Con ict and Investment

Transcription

1 Con ict and Investment Tim Besley, Hannes Mueller and Prakarsh Singh Note Prepared for the IGC Workshop on Fragile States St Anne s College, Oxford (July 6th-7th, 2011) 1 Introduction This note provides a framework and discussion of our knowledge about the link between con ict and investment. The impact of con ict on investment is bound to vary from place to place and time to time. Hence, there is a need to have an approach which gives a clear sense of the di erences across con icts and the type of investment. The idea of an average con ict is therefore unlikely to help us understand about the heterogeneity. Also, since we do not know about the e ects of con icts that are yet to end and those that might yet occur, it would be useful to have models that can be calibrated out of sample. This means having an approach with theoretical foundations which expose the core parameters of interest. These can then be used to explore counterfactuals. It is a fairly well-established (and unsurprising) fact that output falls after the outbreak of violence. Persistent e ects of violence on output are likely to come from the destruction of current assets and falls in investment which reduces future asset creation. Understanding the timing and duration of economic responses to con ict therefore has to have a model of investment at its heart. This note presents a few signposts towards such an approach to understanding the link between con ict and investment. We begin from a simple theoretical model and discuss in detail how assessments about the duration of con ict play a key role in understanding the investment response in theory and practice. The model that we propose considers the switch from con ict to peace as a Markov Chain with parameters which can be studied empirically. We will also discuss how micro and macro facts can be matched beginning from an underlying theoretical model. We illustrate our method at work using micro-data from Punjab. The rest of this note is structured as follows. In the next section, we 1

2 discuss some of the background literature which is now quite extensive. We then present our conceptual framework. Section four applies this to data on investments by farmers in Punjab during a period of rural con ict and shows how persistence matters crucially to the investment response. Section ve lays out the research agenda and the link to policy. 2 Background There is now increasing evidence on the negative impact of violent con- ict on output and growth. According to Cerra and Saxena (2008), the start of a civil war is typically associated with an 18 percent output contraction. 1 This compares to an 8 percent drop which is typically seen at the onset of a banking crisis. This validates the obvious point that con ict is material in the fate of an economy. A civil war with an average 7-year duration is correlated with a 15 percent decline in household income and a 30 percent increase in the incidence of absolute poverty and a civil war of ve years appears to decrease annual growth rate by 2 percent (Collier, 1999; Hoe er and Reynal-Querol, 2003). Terrorism also seems to have long-lasting economy-wide costs. Abadie and Gardeazabal (2003) estimate a negative impact of around 10 percent on GDP per capita between Blomberg et al. (2004) compare consequences of terrorism with external or civil wars. On average, terrorism s negative e ect on growth is smaller and less persistent than that of external or civil wars. 3 Terrorism is also associated with a redirection of economic activity away from investment spending and towards government spending. In fact, the channel through which output is reduced in Israel is that of a decline in private investment during terrorism (Eckstein and Tsiddon, 2004; Fielding, 2003). It appears that a decline in investment is the main channel through which con ict a ects output and growth. So far, most research has focused on using stock markets as a way to understand this link. Although, prolonged exposure to terrorism does not seem to desensitize stock and foreign exchange markets in Israel (Eldor and Melnick, 2004), 1 Cerra and Saxena (2008) analyze the impact of civil war but treat every civil war year as a separate event. Mueller (2011) derives the number of 18 percent from focusing on civil war starts which makes the number comparable to Banking crisis starts analyzed in Cerra and Saxena. 2 They create a synthetic control for the Basque country that takes other regions of Spain in a combination to generate the same pattern of growth before onset of con ict as the Basque country. 3 Several studies nd no long-run e ect of bombings (as opposed to terrorism or civil wars) on income levels in Vietnam, Japan and West Germany (Miguel and Roland, 2005; Davis and Weinstein, 2002; Brakman, Garretsen and Schramm, 2004). 2

