My outcomes of interest are individual income, usual weekly hours worked, industry employment rate and welfare income received. This allows me to account for all sources of reduced individual welfare that could occur as a result of this negative market shock. When appropriate, these outcomes were logarithm transformed so as to capture the relative impact when looking at an industry with a small percentage of the population. I utilized the North American Industrial Classification System codes that were provided by the ACS survey in order to define my own classifications of individuals’ labor industry to better fit the purpose of my study. These redefined industry classifications are used both as a binary outcome measure of the employment rate and as categorical regressors to look at differences in income and hours worked. Prior literature suggests a spillover impact between closely related labor industries . I redefined both closely related and distant occupations of interest into six groups as follows: agriculture, food product manufacturing, food product wholesale, local food industry, transportation, and construction. The food manufacturing category includes processing or production of animal food, produce, grain, sugar, bakery goods, dairy and animal byproduct. The food wholesale industry refers to groceries, farm products, and farm supplies. The local food industry category consists of grocery stores and restaurants. These categories are designed so that primary impacts show in the closely related categories of food wholesale or manufacturing that directly purchase for input the agricultural sector’s output.
This reasoning follows selected industry classifications utilized by Hornbeck and Keskin to observe agricultural spillover. The distant industries such as transportation and construction,10 liter drainage pot among others not included in this study, should experience the smallest or no impact. This follows the supply chain input-output process model documented by Albino et al. . To define my treatment group and time horizon, I looked at the hydrological impacts of the drought. An estimated 90% of drought-related farmland fallowing was restricted to the Tulare Lake Basin and some parts of the San Joaquin River Basin ensuring these areas sustained the greatest direct income differences due to drought . I defined San Joaquin and Tulare County as treatment counties because they comprise the majority of impacted land area both in terms of fallow unusable land and the need for extensive groundwater pumping. All remaining 16 Central Valley counties are set as the control group due to their compositional similarities without extreme drought impact. Although the drought began in 2012, water supplies reached a low in 2014 and 2015 . The event time period is set as 2015 to 2017 to capture the fallout. This paper uses an interaction term between the treatment counties and drought horizon so that the impact of the drought is captured from the coefficient on the interaction. The difference in difference method also relies on an assumption that there is not frequent mobility between the two groups so that accurate values for income and employment changes can be reported. This is assured by checking population in my treatment and control counties in 2012 and 2016. I observe less than a 1% increase in population in my control counties, which have a much larger area, and a 0.04% increase in population in my treatment counties. This does not suggest that there was significant mobility between these two groups or outside of these areas. Table 1 reports summary statistics in 2012 for both heavily impacted counties 7 by measure of extensive fallowing and other Central Valley counties.8 This is to show that both groups follow a similar parallel growth trend in 2012 before the onset of the drought. There are level differences in the amounts of agricultural and food wholesale workers in the pre-drought time period, however, the two groups have followed a similar path over time. This satisfies the parallel trend assumption necessary for performing a difference in difference regression to test outcome differences.When considering the impact of an event with a time horizon and units that were impacted differently, using a difference-in-difference approach is standard. I interact the time horizon of the drought with the counties that experienced the harshest impact. Although ACS survey data is commonly used for such analysis, there are some key limitations in my application. One of the limitations of the study is that the difference in difference regression technique relies upon the parallel trend assumption. While we can verify this trend graphically [Appendix: Figure 1], it has the potential to lend imprecise results. Although ACS survey data is commonly used for such analysis, there are some key limitations in my application. Any study focusing on relatively removed phenomena such as drought faces difficulties with attrition. For survey privacy purposes individuals cannot be linked over time, so I am prevented from using synthetic control to perfect my control group and better ensure both county groups follow parallel trends. I counter this by performing robustness tests by both changing the time horizon for the drought treatment and changing the combination of control groups. These regressions were also run on an interaction with gender and race 10 to ensure the reliability of my findings on spillover within all groups.Additionally, without data that extends to the time period after the 2012 to 2016 drought it is impossible to analyze long-run spillover effects. This study is limited to data from the years 2006 to 2017. The addition of a hydrological framework would also help to improve the study. I use a difference in difference ordinary least squares regression to analyze the variation in outcomes for individuals living in severely impacted counties with those living in similar Central Valley counties that evaded the impact of the 2012 to 2016 drought. For empirical specifications, outcome Y for individual i is regressed on fixed time period effects δt , status as an individual in a highly impacted county Dri , and an interaction where α represents the causal impact of the drought on an individual. Based on earlier