a grid cell with a value of zero can be interpreted as an area with zero nighttime light. On the other hand, the value of 63 is the saturation value and indicates the brightest area for each year.14 We aggregate these gridded data to the tehsil level to match the geo-spatial resolution of the MICS data. While there are problems using these data as indicators of economic activity in Pakistan given the pervasive load-shedding and potential for politically driven electricity allocations (see e.g. Paik and Shapiro, 2013), we still expect that if the floods had substantial adverse local economic impacts those would show up in the difference between 2009 and 2012 illumination levels in flood-affected areas.

2.2 Population Distribution

Our spatial population data was taken from Oak Ridge National Laboratory’s (2008) Landscan dataset, which provides high-resolution (30 arc-second grid (approximately 1 km2)) population distribution data for 2011. The digital administrative boundary data of all tehsils and electoral

14While the digital numbers are relative values and thus are not perfectly comparable between two years, we follow standard practice in assuming that the calibration problem is in fact not crucial.

9 constituencies for the national and provincials assembly were developed from constituency-level maps available from the Electoral Commission of Pakistan (ECP) website.

2.3 Turnout and Voting Patterns

We collected constituency-level election data for the 2002, 2008, and 2013 National and Provincial

Assembly elections. The Pakistan National and Provincial Assemblies combine members elected in single-member first-past-the-post elections at the constituency level (272 for the National As- sembly and 577 for the Provincial Assemblies in the four main provinces) plus a number of seats reserved for women and minorities (70 in the national assembly) that are allocated among parties according to a proportional representation scheme. Most candidates align with a party but some run as independents and affiliate with a party for coalition formation purposes after the election is complete. Candidates in the 2013 election campaigns combined appeals to national issues and party platforms with locality-specific appeals and promises of patronage, with the mix varying by candidate. In prior campaigns the appeals are commonly understood to have been more localized.

The Election Commission of Pakistan (ECP) website lists by-election results when a constituency has changed hands due to the death of its representative or to him/her shifting to a new job. As by-election contests are likely to have a different dynamic than those held during the general election, we replaced as many of the by-election results as possible with the original results from the main polling date. For the National Assembly election, we managed to find general election data on all contests except for five constituencies in 2008. For the Provincial Assembly election, we were able to obtain polling results from the general elections for all by-elections in 2002 and 2008. For each constituency we recorded turnout as well as the vote shares for all candidates on the ballot.

2.4 Behavioral and Attitudinal Outcomes at the Individual Level

We conducted an original survey, which included a large sample (n=13,282) of respondents in the four largest provinces of Pakistan (Balochistan, KPK, Punjab, and Sindh). We collected district- representative samples of 155-675 households in 61 districts with a modest over-sample in heavily

flood-affected districts as determined by our spatial flood exposure data. We sampled 15 districts in Balochistan, 14 in KPK, 12 in Sindh, and 20 in Punjab to ensure we covered a large proportion

10 of the districts in each province. Within each province we sampled the two largest districts and then chose additional districts using a simple random sample. The core results below should be taken as representative for our sample which, while large, does overrepresent Pakistanis from the smaller provinces. Weighted results using either sample weights calculated from Landscan gridded population data or those provided by the Pakistan Federal Bureau of Statistics are substantively and statistically similar.

Our Pakistani partners, SEDCO Associates, ﬁelded the survey between January 7 and March

21, 2012. Our overall response rate was 71%, with 14.5% of households contacted refusing to complete the survey and 14.5% of the targeted households not interviewed because no one was home who could take the survey. This response rate rivals those of high-quality academic surveys in the United States such as the American National Election Study (ANES).

We measure two outcomes that tap civic engagement: political knowledge and attitudes towards using violence as a means of achieving political ends. We also capture self-reported household income, expenditures, and an additive index of 18 non-farm assets that households can own

(televisions, cassette players, radios, etc.) as additional measures of economic well-being.

Behavior: Political Knowledge

We construct a measure of political knowledge using a set of binary questions. To tap awareness of political issues, we asked respondents whether they were aware of four policy debates that were prominent in early-2012: whether to use the army to reduce conﬂict in Karachi; how to incorporate the FATA into the rest of Pakistan; what should be done to resolve the disputed border with

Afghanistan; and whether the government should open peace talks with India. We also asked six questions about various political oﬃces and scored whether respondents correctly identiﬁed the following: who led the ruling coalition in Parliament (the PPP); and the names of the President,

Prime Minister, Chief Minister of their Province, Chief of Army Staﬀ, and Chief Justice of the

Supreme Court. Appendix B provides full question wordings for all these items.

Following Kolenikov and Angeles (2009), we conduct principal component analysis on the poly- choric correlation matrix of these items and use the ﬁrst principal component as our measure of political knowledge. That component accounts for 49.15% of the variance in the index of ten components, suggesting it does a good job of capturing our underlying concept.

11 Critically, the political knowledge index should be considered a behavioral outcome; it is not something that can be faked. Respondents either know these facts or they do not and thus differences must either reflect some real investment in acquiring political knowledge or some background trait that is correlated with both being hit by the floods and political knowledge. Our core estimating equation for individual-level outcomes includes controls for literacy, numeracy, age, education, gender, and head of household status to minimize the latter possibility, making a behavioral interpretation of the knowledge variable more tenable.

Attitudes: Support for Aggressive Political Action

In order to measure support for aggressive political action we use a vignette experiment. This approach circumvents three main challenges to measuring political attitudes. First, respondents may face social desirability pressures to not explicitly support particular views (e.g., aggressive civic protests). Second, concepts such as the political efficacy of aggressive protests are not easily explainable in standard survey questions but can be illustrated with examples. Third, respondent answers to direct questions may not be interpersonally comparable (King et al., 2004). To overcome these challenges, we wrote two vignettes describing concrete (but fictional) examples of two different ways of getting the government’s attention: peaceful petition or aggressive protest. Respondents were randomly assigned to receive one of the vignettes before answering the same two survey questions on how effective they think the chosen method is and whether they approve of it.

More speciﬁcally, the vignette experiment works as follows. At the primary sampling unit (PSU) level, respondents are randomly assigned to read one of the following two vignettes:

Peaceful Petition. Junaid lives in a village that lacks clean drinking water. He works

with his neighbors to draw attention to the issue by collecting signatures on a petition.

He plans to present the petition to each of the candidates before the upcoming local

elections.

Aggressive Protest. Junaid lives in a village that lacks clean drinking water. He works

with his neighbors to draw attention to the issue by angrily protesting outside the oﬃce

of the district coordinating oﬃcial. As the government workers exit the oﬃce, they

threaten and shove them.

12 Following the vignette respondents are asked the following:

• Eﬀectiveness. “How eﬀective do you think Junaid will be in getting clean drinking water for

his village?” (response options: “extremely effective,” “very effective,” “moderately effective,”

“slightly eﬀective,” “not eﬀective at all”)

• Approval. “How much do you approve of Junaid’s actions?” (response options: “a great deal,”

“a lot,” “a moderate amount,” “a little,” “not at all”)

Our vignette experiment is intended to convey a clear diﬀerence in the aggressiveness with which citizens petition for government services.15 To maximize its power we jointly varied three elements of how the citizens in the vignette engaged with the government: immediacy, target, and method.

The next election was expected to happen in 2013 when we did our survey, so the Peaceful Petition vignette clearly conveys a demand that will be delayed, while the Aggressive Protest one describes something that could happen right now. With respect to target, the district coordinating oﬃcial

(DCO) is the relevant oﬃcial for drinking water issues, but he/she is an appointed bureaucrat who reports to the Chief Minister of the province. If citizens go to the DCO he/she can do something right away, whereas local politicians have to go to the senior party leadership of the party in power in their province who may then reach back down to the DCO. The Peaceful Petition vignette thus conveys a situation where the response to citizen action will be indirect at best, while the Aggressive

Protest one portrays citizens going right to the oﬃcial who has to execute on the ground. On the method front, holding a protest that turns violent is clearly more aggressive than signing a petition.

Thus while our compound treatment does not allow us to distinguish which of the three elements was critical, it provides a clear diﬀerence in how aggressive the citizens’ approach is understood to be in the Pakistani context.

Our sample is well balanced across conditions in the vignette experiment on a broad range of geographic, demographic, and attitudinal variables, as Figure 3 shows. The diﬀerence in means between the groups within a region therefore provides an estimate of how eﬀective/acceptable citizens think the use of aggressive action is to pressure their political representatives, which we referred to earlier as a form of unconventional politics.

INSERT FIGURE 3 HERE 15We thank Rashad Bukhari, Ali Cheema, and Basit Zafar for useful comments on this vignette experiment.

13 2.5 Measuring Flood Impact

Figure 1 shows the combined maximum ﬂood extents in 2010 and 2011 overlaid on a map of

Pakistan. The 216 tehsils in which we surveyed are highlighted in grey.16 As one can see the surveyed areas include tehsils that were severely impacted, tehsils that had only small portion of their territory affected, and those that were spared by the floods. As described below, we exploit the randomness inherent in flood exposure to identify the impact of the 2010 and 2011 floods on the variables described in the previous section.

We measure flood exposure with two different sets of variables: (1) objective measures based on geo-spatial data; and (2) subjective measures asked of respondents as part of the survey. We also constructed two measures of ex ante flood risk.

Objective Exposure

Geo-spatial data on the 2010 and 2011 ﬂoods come from the United Nations Institute for Training and Research’s (UNITAR) Operational Satellite Applications Program (UNOSAT), which provides imagery analysis in support of humanitarian relief, human security, and strategic territorial and development planning (United Nations Institute for Training and Research, 2003). For the 2010

ﬂood UNOSAT provided a time series of satellite data recorded between the end of July and mid

September. For the 2011 flood in Sindh province UNOSAT provided a map of the standing flood waters following heavy monsoon rain between mid-August and early October. Overlapping the various images allowed us to generate a layer of maximal flood extent in the two years prior to the survey data collection at the beginning of 2012. These data combine multiple different sources and are the most precise data we know of on those floods, providing estimates of flood extent at

100m×100m resolution.

We calculate objective measures of ﬂood exposure for each of the 409 tehsils, each of the

272 single-member constituencies for the national assembly, and each of the 577 single-member provincial assembly constituencies in two ways: the percent of area flooded and the percent of population exposed. As each method of assessing flood impacts entails some error we report all results for multiple different measures.17

16The tehsil is the third level administrative unit in Pakistan, below provinces and districts. 17Based on the population measure we also created two dichotomous variables: one indicating whether at least

14 These objectively calculated variables underestimate the floods’ impacts in steep areas where the flood waters did not spread out enough to be identified with overhead imagery but where contemporaneous accounts clearly show there was major damage in river valleys. In Upper Dir district in KPK, for example, the UNOSAT data show no flooding but contemporaneous media accounts and survey-based measurements clearly indicate the floods did a great deal of damage to structures that were placed well above the normal high-water mark but still close to rivers (see e.g., Agency for Technical Cooperation and Development, 2010). Note that under the null that the floods had no impact on citizen attitudes this kind of measurement error will attenuate our estimate of flood impacts because we are counting places as having low values on the treatment where the floods actually had substantial effects.

Subjective Reports

To exploit variation in ﬂood impacts at the household level, we also asked respondents how the

ﬂoods impacted them personally. In analyzing our survey data we use the following question to measure respondents’ subjective assessments of ﬂood damage:

“How badly were you personally harmed by the ﬂoods?” (response options: “extremely

badly,” “very badly,” “somewhat badly,” “not at all”)

In addition to treating these responses as an ordinal variable, we created two dichotomous variables of subjective flood exposure: one indicating whether a household was at least “somewhat badly” affected and the other indicating whether a household was at least “very badly” affected.

Responses to this question correlate well with other self-reported measures. We asked respondents to rate how much money they lost as a result of the ﬂoods on an ordinal scale: less than 50k

Pakistani Rupee (Rs.), 50k Rs. to 100k Rs., 100k-300k Rs., and more than 300k Rs. The Pearson correlation between that loss and the subjective measure above is quite high (r = .73). We further measured the relationship between self-reported ﬂood impacts and three measures of current economic outcomes: an asset index constructed from the household’s possession of 24 goods not speciﬁc

5% of a tehsil’s or an electoral constituency’s population was exposed to the flood (roughly the 50th percentile of all tehsils and electoral constituencies exposed) and the other indicating whether at least 40% of a tehsil’s or or electoral constituency’s population was affected by the floods (roughly the 90th percentile of all tehsils and electoral constituencies exposed). We employ these measures in the survey-based regressions to capture potential threshold effects.

15 to agricultural production (cell phones, chairs, televisions, motorcycles, etc.), monthly household income, and monthly household expenditures.18 All three are negatively correlated with self-reported

flood harm, even after controlling for a rich set of variables related to economic outcomes including tehsil fixed-effects, education, gender, literacy, numeracy, and whether the respondent was the head of the household. A one-point increase in the four-point self-reported harm scale is associated with a .39 standard deviation reduction in the asset index, a .14 standard deviation reduction in log monthly income, and a .10 standard deviation reduction in log monthly expenditures.

In the short run, we would expect the floods to have a significantly greater destructive impact on household assets than human capital. The comparatively larger reduction in household assets is thus just what we should expect if self-reported flood exposure is honest and accurate. Over the

1.5 years between the flood and our survey we could expect that rebuilding income would be easier than regaining all pre-flood assets, especially given the absence of flood insurance and the bumper crop Pakistan experienced following the floods which stabilized the rural economy (Looney, 2012).

Ex Ante Flood Risk

We use two sources for data for ex ante ﬂood risk: empirical risk based on medium-resolution historical data generated from overhead imagery; and estimated risk based on hydrological models.

The historical data were developed by the International Water Management Institute and provide estimates of annual ﬂooding at 500m×500m resolution using MODIS Terra Surface Reﬂectance

Data from 2000 to 2009 (Amarnath, 2013). While not constructed with as many inputs or as much post-processing as the UNOSAT data, we believe they nevertheless provide a reasonable estimate of historical ﬂood exposure. The risk data were developed for the UN Environmental Program

(UNEP) and combine data on historical disasters, ground cover, rainfall, soil type, and topography to estimate ﬂood risk on a 1 (low) to 5 (extreme) scale for 10km×10km grid cells worldwide.19

3 Research Design

This section describes our identiﬁcation strategy and the statistical models we estimate for both aggregate- and individual-level outcomes. We begin by discussing the randomness of ﬂood expo-

18We omit assets related to agricultural production as many aid eﬀorts provided help rebuilding those assets. 19For details on the methodology see Herold and Mouton (2011).

16 sure conditional on controls.20 We then discuss our approach to estimating the ﬂood’s impact on aggregate-level outcomes—economic indicators from prior surveys and remote sensing as well as electoral behavior—and then how we modify that strategy to predict individual outcomes— economic indicators in our survey, expressed support for aggressively demanding government services, and political knowledge.

3.1 Conditional Independence

Our identification strategy relies on the observation that whether and how much any individual or region was affected by the floods had a large random component to it due to topographical factors, levy breaks, and strategic dam destructions which had unpredictable consequences on subsequent

ﬂows (e.g., Waraich, 2010). Once we control suitably for observables that could have been used to predict ﬂood exposure and so might have impacted economic outcomes or settlement patterns—risk estimates based on topography and hydrology, distance to major rivers, elevation, and historical

flooding over the previous decade—the remaining variance in flooding should be conditionally independent with respect to other factors influencing the outcomes we study.

To provide initial evidence on this score Figure 4 plots average ﬂood risk vs. observed exposure in

2010-11. Each column reports a different level of geographic aggregation: tehsils; national assembly constituencies, and provincial assembly constituencies. The top row shows exposure measured in terms of proportion of area exposed and the bottom row shows the proportion of the population exposed. Across all six scatter plots it is clear that there is tremendous variance in flood exposure at all but the lowest levels of flood risk. As one would expect from these plots, only 10-12% of the variance in the proportion of the population exposed in 2010-11 could have been predicted with a cubic polynomial model of estimated flood risk.

INSERT FIGURE 4 HERE

Historical measures of ﬂooding are more strongly correlated with ﬂooding in 2010-11, but still

20We use the term randomness to indicate that the conditional independence assumption is met.

17 leave a great deal of variation unexplained. To show this we estimate

Fi = α + BPi + i, and (1)

Fi = α + BΠi + i, (2)

where Fi is the proportion of the population exposed the floods in 2010-11, Pi is a vector of the proportion exposed in each year from 2000 to 2009, and Πi is that same vector of historical exposure as well as squared and cubic terms of each year’s exposure. Even the extremely flexible cubic polynomial in past flood exposure explains only 51% of the variance in the proportion of the population exposed at the Provincial Assembly constituency level, 61% at the National Assembly constituency level, and 72% at the tehsil level. As we will see in the robustness checks, controlling for past flood exposure using the last decade of floods does little to alter our results as we are identifying off of that residual variance in our main specifications.

