Identification of the productivity effects of ethnic diversity is complicated by the fact that individuals typically sort into joint production, or are assigned to production units so as to maximize productivity. Any third factor that influences both a team’s productivity and its ethnicity configuration will induce spurious correlation between team output and diversity. The plant I study is ideal for analyzing the impact of ethnic diversity on productivity because of its position rotation system. The supervisors described the system as follows. Workers returning from leave were assigned to open positions in the order in which they arrived at the plant in the morning. Supervisors would start in one corner of a packing hall and work their way through open positions row by row. A priori it is difficult to see how such an assignment system could lead to systematic correlation between the chacteristics of the workers in a team. The team ethnicity configuration classification I use is depicted in figure 1b. With 46.10 percent Kikuyu and 53.90 percent Luo workers, 25.46 percent of teams should be ethnically homogeneous, 24.85 percent vertically mixed and 49.69 percent horizontally mixed, if assignment was random. Appendix figure 1 displays the distribution of co-workers’ tribe across Kikuyu and Luo suppliers, vertical grow system during each of the three periods. It is clear that workers are not assigned to, or sort into, teams based on ethnicity.
A possible concern is that the underlying productivity of workers that end up in homogeneous teams may nevertheless differ from that of workers in diverse teams, for reasons unrelated to ethnicity itself. Suppose that individuals are equally productive in homogeneous and diverse teams but prefer interacting with coethnics, as in Becker . In that case it may for example be that supervisors assign well-liked, high-productivity workers to desirable homogeneous teams. Appendix figure 2 displays the distribution of workers’ gender, years of education and years of experience across homogeneous, horizontally mixed and vertically mixed teams, during each of the three sample periods. The distributions are essentially identical. A formal test of quasi-random assignment is in table 2. The matrices in the table display the characteristics, tribe×gender×past productivity, of one worker in the row dimension, and those of another worker in the team in the column dimension. The proportion of teams observed in a given cell is shown, as well as the proportion expected under the null hypothesis of independence between the row worker’s characteristics and the column worker’s characteristics. Because the worker rotation system leads to complex temporal correlation in team composition and output, the assumptions required for validity of Pearson’s chi-square tests would be violated if all data was used. A periodical “snapshot” of data is thus used in the table: team compositions on the first day of every month. For the same reason, productivity is measured by a worker’s average output in month t−2.
In the context of the plant I study, quasi-random assignment is less surprising than one might think. Supervisors had little incentive to attempt to optimize team assignment, and little ability to do so given their limited knowledge of worker characteristics and the plant’s leave and rotation system. Managers appeared to be unaware of systematic differences in output across teams of different ethnicity configurations during the first year of the sample period, their limited attention to the packing plant perhaps due to labor costs making up a relatively low proportion of flower farms’ total costs . To alleviate any remaining concerns about systematic team assignment, individual fixed effects are used in the main regressions of the paper.In the context of the plant, the productivity effect of ethnic diversity can be identified by comparing the output of teams of different ethnicity configurations. I begin by focusing on the first year of the sample period, when processors were paid based on own output, and before conflict began. The histogram in figure 5 displays mean output by team ethnicity configuration in 2007, distinguishing between teams with Kikuyu and Luo suppliers. Confidence intervals are shown but are narrow. The magnitudes in the histogram are in the notes to the figure, along with the standard errors. Note first that there are no significant differences between teams with Kikuyu and Luo suppliers.
Most importantly, all-Kikuyu teams are on average as productive as all-Luo teams. Given the nature of work at the plant, this is arguably unsurprising. Focusing instead on output differences that point to discriminatory behavior, it is also the case that the output gap between Kikuyu-Luo-Luo and all-Kikuyu teams is not significantly different from the output gap between Luo-Kikuyu-Kikuyu and all-Luo teams. The same is true for the gap in output between homogeneous and horizontally mixed teams. The evidence in figure 5 thus suggests that Kikuyu and Luo workers are of similar ability and equally discriminatory on average. These results enable a more concise presentation of the evidence to follow. In the remainder of the paper, I do not distinguish between specific ethnic groups and instead focus on the relation between the ethnic backgrounds of workers in a team. It is clear in figure 5 that team output is highest in homogeneous teams and lowest in vertically mixed teams, with output in horizontally mixed teams falling in between the two. The distribution of team and processor output in teams of different ethnicity configurations is displayed in figure 6. Notably, the density of output for coethnic processors in horizontally mixed teams is shifted to the right of that in homogeneous teams. Conversely, the density of output for non-coethnic processors in horizontally mixed teams is shifted to the left of that in vertically mixed teams. The distributions appear close to normal. Regression results corresponding to figure 5 are in table 4. The effects are very precisely estimated. Including individual fixed effects in the regressions has little influence on the results, as expected given quasi-random assignment to teams. The output of processors in vertically mixed teams is eight and a half percent lower than that of processors in homogeneous teams, an output gap that is also reflected in the total output of vertically mixed teams. As predicted by the model, upstream workers discriminate against non-coethnics downstream by under supplying them, it appears. Such discrimination lowers final output. The results in table 4 also indicate that suppliers discriminate horizontally. It is important to distinguish between the two processors in horizontally mixed teams. The output of the non-coethnic processor is eighteen percent lower than that of processors in homogeneous teams, and nine percent lower than that of processors in vertically mixed teams. The output of the coethnic processor is seven percent higher than that of processors in homogeneous teams. That processor output is lower if the other processor is of the same ethnicity as the supplier points to horizontal favoritism, cannabis grow supplies as predicted by the model. As Becker emphasized, favored workers benefit from discrimination against non-favored workers. In some situations, such benefits may give favored individuals an incentive to maintain ethnic divisions in society. Recall that the output loss from horizontal discrimination will depend on the relative productivity of favored and non-favored downstream workers. In the context of the farm, the two ethnic groups are similarly-sized, and we saw above that Kikuyu and Luo workers appear to be of similar ability on average. In such a situation, the output of vertically mixed teams is expected to be lower than that of horizontally mixed teams, which is what we see in table 4. Although vertically mixed are in aggregate four percent less productive than horizontally mixed teams, the lowest output processors are found in horizontally mixed teams. Even if the impact of horizontal discrimination on total output is limited when workers of different ethnic groups are of similar ability, the distribution of output across downstream workers is significantly affected.