explanation of the high costs of land fallowing, Dri is a dummy variable equal to 1 if individual i resides in Tulare or San Joaquin County. Dri is set equal to 0 if the individual resides in any other Central Valley county. This variable serves to represent residence in an area highly impacted by drought. Because water supplies hit a low in 2014, indicator variable P ostt is a binary, set equal to 1 if the individual observation falls after 2014. P ostt is equal to 0 if the observation is from 2014 or prior. The models include other individual-specific control predictors: Fi is a dummy equal to 1 if the individual is female, EDi controls for differences in education, Ai controls for age, and Ri controls for race-related differences in outcomes. My first attempt Equation 1 utilizes female and race base controls. Equation 2 makes the addition of education and age as further controls. For notational simplicity, all four control predictors are aggregated into the variable X in the second form of Equation 2. This notation is used going forward. I again estimated these interaction regressions with income,25 liter pot employment rate and usual hours worked per week as the outcomes. In each regression other than employment rate, I limited the sample pool to one of my 6 specified industry categories to separate the impact and isolate spillover impact. Both triple interacted regressions, female and Hispanic, serve to test the validity of my results and find if any individual group faced greater impact. This model has foundations on a similar difference in difference regression run by Hornbeck and Keskin when analyzing the windfall gains of the Ogalla aquifer on counties with varying degrees of exposure. They followed a similar strategy of comparing non-agricultural industries within each county to observe the spillover in productivity seen from an agricul- tural stimulus. It is a widely accepted model for evaluating the impact of a sudden change that is applied to a subset of units observed. I circumvent typical biases by running robustness tests and including fixed effects to control for error in my model. These equations use a time effect variable δt to control for the invariant presence of time trends in the data. Other similar models utilize state, time and group fixed effects . Because my study is confined to a region within California, there is no need for state fixed effects. The difference in difference equation I estimate utilizes differences between counties’ outcomes based on locational effects of drought. County cluster effects are typically used on data sets with more groups and in this case could potentially remove the effect that is isolated by my regression. To ensure that the populations had normal distributions, I performed iterations of my regressions with bootstrapped error and found no significant differences from previously estimated outcomes. All results are reported with robust error to account for serial autocorrelation. I believe that the other Central Valley counties provide the best approximation of parallel pre-drought trends in income and employment. The region has a similar composition in the sector, crop type, and incomes [Table 1]. Although synthetic control was not available to me with this data set, I ran my regressions on different treatment groups and with an earlier time period to ensure robust results. The model and question face limitations of data that do not allow for complete, precise identification of the causal impact of drought. Further analysis of a greater range of data would be necessary to confirm these results definitively.This section reports estimates of the impact of the 2012 to 2016 California drought on income, hours worked, and employment rate. I breakdown regressions by industry ranging from sectors closely related to agriculture to more distant industries to look for a heterogeneous change in welfare. Table 2 reports results obtained from estimating Equations 1 and 2 for only the agricultural industry. This table reports α, from my difference in difference interaction, and its standard error to represent the impact the drought had on each outcome. This table reports the base model controlling only for gender and race in column 1. Column 2 reports the model from equation 2, adding in additional controls for education and age. I use multiple outcomes as dependent: employment rate, hours worked weekly, log income,and income. The percent change listed is the change in employment in each regression scaled by the population. The individual industry regressions for employment are run on binary indicators for each industry. The interpretation is that this coefficient shows a change in individuals categorized within an industry each year. Any change comes from individuals either switching to a different industry or losing their jobs and becoming unemployed.12 Each time I regressed on income instead of the sector binary mentioned above I limited the sample to each industry to see the impact in that industry.13 The first row in Table 2 reports the decreases seen in agricultural employment as a direct impact of the drought. When scaled, we observe an estimated 7% reduction in the agricultural employment rate in column 1. This comes with a significant 9.3% reduction in agricultural income. The average hours worked per week does not drop by a large or significant amount. This means that when faced with drought and land fallowing, many agricultural workers living in the counties heavily impacted by the drought lost jobs or had large reductions in income. When observing the changes from column 1 to column 2 when adding controls for education and age, we can see these results become more significant and negative. The agricultural employment drops to -9% and income for agricultural workers, all else constant, drops by 11.3%. In both regressions we do not observe any significant level declines in in-come, however this is due to a reduction in growth rates of the agricultural industry due to drought, rather than absolute declines in productivity. In both we also see no significant change in hours worked implying workers left the industry or faced pay cuts rather than fewer hours. Table 3 reports estimates obtained from the equations including a triple interaction .