3.2 Aggregate-Level Outcomes: Economic Indicators and Electoral Behavior

Our identification strategy at the constituency level relies on a combination of fixed-effects and constituency-level geographic controls to isolate the impact of local variation in flood intensity on electoral turnout. In doing so, we need to control for a range of locality-specific confounders. We might worry, for example, that it is easier for politicians to deliver patronage to constituencies close to rivers (which are most likely to be flooded) through a combination of water management projects and prior flood relief, making them more likely to turn out. To avoid confounding flood exposure with more fixed characteristic of constituencies we estimate the following three equations:

y2013 = α + β1Fi + γd + BXi + i (3)

∆y2013,2008 = α + β1Fi + γd + BXi + i (4)

∆y2013,2008 − ∆y2008,2002 = α + β1Fi + γd + BXi + i, (5)

where Fi is a measure of ﬂood impact and ∆Xi is a vector of geo-spatial controls (i.e., distance to major river from the constituency centroid, a dummy for constituencies bordering major rivers,

18 mean constituency elevation, standard deviation of constituency elevation, and estimated flood risk) plus the proportion of population affected in the smaller 2012 floods that occurred after our survey data collection but before the 2013 general elections. γd is a unit fixed-effect for the division, a defunct administrative unit that was larger than the district but smaller than the province. We control for the 27 divisions instead of districts because outside of Punjab, National Assembly constituencies are often aligned with district boundaries or contain multiple districts. We also estimate the same equations using turnout in the provincial elections which have roughly twice as many constituencies. The geo-spatial controls plus division fixed-effects account for 72.2% of the variance in the area flooded among National Assembly constituencies and 72.4% of the variance in population affected (and slightly less for the Provincial Assembly constituencies at 64.2% and 66.2%, respectively). We cluster standard errors at the district level to account for the high probability that the cross-constituency variance in turnout changes are different across districts as campaign activities in the election were mostly managed at the district level.

Our estimate of β1 identifies the impact of the floods to the extent that how exposed one was to the floods depended on factors orthogonal to pre-existing political factors once we condition on district-specific traits and the geographic controls. That condition seems likely to be met given two facts. First, some areas were flooded due to unanticipated dam/levee failures (some intentional by upstream and downstream land owners, others not) (e.g., Waraich, 2010). Second, the two most readily observable indicators of flood risk are distance to major rivers, elevation, and pre-existing risk assessments. We surely account for a large portion of the residual within-division differences between those who live in a flood plain and those who do not by controlling for the linear impact of those variables.

Subject to our identifying assumptions, the Equation 4 measures whether there are any sys- tematic changes at the constituency level between 2008 and 2013 associated with ﬂood exposure.

Equation 5 checks whether the trends in turnout shift differentially in flood-affected constituencies.

Since the main threat to identification here comes from location-specific trends in the political envi- ronment that might be correlated with proximity to rivers, and not to time-invariant district-level political factors which we could account for with fixed effects, Equation 5 is our preferred speci-

ﬁcation. Subject to the assumption that there was no major ﬂood event between 2002 and 2008

(which there was not), it effectively differences out all constituency-specific trends. In all tables we

19 report results for Equation 3 for completeness, though we prefer the other two as the differencing removes constituency-specific factors more effectively.

We further demonstrate that: (1) our main turnout results are robust to two matching approaches, traditional nearest neighbor matching on propensity to be ﬂooded based on historical

ﬂood exposure (Abadie and Imbens, 2011) and a support-vector machine based method designed to isolate the largest matched sample of the data with continuous treatment variables (Ratkovic,

2012); (2) selection on unobservables would have to be quite large to fully account for the main effect (Altonji, Elder and Taber, 2005); and (3) the results are consistent across subsets of the data over which the presence of likely confounders should vary a great deal. Specifically, when we compare places that were similarly likely to be affected by the 2010 flooding due to proximity to rivers, some of which were badly damaged and others spared, our core results remain consistent.

3.3 Individual-Level Outcomes: Economic Indicators, Vignettes Experiment, and Political Knowledge

Our identification strategy at the individual level is similar to that at the aggregate level in that we use district fixed-effects and respondent-level controls to isolate the effect of local variance in

ﬂood impact that is unrelated to average ﬂood risk.

For the economic indicators and the political knowledge index our estimating equation is a

ﬁxed-eﬀect regression

Yi = α + β1Fi + γd + BXi + i, (6)

where Fi is one of our flood exposure measures, γd is a district fixed-effect for objective flood treatments and a tehsil fixed-effect for subjective flood exposure measures, and Xi is a vector of demographic and geographic controls to further isolate the impact of idiosyncratic flood effects by accounting for the linear impact of those variables within tehsils. Geographic controls that enter at the tehsil level include the same controls as in the aggregate regressions, just calculated for the tehsil. Respondent-level demographic controls include gender, a head of household dummy, age, a literacy dummy, a basic numeracy dummy (measured by a basic mathematical task), and level of education. We cluster standard errors at the PSU level.

20 For the vignette experiment our measurement approach leverages a difference-in-difference es- timator to answer the following question: given that people are generally opposed to the aggressive vignette, is the difference in reactions between the aggressive and peaceful vignettes smaller for people in areas exposed to the flooding? To answer that question we need to control for a range of potential confounders. For example, we might worry that in districts close to rivers (which are most likely to be flooded), people are generally more accepting of aggressive action because that location is less desirable and so people living there tend to be poorer and more marginalized. It actually seems unlikely that land near the rivers is undesirable; it is more fertile and population density is substantially higher near major rivers. However, that then raises the concern that people more likely to be affected would tend to be wealthier and less marginalized. In either case, we risk confounding flood exposure with a more fixed characteristic of the region and the people who reside there, a challenge given that we have survey data from a single cross section.

We therefore estimate the following as our core speciﬁcation for analyzing the vignette experiment:

Yi = α + β1Vi + β2Fi + β3(Vi × Fi) + γd + BXi + i, (7)

where i indexes respondents, Yi represents a response to either the eﬀectiveness or approval variable,

Vi is an indicator for whether an individual received the protest vignette, Fi represents a respondent’s flood exposure (either objective or self-reported), γd represents a district fixed effect for objective flood treatments and a tehsil fixed-effect for subjective flood exposure measures intended to capture regional differences in baseline propensities to express approval or perceived effectiveness, and Xi is a vector of demographic and geographic controls to further isolate the impact of idiosyncratic flood effects. We again cluster standard errors at the PSU level since that is the level at which the vignette was randomized.21

The estimate of β3 in these equations isolates the causal impact of the flood to the extent that: (1) which vignette a respondent got was exogenous to their political attitudes; and (2) how exposed one was to the floods depended on factors orthogonal to pre-existing political factors once we condition on district-specific traits and the geographic controls. The first condition is true due

21Results are robust to clustering at the district level to account for the high possibility that the variance attitudes is highly correlated within districts as well as within PSUs.

21 to random assignment of the survey treatment. The second condition is likely to be met for the reasons outlined above. At the individual level, we can also estimate the model with PSU fixed- effects, thereby exploiting variation in flood effects at the household level. Both our discussions with those involved in flood relief and surveys done to assess post-flood recovery needs show there could be huge variation in damages suffered at the household level that were not anticipated (Kurosaki et al., 2011), likely due to minor features of topography that impacted flow rates, how long areas were submerged, and so on. All the individual level results pass the same checks to enhance the plausibility of our claim that we have isolated the impact of the exogenous component of the floods, including: showing that selection on unobservables would have to be unrealistically large relative to observables to fully account for the result; and restricting the analysis to tehsils alongside rivers

(or the ones bordering them) as doing to eﬀectively compare places that were roughly similar ex ante in terms of the threat of ﬂood exposure.

4 Economic Impact of the 2010-11 Floods

We begin our main analysis by showing that the 2010-11 floods had little economic impact one or two years after they occurred. As discussed in the background section, the international, national, and local responses were quite effective in the immediate aftermath of the floods. More importantly, once the initial response was complete the government of Pakistan embarked on very effective recovery programs in both 2010 and 2011. The Government of Pakistan and the UN Food and

Agriculture Program managed a seed distribution effort which ensured that farmers in affected areas did not miss the important winter wheat (rabi) planting season. The FAO distributed wheat and vegetable seed packets to over 480,000 flood-affected households following the 2010 floods and provided assistance to 290,000 livestock owners (Food and Agriculture Organization of the United

Nations (FAO), 2011). A number of relief organizations also supported eﬀorts to repair irrigation structures by paying laborers for their time restoring small-scale on-farm irrigation channels.

At the individual level estimating Equation 6 we find that objective measures of flood exposure had no impact on self-reported income or expenditures in 2012 and only a small negative effect on assets. Figure 5 plots the marginal impact of two measures of flood exposure (area of tehsil exposed and proportion of tehsil population exposed) on an index of 18 non-farm assets, self-

22 reported monthly income, and self-reported household expenditures in the month before the survey.

The only variable where the floods seem to have a marked negative impact is on household assets among farming families. Appendix Tables A.2 and A.3 report the full results as well as the impact of self-reported flood exposure. There we find that self-reported flood exposure is strongly negatively correlated with assets and income (both statistically and substantively) for both farmers and non- farmers, but that the conditional correlation with expenditures is much noisier. So, while it may be that those individuals particularly hard-hit by the floods suffered enduring economic damage, there is little evidence that the economy as a whole was still suffering in January and February

2012 in ﬂood-aﬀected regions relative to otherwise comparable tehsils.

INSERT FIGURE 5 HERE

While our survey only provides a cross-section, the DMSP-OLS data on night lights and the

Punjab MICS provide ways to measure economic activity before and after the floods. As we can see in Figure 6 there is little evidence the floods led to enduring economic changes.22 The left-most column shows that while areas which would be affected by the floods in 2010-11 have lower illumination (conditional on controls) than other areas in 2009 and 2012 (one year before the floods vs. one year after), those differences are consistent over time and the point estimate on the difference-in-differences is very close to zero. Turning to the individual-level data from a tehsil-representative sample of Punjab in 2007 and 2011, there is little evidence of differential change across the three variables captured identically in both surveys: a household wealth index, the individual-level employment rate, or the individual-level unemployment rate.23 All of the point estimates on the difference-in-difference are close to zero or positive in the case of the employment rate. Income measured pre-flood shows no statistically significant difference across the eventual

ﬂood gradient nor do expenditures captured in a full-scale consumption module during the post-

flood survey.24 While only the left-most column of Figure 6 reflects a national-level outcome, the fact that neither our survey nor the massive MICS survey in Punjab show consistent differences in economic activity across the flood gradient suggests that the government and international response

22Appendix Tables A.4-A.7 provide full regression results as well as replicating the analysis using robust regression to ensure they are not driven by high-leverage observations. 23These are calculated as employed/population or employed/(employed + unemployed and looking for work), respectively. 24Not the 2011 ﬂoods had little impact in Punjab so the 2011 MICS is almost entirely post-treatment.

23 and recovery eﬀorts worked quite well.

INSERT FIGURE 6 HERE

5 Political Impact

Although we find little evidence of persistent economic consequences from the 2010-11 floods, at least in the aggregate, we do find substantial evidence of political effects. Flood exposure appears to have led to increased turnout in both NA and PA elections, as Figure 7 demonstrates. The

first row of the figure shows the marginal effects of flood exposure (measured in terms of area or population) on turnout in 2013 at the NA- and PA-constituency levels estimated using Equation 3.

The second and third rows show the same marginal eﬀects on trend in turnout from 2008 to 2013

(Equation 4) and on the change in that trend (Equation 5). The left-most column reports these results for the full sample, the middle column shows the results for constituencies neighboring major rivers, and the right-hand column shows them for constituencies bordering those.25 As is clear from the left-hand column turnout is generally increasing in ﬂood exposure and the eﬀect becomes more pronounced going from levels to trends to changes in trends.

INSERT FIGURE 7 HERE

The results are quite large substantively. A one standard deviation increase in the proportion of the population aﬀected by the 2010-11 ﬂoods (.127) predicts 2.1 percentage point higher turnout in

2013 vs. 2008 in the National Assembly election, which is almost a .25 standard deviation increase in the outcome, and a 3.9 percentage point difference in the trend in turnout, roughly a .28 standard deviation increase. These effects are in line with the effects observed in get-out-the-vote campaigns in the United States. Green, Gerber and Nickerson (2003), for example, found that concerted door- to-door canvassing efforts in six sites (areas ranging from 8,000-43,000 people) yielded an average turnout increase of 2.1 percentage points.

In many countries, voting down ballot is an indicator of intensity of civic engagement as it requires more time in the voting booth and it is harder to collect information on candidates for

25Appendix Tables A.8-A.11 provide the full regression results behind the ﬁgure as well as showing the robust regression results.

24 less prominent oﬃces. In the Pakistan context, voters received two ballots, one for the National

Assembly and one for the Provincial Assembly, and turnout in the PA election was 2 percentage points lower (54.1% vs. 56.1%). Examining the impact of floods on PA turnout may thus yield a better signal of the flood’s affects on civic engagement. Moreover, outside of Punjab the National

Assembly constituencies are often quite large geographically. PA constituencies are typically much smaller; there are roughly twice as many across the country. We therefore test the impact of the

ﬂoods on turnout for PA constituencies as well. As Appendix Figure A.1 shows, there is substantial variance in ﬂood impacts at the PA constituency level that is masked at the NA constituency level.

Given that we have more observations in the PA elections, it should not be surprising that the results are statistically stronger. A one standard deviation increase in the proportion of the population affected by the 2010-11 floods (.146) predicts a 1.6 percentage point increase in turnout from 2008 to 2013, which is almost a .18 standard deviation increase in the outcome and a 3.6 percentage point difference in the trend in turnout, roughly a .24 standard deviation increase. On average, the floods increased turnout in the 2013 PA election by 1.5 percentage points and based on our estimates the flood accounts on average for 14.65% of the turnout increase (with a 95% confidence interval of 6.3% to 22.9%).

Of course, a concern with these results is that since our fixed effects are not entering at the unit level some of the differences that we see in turnout may reflect enduring differences between those who live in flood plains and those who do not. Though this seems unlikely to be the case given our controls, we re-estimate our core specification on the subset of constituencies that border major rivers in the middle column.26 The results in the middle column are thus identified by differences in the specific course of the 2010-11 floods among a set of places that are all unambiguously within potential floodplains. Comparing what we can think of as lucky constituencies (i.e., those neighboring rivers, but not flooded) with the unlucky ones (i.e., those neighboring rivers that were hit by the flood) produces similar estimates of the floods’ impact on voting to what we saw in the full sample. The next set of constituencies out from major rivers provide another apt comparison, what we can think of as normal tehsils (i.e., those which would not have expected to be flooded and were not) from surprised ones (i.e., those which were flooded and would not have expect to be

26Major rivers include the Indus river and its arms (i.e., Chenab, Jhelum, Kabul, Ravi, Soan, and Sutlej).

25 given that they did not border a river).27 Within this subset the difference in levels disappears, but the difference in trends is more than twice as strong and the change in trends stronger still compared to those constituencies bordering major rivers. These last two findings provide strong evidence that the main result is not driven by the proximity to rivers or something unusual about populations who live in flood plains.

6 Robustness

How robust are the turnout results? We address this concern in several ways.

6.1 Controlling for Empirical Risk

An obvious concern is that we may not have adequately accounted for selection of specific populations into high-risk areas, and so what we are capturing is that populations who were particularly likely to turn out in 2013 for other reasons were also more likely to settle in risky regions. We address that concern first by controlling for empirical flood risk based on the last decade’s flood history. To calculate a univariate measure of empirical risk we take the predicted values from equation 1 which predicted the proportion of the population exposed to flooding in 2010-11 as a linear function of the proportion effected in each of the last 10 years.28 This estimate is akin to what a reasonable person living in the area might do by taking a weighted average of past flood events.

Appendix Tables A.12 and A.13 show that controlling for this measure of empirical risk makes the results a bit stronger for the NA and PA elections, respectively. Each table reports the results of our baseline turnout regressions for our two standard measures of flood exposure (columns 1 and 5). We then sequentially add the UNEP measure of flood risk in linear (columns 2 and 6) or cubic polynomial fashion (columns 3 and 7). Finally, we restrict the analysis to the region of common support on this empirical measure of flood risk. None of these changes substantially alter the results. 27Given the geography of Pakistan it is not possible to conduct this test in a meaningful way for one more constituency removed from the rivers. 28Note, the empirical risk is distinct from the UNEP estimated flood risk which we control for in our baseline specification. The two correlate at ρ = .23 for NA constituencies and ρ = .26 for PA constituencies.