3 expectations may play a key role in determining the extent of decline in investment and these expectations are likely to be dependent on the region-speci c history of violence. 4 Besley and Mueller (2011) estimate the impact of the peace process in Northern Ireland on house prices. After establishing a negative correlation between killings and house prices, the parameters of a Markov switching model with con ict and peace as latent states are estimated. The model then allows estimation of the size of the peace dividend as captured in house price changes. The methodology applied by Zussman, Zussman and Nielson (2008) is a search for breaking points in Israel s stock market index. The largest breaks are then matched with con ict events and they nd a sharp drop of 22 percent in market value resulting from the outbreak of violence (Intifada) in 2000 and an increase of 25 percent in market value resulting from the adoption of a major peace initiative in Thus, one can use data on asset prices to estimate either costs of war or peace dividends. However, extrapolating results from one con ict to another is not at all straightforward. At the very least, it requires some kind of structural economic model. Below, we discuss how this can be developed for the purposes of studying investment. Although several studies have analyzed macroeconomic impacts of con ict, there are problems associated with causal identi cation. 5 There has been a surge in microeconomic studies of civil wars on human capital accumulation. There appear to be immediate adverse e ects on educational attainment post-con ict in Rwanda and Peru (Akresh and de Walque, 2009; León, 2009). 6 Negative e ects of civil con ict on health outcomes have been found in Colombia, Burundi and Iraq (Camacho, 2008; Bundervoet et al., 2009; Guerrero-Serdán, 2009). 7 However, there is very little evidence on investments by rms during and after con ict. Micro-level evidence not only allows us to get cleaner estimates and explore heterogeneous e ects empirically, but it may crucially allow the estimation of a structural model linking con- 4 There may not necessarily be a decline in investment. Guidolin and La Ferrara (2004) present a case of diamond-mining rms in Angola that show a loss in value when the civil war comes to a sudden end in Confounding biases include those of omitted variables (such as institutional capacity, political economy, geography and culture) as well as reverse causality from investment to con ict. 6 There are negative consequences of civil wars on educational attainment of women in Tajikistan and Guatemala (Shemyakina, 2006; Chamarbagwal and Moran, 2011). 7 For identi cation, these studies use di erence-in-di erences methodology that relies on the assumption that con ict a ected areas would have the same trend for investment as peaceful areas if there was no con ict. This may not necessarily hold in practice. 3

4 ict and investment. Post-election violence in Kenya appears to reduce export volumes of ower rms by 38 percent (Ksoll et al., 2010) by a direct impact on worker attendance. 8 Singh (2011) provides evidence for a particular investment channel through which sustained terrorism in rural areas may a ect growth in developing countries. Micro-level data from agricultural surveys during the period of terrorism in Punjab, India shows that a major terrorist incident in a district in a year reduces long-term xed investment by around 17 percent for an average farmer but e ects are muted for short-term investment. This results in a farmer losing close to 4 percent of his income annually because of the insurgency between and there are heterogeneous e ects by income and location of farmer. To sum up, the range of negative impacts on output and growth seem to be driven by a decline in investment. The empirical estimates for the decline in investment have been found through stock market reactions, asset prices, years of schooling and health indicators. These estimates are likely to di er by type of violence, geographical location, persistence of con ict and horizon of investment. We have a paucity of rm-level data during or immediately post-con ict that removes biases or sheds light on mechanisms and there is no work that links these results in a structural model to facilitate their comparison. 3 Conceptual Framework We discuss this in two parts. First, we lay out a simple investment model. Second, we introduce the Markov switching model. 3.1 A Model of Investment The standard economic approach to investment can be modi ed to think about how state fragility, and con ict in particular, has a bearing on investment decisions. The key idea is that investment should weigh up the costs and bene ts over the economic horizon of the investment. To x ideas, consider a standard capital investment then the value of investment should be re ected in the marginal product of capital over the period of the project. That marginal product could be a ected by con- ict in a variety of ways and, in principle, one could study and model these di erent channels. However, to keep things simple, imagine that con ict can be thought of in terms of an uncertain "transactions cost" which reduces the value of investing. This cost could be due to disruption in the supply chain, the destruction of assets and/or the di culties in 8 Deininger (2003) nds that civil strife, in contrast to theft, reduces investment and non-agricultural enterprise startups in Uganda. 4