Suppose, for purposes of illustration, that in the absence of misallocation of roses across the two processors in a team, the output of a coethnic processor in a horizontally mixed team would be equal to that of a processor in a homogeneous team. Similarly, suppose that in such a scenario the output of a non-coethnic processor in a horizontally mixed team would be equal to that of a processor in a vertically mixed team. In that case we can decompose the output gap between homogeneous and horizontally mixed teams: 14 percent would be due to the effect of horizontal misallocation and 86 percent due to vertical misallocation. While the magnitude of the “misallocation multiplier” associated with horizontal discrimination will depend on the relative productivity of those being favored and those being discriminated against, generally speaking intermediate goods note being passed downstream will tend to lower final output more than intermediate goods being “invested” in a less productive downstream producer.. The analysis explores what happens when a worker is replaced by another worker of the same productivity tercile but the other ethnicity, controlling for pair fixed effects for the pair of workers that remain in the team before and after the switch. Note that there is no significant change in output when the outgoing and incoming worker are of the same ethnic group: worker switches do not in themselves affect the productivity of a team. In columns 1 and 3, the output of an unswitched processor is regressed on dummies for the change in team ethnicity configuration when a supplier or processor of productivity comparable to the replaced worker joins the team. For clarity, I lay out the effects for a worker in processor position 1 . The output of a processor 1 who is of the same ethnic group as the supplier increases by five percent when a processor 2 of the other ethnic group replaces a comparably productive processor 2 of the supplier’s ethnic group. When a supplier who is not of processor 1’s ethnic group replaces a comparably productive supplier of processor 1’s ethnic group, processor 1’s output falls by nine percent if the two processors are of the same ethnic group. If instead processor 2 is of the incoming supplier’s ethnic group, processor 1’s output falls by 25 percent. The output of a processor 1 who is not of the supplier’s ethnic group increases by nine percent if a processor 2 of processor 1’s ethnic group replaces a comparably productive processor 2 of the supplier’s ethnic group. The estimates for team output in columns 3 and 4 are similar, output falling by five percent when a team goes from being homogeneous to horizontally mixed due to a worker switch, by nine percent when a team goes from being homogeneous to vertically mixed, and by four percent when a team goes from being horizontally to vertically mixed. Comparing teams that share the workers in two positions and the productivity tercile of the worker in the third position thus yields similar estimates to comparing all teams of different ethnicity configurations, providing reassurance that the estimates in table 4 represent the causal effect of ethnic diversity. If the estimates in table 4 were due in part to for example non-random assignment to teams or differences in ability across the two ethnic groups interacting with non-linear complementarities in the production function, then controlling for pair fixed effects and the third worker’s productivity tercile should lead to different estimates. Figure 7 depicts the temporal response of team output to the “event” of a worker substitution leading to a change in a team’s ethnicity configuration. Panels A – C plot the dynamic response of the first difference of output to a change in a team’s ethnicity configuration, and panels D – E the cumulative response over time. The decrease in output when a team “becomes mixed” is apparent. The first differenced response occurs almost entirely on the first day after the switch: the difference in output between homogeneous and diverse teams is relatively constant through teams’ duration.The tribal categorization used here is meaningful. Recall that this paper distinguishes primarily between workers designated as belonging to the Luo and Kikuyu tribal blocs. Categorization was on the basis of political alliances and relations between specific tribes. 86 percent of the sample belongs to three tribes: the Kikuyu , Luo and Luhya . I now consider sub-samples of teams in which workers belong to two specific tribes, focusing on the Kikuyu – Luo, Kikuyu – Luhya, and Luo – Luhya sub-samples. The Luo and Luhya tribes are categorized as belonging to the “Luo” ethnic group in this paper.