26 6.2 Matching on Historical Flood Exposure and Sensitivity Analysis

We also conduct a more formal matching exercise on objectively-measured ﬂood history in three ways. First, we categorize districts as aﬀected if they had any population at all exposed in the 2010-

11 floods and then match affected to unaffected units using the Mahalanobis distance between their

ﬂood histories (i.e., the proportion of the unit’s population exposed to ﬂooding in each year from

2000 to 2009).29 Second, we conduct nearest neighbor matching on aﬀected using the empirical ﬂood history constraining pairs to be matched exactly within provinces (Abadie and Imbens, 2011).30

Third, because our treatment variable is continuous, we us the support-vector machine method introduced by Ratkovic (2012) to identify the largest subset of the data that is balanced on each element of the empirical ﬂood history.31 We then regress turnout on the 2010-11 ﬂood exposure within the matched subset, clustering standard errors at the district level.32

INSERT TABLES 1 AND 2 HERE

As Table 1 shows for the NA constituencies and Table 2 shows for the PA constituencies, the matching results are broadly consistent with our core specification. At the NA level, there is little consistent difference in 2013 turnout between affected and unaffected constituencies, the difference is negative for Mahalanobis matching, negative for nearest neighbor matching, and essentially zero for SVM matching. Both the trend in turnout and the change in the trend are strongly positively correlated with flood exposure across all our matching estimates for the NA constituencies. Within the matched subsample of 56 constituencies identified by the SVM algorithm, a one standard deviation increase in percent of population exposed in that sample (.163) predicts a 3.2 percentage point increase in voter turnout from 2008 to 2013, a .6 standard deviation movement within that sample. At the PA constituency level the matching results are weaker, both substantively and statistically. Neither the Mahalanobis nor nearest neighbor matching results for the change from

2009 to 2013 are signiﬁcant. The change in trend, however, is in the expected direction for those

29Standard errors are calculated per Abadie and Imbens (2006) using Leuven and Sianesi (2003). 30The intuition is that because the PDMAs differed in their capacity an unaffected constituency in Sindh would make a poor counterfactual for an affected constituency in Punjab even if they had very similar past exposure. 31This technique flips traditional propensity score matching around by using an extremely flexible machine learning algorithm to classify units. The balanced subsample is identified by selecting units close enough to the decision boundary that their assignment status cannot be distinguished from sampling noise. The algorithm outperforms most commonly used methods on several canonical datasets (Ratkovic, 2012). 32Clustering at the division level yields similar results.

27 estimators and the SVM results are consistent with our regression results for both the trend and the change in trend. At the PA level a one standard deviation increase in percent of population exposed in the 141 matched constituencies (.118) predicts a 1.8 percentage increase in voter turnout from 2008 to 2013, a .2 standard deviation movement within that sample. Overall the matching analysis increases our conﬁdence in the main result.

As an additional robustness check we conduct the sensitivity analysis suggested by Altonji,

Elder and Taber (2005) and show that a potentially omitted variable to fully explain the effect of the flooding on the change in NA turnout from 2008 to 2013 its impact would have to be 2.5 times as large as the bias removed by adding our vector of geospatial controls on top of district fixed effects and 7 times as large for the PA results.33

6.3 Is This Just a Compositional Eﬀect?

An immediate concern with any analysis of the impact of a natural disaster which are not based on panel data is that we may simply be picking up a compositional effect. If people who moved out after the floods were systematically less likely to vote than those who stayed put (or moved in), then the changes we are attributing to the flood’s impact on individual civic engagement could actually be an artifact of those migration decisions. There in no evidence in surveys designed to study migration that there were significant permanent population shifts in Pakistan due to the

2010-11 floods, either to or from flood-affected districts (Mueller, Gray and Kosec, 2013). Less than 2% of those reporting their village was hit in the 2010 or 2011 floods in the Mueller, Gray and

Kosec (2013) nationally representative panel study were living in a diﬀerent village than in 2001.34

Fortunately, our survey provides some evidence on migration, allowing us to assess if compositional effects are driving our results. A number of respondents reported suffering from flood damage who lived in places that were not affected by the floods in 2010 or 2011 according to a combination of UNOSAT data and maps from the Pakistan National Disaster Management Agency. Of the 1,201 respondents who reported being hurt “extremely badly” by the floods, only 75 lived in a tehsil that was not hit by the flood. Of the 2,360 who reported being “very badly” or “extremely badly” hit,

33Appendix Table A.14 provides results for the full set of outcome variables—levels, trends, and changes in trends— across both measures of ﬂood exposure—area exposed and population exposed—for both elections—NA and PA. For both trends and changes in trends all the ratios are well away from the interval [−1, 1]. 34Private communication with the authors.

28 only 170 live in a tehsil not hit. These numbers are inconsistent with massive outmigration from

ﬂood-aﬀected areas.

If we assume that all those reporting any damage who live in unaffected districts migrated because of the flood, then we estimate that 4.6% of the population in unaffected districts are migrants from the flood-affected regions and that a total of 2.05% of Pakistan’s population migrated as a result of flood damage. This is surely an overestimate as many of those who report being affected but live in districts with no flooding either moved for other reasons, are referring to damage suffered by kin, or answered based on damage suffered from monsoon rains in the summer of 2010.

Still, we can use our estimates of migration to benchmark the diﬀerence in turnout attributable to the impact of the ﬂood.

The simplest way to do so is to estimate the migration rates for the 61 districts in our survey

(recall the sample was designed to be district representative) and include the estimates of the proportion of migrants in a district to our National Assembly turnout regressions. If people who moved out were less likely to vote, then we should see a negative conditional correlation between number of migrants in unaffected communities and the change in turnout from 2008. As Panel A of Appendix Table A.15 shows, we do not observe this. Instead, the coefficients on migration is positive and insignificant in all regressions. We can also estimate our baseline turnout regression for the 157 NA constituencies that we estimate did not receive any migrants (i.e., they are in districts we surveyed that were either clearly hit by the floods or that had no one report flood exposure).

As Panel B shows the core results remain substantially unchanged within that sub-sample. Panel

C repeats the analysis of Panel B for the 378 PA constituencies that were either hit or reside in a district where we found no migration. The results remain robust in that subsample.

We should close this section by noting that the particulars of Pakistan’s voting system also make it unlikely that compositional changes are driving the results. The major door-to-door voter registration effort by the Electoral Commission of Pakistan for the 2013 election occurred from Au- gust 22 to November 30, 2011 (mostly after the 2011 floods). Voters were registered at the address on their national identity card and anyone not home during the door-to-door drive could register until March 22, 2013 at their local electoral commission office by providing a national identity card.

Because changing the address on one’s national identity card is a relatively cumbersome process

(it requires visiting an oﬃce with either proof of property ownership or a certiﬁcate from a local

29 government representative), many people choose to vote where they were registered rather than shifting that address. This registration process means that if those who moved out were dispro- portionately inclined not to vote, then their registration would likely remain in ﬂood-aﬀected areas

(there would be, after all, no reason for them to shift their registration if they do not plan to vote).

That would introduce a downward bias to our estimates of turnout, therefore making our main estimates of the eﬀorts of the ﬂoods on turnout conservative.

Overall, there is little evidence that we are detecting a compositional eﬀect, though we cannot rule it out without better information on migration patterns. As we will see below, the attitudinal changes related to ﬂood impacts are consistent across both objectively-measured and self-reported

ﬂood exposure, suggesting again that we are not simply capturing the impact of diﬀerential migration.

7 Potential Mechanisms

So what is driving the increased turnout in flood-affect constituencies? This section shows first that there is little evidence to suggest these results are driven by citizens rewarding aid spending. We then turn to evidence that citizens’ views of civic engagement change in ways that have behavioral consequences at the individual level.

7.1 Does Increased Turnout Reﬂect a Reward for Relief and Recovery Eﬀorts?

First, we find only modest evidence that voters rewarded the government for what was generally considered to be an effective response (particularly compared to previous floods in Pakistan). At

first glance, the incumbent Pakistan Peoples’ Party (PPP) seems to have done better in flood- affected areas. They lost massively in the 2013 election, losing 16.5 percentage points of vote share from 2008 in the average constituency, but their loses were roughly 25% smaller in flood-affected constituencies on average. However, once we break the results down by province, a much more complicated picture emerges, as Figure 8 highlights by plotting the mean and 95% confidence interval for each party’s vote share nationally and by province in each election. All the endogeneity concerns which lead us to need to model turnout also apply to vote shares, of course, and when we turn to conditional correlations between the intensity of floods and party performance, an

30 ambiguous story emerges.

INSERT FIGURE 8 HERE

Once we control for potential confounders using our standard estimating equation it appears that the two main enduring national-level parties which were in government at the national level

(PPP) or provincial level (PPP in Sindh and Balochistan, PML-N in Balochistan) when the floods hit gained vote share disproportionally in flood-affected areas relative to 2008, as Table 3 shows.

While the statistical strength of the results are inconsistent across levels of aggregation—the conditional correlation between the combined national party vote share and area ﬂooded is statistically signiﬁcant at the 10% level in the NA elects but not the PA ones, and the opposite is true for the conditional correlation with population exposed—they are substantively meaningful. A one standard deviation increase in the proportion of the population exposed at the PA level (.146) predicts a

3.5 percentage point increase in combined national party vote share, a .12 standard deviation shift.

Breaking these results down by party, however, shows that the effects are quite noisy, rarely reach traditional levels of statistical significance, and flip sign in one specification, as Table 4 highlights.

INSERT TABLES 3 AND 4 ABOUT HERE

Second, there was no obvious impact of aid spending on turnout. To measure aid spending we used district-level data on aid disbursements provided by the National Disaster Management

Authority. Using these data we constructed three measures of aid disbursements: the standardized amount of food relief in 1,000 MT units, an additive index of the standardized values for eight categories of shelter relief, and an additive index which combines the standardized amount of food relief with the shelter index.35

The distribution of this aid varied quite a bit across levels of ﬂooding. To see how much we estimate:

yi = βpi + π + i (8)

where each relief variable yi is a function of the proportion of the population exposed pi and a set of province ﬁxed eﬀects π to account for the fact that each province had its own disaster management

35The eight categories of shelter aid were tents, tarpaulins, ropes, toolkits, blankets, kitchen kits, bedding, and plastic mats. The data record the count of each distributed by the district level.

31 authority.36 There was substantial variance in the distribution of aid across all levels of flood effects, as Figure 9 shows by plotting the residuals from those regressions on the proportion of the population exposed for NA (top panel) and PA constituencies (bottom panel). These residuals can be thought of as representing a combination of the generosity of government aid provision to each area and the intra-district variance in flood impact. Our aid data are at the district level and so may mask intra-district variation in disbursements that are correlated with constituency-level flood exposure.37

INSERT FIGURE 9 HERE

Using the relief data we examine how our baseline results on the relationship between the proportion of the population affected and turnout varies if we: (1) control directly for relief efforts at the district level; and (2) control for the residuals which represent, in part, the excess aid distribution to particular places, perhaps as an effort to earn citizen loyalty. Since relief efforts are obviously quite correlated with the proportion of the population affected (r2 = .25 in a regression of food relief on population exposed without province fixed effects for NA constituencies and .42 with them), we should expect that the coefficient on population exposed would be attenuated when we control for our relief variables. If excess aid distributions are driving the results, the estimated impact should also be attenuated when we control for the residuals from equation 8. If, however, something other than gratitude for immediate aid is driving the effect, then controlling for those residuals should have no impact on our estimates.

INSERT TABLES 5 AND 6 ABOUT HERE

We find little evidence that relief aid is the mechanism underlying our results. In Table 5 for NA constituencies and Table 6 for PA constituencies we see that across all three measures of relief controlling for the level of relief attenuates the estimated impact of the floods (columns 2, 4, and 6) compared to our baseline (column 1), but controlling for excess aid distributions leaves the results unchanged or makes them a bit stronger (columns 3, 5, and 7). We also find no difference in responses across the flood gradient to a question asking respondents “In your opinion, did the

36The capabilities of the district offices varied widely according to most accounts. 37Higher resolution aid data were collected during the recovery effort but were erased when the NDMA changed its website and computer systems in early-2013. Personal communication with NDMA officials, November 15, 2013.

32 government do a good or bad job in responding to the ﬂoods after they occurred?” on a four- point scale ranging from “very bad” to “very good” with no midpoint so that respondents were forced to assign a direction to their views of the government response.38 It therefore seems unlikely that gratitude for preferential aid spending in the immediate aftermath of the ﬂoods is driving the results.

7.2 Does the Increased Turnout Reﬂect Deeper Changes in Civic Engagement?

We provide two pieces of evidence that the results are driven by a change in the importance individuals attach to inﬂuencing government action.

7.2.1 Behavioral Changes

We find clear evidence of behavioral change among flood-affected individuals in so far as they seem to have invested more in acquiring political knowledge, as Table 7 highlights. Our main index of political knowledge is increasing across a range of objective and subjective measures of flood impact and highly statistically significantly so in 5 of 7 measures. The effects are modest in magnitude.

A one standard deviation increase in the proportion of the population aﬀected by the ﬂoods in the surveyed teshils (.136) predicts a .061 increase in the knowledge index, a move of .08 standard deviations. In order to ensure that the results are not an artifact of the PCA procedure, panel

B reports the same regressions with a simple additive index of the 10 knowledge questions. The results are substantively similar.

INSERT TABLE 7 ABOUT HERE

Clearly citizens hit hard by the ﬂoods invest more in acquiring political knowledge.

7.2.2 Attitudinal Changes

Using a vignette experiment, we also find that citizens exposed to floods are significantly more supportive of aggressive approaches towards demanding public services and believe them to be more effective than their non-exposed counterparts (see Table 8). The first coefficient in each model, β1 from Equation 7, measures the difference in the outcome variables between the aggressive and

38Results available on request.

33 peaceful vignettes for people who scored a zero on the flood exposure measures. As shown in the top row of Panel A, those who received the aggressive vignette but did not experience flooding rated the effectiveness of Junaid’s actions between .2 and .25 units lower in the violent vignette on a 0-1 scale controlling for a broad range of geographic and demographic controls, which is a movement of more than .5 standard deviations in all models. That effect is consistent across various objective (columns 1-4) and self-reported measures (columns 5-7) of flooding. We find results of similar magnitude for the approval dependent variable (see Panel B).

INSERT TABLE 8 ABOUT HERE

Across a broad range of observational and self-reported measures, however, exposure to the flood substantially and significantly lessened this disapproval. The coefficient on the third variable in each model, β3 from 7, indicates the moderating impact of flood exposure on the effect of the aggressive vignette. A one standard deviation movement in the proportion of the population exposed in tehsils with non-zero flood exposure (.17) corresponds to a .16 standard deviation increase in perceived effectiveness of the violent action and a .18 standard deviation increase in approval for the aggressive approach.

To benchmark these results consider the relationships between the vignette response and gender.

Existing research has shown that men are on average more likely to have more aggressive attitudes in the context of normal social relations (Funk et al., 1999) and tend to have more extreme views in some political settings involving violent contestation (Jaeger et al., 2012). The difference in perceived effectiveness of the aggressive vignette between men and women is approximately .08, which equates to a .2 standard deviation movement in effectiveness, and the difference in approval is of similar size (i.e., .07). The difference between those affected by the flood and those who were not in terms of approval is thus slightly smaller than the gender difference in the approval of aggressive action. The gender difference in perceived effectiveness and approval across the two conditions is even smaller, roughly .06 for effectiveness and .04 for approval, both of which are substantially smaller than all the flood coefficients. Drawing on prior work we can also compare the flood affects to differences across attitudes towards Islamist militants’ political positions. As in Fair, Malhotra and Shapiro (2012) we measured individuals’ support for five political positions espoused by militant Islamist groups and combined these in a simple additive scale ranging from 0

34 to 1. Moving from 0 to 1 on this scale equates to a .21 increase in approval for the violent vignette and a .11 increase in effectiveness. The impact of a one standard deviation move in flood exposure is similar in terms of approval of the aggressive vignette to the movement in support for militant groups when someone goes from agreeing with none of the Islamist policy positions to agreeing with all five (and is much larger on the effectiveness measure), which indicates a substantively significant shift.

Interestingly, the results are not a proxy for satisfaction with flood relief. We asked respondents “In your opinion, did the government do a good or bad job in responding to the floods after they occurred?” on a four-point scale ranging from “very bad” to “very good” with no midpoint so respondents were forced to assign a direction to their views of the government response. Re- spondents’ feelings about the violent vignette are not consistently correlated with how respondents believe the government did in responding to the floods. For some measure of flood effects the difference-in-difference is larger among the 6,158 respondents who felt the government did a poor job of responding to the floods (about 50% of the sample), while for others it is higher among the

6,149 respondents who felt the government did a good job. Clearly we cannot interpret the lack of a difference as falsifying a causal relationship between the quality of government response and attitudes on the vignette. Individuals who rate the government response poorly may do so because they have some unobservable difference that also makes them more approving of violent protests to gain political services. Nevertheless, the fact that there is no consistent correlation suggests that the floods affected attitudes (and observable measures of civic engagement as we will see) through some channel other than satisfaction with government performance.

The finding that flood victims approve of violent protests and believe they are more effective in getting a government response is quite robust. One might be concerned, for example, that there is unobserved heterogeneity between tehsils which is driving these results. To account for this possibility and to exploit the substantial within-village variation noted by many observers (Kurosaki et al., 2011), we estimated the impact of self-reported flood measures on the vignette experiment including tehsil fixed effects. As shown in columns 5-7 of Table 8, the results on self-reported flood effects actually become stronger once we account for tehsil-level variance in flood impacts.