5 hiring labor. This approach could also include intensi cation of corruption or diversion of resources into military spending. It is not important to be speci c. We measure this in proportion to the value of output like we would with a proportional tax. So write: marginal value of investing = (1 transaction cost) (1) marginal value product of capital. A rational investor will not typically know the disruption that he/she will face and hence must form expectations about this given the history of con ict. So being able to say something about the impact of violence/con ict on investment requires an underlying model of expectations formation something we will look at in the next section. The investor will invest to a point where the marginal cost and the expected bene t of investment are equal, i.e. E(marginal value of investing) marginal cost. (2) The reason for the inequality is to encompass the possibility that investment may not be worthwhile at all. If we insert equation (1) in the inequality in (2) it immediately follows that the expected development of transaction costs due to violence will a ect investments. In principle, this framework can apply to any kind of investment decision in either physical or human capital. The time horizon may vary in each case and whether investment decisions are reversible will also have a bearing on the value of investing. Given this observation, we would expect investment decisions based on this kind of simple calculation to vary by type of investment. To make this kind of theory operational requires a number of components some knowledge of technological knowledge of costs and bene ts and a measure of the way that di erent kinds of con ict a ect transactions costs. The challenge lies in being able to measure this (typically in micro) data to build a model where the e ects of policy interventions can be calibrated. The dynamics of investment will come in particular from changing expectations about the outlook for violence and the disruption that it brings. 3.2 Con ict Persistence Investments are actions taken for a future bene t and therefore critically depend on expectations. An analysis of the impact of violence on investments has to rely on a model of how present violence maps into future violence. In particular, the persistence of peace and con ict will in uence how economic actors react to events like the outbreak of violence 5

6 or cease- re agreements. A given outbreak of violence, for example, will have a much stronger impact on investments if violence is regarded as persistent than if the outbreak is regarded as a one-o event. One simple way to model the role of persistence for investment is to assume that violence in a region follows a so-called Markov Chain. A Markov Chain in its simplest form is a statistical process that describes the switching back and forth between two "states". In the application to con ict these two states can be thought of as con ict and peace, each associated with a di erent level of violence. The Markov Chain model delivers a very simple description of persistence because it is completely described by two probabilities - the likelihood that a period of peace is followed by another period of peace and the likelihood that a period of con ict is followed by another period of con ict. Tables 1 and 2 illustrate the di erence between a persistent and a non-persistent Markov Chain process with two examples. The tables show the probabilities of peace and con ict in year t + 1 conditional on the state in year t. Table 1 describes a relatively persistent violence process. The probability that con ict this year is followed by con ict next year is 80 percent whereas the probability that peace is followed by peace is 90 percent. This contrasts with the second example in table 2 in which con ict follows con ict with a likelihood of only 55 percent and peace follows peace in only 51 percent of the cases. Table 1: Persistent Violence Process con ict in year t peace in year t con ict in year t % 10% peace in year t % 90% Table 2: Non-persistent Violence Process con ict in year t peace in year t con ict in year t % 49% peace in year t % 51% The role of persistence in shaping expectations is easily shown. Assume, for instance, that in each of the two examples investors observe a transition from con ict to peace. In the persistent process economic agents will adjust their expectations on future violence drastically if they see that violence is fading. The bottom row in table 1 shows that the likelihood of peace tomorrow rises from 20 percent to 90 percent if the present state changes from con ict to peace. In the second example economic agents only marginally adjust their expectations. The bottom row of table 2 shows that the probability of peace tomorrow rises from 6

