UNEVEN COTTON OUTCOMES IN COLONIAL AFRICA: AN UNRESOLVED PUZZLE

While the actual importance of empire cotton to European textile industries is disputable, it is beyond doubt that there was a widespread and persistent conviction in colonial circles that cotton should be exported from tropical Africa. British commentators hyperbolically branded West Africa as Lancashire’s potential sole supplier (1871, Dawe Reference Dawe1993, p. 24), “future salvation” (1904, in Ratcliffe Reference Ratcliffe1982, p. 113), and the “new Mecca” (1907, in Hogendorn Reference Hogendorn, Isaacman and Robert1995, p. 54). Similar sentiments were present in French and other colonial circles (Marseille Reference Marseille1975; Pitcher Reference Pitcher1993; Roberts Reference Roberts1996, pp. 60–75; Sunseri Reference Sunseri2001). After 1900, in the context of rising raw cotton prices, such sentiments began to be backed up with concerted efforts to stimulate cotton export from Europe’s African possessions. To that aim, textile lobbies throughout Europe established organizations receiving substantial financial and moral support from their governments (Dawe Reference Dawe1993, pp. 33–179; Roberts Reference Roberts1996, pp. 34–5; Robins Reference Robins2016, pp. 72–115). In the long run, France proved the most determined (Bassett Reference Bassett2001, pp. 90–4), and cotton promotion even remained a key feature of France’s involvement in West Africa long after the end of colonial rule (Lele, van de Walle, and Gbetibouo 1989).

Measured against the ambitious rhetoric and persistent efforts, cotton imperialism in Africa was far from successful. Footnote ² Sub-Saharan Africa’s share of global cotton production (excluding communist countries) rose from a mere 0.3 percent in the 1910s to a still-modest 5 percent during the decade after WWII (Dawe Reference Dawe1993, p. 431). Even at the end of the colonial period, cotton accounted for less than 9 percent of tropical Africa’s exports, ranking behind coffee, copper, cocoa, and peanuts (Hance, Kotschar and Peterec 1961, p. 491). After half a century of disappointment, numerous colonial officials concluded with resignation that their attempts to entice African farmers to grow the crop had failed (Robins Reference Robins2016, p. 29; Bassett Reference Bassett2001, p. 81; Likaka Reference Likaka1997, p. 89). When investments did not yield the hoped-for returns, colonial governments often responded with coercive measures. In French Equatorial Africa, the Belgian Congo, and Portuguese Mozambique, where marketing was monopolized by concessionary companies and cotton was often grown in collective fields, substantial coercion was applied, at the expense of farmers’ freedom, food security, and income (Likaka Reference Likaka1997; Isaacman Reference Isaacman1996; Beckert Reference Beckert2014). Numerous scholars identify coercion as the key determinant of cotton adoption in Africa (Bassett Reference Bassett2001, p. 7; Beckert Reference Beckert2014, p. 373). According to Isaacman and Roberts (1995, p. 29), cotton not only became Africa’s “premier colonial crop,” but also its “premier forced crop.”

Did heterogeneous cotton adoption outcomes indeed correlate closely with colonial investment and coercive effort? For 20 African colonies where serious attempts were undertaken to expand output, Figure 1 gives the average cotton output during the 1930s and 1950s. The take-off in the Anglo-Egyptian Sudan was exceptional, relying on the capital-intensive “Gezira” irrigation scheme. Meanwhile, Africa’s largest and earliest rainfed cotton take-off took place in Uganda, which alone produced 47 percent of all rainfed cotton in tropical Africa in the 1930s and 26 percent in the 1950s (Figure 1). The scope of Uganda’s take-off becomes clear when we compare it with Egypt, by far the continent’s largest cotton producer: Ugandans produced 56.5 percent of the cotton of their Egyptian ounterparts per capita, even though Egypt was highly specialized in cotton production, had a much longer history of commercial cultivation, and had an advanced irrigation infrastructure (De Haas Reference De Haas2017b; Karakoç and Panza 2021).

Figure 1 COTTON OUTPUT IN 20 AFRICAN COLONIES IN THE 1930S AND 1950S

Sources: See Online Appendix A.1.1.

Despite its singular importance, Uganda barely features in the literature on cotton imperialism in Africa (Isaacman and Roberts 1995; Beckert Reference Beckert2014). When it does, the focus is placed on Buganda, a prominent kingdom in the southern part of the country, and emphasis is put on its purportedly exceptional characteristics—centralized pre-colonial institutions, fertile soils, and reliance on bananas (a relatively undemanding crop) as the staple food (Austin Reference Austin2008, p. 597, 601; De Haas Reference De Haas2017b, p. 610; Elliot Reference Elliot, Jones and Woolf1969, p. 136–7; Isaacman and Roberts 1995, p. 23). However, this Buganda-focus is not justified by the actual spatial pattern of cotton adoption in Uganda; although Buganda was the first area in which cotton took off, within a decade most cotton was grown in the northern and eastern provinces that did not share most of Buganda’s distinctive “cotton-prone” features (De Haas and Papaioannou 2017; Vail Reference Vail1972; Tosh Reference Tosh1978; Wrigley Reference Wrigley1959). In other words, Uganda’s cotton take-off cannot simply be reduced to Buganda as “an exception that proves the rule.”