It appears clear that citizens hit hard by the ﬂoods developed more assertive attitudes about demanding government services and invest more in acquiring political knowledge. Both changes

35 should shift politicians’ incentives. Notice too that in equilibrium those changes can alter service delivery without aﬀecting vote shares. It is easy to imagine a situation in which an exogenous increase in citizen attention leads politicians on all sides to exert increased eﬀort, leading to increased turnout but leaving equilibrium vote shares unchanged.

8 Conclusion

We have shown that in the case of a massive natural disaster in which the government (and international) response was very effective at ameliorating economic impacts one to two years on, there was still a lasting political impact. Turnout and civic engagement increased in flood-affected districts compared to similar unaffected districts. These effects were particularly strong in the subset of places that had a low ex ante risk of being flooded (i.e., those places that were genuinely surprised by the flood).

Examining underlying mechanisms, we found little evidence these changes reflect citizens rewarding politicians for delivering large amounts of aid. Instead, the evidence points to changes in political attitudes. Citizens exposed to the flood know more about politics, reflecting a greater investment in acquiring political knowledge, and developed more assertive attitudes about demanding government services. We argue that this evidence points towards an informational channel: flood exposure highlights the importance of a responsive government and affects citizens’ belief about how likely they are to require government assistance in the future. Flood exposure therefore provides incentives to invest in political knowledge and become more politically active in order to ensure a responsive government. Both changes should shift politicians’ incentives. The implication is that natural disasters can alter the political equilibrium even when effective government actions forestalls major economic consequences.

These results should call into question the interpretation of a broad set of papers that use natural disasters as a source of variation in economic conditions that is plausibly exogenous to political factors. The economic impact of disasters can obviously be highly contingent on government response, yet even when it is eﬀective at reducing those impacts we can still see large changes in citizens’ political attitudes and behavior. Thus, the exclusion restriction in a number of recent papers on political liberalization and democratic transition is clearly violated in at least one case

36 and almost certainly in many others. These results buttress the rich literature in political science showing that disasters can have complicated political eﬀects that are often quite divorced from economic impacts.

All of this is good news for policy makers worried that natural disasters in weakly institu- tionalized countries will undermine democratic institutions. Exposure to natural disasters might actually highlight the necessity of governmental services and strengthen citizen’s willingness to demand government responsiveness and increase their engagement in democratic process. Future research needs to probe this possibility explicitly in three ways. First, it is important to get direct evidence of the proposed informational mechanism. We have recently completed a survey that captures how expectations about needing government services changed across the flood boundary, which should provide tighter evidence on that score. Second, it is important to assess whether the increase in political engagement we identify actually resulted in policy changes. In future work we plan to examine differences in the provision of local goods and services across the flood boundary.

Finally, it will be important to assess how long these eﬀects last. What we know now is that the

floods had enduring effects two years on, but evidence from work in Europe suggests we may expect effects long after that (Bechtel and Hainmueller, 2011).

Our results also speak to three additional literatures. First, this paper advances the broad literature on the political impact of natural disasters on two distinct fronts. Researchers studying whether voters objectively evaluate politicians’ performance and respond accordingly have become increasingly interested in how voters react to natural disasters because their occurrence is exogenous to politicians’ actions.39 Achen and Bartels (2004), for example, ﬁnd that extreme droughts and

ﬂoods cost incumbent U.S. presidents about 1.5 percentage points of the popular vote. They theoretically interpret this empirical pattern as an example of “blind retrospection,” or voters failing to hold leaders accountable only for conditions over which they have direct control and responsibility.

Healy and Malhotra (2010) similarly find that heavy damage from tornadoes decreases presidential vote share by about 2 percentage points. Subsequent studies have found that voters may in fact be reacting to the government’s actions in responding to the disaster, suggesting that citizens sanction elected officials only when they fail to adequately address the negative effects of disasters (Gasper and Reeves, 2011). For instance, politicians are rewarded for providing relief payments in the wake

39Though not their consequences, of course.

37 of a disaster (Healy and Malhotra, 2009; Bechtel and Hainmueller, 2011; Cole, Healy and Werker,

2011) or even for providing distributive spending under the guise of relief eﬀorts (Chen, 2013).

With a few exceptions this research is built on studies of the effect of natural disasters in advanced developed democracies, a gap we help to fill by focusing on a young developing democracy with weak democratic institutions and active militant groups.40 In addition, the vast majority of these studies focus on the political outcome of this exogenous event (e.g., Lay, 2009; Reeves, 2011; Velez and Martin, 2013) and not the underlying causal mechanism. Ours is the first study we know of which measures political attitudes in the period between a highly salient natural disaster and a subsequent election. This enables us to measure how the flood impacted attitudes relevant to interpreting the electoral outcome in the disaster’s aftermath. Consequently, we shed light on the mechanism through which disasters affect elections.

Second, we provide valuable evidence on the question of what drives governments from patron- client systems—which focus on providing targeted benefits to supporters at the cost of services with larger collective benefits—to programmatic systems focused on effective service provision.

Most work on subject has focused on elite bargaining and has left unexamined how changes in citizens’ preferences impact elite incentives (e.g., Shefter, 1977; Acemoglu and Robinson, 2012).

Yet, as Besley and Burgess (2002) show theoretically and empirically, more informed and politically active electorates create strong incentives for governments to deliver services.41 The evidence from

Pakistan, a country long considered a stronghold of patronage politics, suggests that exogenous events can create just such changes in the electorate. Situated on a range of geo-political fault lines and with a population of more than 180 million people, scholars and analysts variously describe

Pakistan as a struggling military-dominated democracy, a revisionist nuclear power locked in a security competition with a nuclearized India, and a failed or failing state (Ziring, 2009; Narang,

2013; Shah, 2013). The 2013 elections there were the ﬁrst time a freely elected parliament served its term and handed power to another elected government. It was, in other words, the ﬁrst time that voters were able to punish/reward politicians in something even remotely close to the democratic ideal.42 Those hit hard by a natural disaster years before the election used that opportunity to

40The main exceptions are Cole, Healy and Werker (2011); Gallego (2012); Remmer (2013). 41Pande (2011) provides a review of experimental evidence showing that providing voters with information improves electoral accountability. 42The latter was a result of the galvanized public, as evidenced by unprecedented and diverse (gender, age, education) voter turnout, the rise of a competitive third party, and numerous highly contested races in previous one-party

38 vote at higher rates and to reward a government that exceeded almost all expectations about how well it would respond (relative to the national trend, of course). Such changes augur well for the emergence of programmatic politics and suggest the theoretical literature should explicitly consider the conditions under which such changes lead to diﬀerences in actual policies.

Finally, our results are relevant to the emerging academic literature on the impact of natural disasters on conflict and to government decision makers planning disaster response. Scholars in this literature typically find a positive relationship between natural disasters and conflict (see e.g.,

Miguel, Satyanath and Sergenti, 2004; Brancati, 2007; Ghimire and Ferreira, 2013), though there are exceptions (Berghold and Lujala, 2012). These ﬁndings worry many as climate change is predicted by most models to lead to a long-run increase in the incidence of severe weather-related disasters

(Burke, Hsiang and Miguel, 2013). The evidence from Pakistan suggests that eﬀective response to such disasters can mitigate such negative political consequences. In this case, the international community provided a great deal of post-disaster assistance which the state eﬀectively coordinated.

The net result was a differential increase in civic engagement by citizens in flood-affected regions.

The results thus provide micro-level evidence that aid in the wake of natural disasters can turn them into events which enhance democracy, a possibility consistent with the cross-national pattern identified in Ahlerup (2011) who finds that natural disasters are correlated with democratization in countries that are substantial aid recipients.43 Overall, our findings suggest enhanced investments in helping poor countries respond well to natural disasters could yield long-run political gains in addition to their obvious economic value.

strongholds. 43Note, this interpretation is not consistent with the theoretical arguments that aid ﬂows increase the risk of conﬂict by increasing the rents to be captured from control of the government (Besley and Persson, 2011).

39 Figures

Figure 1: Maximal Composite Flood Extent in 2010 and 2011 and Surveyed Tehsils

40 Figure 2: Standardized Impact of Floods in Pakistan 1975 – 2012 Pakistan Floods 1975−2012 Source: International Disaster Database (EM−DAT) 4 3 2 Killed 1 Affected 0 z−score killed/affected 1975 1980 1985 1990 1995 2000 2005 2010

Source: Dartmouth Flood Observatory (DFO) 4 3 2 Death 1 Displaced 0 1975 1980 1985 1990 1995 2000 2005 2010 z−score deaths/displaced Year

41 Figure 3: Balance of Covariates for the Vignette Experiment

●● % Area Flooded

●● % Population Exposed

●● Population Exposed > Median

●● Pop Exposed > 90th Percentile

●● Flood Exposure (4pt Index)

●● Affected

●● Very Bad | Extremely Bad

●● UNEP Flood Risk (5pt Index)

●● Distance to Nearest River (100km)

●● Neighboring River

●● Std. Dev. of Elevation (1000m)

●● Mean Elevation (1000m) ●● Treatment Control

●● Household Head

●● Male

●● Age

●● Can Read

●● Can Do Basic Math

●● Education (6pt Index)

●● Sunni

●● Support for Democracy (Index)

●● Religious Knowledge (Index)

●● Religious Practice (Index)

●● Jihadi (Index)

0.0 0.3 0.6 0.9 Mean

42 Figure 4: Scatter Plots of UNEP Flood Risk versus Eﬀective Flood Exposure 2010/11

43 Figure 5: Impact of Flooding on Assets, Income, and Expenditures in 2012 for All Citizens and Farmers and Non-Farmers Assets Income Expenditures

0.2

● ●● ●

0.0

● ● ● ●● ● ●● ●● ●●

% Area ●● Flooded

●● % Population −0.2 Exposed

Regression Coefficient with Confidence Intervals −0.4

−0.6

All Farmers Non All Farmers Non All Farmers Non Farmers Farmers Farmers

44 Figure 6: Impact of Flooding on Night Lights per Capita (2009 and 2012) and Various Economic Indicators in Punjab (2007/08 and 2011)

Night−Light Emission Wealth Employment Unemployment Income Expenditures 1.0 Pre−Flooding (2007/08)

0.5

●●

0.0 ●● ●● ●● ●●

−0.5

1.0 Post−Flooding (2011)

0.5

●●

●● 0.0 ●● ●● ●●

−0.5

1.0 Trend (2011−2007/08)

0.5 Standardized Regression Coefficient with Confidence Intervals Standardized

●●

0.0 ● ● ● ●● ●

−0.5

●● % Area Flooded % Population Exposed

45 Figure 7: Impact of Flooding on Turnout in the National and Provincial Assembly Elections (2013, 2008, and 2002) in All Constituencies and Two Diﬀerent Subsets

Constituencies Bordering Constituencies Next to those All Constituencies Major Rivers Bordering Major Rivers

0.8 Level 0.4

● ● ● ● ● ● ● ● ●● 0.0

●●

−0.4

0.8

Trend % Area 0.4 Flooded ●● ● % Population ● ●● Exposed ●● ●● ●● ●● 0.0

−0.4 Regression Coefficient with Confidence Intervals

0.8 ●● Change in Trend

●●

0.4 ●● ●●

●● ●● 0.0

−0.4 National Provincial National Provincial National Provincial Assembly Assembly Assembly Assembly Assembly Assembly

46 Figure 8: Major Party Vote Shares in Constituencies by Flood Exposure 47 Figure 9: Residual Plots of Immediate Relief Controlling for Percent of Population Exposed Standardized Additive Index of Standardized Additive Index of 1000 MT of Food Relief Shelter Relief Food and Shelter Relief 10

● National Assembly ● 5 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ●● ● ● ● ●● ● ●●● ● ●●● ●●●●●● ● ● ● ● ●● ●● ● ● ●●● ● ● ●● ● ● ● ●● ●●● ● ● ● ●●●●● ● ● ●● ● ● ●●● ● ● ● ●●●● ●●● ● ● ● ●● ● ● ●●● ● ● ●●●●●●●●●● ●●●●● ●● ●● ● 0 ●●● ●●● ● ●● ●●●● ●●● ● ● ● ●●●●●●● ● ● ● ● ●● ●● ●●●●●●●● ●●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●

−5 ●

−10 ● ●● ● 10 ● ●

●● ●

● ● ●● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● Provincial Assembly ● ● ● 5 ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ●● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ●●● ●●● ● ●●● ● ● ●● ●● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●●●●● ● ● ● ● ● ●●● ●● ● ● ● ● ●●● ●● ● ● ● ● ●●●● ●●●● ● ● ●●● ● ● ● ●●●● ● ●● ●●●●●●●●●●●●●● ●● ● ● ●● ● ● 0 ●●●● ●●●●●●●●●●●● ● ● ●●●●●●●● ● ● ●●●● ● ●●●●●●●●●● ●●●●● ●●●●●●●● ● ● ● ● ● ● ● ● ●● ● ●●●● ● ●●●●●● ● ●● ●●● ● ● ● ●● ●● ● ● ●●● ● ●● ● ● ● ● ● ● ● ●●●● ● ●●●● ● ● ● ● ● ● ● ●● ● ● ●● ●●●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ●● ●●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● −5 ●● ● ● ● ● Relief | Proportion of Population Affected and Province Fixed−Effects and Province Affected Relief | Proportion of Population ● ●

● −10 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Proportion of Population Exposed

48 Tables

Table 1: Robustness of Flood Effect on National Assembly Election Turnout Matching on Flood History (2000-2009) (1) (2) (3) (4) (5) Mahalanobis Nearest Neighbor SVM Matching Flood % Area % Population Treatment: Affected Affected Affected Flooded Exposed Panel A: Turnout 2013 (mean=0.56, sd=0.09) Treated - Controls -0.046*** (0.016) Flood Treatment 0.026* 0.008 0.011 0.000 (0.015) (0.014) (0.032) (0.045) R-Squared – – 0.005 0.002 0.000 Observations 246 246 56 56 56 Clusters – – 38 38 38 Panel B: Turnout Trend (2013-2008) (mean=0.11, sd=0.08) Treated - Controls 0.037* (0.020) Flood Treatment 0.035** 0.081*** 0.151** 0.194** (0.017) (0.022) (0.063) (0.075) R-Squared – – 0.207 0.141 0.188 Observations 246 246 56 56 56 Clusters – – 38 38 38 Panel C: Change in Turnout Trend ((13-08)-(08-02)) (mean=0.09, sd=0.14) Treated - Controls 0.074** (0.034) Flood Treatment 0.080*** 0.149*** 0.273*** 0.313*** (0.028) (0.038) (0.096) (0.110) R-Squared – – 0.251 0.162 0.172 Observations 246 246 56 56 56 Clusters – – 38 38 38 Notes: Exposure measures calculated at the constituency level. Mahalanobis and Nearest Neighbor matching conducted on a binary exposure variable indicated whether any of a constituency’s population was exposed in the 2010-11 floods using the vector of population exposure estimates for 2000 to 2009. SVM matching conducted using the same covariates on both the binary exposure measure and two continuous exposure measures. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors for Mahalanobis matching calculated per Abadie and Imbens (2006). Nearest neighbor matching results for ATE presented with standard errors calculated per Abadie and Imbens (2011). Standard errors for regressions on the matched subset identified by the SVM are clustered at the district level.

49 Table 2: Robustness of Flood Effect on Provincial Assembly Election Turnout Matching on Flood History (2000-2009) (1) (2) (3) (4) (5) Mahalanobis Nearest Neighbor SVM Matching Flood % Area % Population Treatment: Affected Affected Affected Flooded Exposed Panel A: Turnout 2013 (mean=0.56, sd=0.09) Treated - Controls -0.037* (0.022) Flood Treatment 0.003 -0.023 0.062** 0.017 (0.015) (0.017) (0.030) (0.036) R-Squared – – 0.014 0.021 0.001 Observations 560 560 141 141 141 Clusters – – 65 65 65 Panel B: Turnout Trend (2013-2008) (mean=0.11, sd=0.08) Treated - Controls -0.014 (0.022) Flood Treatment 0.003 0.041* 0.123** 0.094* (0.015) (0.022) (0.046) (0.052) R-Squared – – 0.049 0.085 0.046 Observations 560 560 141 141 141 Clusters – – 65 65 65 Panel C: Change in Turnout Trend ((13-08)-(08-02)) (mean=0.09, sd=0.14) Treated - Controls 0.034 (0.032) Flood Treatment 0.039* 0.079** 0.167* 0.118 (0.024) (0.039) (0.090) (0.114) R-Squared – – 0.062 0.053 0.025 Observations 560 560 141 141 141 Clusters – – 65 65 65 Notes: Exposure measures calculated at the constituency level. Mahalanobis and Nearest Neighbor matching conducted on a binary exposure variable indicated whether any of a constituency’s population was exposed in the 2010-11 floods using the vector of population exposure estimates for 2000 to 2009. SVM matching conducted using the same covariates on both the binary exposure measure and two continuous exposure measures. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors for Mahalanobis matching calculated per Abadie and Imbens (2006). Nearest neighbor matching results for ATE presented with standard errors calculated per Abadie and Imbens (2011). Standard errors for regressions on the matched subset identified by the SVM are clustered at the district level.