7 45 percent to 51 percent with a change from con ict to peace. Clearly, persistence makes a signi cant di erence for how expectations change with a transition from con ict to peace. The Markov Chain model can be applied to both time series of violence and pre-existing codings of con ict and peace. To illustrate the application we rst apply the process to the existing codings of civil war available in the Correlates of War dataset provided in Sarkees (2000). Table 3 shows data for two countries, Zimbabwe and Nigeria, for the period 1960 to In the sample period both Nigeria and Zimbabwe su ered a similar number of years in civil war according to the correlates of war coding. But the pattern of war and peace is very di erent. Whereas Nigeria dips in and out of civil war several times, Zimbabwe su ered one continuous eight-year period of civil war. As can be seen at the bottom of table 3 this di erence in con ict character is captured by di erences in persistence calculated in the Markov Chain model. 9 Nigeria clearly features less persistence of peace (despite having less civil war years) due to renewed outbreaks of civil war in the 1980s. The persistence of civil war, on the other hand, is much higher in Zimbabwe because of its continuous civil war episode in the 1970s. This di erence is disregarded by studies that treat the civil war data as a collection of independent 0/1 indicators without looking at the connection of these indicators across time. Hence, the Markov Chain model o ers a simple way to categorize con icts regarding their persistence of war and peace. We show next why modelling persistence in this way is useful for the analysis of the impact of con ict on investments Con ict Persistence and Investment We use the example of a violent con ict in the state of Punjab in India during the 1980 s to illustrate how measures of persistence can help us 9 The calculations are simple. For example, the likelihood that peace persists in Nigeria is calculated by dividing the number of transitions from peace to peace by the total number of peace years, 28=31 90%: 10 Some caveats to the methodology are worth mentioning. First, Markov Chain estimates can depend on what period one looks at. In Northern Ireland this did not seem too much of a problem - results were similar when we ran the EM algorithm over the whole period of violence. But, in general this could be an issue. Second, changes in probabilities are not in the model. For instance, we cannot capture peace that is perceived as more and more persistent. In principle, we could look at this by re-estimating the Markov chain parameters. Third, an asymmetry between cost at start and bene t at the end of con ict is not captured by the Markov Chain method. Given the cross-country evidence in Cerra and Saxena (2008) this might be unrealistic. The loss at the beginning might be bigger than the gain at the end. 7

8 Table 3: Illustration of Persistence Calculations with Correlates of War Data Prob(conflict follows conflict): Prob(peace follows peace): Zimbabwe Nigeria % 57% 97% 90% Source: Correlates of War Intra-State War Data,

9 understand the reaction of investments to violence. The rural insurgency which started as a separatist movement in 1981 haunted the region until 1993 claiming more than 20,000 lives. The militant groups employed terrorism against civilians in order to try and achieve political gains. 11 We use data on major terrorist killings at the district level as our measure of con ict and match a Markov Chain to each violence process. Figure 1 displays both the number of deaths due to terror attacks in the Amritsar district and the estimated number of deaths derived after matching the Markov Chain process to the raw data using the EM-algorithm by Hamilton (1990). 12 The gure illustrates that as soon as violence breaks out the algorithm estimates the district to be in con ict and once violence stops in 1993 the district is estimated to be in peace. As explained above, the Markov Chain model relies on two persistence probabilities to describe con ict episodes. The estimates for Amritsar are summarized in table 4. As can be seen in the table both peace and con ict were highly persistent in the district. Con ict years followed con ict years with a likelihood of 88 percent while peace followed peace with a likelihood of 94 percent. These estimates of persistence can be 11 Punjab was classi ed by UCDP as a state undergoing civil war between Also, there was a sharp decline in the agricultural growth rate from 6% to 2% between See Singh (2011) for a description of the con ict and data. 12 As opposed to the correlates of war data, here, we do not use pre-de ned states. The algorithm estimates the probabilities of being in con ict or peace jointly with the state-dependent parameters of the model. For a description, see Besley and Mueller (2011). 8

