The potential outcomes definition of causality says that a causal effect is the difference between an actual outcome, and how it would have been if and only if the causal agent had been different in some specific way (Gerber and Green 2012). While this effect is fundamentally unobservable, it can be shown that the average causal effect between larger groups can be estimated. One identification strategy with a proven track record is randomized experiments. The randomization process ensures that the expected average outcomes between groups decided by the randomized variable are equal. Any differences in outcomes must be attributed to the difference in treatment in these groups.
Experiments are commonly designed by the researcher. However, it is also possible to find naturally occurring experiments or naturally occurring processes that can arguably be treated as experiments (see Gerber and Green 2012:15ff). Nonetheless, regardless of what kind of experiment you have, the same criteria must be fulfilled. Most important of these is a randomized treatment assignment: Assignment must be stochastic and with known probabilities. While you generally would know these two criteria to be correct when designing your own experiment, in any natural experiment, we must somehow discover these properties.
Previous literature has tried to identify a causal effect of climate variability of conflict through controlling for (un)observables through a panel structure. However, we believe that a more promising identification strategy is the randomized experiment, as it is based on a much simpler statistical procedure (in principle no need for controls), and relies more on a sound research design. Our central argument is that precipitation, treated in a particular way, can be seen as a randomized treatment. Further, we describe how we get to the naturally occurring randomization process and discuss the two additional core assumptions in experiments: non-interference and excludability.
Randomization of treatment
The definition of a randomized treatment is that treatment is assigned with a known probability. The problem with experiments where nature randomizes and assigns treatment is that we do not know the probability of treatment assignment. Thus, we need to approach natural experiments through finding processes where the treatment can be argued to be assigned with equal probabilities. With the case of precipitation, it is clear that the probability distribution of precipitation varies across space and over different times of the year. It rains more some months of the year than others and more in some areas than in others. It is only for specific area-month sets (e.g., Darfur, in January between 1948 and 2013) that it is convincing to assume equal probability of precipitation.
Following this assumption, we could do an experiment, using monthly average precipitation within one area as our causal variable. However, a problem arises when we try to combine that experiment with experiments from other area-months. To find an unbiased causal estimate in such an analysis, we have to know the probability distribution from which precipitation was assigned for each area-month and incorporate each such distribution into our analysis.
The Standardized Precipitation Index (SPI) tries to fix this exact problem (McKee et al. 1993; Guttman 1999). SPI is estimated through fitting the probability distribution of precipitation for a given area-month and then standardizes that fit (with expected mean 0 and a standard deviation of 1). The SPI should be interpreted as the normality of observing the given precipitation in that area for that month. Large negative numbers mean it was abnormally dry, and vice versa for large positive numbers.
In principle, if we select −1.5 SPI or lower to be our treatment assignment variable, then the probability of treatment should be expected to be equal for all area-months, meaning that we could treat all area-months as a large pooled experiment. However, SPI normalization is not perfect; some area-months are vastly more likely to have an SPI of −1.5 than others. For instance, in a very dry area such as Darfur, where the normal outcome during the driest months of the year is near zero precipitation, large negative SPI values for those months are not observed at all because they are not possible. Therefore, even after standardization, we need to take the blocked randomization into account when doing our analysis. We do it through adding area-month fixed effects.
Fjelde and von Uexkull (2012) also use SPI and employ an administrative unit-year design. However, rather than estimating SPI for each unit-year, they apply the annualized discrete coding used in PRIO-GRID (Tollefsen et al. 2012). This means that they have no means to correctly control for different treatment probabilities, which should be expected to lead to biased results.
Our research design needs 12 fixed effects for each areal unit; hence, we need to ensure that the probability of observing a conflict event within each unit is fairly large. Therefore, rather than using administrative units, or grid-cells, we identified six spatially segregated areas.