The case of Uganda also does not conform with the idea that African cotton output was a direct corollary of targeted colonial effort. Uganda’s cotton sector did not receive a particularly large colonial investment. Unlike the railroad linking Kano to Nigeria’s coast in 1912 (Hogendorn Reference Hogendorn, Isaacman and Robert1995, pp. 54–6), the “Uganda railway” which unlocked Uganda to global commodity markets was not built with the prospect of cotton exports in mind (Ehrlich Reference Ehrlich1958, pp. 49, 63–8; Wrigley Reference Wrigley1959, pp. 13–5). Similarly, the British Cotton Growing Association focused its efforts on Nigeria and was initially only peripherally involved in Uganda. Tellingly, British cotton advocate J. A. Todd (Reference Todd1915, p. 170), reflecting decades of optimism about West Africa’s cotton potential, asserted that “the possible cotton crop of Nigeria is about 6,000,000 bales of 400 pounds.” About Uganda, in contrast, he stated that “it is doubtful whether any really large quantity of cotton, more than, say, 100,000 bales per annum, is likely to be raised for a good many years to come” (Todd Reference Todd1915, p. 170). In reality, Nigeria came to export 70,000 bales annually between 1920 and 1960 (1.2 percent of Todd’s projected amount), compared to 260,000 bales from Uganda (260 percent).

Finally, while Uganda’s cotton take-off certainly involved colonial coercion, this cannot explain why Uganda’s farmers adopted cotton more readily than their counterparts in most other places. Indeed, Uganda did not implement many of the repressive features that characterized the cotton regimes elsewhere in central Africa, a fact that was even observed by a contemporary Belgian colonial official in the Congo: “we have failed to make the crop as popular here as in Uganda. The remuneration is inadequate, and the Blacks [sic] are growing the crop only under the pressure of the administration” (Likaka Reference Likaka1997, p. 89).

SEASONALITY AND COTTON OUTPUT ACROSS TROPICAL AFRICA

It is clear that much of the heterogeneity in African cotton adoption remains unexplained by variations in colonial investment and coercion. Can a consideration of agricultural seasonality which, as hypothesized by Tosh (Reference Tosh1980) and others, crucially shaped Africans’ labor allocation to cash-crop production, bring us closer to understanding why cotton was widely adopted in some places but not others? A first step towards answering this question is to simply explore the association between the length of the agricultural growing period and rainfed cotton output across colonial Africa. Areas with no or brief agricultural cycles (four or fewer wet months per year), including large swaths of the steppe and desert, faced the most restrictive seasonality constraint. We expect them to barely cultivate any cotton unless irrigated. Most farmers in tropical Africa operated in savanna areas with longer but still single growing cycles (five to eight wet months). Such areas tended to be highly suitable for rain-fed cotton, but we expect output to be constrained by seasonal labor bottlenecks. Some areas were endowed with two growing cycles (nine wet months or more). Such conditions enabled farmers to smooth their labor inputs intra-annually, attenuating the tradeoff between food crops and cotton. If seasonality mattered, we should expect the largest involvement in cotton production here.

To explore this hypothesized relationship between cotton output and the number of wet months, I first impose a one-by-one-degree grid on tropical Africa. For each grid cell, I calculate the average annual number of months with average rainfall over 60 millimeters (the standard threshold for a “tropical wet month” in the Köppen climate classification) during the period 1901–1960, using gridded rainfall data (data sources are listed under Table 1). Next, I sort all grid-cells in mainland tropical Africa into 12 “rainfall zones” by their number of tropical wet months (Table 1, Column (1)). I then calculate the percentage share that each zone contributed to tropical Africa’s total rainfed cotton output, based on a map showing the sub-national distribution of cotton output in 1957 (Column (2)).

Table 1 RAINFED COTTON OUTPUT AND ENDOWMENTS OF LAND, POPULATION, RAILROADS, AND COTTON YIELD POTENTIAL C. 1960, BY RAINFALL ZONE (SHARES OF TROPICAL AFRICAN TOTAL)

Notes: Columns (2) to (6) each add up to 100 percent. Columns (7) to (10) average out at 1.0. A “wet month” is any month of the year with over 60 millimeters of rain on average in the period 1901–1960, measured per 1 degree by 1 degree grid cell. For the definition of “tropical Africa” see footnote 1.