50 Table 3: Combined Vote Share of the PPP and PML(N) in the Pakistani National Assembly and Provincial Assembly Elections (1) (2) (3) (4) National Assembly Provincial Assembly Flood % Area % Population % Area % Population Treatment: Flooded Exposed Flooded Exposed Panel A: PPP and PML(N) Combined Vote Share 2013 (mean=0.47, sd=0.24) (mean=0.40, sd=0.23) Flood Treatment 0.058 0.178 0.026 0.008 (0.164) (0.200) (0.075) (0.084) R-Squared 0.591 0.593 0.497 0.497 Observations 249 249 565 565 Clusters 93 93 109 109 Panel B: Trend of PPP and PML(N) Combined Vote Share (2013-2008) (mean=-0.06, sd=0.21) (mean=-0.05, sd=0.22) Flood Treatment 0.282* 0.215 0.144 0.244** (0.149) (0.210) (0.089) (0.101) R-Squared 0.271 0.263 0.159 0.165 Observations 249 249 565 565 Clusters 93 93 109 109 Panel C: Change in Trend of PPP and PML(N) Combined Vote Share ((‘13-08)-(‘08-02)) (mean=-0.22, sd=0.35) (mean=-0.17, sd=0.35) Flood Treatment 0.602** 0.412 0.159 0.357** (0.244) (0.320) (0.134) (0.154) R-Squared 0.241 0.227 0.207 0.213 Observations 249 249 565 565 Clusters 93 93 109 109 Notes: Exposure measures calculated at the constituency level. All regressions include division fixed effects, geographic controls at the constituency level including UNEP flood risk, distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors are clustered at the district level.

51 Table 4: Impact of the 2010/11 Flood in the PPP’s and PML(N)’s Vote Shares in the National and Provincial Assembly Elections (1) (2) (3) (4) (5) (6) (7) (8) National Assembly Provincial Assembly % Area Flooded % Population Exposed % Area Flooded % Population Exposed PPP PML(N) PPP PML(N) PPP PML(N) PPP PML(N) Panel A: Vote Shares in 2013 Flood Treatment 0.208 -0.150 0.147 0.031 0.078 -0.053 0.005 0.003 (0.132) (0.093) (0.147) (0.113) (0.064) (0.046) (0.062) (0.059) R-Squared 0.680 0.783 0.675 0.781 0.644 0.646 0.642 0.645 Observations 249 249 249 249 565 565 565 565 Clusters 93 93 93 93 109 109 109 109 Panel B: Trend in Vote Shares (2013-2008) Flood Treatment 0.161 0.121 -0.015 0.230 0.060 0.084 0.136** 0.109 (0.141) (0.117) (0.147) (0.160) (0.072) (0.055) (0.063) (0.074) R-Squared 0.191 0.288 0.184 0.292 0.233 0.124 0.238 0.125 Observations 249 249 249 249 565 565 565 565 Clusters 93 93 93 93 109 109 109 109

52 Panel C: Change in Trend of Vote Shares ((13-08)-(08-02)) Flood Treatment 0.396* 0.206 0.073 0.339 -0.008 0.167* 0.174 0.182* (0.208) (0.177) (0.191) (0.251) (0.106) (0.084) (0.113) (0.100) R-Squared 0.146 0.229 0.134 0.231 0.155 0.125 0.159 0.125 Observations 249 249 249 249 565 565 565 565 Clusters 93 93 93 93 109 109 109 109 Notes: Exposure measures calculated at the constituency level. All regressions include division fixed effects, geographic controls at the constituency level including UNEP flood risk, distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors are clustered at the district level. Table 5: Impact of Immediate Flood Relief on the Turnout Effect in National Assembly Election (1) (2) (3) (4) (5) (6) (7) Relief Food Food Relief Shelter Shelter Relief Total Total Relief Controls: Baseline Relief Residuals Relief Residuals Relief Residuals Panel A: Turnout 2013 (mean=0.56, sd=0.09) % Population Exposed 0.081* 0.131** 0.056 0.082** 0.087** 0.105** 0.074 (0.042) (0.056) (0.055) (0.038) (0.044) (0.045) (0.051) Relief (Residuals) -0.013 -0.014 0.001 0.001 -0.003 -0.004 (0.009) (0.009) (0.006) (0.006) (0.005) (0.005) R-Squared 0.614 0.620 0.622 0.609 0.609 0.610 0.610 Observations 246 246 246 243 243 243 243 Clusters 92 92 92 90 90 90 90 Panel B: Turnout Trend (2013-2008) (mean=0.11, sd=0.08) % Population Exposed 0.159** 0.153* 0.163** 0.093 0.174** 0.109 0.182** (0.075) (0.078) (0.075) (0.069) (0.072) (0.074) (0.079) Relief (Residuals) 0.003 0.002 0.020*** 0.020*** 0.008* 0.008* (0.007) (0.007) (0.007) (0.007) (0.005) (0.005)

53 R-Squared 0.326 0.327 0.327 0.354 0.354 0.343 0.342 Observations 246 246 246 243 243 243 243 Clusters 92 92 92 90 90 90 90 Panel C: Change in Turnout Trend ((2013-2008)-(2008-2002)) (mean=0.09, sd=0.14) % Population Exposed 0.317** 0.301** 0.325** 0.190 0.337** 0.218 0.352** (0.139) (0.144) (0.142) (0.127) (0.138) (0.136) (0.152) Relief (Residuals) 0.007 0.004 0.036*** 0.035*** 0.016* 0.015* (0.013) (0.013) (0.012) (0.013) (0.008) (0.008) R-Squared 0.292 0.293 0.293 0.319 0.318 0.308 0.307 Observations 246 246 246 243 243 243 243 Clusters 92 92 92 90 90 90 90 Notes: Exposure measures calculated at the constituency level. All regressions include division fixed effects and geographic controls at the constituency level, including UNEP flood risk, distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors are clustered at the district level. Table 6: Impact of Immediate Flood Relief on the Turnout Effect in Provincial Assembly Election (1) (2) (3) (4) (5) (6) (7) Relief Food Food Relief Shelter Shelter Relief Total Total Relief Controls: Baseline Relief Residuals Relief Residuals Relief Residuals Panel A: Turnout 2013 (mean=0.54, sd=0.10) % Population Exposed 0.066** 0.069*** 0.063** 0.066*** 0.071** 0.068*** 0.069** (0.026) (0.026) (0.030) (0.023) (0.029) (0.024) (0.031) Relief (Residuals) -0.002 -0.003 0.002 0.001 0.000 -0.000 (0.007) (0.007) (0.005) (0.005) (0.003) (0.004) R-Squared 0.613 0.613 0.613 0.613 0.613 0.613 0.613 Observations 556 556 556 546 546 546 546 Clusters 109 109 109 106 106 106 106 Panel B: Turnout Trend (2013-2008) (mean=0.10, sd=0.09) % Population Exposed 0.093** 0.084* 0.098** 0.063 0.093** 0.064 0.095** (0.042) (0.044) (0.042) (0.047) (0.041) (0.048) (0.042) Relief (Residuals) 0.006 0.005 0.011 0.011 0.006 0.005 (0.008) (0.008) (0.008) (0.008) (0.005) (0.005)

54 R-Squared 0.316 0.317 0.317 0.325 0.324 0.323 0.322 Observations 556 556 556 546 546 546 546 Clusters 109 109 109 106 106 106 106 Panel C: Change in Turnout Trend ((2013-2008)-(2008-2002)) (mean=0.07, sd=0.15) % Population Exposed 0.171** 0.143 0.185** 0.111 0.160** 0.107 0.167** (0.085) (0.090) (0.085) (0.097) (0.078) (0.096) (0.079) Relief (Residuals) 0.018 0.014 0.019 0.018 0.012 0.010 (0.014) (0.013) (0.015) (0.015) (0.009) (0.009) R-Squared 0.275 0.280 0.278 0.289 0.288 0.289 0.287 Observations 556 556 556 546 546 546 546 Clusters 109 109 109 106 106 106 106 Notes: Exposure measures calculated at the constituency level. All regressions include division fixed effects and geographic controls at the constituency level, including UNEP flood risk, distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors are clustered at the district level. Table 7: Impact of Flooding on Political Knowledge (1) (2) (3) (4) (5) (6) (7) Flood % Area % Population Pop. Exposed > Pop. Exposed > Self-Reported Self-Reported Self-Reported Treatment: Flooded Exposed Median 90th Percentile Flood Exposure Affected Very|Extremely Bad Panel A: Index of Political Knowledge (Principal Component Analysis) (mean = 2.024, sd = 0.741) Flood Treatment 0.158 0.452*** 0.073 0.213*** 0.181*** 0.125*** 0.078*** (0.134) (0.136) (0.058) (0.058) (0.050) (0.030) (0.028) R-Squared 0.375 0.376 0.375 0.376 0.408 0.409 0.407 Observations 12178 12178 12178 12178 12178 12178 12178 Clusters 1097 1097 1097 1097 1097 1097 1097 Panel B: Index of Political Knowledge (Additive) (mean = 0.673, sd = 0.227) Flood Treatment 0.050 0.139*** 0.020 0.063*** 0.046*** 0.031*** 0.020** (0.039) (0.041) (0.017) (0.018) (0.015) (0.009) (0.008) R-Squared 0.380 0.381 0.380 0.381 0.411 0.412 0.411 Observations 12178 12178 12178 12178 12178 12178 12178 Clusters 1097 1097 1097 1097 1097 1097 1097 Notes: Exposure measures calculated at the tehsil level in columns (1) - (4) and the individual level in columns (5) - (7). All regression include; district fixed effects (columns (1) - (4)) or tehsil fixed effects (columns (5) - (7)); geographic controls at the tehsil level (columns (1) - (4)), including UNEP flood risk, 55 distance to major river, dummy for tehsils boardering a major river, std. dev. of the tehsil’s elevation, and mean tehsil elevation; and demographic controls including gender, a head of household dummy, age, a literacy dummy, a basic numeracy dummy, education, and a Sunni dummy. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors are clustered at the primary sampling unit. Table 8: Impact of Flooding on Approval and Perceived Efficacy of Violent Protest (1) (2) (3) (4) (5) (6) (7) Flood % Area % Population Pop. Exposed > Pop. Exposed > Self-Reported Self-Reported Self-Reported Treatment: Flooded Exposed Median 90th Percentile Flood Exposure Affected Very|Extremely Bad Panel A: Effectiveness of Junaid’s/Mahir’s Actions (mean = 0.65, sd in non-violent = 0.30, sd in violent = 0.38) Violent Vignette -0.213*** -0.214*** -0.218*** -0.203*** -0.256*** -0.224*** -0.215*** (0.018) (0.018) (0.019) (0.016) (0.027) (0.019) (0.018) Flood Treatment 0.308*** 0.095 0.051 -0.041 0.068** 0.019 0.043** (0.117) (0.110) (0.043) (0.047) (0.034) (0.021) (0.020) Violent × Flood 0.196** 0.283*** 0.086*** 0.122** 0.138*** 0.080*** 0.063** (0.086) (0.096) (0.033) (0.058) (0.046) (0.029) (0.029) R-Squared 0.240 0.237 0.236 0.233 0.306 0.305 0.305 Observations 11963 11963 11963 11963 11963 11963 11963 Clusters 1094 1094 1094 1094 1094 1094 1094 Panel B: Approval of Junaid’s/Mahir’s Actions (mean = 0.62, sd in non-violent = 0.31, sd in violent = 0.39) Violent Vignette -0.226*** -0.229*** -0.229*** -0.217*** -0.283*** -0.241*** -0.230*** (0.019) (0.018) (0.019) (0.017) (0.028) (0.020) (0.019) Flood Treatment 0.203 -0.013 0.052 -0.122** 0.004 -0.025 0.016 56 (0.129) (0.117) (0.043) (0.048) (0.038) (0.023) (0.022) Violent × Flood 0.209** 0.326*** 0.084** 0.179*** 0.182*** 0.107*** 0.093*** (0.088) (0.098) (0.034) (0.056) (0.050) (0.031) (0.031) R-Squared 0.256 0.255 0.255 0.254 0.322 0.322 0.322 Observations 11963 11963 11963 11963 11963 11963 11963 Clusters 1094 1094 1094 1094 1094 1094 1094 Notes: Exposure measures calculated at the tehsil level in columns (1) - (4) and the individual level in columns (5) - (7). All regression include; district fixed effects (columns (1) - (4)) or tehsil fixed effects (columns (5) - (7)); geographic controls at the tehsil level (columns (1) - (4)), including UNEP flood risk, distance to major river, dummy for tehsils boardering a major river, std. dev. of the tehsil’s elevation, and mean tehsil elevation; and demographic controls including gender, a head of household dummy, age, a literacy dummy, a basic numeracy dummy, education, and a Sunni dummy. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors are clustered at the primary sampling unit. References

Abadie, Alberto and Guido Imbens. 2006. “Large sample properties of matching estimators for average treatment eﬀects.” Econometrica 74(1):235–267.

Abadie, Alberto and Guido Imbens. 2011. “Bias-Corrected Matching Estimators for Average Treat- ment Eﬀects.” Journal of Business and Economic Statistics 29(1):1–11.

Acemoglu, Daron and James Robinson. 2001. “A Theory of Political Transitions.” American Eco- nomic Review 91(4):938–963.

Acemoglu, Daron and James Robinson. 2006. Economic Origins of Dictatorship and Democracy. Cambridge University Press.

Acemoglu, Daron and James Robinson. 2012. Why Nations Fail. Crown Publishing.

Achen, Christopher H. and Larry M. Bartels. 2004. Blind Retrospection: Electoral Responses to Drought, Flu, and Shark Attacks. Working paper Department of Politics, Princeton University.

Agency for Technical Cooperation and Development. 2010. Rapid Needs Assessment: Upper Dir District — KPK Province. Report ACTED.

Ahlerup, Pelle. 2011. Democratization in the Aftermath of Natural Disasters. Working paper University of Gothenburg.

Ahmed, Issam. 2010. “Pakistan Floods: How new networks of Pakistanis are mobilizing to help.”.

Altonji, Joseph G, Todd E Elder and Christopher R Taber. 2005. “Selection on Observed and Unobserved Variables: Assessing the Eﬀectiveness of Catholic Schools.” Journal of Political Economy 113(1):151–184.

Amarnath, Giriraj. 2013. Pakistan Flood Inundation Mapping. Documentation report International Water Management Institute.

Andrabi, Tahir and Jishnu Das. 2010. In Aid We Trust: Hearts and Minds and the Pakistan Earthquake of 2005. Working paper World Bank.

Bechtel, Michael M. and Jens Hainmueller. 2011. “How Lasting Is Voter Gratitude? An Analysis of the Short- and Long-Term Electoral Returns to Beneﬁcial Policy.” American Journal of Political Science 55(4):851–867.

Berghold, Drago and P¨aiviLujala. 2012. “Climate-related disasters, economic growth, and armed civil conﬂict.” Journal of Peace Research 49(1):147–162.

Berrebi, Claude and Jordan Ostwald. 2011. “Earthquakes, hurricanes, and terrorism: do natural disasters incite terror?” Public Choice 149:383–403.

Besley, Timothy and Robin Burgess. 2002. “The Political Economy of Government Responsiveness: Theory and Evidence from India.” Quarterly Journal of Economics 117(4):1415–1451.

Besley, Timothy and Torsten Persson. 2010. “State Capacity, Conﬂict, and Development.” Econo- metrica 78(1):1–34.

57 Besley, Timothy and Torsten Persson. 2011. “The Logic of Political Violence.” Quarterly Journal of Economics 126(3):1411–1445.

Brancati, Dawn. 2007. “Political aftershocks: The impact of earthquakes on intrastate conﬂict.” Journal of Conﬂict Resolution 51(5):715–743.

Br¨uckner, Markus, Antonio and Andrea Tesei. 2012. “Oil Price Shocks, Income, and Democracy.” Review of Economics and Statistics 94(2):389–399.

Br¨uckner, Markus and Antonio Ciccone. 2011. “Rain and the Democratic Window of Opportunity.” Econometrica 79(3):923–947.

Burke, Marshall B., Edward Miguel, Shanker Satyanath, John A. Dykema and David B. Lobell. 2009. “Warming increases the risk of civil war in Africa.” Proceedings of the National Academy of Sciences 106(49):20670–20674.

Burke, Marshall, Solomon M. Hsiang and Edward Miguel. 2013. Quantifying the Climatic Inﬂuence on Human Conﬂict, Violence and Political Instability. Working paper University of California and Princeton University.

Burke, Paul J. and Andrew Leigh. 2010. “Do Output Contractions Trigger Democratic Change?” American Economic Journal: Macroeconomics 2(4):124–57.

Center for Research on the Epidemiology of Disasters (CRED). 2013. “the International Disaster Database (EM-DAT).” Electronic resource. Available at http://www.emdat.be/. Chaney, Eric. Forthcoming. “Revolt on the Nile: Economic Shocks, Religion and Political Inﬂu- ence.” Econometrica .

Chen, Jowei. 2013. “Voter Partisanship and the Eﬀect of Distributive Spending on Political Par- ticipation.” American Journal of Political Science 51(1):200–217.