10 combined with the information displayed in gure 1 to derive the expected value of violence at each point in time. Put simply, the expected value of violence is a combination of the level violence in a given year with the estimate of how persistent this violence is. The more persistent the con ict, the stronger do changes in violence in a given year a ect expectations about the future level of violence. 13 Table 4: Terror in Amritsar con ict in year t peace in year t con ict in year t % 6% peace in year t % 94% In order to illustrate how persistence a ects investments we combine the derived expected violence with actual investment data from the Punjab region. More precisely, we regress information on investments in tube-wells by farmers in Punjab from 1978 to 1990 on our expected violence estimates. 14 An application of these regression results to Amritsar indicates that farmers in the district decreased their investments by 448 rupees on average when violence started. This is quite a large e ect given that the mean and standard deviation of investments were 1086 and 733 rupees respectively, i.e. the outbreak of the con ict led to a fall in investment of around 40 percent. Figure 2 highlights the role of persistence for this estimate of the investment contraction. The graph shows the estimated reaction of investments at di erent levels of persistence holding all other parameters xed. The vertical axis measures the extent to which farmers had reduced their investments if con ict had been regarded as more or less persistent than 88 percent. As can be seen, investment is a ected quite strongly by the level of persistence. In particular, if farmers had believed that con ict would persist with only 30 percent they would have decreased their investments by only 130 rupees on average. If farmers had regarded con ict as certainly persisting, investments would have declined by more than 900 rupees. It should be noted that most other districts in the Punjab region had much less persistent con ict episodes with a persistence of con ict 13 An assumption behind this methodology is that the Markov Chain parameters are used "as if" economic agents knew them. This assumption is justi ed if local investors have a notion of how long con ict and peace will last, i.e. if they observe the level of tensions and con ict potential on the ground. 14 In our regression we control for time and district xed e ects. In addition, the results are robust to a set of farm level covariates and district-speci c time trends. 9

11 between 30 and 60 percent. Not surprisingly, our estimates of damage to investments is much lower for these districts. The Markov Chain persistence estimates therefore adds a way to describe heterogeneity of violence in the di erent districts which cannot be captured otherwise. 5 Discussion and Directions for Future Research The approach suggested here provides a portable framework for thinking about investment and con ict in a wide variety of situations. It also provides a way forward for the research agenda. One of the main advantages of the structure lies in giving a sense of the timing of investment returns to establishing peace. If this structure is matched with a better understanding of con ict dynamics and the evolution of expectations it will lead to a much more realistic estimate of the economic reactions to con ict. A key aspect of the approach is the way that it permits us to think about persistence in establishing peace. It is hard to assess whether economic actors have faith in the e orts into paci cation (that have been) via military intervention or otherwise to convince investors to commit. Whether investors expectations are well modeled by a Markov Chain is moot. And, to date, there has been little work directly surveying expectations about con ict persistence. One direction for future research e orts is to engage in more survey work on expectation formation. This would create scope to examine how objective measures based on data on violence agree or disagree with more subjective indicators. Perhaps the biggest issue which lies some distance away is trying 10

12 to study how con ict and investment interact when investment itself provides a stake in peace. This is particularly interesting in the context of decisions to invest in xed assets such as housing and land. Increases in the value of such assets provides a tangible stake in the maintenance of peace. There may even therefore be a case for policy to encourage such investments if the externality associated with the maintenance of peace is not being internalized by investors. For an approach along these lines to be convincing empirically it would be necessary to have an approach where there is a feedback loop from investment to future violence. This could be speci ed and calibrated as part of the model. Such an approach should be useful in planning strategies for economic recovery in fragile states and being realistic about the timing of recovery. The approach taken so far has not considered the complementarities between di erent kinds of investments which could also be important. An obvious example is the importance of reinforcing public and private investments. However, public investment is likely only to be e ective in promoting private investment when there is reasonable optimism about the continuation of peace. The model above could be used to simulate the e ect of investments with di erent return structures as a function of violence persistence and hence could be used as a tool in policy analysis. Although the literature on con ict on investment is growing, we are still a good way short of understanding all the issues. We view an essential feature of the next phase of work to provide a stronger link between theory and empirics. Such approaches have a better prospect at doing justice to the heterogeneity in country experiences and in predicting out of sample. From a data collection point of view, the main dividends are likely to come from having more micro-data before and after the onset of con ict. But since con ict may itself disrupt the process of data collection, this is often demanding. 11