The areas we selected are by no means a random sample, rather we have selected areas were we believe we might observe the hypothesized effect. All the areas had a large number of communal conflict events between 1989 and 2013, and the livelihoods in these areas are likely to be affected by climate variability. Since we did not select randomly, we must be cautious about generalizing. Our areas are the Nigeria highland plateau, Rift Valley in Kenya including neighboring areas in Uganda, Darfur in Sudan, and three areas in India (Telangana, Bihar, and Manipur/Nagaland) (Map 1).
Non-interference
Non-interference is the assumption that the treatment or outcome of another unit is independent of the treatment of any other unit (Gerber and Green 2012: 43f).
A popular approach in the climate-conflict literature has been to use grid-cells or administrative units and add lags to control for spatial dependencies (e.g., Fjelde and von Uexkull 2012). However, this type of control does not fix the problem that treatment is being applied non-randomly over a larger area. By treating each grid-cell as independent, the estimate of the standard error will be lower than it should. The modelling solution to this would be to somehow cluster cells depending on the monthly weather patterns. Another solution is to add fixed effects on each area-month, as we already do to ensure a randomized treatment assignment.
Rather than modelling spatial dependence, we have selected areas that contain whole conflicts. Our assumption here is that communal conflicts are local phenomena, between relatively stationary groups. By increasing the area size and selecting areas containing whole groups and conflicts, we reduce the likelihood that outside interference affects the outcomes.
Another source of interference in our setup is temporal dependence. It might be that past conflicts that partly were a result of droughts increase the likelihood of subsequent conflicts. If this is the case, we will overestimate the effect of the treatment. A problem with our setup is that the treatment is applied over a longer period of time (1, 3, 6, 12, and 24 months). If we control for conflict events that happened in the period after treatment is being applied, we might inadvertently control for outcomes that happened due to the treatment. Our solution is to control for conflict events before treatment is being applied (2 months before measuring conflict events when using 1-month SPI [since we measure outcome 1 month after the last treatment month], 4 months before when using 3-month SPI, etc.). The control we use is log(n + 1) of the number of events (n) that happened in the 6- and 12-month periods before starting treatment.
Excludability
Excludability refers to the point that random assignment should only affect the outcome through the assigned treatment, and not something else in addition (Gerber and Green 2012: 39ff). To discuss excludability, we must be clear about what the treatment is. While the treatment assignment variable (Z) is the SPI, our treatment of interest is not SPI, rather it is scarcity of resources, lost revenues, and lost livelihoods that may result from the extreme weather (X). One problem arising from this is “non-compliance,” simply the point that some individuals who are assigned to unusually dry weather “fail” to lose their livelihoods or vice versa. “Non-compliance” should be expected to bias the estimate of the average treatment effect of X on communal conflict toward zero.
Further, we cannot exclude the possibility that the government or NGOs provide assistance whose intensity is correlated with the treatment assignment variable (Z). We can imagine that the effect of SPI on conflict approaches zero if the government has sound insurance policies for all farmers who lose their income and investments under floods and droughts, if food prices are subsidized when prices surge due to droughts, and if markets are well connected so that the needed resources are available to the population. This scenario would bias our estimate toward zero as assistance is increased.
Another scenario is that both droughts and floods change the terrain, which could make fighting more or less difficult (Miguel et al. 2004). If roads are impossible to travel due to floods, then the likelihood of conflicts might go down. However, floods could make it more difficult to pursue attackers with heavy weaponry such as government forces, possibly making floods a good cover to initiate a raid.
We have no way of controlling for these influences in our current setup. Thus, what we estimate is the so-called intention-to-treat effect (ITT) (Gerber and Green 2012: 138f). The ITT effect is the causal effect of experiencing unusual precipitation on the likelihood of a communal conflict. Not being able to estimate the average treatment effect of resource scarcity, however, means that we must be careful in our interpretation of the ITT. In the following section, we describe the data and the statistical interference method that we will use to test the relationship between communal conflict and climate variability.