Cole, Shawn, Andrew Healy and Eric Werker. 2011. “Do voters demand responsive governments? Evidence from Indian disaster relief.” Journal of Development Economics 97:167–181.

Dartmouth Flood Observatory (DFO). 2013. “Global Active Archive of Large Flood Events.” Electronic resource. Available at http://www.dartmouth.edu/~floods/Archives/index.html. Fair, C. Christine, Neil Malhotra and Jacob N. Shapiro. 2012. “Faith or Doctrine: Religion and Support for Political Violence in Pakistan.” Public Opinion Quarterly .

Food and Agriculture Organization of the United Nations (FAO). 2011. “Pakistan Floods: One Year On, From Relief to Recovery.”.

Funk, Jeanne B., Robert Elliott, Michelle L. Urman, Geysa T. Flores and Rose M. Mock. 1999. “The Attitudes Towards Violence Scale: A Measure for Adolescents.” Journal of Interpersonal Violence 14(11):1123–1136.

Gallego, Jorge A. 2012. Natural Disasters and Clientelism: The Case of Floods and Landslides in Colombia. Working paper Wilf Family Department of Political Science.

Gasper, John T. and Andrew Reeves. 2011. “Make It Rain? Retrospection and the Attentive Electorate in the Context of Natural Disasters.” American Journal of Political Science 55(2):340– 355.

58 Ghimire, Ramesh and Susana Ferreira. 2013. Economic Shocks and Civil Conﬂict: The Case of Large Floods. Working paper University of Georgia.

Ghosh, Biswajit. 2009. “NGOs, Civil Society and Social Reconstruction in Contemporary India.” Journal of Developing Societies 25(2):229–252.

Green, Donald P., Alan S. Gerber and David W. Nickerson. 2003. “Getting Out the Vote in Local Elections: Results from Six Door-to-Door Canvassing Experiments.” Journal of Politics 65(4):1083–1096.

Gronewold, Nathanial. 2010. “What Caused the Massive Flooding in Pakistan.”.

Halvorson, S. J. and Jennifer Parker Hamilton. 2010. “In the aftermath of the Qa’yamat: the Kashmir earthquake disaster in northern Pakistan.” Disasters 34(1):184–204.

Haque, Shamsul. 2004. “Governance based on partnership with NGOs: implications for development and empowerment in rural Bangladesh.” International Review of Administrative Sciences 70(2):271–290.

Healy, Andrew and Neil Malhotra. 2009. “Myopic Voters and Natural Disaster Policy.” American Political Science Review 103(387-406).

Healy, Andrew and Neil Malhotra. 2010. “Random Events, Economic Losses, and Retrospec- tive Voting: Implications for Democratic Competence.” Quarterly Journal of Political Science 5(2):193–208.

Herold, C. and F. Mouton. 2011. “”Global ﬂood hazard mapping using statistical peak ﬂow estimates.” Hydrology and Earth System Sciences Discussions 8:305–363.

Jaeger, David, Esteban Klor, Sami Miaari and M. Daniele Paserman. 2012. “The Struggle for Palestinian Hearts and Minds: Violence and Public Opinion in the Second Intifada.” Journal of Public Economics 96(3-4):354–368.

Khan, Riaz and Roshan Mughal. 2010. “Floods ravage NW Pakistan, kill 430 people.” Associated Press (AP).

King, Gary, Christopher J. L. Murray, Joshua A. Salomon and Ajay Tandon. 2004. “Enhancing the Validity and Cross-Cultural Comparability of Measurement in Survey Research.” American Political Science Review 98(1):191–207.

Kirsch, Thomas D., Christina Wadhwani, Lauren Sauer, Shannon Doocy and Christine Catlett. 2012. “Impact of the 2010 Pakistan Floods on Rural and Urban Populations at Six Months.”.

Kolenikov, Stanislav and Gustavo Angeles. 2009. “Socioeconomic Status Measurement with Dis- crete Proxy Variables: Is Principal Component Analysis a Reliable Answer?” Review of Income and Wealth 55(1):128–165.

Kurosaki, Takashi, Humayun Khan, Mir Kala Shah and Muhammad Tahir. 2011. Natural Disasters, Relief Aid, and Household Vulnerability in Pakistan: Evidence from a Pilot Survey in Khyber Pakhtunkhwa. Primced discussion paper series, no. 12 Hitotsubashi University.

Lay, J Celeste. 2009. “Race, Retrospective Voting, and Disasters The Re-Election of C. Ray Nagin after Hurricane Katrina.” Urban Aﬀairs Review 44(5):645–662.

59 Leuven, Edwin and Barbara Sianesi. 2003. “PSMATCH2: Stata module to perform full Ma- halanobis and propensity score matching, common support graphing, and covariate imbalance testing.”. URL: http://ideas.repec.org/c/boc/bocode/s432001.html

Looney, Robert. 2012. “Economic Impact of the Floods in Pakistan.” Contemporary South Asia 20(2):225–241.

Maskrey, Andrew. 2011. Global Assessment Report on Disaster Risk Reduction. Report United Nations.

Miguel, Edward, Shanker Satyanath and Ernest Sergenti. 2004. “Economic Shocks and Civil Con- ﬂict: An Instrumental Variables Approach.” Journal of Political Economy 112(4):725–753.

Mueller, Valerie, Clark Gray and Katrina Kosec. 2013. Long-Term Migration and Weather Anoma- lies in Pakistan. Working paper International Food Policy Research Institute.

Narang, Vipin. 2013. “Posturing for Peace? Pakistan’s Nuclear Postures and South Asian Stability.” International Security 34:38–78.

National Disaster Management Agency. 2011. Pakistan Floods 2010: Learning from Experience. Report NDMA.

Nel, Philip and Marjolein Righarts. 2008. “Natural Disasters and the Risk of Civil Conﬂict.” International Studies Quarterly 52(1):159–185.

Oak Ridge National Laboratory. 2008. “LandScan Global Population Database.” Electronic resource. Available at http://www.ornl.gov/landscan/. Office for the Coordination of Humanitarian Affairs (OCHA). 2010. Pakistan: Floods Relief and Early Recovery Response Plan. Flash appeal United Nations. URL: http://www.unocha.org/cap/appeals/revision-pakistan-floods-relief-and-early-recovery- response-plan-november-2010

Paik, Christopher and Jacob N. Shapiro. 2013. Income or Politics? A Study of Satellite Streetlight Imagery with Application in Pakistan. Technical report Princeton University.

Pande, Rohini. 2011. “Can Informed Voters Enforce Better Governance? Experiments in Low- Income Democracies.” Annual Review of Economics 3:215–237.

Rahman, Sabeel. 2006. “Development, democracy and the NGO sector theory and evidence from Bangladesh.” Journal of Developing Societies 22(4):451–473.

Ramsay, Kristopher W. 2011. “Revisiting the Resource Curse: Natural Disasters, the Price of Oil, and Democracy.” International Organization 65(3):507–529.

Ratkovic, Marc. 2012. Identifying the Largest Balanced Subset of the Data Under General Treat- ment Regimes. Working paper Princeton University.

Reeves, Andrew. 2011. “Political disaster: Unilateral powers, electoral incentives, and presidential disaster declarations.” The Journal of Politics 73(04):1142–1151.

Remmer, Karen L. 2013. “Exogenous Shocks and Democratic Accountability: Evidence from the Caribbean.” Comparative Political Studies .

60 Shah, Aqil. 2013. “Constraining consolidation: military politics and democracy in Pakistan (20072013).” Democratization forthcoming.

Shahbaz, Babar, Qasim Ali Shah, Abid Q. Suleri, Steve Commins and Akbar Ali Malik. 2012. Livelihoods, basic services and social protection in north-western Pakistan. Report Overseas Development Institute and Sustainable Development Policy Institute. URL: http://www.odi.org.uk/publications/6756-livelihoods-basic-services-social-protection- north-western-pakistan

Shefter, Martin. 1977. “Party and Patronage: Germany, England, and Italy.” Politics & Society 7(4):403–451.

Slettebak, Rune T. 2012. “Don’t blame the weather! Climate-related natural disasters and civil conﬂict.” Journal of Peace Research 49(1):163–176.

United Nations Institute for Training and Research. 2003. “UNITAR’s Operational Satellite Ap- plications Programme (UNOSAT).” Electronic resource. Available at http://www.unitar.org/ unosat/.

United Nations Oﬃce for the Coordination of Humanitarian Aﬀairs. 2010. Pakistan – Flood – July 2010 Table B: Total Humanitarian Assistance per Donor. Technical report Financial Tracking Service (FTS): Tracking Global Humanitarian Aid Flows http://fts.unocha.org.

UNOCHA. 2013. Pakistan – Flood – July 2010 Table B: Total Humanitarian Assistance per Donor. Technical report Financial Tracking Service (FTS): Tracking Global Humanitarian Aid Flows http://fts.unocha.org.

Velez, Yamil and David Martin. 2013. “Sandy the Rainmaker: The Electoral Impact of a Super Storm.” PS: Political Science & Politics 46(02):313–323.

Waraich, Omar. 2010. “Pakistan’s rich ’diverted ﬂoods to save their land.” The Independent.

Ziring, Lawrence. 2009. Weak State, Failed State, Garrison State: The Pakistan Saga. In South Asias Weak States: Understanding the Regional Insecurity Predicament, ed. T.V. Paul. Stanford CA: Stanford University Press p. 181182.

61 A Appendix Tables and Figures

Table A.1: Summary Statistics of All Covariates

Variable Unit Median Mean Std. Dev. Min Max N Panel A: National Survey 2012 % Area Affected 2010 & 2011 percent 0 0.083 0.157 0 0.928 12994 % Population Exposed 2010& 2011 percent 0 0.059 0.136 0 0.897 12994 Population Exposed > Median dummy 0 0.222 0.416 0 1 12994 Population Exposed > 90th Percentile dummy 0 0.047 0.211 0 1 12994 Flood Exposure (subjective) 4pt Index 0.25 0.385 0.248 0.25 1 12994 Affected (subjective) dummy 0 0.268 0.443 0 1 12994 Very Bad | Extremely Bad dummy 0 0.181 0.385 0 1 12994 UNEP Flood Risk 5pt Index 0.628 0.985 0.982 0 4.048 12994 Distance to Nearest Major River 100 kilometers 0.331 0.899 1.242 0.005 6.856 12994 Tehsil Neighboring Major River dummy 0 0.469 0.499 0 1 12994 Std. Dev. of Elevation 1000 meters 0.029 0.128 0.195 0.002 1.021 12994 Mean Elevation 1000 meters 0.191 0.54 0.724 0.005 4.247 12994 Male dummy 1 0.515 0.5 0 1 12994 62 Head of Household dummy 0 0.373 0.484 0 1 12968 Age 100 years 0.34 0.35 0.119 0.18 0.85 12990 Can Read dummy 1 0.545 0.498 0 1 12931 Can Do Basic Math dummy 1 0.776 0.417 0 1 12367 Education Level 6pt Index 0.167 0.275 0.281 0 1 12889 Sunni dummy 1 0.93 0.256 0 1 12923 Asset Index (PCA) [0,1] 0.5 0.489 0.16 0 1 12994 Income (log) log(1000 Rupees) 9.680 9.679 0.543 6.909 12.612 12277 Expenditure (log) log(1000 Rupees) 9.616 9.502 0.621 0 11.408 12456 Farmer dummy 0 0.092 0.288 0 1 12994 Judgment of Effectiveness dummy 0.75 0.649 0.356 0 1 12732 Judgment of Approval dummy 0.75 0.618 0.369 0 1 12732 Index of Political Knowledge (PCA) Index 2.120 2.024 0.741 0 3.006 12994 Index of Political Knowledge (Additive) Index 0.7 0.673 0.227 0 1 12994 Panel B: Multiple Indicator Cluster Survey (MICS) Punjab 2007/08 and 2011 % Area Affected 2010 & 2011 percent 0.124×10−3 0.064 0.111 0 0.522 118 Area Affected > Median dummy 0 0.246 0.432 0 1 118 Area Affected > 90th Percentile dummy 0 0.042 0.202 0 1 118 % Population Exposed 2010& 2011 percent 0 0.046 0.105 0 0.749 118 Population Exposed > Median dummy 0 0.237 0.427 0 1 118 Population Exposed > 90th Percentile dummy 0 0.042 0.202 0 1 118 UNEP Flood Risk 5pt Index 1.016 1.213 0.957 0 3.717 118 Continued on next page Table A.1 – continued from previous page Variable Unit Median Mean Std. Dev. Min Max N Distance to Nearest Major River 100 kilometers 0.214 0.243 0.157 0.018 0.803 118 Tehsil Neighboring Major River dummy 1 0.746 0.437 0 1 118 Std. Dev. of Elevation 1000 meters 0.007 0.037 0.077 0.002 0.431 118 Mean of Elevation 1000 meters 0.174 0.23 0.187 0.073 1.314 118 Mean of Wealth Index (PCA) 2007/08 Index -0.102 -0.053 0.485 -1.37 1.295 114 Mean Income 2007/08 (log) log(1000 Rupees) 4.884 4.921 0.308 4.375 5.838 114 Employment Rate 2007/08 percent 0.393 0.396 0.038 0.297 0.534 114 Unemployment Rate 2007/08 percent 0.063 0.071 0.031 0.024 0.191 114 Mean of Wealth Index (PCA) 2011 Index -0.019 0.071 0.53 -0.998 1.72 114 Mean Expenditures 2011 (log) log(1000 Rupees) -2.147 -2.305 1.057 -9.478 -0.228 114 Employment Rate 2011 percent 0.417 0.429 0.054 0.327 0.576 114 Unemployment Rate 2011 percent 0.043 0.05 0.03 0.007 0.136 114 Difference in Wealth Indices 07/08-11 0.131 0.124 0.212 -0.507 0.678 114 Difference in Employment Rate 07/08-11 percent 0.029 0.033 0.053 -0.081 0.206 114 Difference in Unemployment Rate 07/08-11 percent -0.017 -0.022 0.03 -0.118 0.05 114 Panel C: Night Lights at the Tehsil Level 2009 and 2012 % Area Affected 2010 & 2011 percent 0.004 0.118 0.188 0 0.928 330 Area Affected > Median dummy 0 0.315 0.465 0 1 330

63 Area Affected > 90th Percentile dummy 0 0.061 0.239 0 1 330 % Population Exposed 2010& 2011 percent 0.003 0.095 0.165 0 0.897 330 Population Exposed > Median dummy 0 0.297 0.458 0 1 330 Population Exposed > 90th Percentile dummy 0 0.058 0.233 0 1 330 UNEP Flood Risk 5pt Index 0.865 1.185 1.107 0 5 330 Distance to Nearest Major River 100 kilometers 0.329 0.822 1.203 0.003 6.856 330 Tehsil Neighboring Major River dummy 0 0.464 0.499 0 1 330 Std. Dev. of Elevation 1000 meters 0.019 0.141 0.21 0 1.021 330 Mean of Elevation 1000 meters 0.184 0.533 0.748 0.004 4.247 330 % Population Exposed 2012 percent 0 0.02 0.067 0 0.504 330 Night Lights per 100 citizens 2009 Index 2.518 2.459 1.589 0 11.297 328 Night Lights per 100 citizens 2012 Index 2.818 2.862 2.063 0 20.368 328 Difference in Night Lights 09-12 Index 0.209 0.403 0.887 -2.243 9.071 328 Panel D: National Assembly Constituencies % Area Affected 2010 & 2011 percent 0 0.082 0.15 0 0.79 249 % Population Exposed 2010 & 2011 percent 0 0.065 0.127 0 0.61 249 Affected (NDMA) dummy 1 0.51 0.501 0 1 249 UNEP Flood Risk 5pt Index 0.995 1.323 1.173 0 5 249 Distance to Nearest Major River 100 kilometers 0.260 0.482 0.715 0.005 5.517 249 Constituency Neighboring Major River dummy 1 0.522 0.501 0 1 249 Std. Dev. of Elevation 1000 meters 0.007 0.093 0.198 0 1.097 249 Mean of Elevation 1000 meters 0.179 0.351 0.575 0.004 3.922 249 Continued on next page Table A.1 – continued from previous page Variable Unit Median Mean Std. Dev. Min Max N % Population Exposed 2012 percent 0 0.012 0.052 0 0.489 249 Turnout 2013 percent 0.5767 0.561 0.085 0.296 0.848 246 Trend in Turnout (2013-2008) percent 0.114 0.113 0.079 -0.114 0.372 246 Change in Turnout Trend ((13-08)-(08-02)) percent 0.083 0.086 0.141 -0.371 0.638 246 Vote Share of PPP and PML(N) percent 0.487 0.468 0.242 0.008 0.976 249 Trend in Vote Share of PPP and PML(N) percent -0.064 -0.055 0.212 -0.489 0.793 249 Change in Vote Share Trend of PPP and PML(N) percent -0.203 -0.22 0.353 -1.394 1.394 249 Vote Share PPP 2013 percent 0.069 0.157 0.191 0 0.795 249 Vote Share PML(N) 2013 percent 0.371 0.311 0.233 0 0.731 249 Trend in Vote Share of PPP percent -0.158 -0.165 0.166 -0.59 0.388 249 Trend of Vote Share of PML(N) percent 0.051 0.11 0.194 -0.489 0.669 249 % Population Exposed 2000 percent 0.002 0.02 0.044 0 0.318 249 % Population Exposed 2001 percent 0.002 0.022 0.044 0 0.293 249 % Population Exposed 2002 percent 0.001 0.019 0.042 0 0.298 249 % Population Exposed 2003 percent 0.003 0.025 0.05 0 0.384 249 % Population Exposed 2004 percent 0.002 0.022 0.049 0 0.339 249 % Population Exposed 2005 percent 0.003 0.025 0.052 0 0.332 249 % Population Exposed 2006 percent 0.008 0.034 0.061 0 0.398 249