13 References [1] Abadie, A. and J. Gardeazabal (2003) The Economic Costs of Con- ict: A Case-Control Study for the Basque Country. American Economic Review, 93(1), [2] Akresh, R. and D. de Walque (2010) Armed Con ict and Schooling: Evidence from the 1994 Rwandan Genocide. World Bank Policy Research Paper, [3] Besley, T. and H. Mueller (2011) Estimating the Peace Dividend: The Impact of Violence on House Prices in Northern Ireland. American Economic Review (forthcoming). [4] Blomberg, S., G. Hess and A. Orphanides (2004) The macroeconomic consequences of terrorism. Journal of Monetary Economics, 51, [5] Brakman, S., H. Garretsen and M. Schramm (2004) The Strategic Bombing of German Cities During World War II and Its Impact on City Growth. Journal of Economic Geography, 4(2), [6] Bundervoet, T., P. Verwimp and R. Akresh (2009) Health and Civil War in Rural Burundi. Journal of Human Resources, 44, 2. [7] Busse, M. and C. Hefeker (2007) Political risk, institutions and foreign direct investment. European Journal of Political Economy 23, [8] Camacho, A. (2008) Stress and Birth Weight: Evidence from Terrorist Attacks. American Economic Review, 98(2), [9] Cerra, V. and S. Saxena (2008) Growth Dynamics: The Myth of Economic Recovery. American Economic Review, 98(1), [10] Chamarbagwala, R. and H. Morán (2011) The human capital consequences of civil war: Evidence from Guatemala. Journal of Development Economics 94(1), [11] Collier, P. (1999) On the Economic Consequences of Civil War. Oxford Economic Papers, 51, [12] Coyne, C., G. Dempster and J. Isaacs (2010) Asset Values and the Sustainability of Peace Prospects. The Quarterly Review of Economics and Finance, 50(2), [13] Davis, D. and D. Weinstein (2002) Bones, Bombs, and Break Points: The Geography of Economic Activity. American Economic Review, 92(5), [14] Deininger, K. (2003) Causes and Consequences of Civil Strife: Micro-level Evidence from Uganda. Oxford Economic Papers, 55, [15] Eckstein, Z. and D. Tsiddon (2004) Macroeconomic consequences of terror: theory and the case of Israel. Journal of Monetary Economics, 51(5),

14 [16] Eldor, R. and R. Melnick (2004) Financial markets and terrorism. European Journal of Political Economy, 20(2), [17] Fielding, D. (2003) Modelling Political Instability and Economic Performance: Israeli Investment during the Intifada. Economica, 70 (277), [18] Guerrero-Serdán, G. (2009) The E ects of the War in Iraq on Nutrition and Health: An Analysis Using Anthropometric Outcomes of Children. Royal Holloway mimeo 09/01. [19] Guidolin, M. and E. La Ferrara (2007) Diamonds Are Forever, Wars Are Not: Is Con ict Bad for Private Firms? American Economic Review, 97(5), [20] Guidolin, M. and E. La Ferrara (2005) The Economic E ects of Violent Con ict: Evidence from Asset Market Reactions. Federal Reserve Bank of St. Louis Working Papers, [21] Hamilton, J. D. (1989) A New Approach to the Economic Analysis of Nonstationary Time Series and the Business Cycle. Econometrica, 57, [22] Hoe er, A. and M. Reynal-Querol (2003) Measuring the costs of con ict. Oxford Working Paper. [23] Ksoll, C., R. Macchiavello and A. Morjaria (2010) The E ect of Ethnic Violence on an Export- Oriented Industry. CEPR Discussion Papers, [24] León, G. (2010) Civil Conzict and Human Capital Accumulation: The Long Term E ects of Political Violence in Peru. BREAD Working Paper, 245. [25] Miguel, E. and G. Roland (2010) The Long Run Impact of Bombing Vietnam. Unpublished manuscript, Univ. of California, Berkeley. [26] Mueller, H. (2011) Growth Dynamics: The Myth of Economic Recovery - Comment. IAE mimeo. [27] Sarkees, M. R. (2000) The Correlates of War Data on War: An Update to Con ict Management and Peace Science, 18(1), [28] Shemyakina, O. (2006) The E ect of Armed Con ict on Accumulation of Schooling: Results from Tajikistan. Journal of Development Economics (forthcoming). [29] Singh, P. (2011) Impact of Terrorism on Investment Decisions of Farmers: Evidence from the Punjab Insurgency. LSE mimeo. [30] Zussman, A., N. Zussman and M. Orregaard Nielsen (2008) Asset Market Perspectives on the Israeli-Palestinian Con ict. Economica, 75,