64 % Population Exposed 2007 percent 0.006 0.032 0.06 0 0.351 249 % Population Exposed 2008 percent 0.004 0.025 0.05 0 0.341 249 % Population Exposed 2009 percent 0.004 0.023 0.047 0 0.335 249 Index of Shelter Relief Index -0.315 0.024 0.858 -0.315 6.049 246 Index of Food Relief Index -0.419 0 1 -0.419 5.489 249 Index of Total Relief Index -0.734 0.029 1.675 -0.734 7.641 246 Migration Estimate percent 0.003 0.02 0.034 0 0.128 160 Panel E: Provincial Assembly Constituencies % Area Affected 2010 & 2011 percent 0 0.086 0.167 0 0.912 569 % Population Exposed 2010 & 2011 percent 0 0.067 0.146 0 1 569 Affected (NDMA) dummy 0 0.464 0.499 0 1 569 UNEP Flood Risk 5pt Index 0.920 1.303 1.266 0 5 569 Distance to Nearest Major River 100 kilometers 0.269 0.579 0.893 0.001 6.353 569 Constituency Neighboring Major River dummy 0 0.373 0.484 0 1 569 Std. Dev. of Elevation 1000 meters 0.002 0.026 0.053 0 0.323 569 Mean of Elevation 1000 meters 0.054 0.118 0.182 0.001 1.293 569 % Population Exposed 2012 percent 0 0.013 0.062 0 0.700 569 Turnout 2013 percent 0.563 0.541 0.104 0.012 0.738 560 Trend in Turnout (2013-2008) percent 0.112 0.103 0.094 -0.416 0.506 560 Change in Turnout Trend ((13-08)-(08-02)) percent 0.077 0.072 0.15 -0.544 0.822 560 Vote Share of PPP and PML(N) percent 0.422 0.401 0.225 0 0.98 569 Trend in Vote Share of PPP and PML(N) percent -0.063 -0.046 0.218 -0.807 0.684 569 Continued on next page Table A.1 – continued from previous page Variable Unit Median Mean Std. Dev. Min Max N Change in Vote Share Trend of PPP and PML(N) percent -0.170 -0.17 0.353 -1.294 1.134 569 % Population Exposed 2000 percent 0 0.021 0.065 0 1 569 % Population Exposed 2001 percent 0 0.024 0.067 0 0.924 569 % Population Exposed 2002 percent 0 0.021 0.064 0 0.855 569 % Population Exposed 2003 percent 0.2×10−3 0.027 0.075 0 0.784 569 % Population Exposed 2004 percent 0 0.024 0.075 0 0.997 569 % Population Exposed 2005 percent 0 0.027 0.078 0 1 569 % Population Exposed 2006 percent 0.9×10−3 0.035 0.086 0 1 569 % Population Exposed 2007 percent 0.001 0.035 0.09 0 1 569 % Population Exposed 2008 percent 0.7×10−3 0.027 0.072 0 0.931 569 % Population Exposed 2009 percent 0 0.023 0.062 0 0.725 569 Index of Shelter Relief Index -0.313 -0.003 0.812 -0.313 6.323 559 Index of Food Relief Index -0.416 0 1 -0.461 4.124 569 Index of Total Relief Index -0.773 0.005 1.617 -0.773 7.628 559 Migration Estimate percent 0.003 0.022 0.04 0 0.224 355 65 Table A.2: Impact of Flooding on Assets, Income, and Expenditures 2012 (1) (2) (3) (4) (5) (6) (7) Flood % Area % Population Pop. Exposed > Pop. Exposed > Self-Reported Self-Reported Self-Reported Treatment: Flooded Exposed Median 90th Percentile Flood Exposure Affected Very|Extremely Bad Panel A: Asset Index (mean=-0.49, sd=0.16) Flood Treatment -0.062 -0.062 0.014 -0.017 -0.068*** -0.033*** -0.041*** (0.045) (0.040) (0.016) (0.016) (0.014) (0.008) (0.007) R-squared 0.416 0.416 0.415 0.415 0.471 0.470 0.471 Observations 12132 12132 12132 12132 12132 12132 12132 Clusters 1093 1093 1093 1093 1093 1093 1093 Panel B: Household Income (logged) (mean=9.68, sd=0.54) Flood Treatment 0.046 -0.132 0.084* -0.066 -0.105** -0.070*** -0.053* (0.135) (0.155) (0.045) (0.070) (0.045) (0.025) (0.028) R-Squared 0.251 0.251 0.252 0.251 0.296 0.297 0.296 Observations 11539 11539 11539 11539 11539 11539 11539 Clusters 1080 1080 1080 1080 1080 1080 1080 Panel C: Household Expenditures (logged) (mean=9.50, sd=0.62) Flood Treatment -0.050 -0.001 0.069* 0.012 -0.039 -0.021 -0.035 66 (0.108) (0.118) (0.037) (0.041) (0.040) (0.022) (0.023) R-Squared 0.218 0.218 0.219 0.218 0.265 0.265 0.265 Observations 11684 11684 11684 11684 11684 11684 11684 Clusters 1082 1082 1082 1082 1082 1082 1082 Notes: Exposure measures in columns 1-4 are calculated at the tehsil level and those in columns 5-7 at the individual level. All regressions in columns 1-4 include district fixed effects, geographic controls at the tehsil level, including UNEP flood risk, distance to major rivers, dummy for tehsil’s bordering major rivers, std. dev. of the tehsil’s elevation, and mean constituency elevation, and demographic controls, including gender, a head of household dummy, age, a literacy dummy, a basic numeracy dummy, education, and a Sunni dummy. All regressions in columns 5-7 include tehsil fixed effects and the same set of demographic controls. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors are clustered at the primary sampling unit. Table A.3: Impact of Flooding on Assets, Income, and Expenditures by Farmers and Non-Farmers 2012 (1) (2) (3) (4) (5) (6) (7) Flood % Area % Population Pop. Exposed > Pop. Exposed > Self-Reported Self-Reported Self-Reported Treatment: Flooded Exposed Median 90th Percentile Flood Exposure Affected Very|Extremely Bad Panel A: Asset Index (mean=-0.49, sd=0.16) Flood Treatment -0.043 -0.034 0.021 -0.008 -0.063*** -0.031*** -0.037*** (0.044) (0.038) (0.016) (0.015) (0.014) (0.008) (0.007) Farmer 0.015** 0.012 0.012* 0.004 0.029*** 0.015** 0.017** (0.008) (0.007) (0.007) (0.007) (0.010) (0.007) (0.007) Flood Treatment × Farmer -0.138*** -0.127*** -0.039** -0.046* -0.046** -0.021* -0.034*** (0.035) (0.035) (0.016) (0.024) (0.020) (0.013) (0.012) R-Squared 0.417 0.417 0.416 0.416 0.472 0.471 0.472 Observations 12132 12132 12132 12132 12132 12132 12132 Clusters 1093 1093 1093 1093 1093 1093 1093 Panel B: Household Income (logged) (mean=9.68, sd=0.54) Flood Treatment 0.052 -0.128 0.084* -0.080 -0.096** -0.069*** -0.044 (0.140) (0.153) (0.045) (0.065) (0.045) (0.025) (0.027) Farmer 0.127*** 0.128*** 0.127*** 0.115*** 0.178*** 0.136*** 0.144*** 67 (0.027) (0.027) (0.027) (0.026) (0.038) (0.028) (0.025) Flood Treatment × Farmer -0.123 -0.131 -0.048 0.020 -0.145* -0.062 -0.117** (0.141) (0.135) (0.057) (0.096) (0.076) (0.045) (0.052) R-Squared 0.254 0.255 0.255 0.255 0.300 0.300 0.300 Observations 11539 11539 11539 11539 11539 11539 11539 Clusters 1080 1080 1080 1080 1080 1080 1080 Panel C: Household Expenditures (logged) (mean=9.50, sd=0.62) Flood Treatment -0.063 -0.008 0.061* 0.015 -0.043 -0.025 -0.036 (0.111) (0.120) (0.037) (0.043) (0.041) (0.023) (0.024) Farmer 0.108*** 0.116*** 0.107*** 0.113*** 0.133*** 0.117*** 0.121*** (0.028) (0.028) (0.028) (0.025) (0.040) (0.029) (0.026) Flood Treatment × Farmer 0.015 -0.073 0.002 -0.061 -0.049 -0.017 -0.036 (0.112) (0.102) (0.050) (0.060) (0.075) (0.048) (0.051) R-Squared 0.221 0.221 0.221 0.221 0.268 0.267 0.268 Observations 11684 11684 11684 11684 11684 11684 11684 Clusters 1082 1082 1082 1082 1082 1082 1082 Notes: Exposure measures in columns 1-4 are calculated at the tehsil level and those in columns 5-7 at the individual level. All regressions in columns 1-4 include district fixed effects, geographic controls at the tehsil level, including UNEP flood risk, distance to major rivers, dummy for tehsil’s bordering major rivers, std. dev. of the tehsil’s elevation, and mean constituency elevation, and demographic controls, including gender, a head of household dummy, age, a literacy dummy, a basic numeracy dummy, education, and a Sunni dummy. All regressions in columns 5-7 include tehsil fixed effects and the same set of demographic controls. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors are clustered at the primary sampling unit. Table A.4: Wealth, Income, Employment Rate, and Unemployment Rate in Punjab 2007/2008 (1) (2) (3) (4) Flood Treatment: % Area Flooded % Population Exposed OLS Robust OLS Robust Panel A: Wealth Index (mean=-0.05, sd=0.48) Flood Treatment -0.882 -0.426 -0.867** -0.952** (0.552) (0.471) (0.390) (0.386) R-squared 0.561 0.590 0.564 0.623 Observations 114 114 114 114 Clusters 36 – 36 – Panel B: Income (in 1,000 Ruppees, logged) (mean=4.92, sd=0.31) Flood Treatment -0.250 0.040 -0.075 -0.032 (0.350) (0.416) (0.240) (0.357) R-Squared 0.312 0.402 0.310 0.399 Observations 114 114 114 114 Clusters 36 – 36 – Panel C: Employment Rate (mean=0.40, sd=0.04) Flood Treatment 0.069 0.076* 0.023 0.050 (0.056) (0.043) (0.047) (0.037) R-squared 0.427 0.493 0.418 0.477 Observations 114 114 114 114 Clusters 36 – 36 – Panel D: Unemployment Rate (mean=0.07, sd=0.03) Flood Treatment -0.044 -0.061** -0.039 -0.010 (0.047) (0.030) (0.037) (0.027) R-squared 0.401 0.531 0.402 0.508 Observations 114 114 114 114 Clusters 36 – 36 – Notes: Exposure measures calculated from the Punjab MICS Household Survey 2007/08 at the tehsil level. All regressions include division fixed effects and geographic controls at the tehsil level, including UNEP flood risk, distance to major rivers, dummy for tehsil’s bordering major rivers, std. dev. of the tehsil’s elevation, and mean constituency elevation. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors in the OLS regressions are clustered at the district level.

68 Table A.5: Wealth, Expenditures, Employment Rate, and Unemployment Rate in Punjab 2011 (1) (2) (3) (4) Flood Treatment: % Area Flooded % Population Exposed OLS Robust OLS Robust Panel A: Wealth Index (mean=0.07, sd=0.53) Flood Treatment -1.043** -0.716 -0.834** -0.797** (0.407) (0.466) (0.348) (0.385) R-squared 0.667 0.689 0.666 0.706 Observations 114 114 114 114 Clusters 36 – 36 – Panel B: Expenditures (in 1,000 Ruppees, logged) (mean=0.14, sd=0.13) Flood Treatment 2.127 -0.412 -0.749 -3.431*** (3.478) (1.129) (2.613) (1.004) R-Squared 0.328 0.286 0.316 0.325 Observations 114 114 114 114 Clusters 36 – 36 – Panel C: Employment Rate (mean=0.43, sd=0.05) Flood Treatment 0.039 0.097** 0.125*** 0.124*** (0.059) (0.049) (0.039) (0.041) R-squared 0.612 0.667 0.635 0.676 Observations 114 114 114 114 Clusters 36 – 36 – Panel D: Unemployment Rate (mean=0.05, sd=0.03) Flood Treatment -0.007 -0.001 -0.014 -0.011 (0.022) (0.028) (0.015) (0.024) R-squared 0.701 0.682 0.702 0.683 Observations 114 114 114 114 Clusters 36 – 36 – Notes: Exposure measures calculated from the Punjab MICS Household Survey 2011 at the tehsil level. All regressions include division fixed effects and geographic controls at the tehsil level, including UNEP flood risk, distance to major rivers, dummy for tehsil’s bordering major rivers, std. dev. of the tehsil’s elevation, and mean constituency elevation. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors in the OLS regressions are clustered at the district level.

69 Table A.6: Trend in Wealth, Employment Rate, and Unemployment Rate in Punjab (2011 - 2007/2008) (1) (2) (3) (4) Flood Treatment: % Area Flooded % Population Exposed OLS Robust OLS Robust Panel A: Change in Wealth Index (2011-2007/2008) (mean=0.12, sd=0.21) Flood Treatment -0.161 0.174 0.032 0.292 (0.442) (0.310) (0.377) (0.265) R-squared 0.276 0.256 0.274 0.261 Observations 114 114 114 114 Clusters 36 – 36 – Panel B: Change in Employment Rate (2011-2007/2008) (mean=0.03, sd=0.05) Flood Treatment -0.031 -0.020 0.101** 0.077 (0.088) (0.062) (0.047) (0.054) R-squared 0.382 0.481 0.398 0.484 Observations 114 114 114 114 Clusters 36 – 36 – Panel C: Change in Unemployment Rate (2011-2007/2008) (mean=-0.02, sd=0.03) Flood Treatment 0.037 0.048 0.025 0.004 (0.055) (0.038) (0.038) (0.033) R-squared 0.228 0.235 0.226 0.237 Observations 114 114 114 114 Clusters 36 – 36 – Notes: Exposure measures calculated from the Punjab MICS Household Surveys 2007/2008 and 2011 at the tehsil level. All regressions include division fixed effects and geographic controls at the tehsil level, including UNEP flood risk, distance to major rivers, dummy for tehsil’s bordering major rivers, std. dev. of the tehsil’s elevation, and mean constituency elevation. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors in the OLS regressions are clustered at the district level.

70 Table A.7: Pre-Flood (2009), Post-Flood (2012), and Trend in Night-Light Emissions (2012-2009) (1) (2) (3) (4) Flood Treatment: % Area Flooded % Population Exposed OLS Robust OLS Robust Panel A: Pre-Flood Night-Light Emissions per 100 Citizens (2009) (mean=2.46, sd=1.59) Flood Treatment -1.348* -0.996** -1.310** -0.722 (0.772) (0.462) (0.626) (0.498) R-Squared 0.452 0.637 0.450 0.634 Observations 327 327 327 327 Clusters 113 – 113 – Panel B: Post-Flood Night-Light Emissions per 100 Citizens (2012) (mean=2.86, sd=2.06) Flood Treatment -1.551 -1.296** -1.578* -0.940* (0.976) (0.511) (0.828) (0.552) R-Squared 0.344 0.631 0.343 0.627 Observations 327 327 327 327 Clusters 113 – 113 – Panel C: Trend in Night-Light Emissions per 100 Citizens (12-09) (mean=0.40, sd=0.89) Flood Treatment -0.204 -0.263 -0.268 -0.299 (0.367) (0.202) (0.356) (0.216) R-Squared 0.285 0.386 0.286 0.390 Observations 327 327 327 327 Clusters 113 – 113 – Notes: Exposure measures calculated at the tehsil level. All regressions include division fixed effects and geographic controls at the tehsil level, including UNEP flood risk, distance to major rivers, dummy for tehsil’s bordering major rivers, std. dev. of the tehsil’s elevation, , and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors in the OLS regressions are clustered at the district level.

71 Table A.8: Impact of Flooding on Turnout in the Pakistani National Assembly Elections (1) (2) (3) (4) Flood Treatment: % Area Flooded % Population Exposed OLS Robust OLS Robust Panel A: Turnout 2013 (mean=0.56, sd=0.09) Flood Treatment 0.038 0.015 0.081* 0.045 (0.030) (0.037) (0.042) (0.045) R-Squared 0.611 0.716 0.614 0.715 Observations 246 245 246 245 Clusters 92 – 92 – Panel B: Turnout Trend (2013-2008) (mean=0.11, sd=0.08) Flood Treatment 0.089 0.088 0.159** 0.144** (0.063) (0.054) (0.075) (0.064) R-Squared 0.315 0.309 0.326 0.316 Observations 246 245 246 246 Clusters 92 – 92 – Panel C: Change in Turnout Trend ((13-08)-(08-02)) (mean=0.09, sd=0.14) Flood Treatment 0.219* 0.190** 0.317** 0.271** (0.120) (0.097) (0.139) (0.114) R-Squared 0.284 0.280 0.292 0.288 Observations 246 246 246 246 Clusters 92 – 92 – Notes: Exposure measures calculated at the constituency level. All regressions include division fixed effects and geographic controls at the constituency level including UNEP flood risk, distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors in the OLS regressions are clustered at the district level.

72 Table A.9: Turnout in the National Assembly Elections Among Different Constituency-Subsets (1) (2) (3) (4) (5) (6) (7) (8) Constituencies Neighboring Major Rivers Constituencies Next to Those Neighboring Major Rivers Flood Treatment: % Area Flooded % Population Exposed % Area Flooded % Population Exposed OLS Robust OLS Robust OLS Robust OLS Robust Panel A: Turnout 2013 (mean=0.58, sd=0.07) (mean=0.55, sd=0.09) Flood Treatment 0.044 0.041 0.046 0.038 -0.162*** -0.190* -0.198** -0.220** (0.035) (0.047) (0.043) (0.051) (0.055) (0.102) (0.095) (0.091) R-Squared 0.651 0.658 0.651 0.672 0.777 0.822 0.776 0.917 Observations 129 128 129 129 78 76 78 77 Clusters 61 – 61 – 47 – 47 – Panel B: Turnout Trend (2013-2008) (mean=0.12, sd=0.07) (mean=0.12, sd=0.08) Flood Treatment 0.113 0.130** 0.166* 0.126* 0.291** 0.221 0.299* 0.307 (0.080) (0.060) (0.086) (0.067) (0.113) (0.137) (0.158) (0.235) R-Squared 0.315 0.504 0.327 0.493 0.595 0.738 0.585 0.511 Observations 129 129 129 129 78 77 78 74 Clusters 61 – 61 – 47 – 47 – 73 Panel C: Change in Turnout Trend ((13-08)-(08-02)) (mean=0.09, sd=0.13) (mean=0.09, sd=0.14) Flood Treatment 0.298* 0.234** 0.354** 0.254* 0.571*** 0.594* 0.709*** 0.693 (0.154) (0.116) (0.168) (0.132) (0.139) (0.301) (0.209) (0.429) R-Squared 0.231 0.399 0.236 0.363 0.634 0.567 0.630 0.524 Observations 129 129 129 129 78 77 78 75 Clusters 61 – 61 – 47 – 47 – Notes: Exposure measures calculated at the constituency level. All regressions include division fixed effects and geographic controls at the constituency level, including UNEP flood risk, distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors in the OLS regressions are clustered at the district level. Table A.10: Impact of Flooding on Turnout in the Pakistani Provincial Assembly Elections (1) (2) (3) (4) Flood Treatment: % Area Flooded % Population Exposed OLS Robust OLS Robust Panel A: Turnout 2013 (mean=0.54, sd=0.10) Flood Treatment 0.045* 0.032 0.066** 0.059** (0.026) (0.024) (0.026) (0.026) R-Squared 0.612 0.738 0.613 0.741 Observations 556 556 556 556 Clusters 109 – 109 – Panel B: Turnout Trend (2013-2008) (mean=0.10, sd=0.09) Flood Treatment 0.093*** 0.096*** 0.093** 0.126*** (0.033) (0.033) (0.042) (0.036) R-Squared 0.317 0.361 0.316 0.369 Observations 556 556 556 556 Clusters 109 – 109 – Panel C: Change in Turnout Trend ((2013-2008)-(2008-2002)) (mean=0.07, sd=0.15) Flood Treatment 0.196*** 0.214*** 0.171** 0.263*** (0.058) (0.056) (0.085) (0.061) R-Squared 0.281 0.305 0.275 0.313 Observations 556 556 556 556 Clusters 109 – 109 – Notes: Exposure measures calculated at the constituency level. All regressions include division fixed effects and geographic controls at the constituency level including UNEP flood risk, distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors in the OLS regressions are clustered at the district level.

74 Table A.11: Turnout in the Provincial Assembly Elections Among Different Constituency-Subsets (1) (2) (3) (4) (5) (6) (7) (8) Constituencies Neighboring Major Rivers Constituencies Next to Those Neighboring Major Rivers Flood Treatment: % Area Flooded % Population Exposed % Area Flooded % Population Exposed OLS Robust OLS Robust OLS Robust OLS Robust Panel A: Turnout 2013 (mean=0.57, sd=0.08) (mean=0.55, sd=0.08) Flood Treatment 0.035 0.011 0.080*** 0.063** 0.050 -0.002 0.025 -0.026 (0.040) (0.031) (0.028) (0.031) (0.072) (0.068) (0.087) (0.073) R-Squared 0.577 0.616 0.588 0.622 0.695 0.700 0.694 0.917 Observations 209 209 209 209 180 179 180 179 Clusters 63 – 63 – 64 – 64 – Panel B: Turnout Trend (2013-2008) (mean=0.10, sd=0.08) (mean=0.11, sd=0.09) Flood Treatment 0.108*** 0.137*** 0.070 0.155*** 0.087 0.068 0.246*** 0.231** (0.039) (0.040) (0.052) (0.040) (0.072) (0.097) (0.073) (0.103) R-Squared 0.319 0.383 0.303 0.399 0.443 0.423 0.459 0.445 Observations 209 209 209 209 180 179 180 178 Clusters 63 – 63 – 64 – 64 – 75 Panel C: Change in Turnout Trend ((2013-2008)-(2008-2002)) (mean=0.07, sd=0.13) (mean=0.08, sd=0.14) Flood Treatment 0.209*** 0.276*** 0.094 0.278*** 0.233* 0.057 0.569*** 0.571*** (0.069) (0.071) (0.104) (0.074) (0.138) (0.169) (0.149) (0.176) R-Squared 0.201 0.266 0.171 0.278 0.323 0.390 0.351 0.440 Observations 209 209 209 209 180 180 180 179 Clusters 63 – 63 – 64 – 64 – Notes: Exposure measures calculated at the constituency level. All regressions include division fixed effects and geographic controls at the constituency level, including UNEP flood risk, distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors in the OLS regressions are clustered at the district level. Figure A.1: Proportion of Area Flooded and Population Exposed in National and Provincial As- sembly Constituencies

76 Table A.12: Robustness of Flood Eﬀect on National Assembly Election Turnout Controlling for Empirical Risk (1) (2) (3) (4) (5) (6) (7) (8) % Area Flooded % Population Exposed Controlling for Empirical Risk Controlling for Empirical Risk Common Common Baseline Linear Cubic Baseline Linear Cubic Support Support Panel A: Turnout 2013 (mean=0.56, sd=0.09) Flood Treatment 0.038 0.026 0.022 0.018 0.081* 0.062 0.066 0.039 (0.030) (0.030) (0.029) (0.034) (0.042) (0.043) (0.042) (0.044) R-Squared 0.611 0.615 0.620 0.633 0.614 0.617 0.622 0.634 Observations 246 246 246 228 246 246 246 228 Clusters 92 92 92 87 92 92 92 87 Panel B: Turnout Trend (2013-2008) (mean=0.11, sd=0.08) Flood Treatment 0.089 0.096 0.095 0.094 0.159** 0.182** 0.178** 0.144* (0.063) (0.062) (0.061) (0.067) (0.075) (0.074) (0.073) (0.077) R-Squared 0.315 0.317 0.321 0.345 0.326 0.331 0.333 0.350 Observations 246 246 246 228 246 246 246 228 Clusters 92 92 92 87 92 92 92 87

77 Panel C: Change in Turnout Trend ((13-08)-(08-02)) (mean=0.09, sd=0.14) Flood Treatment 0.219* 0.230* 0.230** 0.261** 0.317** 0.351** 0.339** 0.310** (0.120) (0.118) (0.115) (0.122) (0.139) (0.135) (0.133) (0.138) R-Squared 0.284 0.286 0.289 0.321 0.292 0.296 0.297 0.321 Observations 246 246 246 228 246 246 246 228 Clusters 92 92 92 87 92 92 92 87 Notes: Exposure measures calculated at the constituency level. All regressions include division fixed effects and geographic controls at the constituency level, including UNEP flood risk, distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors are clustered at the district level. Table A.13: Robustness of Flood Effect on Provincial Assembly Election Turnout Controlling for Empirical Risk (1) (2) (3) (4) (5) (6) (7) (8) % Area Flooded % Population Exposed Controlling for Empirical Risk Controlling for Empirical Risk Common Common Baseline Linear Cubic Baseline Linear Cubic Support Support Panel A: Turnout 2013 (mean=0.54, sd=0.10) Flood Treatment 0.045* 0.033 0.033 0.047* 0.066** 0.053* 0.056** 0.057* (0.026) (0.024) (0.024) (0.026) (0.026) (0.027) (0.028) (0.029) R-Squared 0.612 0.613 0.614 0.630 0.613 0.614 0.615 0.630 Observations 556 556 556 521 556 556 556 521 Clusters 109 109 109 106 109 109 109 106 Panel B: Turnout Trend (2013-2008) (mean=0.10, sd=0.09) Flood Treatment 0.093*** 0.091*** 0.092*** 0.100*** 0.093** 0.092* 0.092* 0.077 (0.033) (0.034) (0.033) (0.036) (0.042) (0.048) (0.049) (0.050) R-Squared 0.317 0.317 0.318 0.325 0.316 0.316 0.317 0.320 Observations 556 556 556 521 556 556 556 521 Clusters 109 109 109 106 109 109 109 106

78 Panel C: Change in Turnout Trend ((13-08)-(08-02)) (mean=0.07, sd=0.15) Flood Treatment 0.196*** 0.190*** 0.192*** 0.208*** 0.171** 0.162 0.159 0.138 (0.058) (0.060) (0.060) (0.067) (0.085) (0.101) (0.102) (0.107) R-Squared 0.281 0.281 0.285 0.295 0.275 0.276 0.279 0.284 Observations 556 556 556 521 556 556 556 521 Clusters 109 109 109 106 109 109 109 106 Notes: Exposure measures calculated at the constituency level. All regressions include division fixed effects and geographic controls at the constituency level, including UNEP flood risk, distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors are clustered at the district level. Table A.14: Sensitivity Analysis of Turnout Results (1) (2) (3) (4) Comparison of Main National Assembly Election Provincial Assembly Election Effect to Bias Controls in the Controls in the % Area % Population % Area % Population Accounted for by: Restricted Set: Full Set: Flooded Exposed Flooded Exposed Panel A: Turnout 2013 Geographic Controls District FE Full Set of Controls -0.792 -0.991 -1.627 -1.364 Panel B: Turnout Trend (2013-2008) Geographic Controls District FE Full Set of Controls 45.549 -2.539 3.398 -6.997 Panel C: Change in Turnout Trend ((13-08)-(08-02)) Geographic Controls District FE Full Set of Controls -60.731 -3.634 10.422 -6.087

Notes: Each Cell in the table reports ratios based on the coefficient of the flood treatment from two individual-level regressions. In one, the covariates include the “restricted set” of control variables; call this coefficient βR. In the other, the covariates include the “full set” of controls; call this coefficient βF . In both regressions, the sample sizes are the same. The reported ratio is calculated as: βF /(βR − βF ). Geographic controls at the constituency level include distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation. 79 Table A.15: Flood Impact on Turnout in National and Provincial Assembly Elections Controlling for Migration (1) (2) (3) (4) (5) (6) Outcome Turnout 2013 Turnout Trend Change in Turnout Trend Variable: 2013-2008 (2013-2008)-(2008-2002) Flood % Area % Population % Area % Population % Area % Population Treatment: Exposed Exposed Exposed Exposed Exposed Exposed Panel A: National Assembly Elections Including Migration Estimate Flood Treatment -0.018 0.024 0.051 0.180** 0.173 0.324** (0.036) (0.040) (0.070) (0.089) (0.126) (0.161) Migration 0.121 0.124 0.270 0.287 0.981** 1.010** (0.233) (0.234) (0.211) (0.208) (0.442) (0.444) R-Squared 0.623 0.623 0.451 0.467 0.465 0.475 Observations 158 158 158 158 158 158 Clusters 55 55 55 55 55 55 Panel B: National Assembly Constituencies Without Migration Flood Treatment 0.012 0.048 0.061 0.124* 0.148 0.252* (0.038) (0.046) (0.064) (0.070) (0.124) (0.133) R-Squared 0.715 0.717 0.333 0.347 0.302 0.317 Observations 157 157 157 157 157 157 Clusters 69 69 69 69 69 69 Panel C: Provincial Assembly Elections Including Migration Estimate Flood Treatment 0.022 0.043 0.105*** 0.084 0.218*** 0.139 (0.028) (0.028) (0.038) (0.066) (0.079) (0.142) Migration 0.082 0.084 -0.239 -0.250 -0.093 -0.122 (0.173) (0.172) (0.170) (0.173) (0.326) (0.344) R-Squared 0.649 0.650 0.446 0.441 0.405 0.394 Observations 345 345 345 345 345 345 Clusters 58 58 58 58 58 58 Panel D: Provincial Assembly Constituencies Without Migration Flood Treatment 0.047* 0.071*** 0.073** 0.077* 0.134** 0.124 (0.028) (0.027) (0.033) (0.039) (0.052) (0.075) R-Squared 0.650 0.653 0.318 0.318 0.283 0.280 Observations 378 378 378 378 378 378 Clusters 87 87 87 87 87 87 Notes: Exposure measures calculated at the constituency level. All regressions include division fixed effects and geographic controls at the constituency level, including UNEP flood risk, distance to major river, dummy for constituencies bordering a major river, std. dev. of the constituency’s elevation, and mean constituency elevation, as well as the proportion of constituency population exposed to the 2012 floods. Estimates significant at the 0.05 (0.10, 0.01) level are marked with ** (*, ***). Standard errors are clustered at the district level.

B Knowledge Questions

As noted in the text construct a measure of political knowledge using the following binary questions. Critically, the political knowledge index should be considered a behavioral outcome; it is not something that can be faked. Respondents either know these facts or they do not and thus differences must either reflect some real investment in acquiring political knowledge or some background trait that is correlated with both being hit by the floods and political knowledge. Our core estimating equation for individual-level outcomes includes controls for literacy, numeracy, age, education, gender, and head of household status to minimize the latter possibility, making a behavioral

80 interpretation of the knowledge variable more tenable.

B.1 Policy Debates Prominent in Early-2012 To tap awareness of political issues, we asked respondents whether they were aware of four policy debates that were prominent in early-2012 as follows.

Q540. Have you heard about debates over whether to deploy the Army to deal with violence in Karachi?

(1) Yes (2) No

Q550. Have you heard about the Frontier Crimes Regulation (sarhad main kavanin) and plans to revise it?

(1) Yes (2) No

Q560. Have you heard about discussions between the Governments of Pakistan and Afghanistan to use peace jirgas to resolve their disputes for example the location of the border?

(1) Yes (2) No

Q570. Have you heard about debates over ongoing efforts between the Indian and Pakistani governments to resolve their difference through dialogue?

(1) Yes (2) No

B.2 Political Officeholders To tap knowledge of basic facts about politics we asked six questions about who held various political offices as follows. Note that the following set includes the correct answers for anyone living in one of the four main provinces as well as the Chief Ministers of the other three provinces, who are all well-known political figures. It thus constitutes a relatively discriminate test of how well one knows what positions people who are regularly in the news actually occupy.

Q510. What party heads the ruling Coalition in Parliament?

(1) PPP: Pakistan Peoples Party (2) JI: Jamaat-e-Islami (3) ANP: Awami National Party (4) BNP: Baloch National Party (5) JUI-F: Jamiat Ulema-e-Islam (Fazlur Rehman faction) (6) PML-N: Pakistan Muslim League-Nawaz Sharif (7) PML-Q: Pakistan Muslim League-Qaid (8) PTI- Pakistan Tehreek-e-Insaf (9) MQM-Muttehida Qaumi Mahaz

81 Now were going to ask you some questions about public figures.

[FOR Q515-Q535 READ RESPONDENTS THE FOLLOWING OPTIONS:] 1. Asif Ali Zardari 2. Yusaf Rhaza Gilani 3. Iftikar Chaudry 4. Ashraf Parvez Kiyani 5. Shabaz Sharif 6. Syed Qaim Ali Shah 7. Ameer Haider Khan Hoti 8. Aslam Raisani 9. Altaf Hussain

Q515. Who is the President of Pakistan?

Q520. Who is the Prime Minister of Pakistan?

Q525. Who is the Chief Minister of (INSERT PROVINCE OF RESPONDENT)?

Q530. Who is the Chief Justice of the Supreme Court?

Q535. Who is the Chief of Army Staff?