Sources: Gridded rainfall from Matsuura and Willmot (2018). Rain-fed cotton output from a digitized map by Hance, Kotschar, and Peterec (1961) showing the spatial distribution of Africa’s commodity exports, including cotton. This map’s accuracy has been confirmed by Roessler et al. (2020). Additionally, Online Appendix A.1.3 shows that the country-aggregated output correlates well with the data independently collected for Figure 1. Sub-nationally disaggregated population estimates from Klein Goldewijk et al. (2017), railroads from Moradi and Jedwab (2016), cotton yield potential (“agro-climatic suitability”) from FAO/IISA (2011). Further details in Online Appendix A.1.2 For replication files, see De Haas (2021).

It is important to consider that we may expect more rainfed cotton output in some rainfall zones compared with others simply due to differences in their endowments of land, population, market access (railroads), and agro-climatic cotton yield potential (Columns (3)–(6)). Rather than presenting absolute values, I calculate shares by simply dividing each rainfall zone’s value over tropical Africa’s total so that all rainfall zones jointly add up to 100 percent. For the population shares, I aggregate the population of all constituent grid cells of each rainfall zone. For the landmass shares, I add up the surface area of each rainfall zone’s grid cells. For the railroad shares, I calculate the railway mileage traversing each rainfall zone. For the cotton yield potential, I measure potential yield at the grid-level and add up the values for each rainfall zone. For this purpose, I use the agro-climatic suitability variable from FAO’s Global Agro-Ecological Zones database, assuming rainfed cotton and low-input conditions. This variable considers how favorable local temperature, radiation, and moisture regimes are for crop growth and maturation. Rainfall is factored in, but only to the extent that it is relevant for the growth of the crop itself, not accounting for the seasonality-induced tradeoffs and constraints faced by the farmer who has to cultivate it.

As Columns (2) to (6) of Table 1 show, very little rainfed cotton was produced in areas with zero to four rainy months (1.9 percent of tropical Africa’s total rainfed cotton output), despite harboring a substantial share of tropical Africa’s landmass (40.1 percent), population (18.2 percent), railroads (21.3 percent), and cotton yield potential (22.4 percent). In areas with five to eight rainy months per year, cotton output was close to the share of landmass, population, railroads, and cotton suitability. In areas with 9 to 12 rainy months, cotton output (41.9 percent) substantially exceeded the shares in tropical Africa’s total landmass (17.1 percent), population (25.5 percent), railroad infrastructure (13.3 percent), and yield potential (16.8 percent).

To further facilitate comparison between the different rainfall zones, I divide cotton output shares by the landmass (Column (7)), population (Column (8)), railroad (Column (9)), and yield potential (Column (10)) shares. The resultant measure takes a value below one if a specific rainfall zone generated less cotton output than we might expect based on its share of tropical Africa’s endowments of landmass, population, railroad, and cotton yield potential. If the value is above one, it generated a larger cotton output share than expected based on these relevant endowments. For example, Column (10) shows that areas with 9 to 12 wet months contributed 2.49 times as much cotton to Africa’s aggregate cotton output as we should expect based on its yield potential. Areas with 5 to 8 wet months contributed 0.86 times the expected output. This allows us to infer that the output per “unit of yield potential” was almost three (2.49/0.86 = 2.90) times higher in areas with two growing cycles than in areas with a single growing cycle.

The above exercise does not have immediate causal implications and does not elucidate any of the mechanisms by which seasonality affected cotton output. However, the fact that substantially more cotton output was generated in areas with two agricultural growing cycles per year than its relevant resource and infrastructure endowments would “predict,” is consistent with our argument that seasonality mattered for cotton (non-)adoption. It also demonstrates that agro-climatic suitability alone poorly explains output. To assess more precisely how rainfall seasonality was linked to cotton output, I now zoom in on two regions, the interior savanna of FWA and Uganda, which were similar in many respects, but had very different rainfall distributions, as well as opposing cotton adoption outcomes.

CONTRASTING CASES: UGANDA AND FRENCH WEST AFRICA

Cotton exports from landlocked Uganda took off soon after the completion of a coastal railway in 1902 (Ehrlich Reference Ehrlich, Harlow and Chilver1965). During the 1920s, output further accelerated and expanded spatially. By the 1950s, cotton production had diffused to a diverse region of ca. 135,000 hectares, inhabited by approximately 4.5 million people (Uganda Protectorate 1961). During the 1950s, smallholder-grown coffee took over as Uganda’s most valuable export crop, but cotton production was maintained on a large scale and continued to be the sole major cash crop in most regions (De Haas Reference De Haas2017b). Uganda’s cotton exports declined precipitously during the 1970s under the Amin regime, in the context of collapsing institutions and the expulsion of Uganda’s commercially important Asian minority (Jamal Reference Jamal1976). Cotton production recovered somewhat later on but never bounced back to pre-collapse levels.

Meanwhile, French efforts to export cotton from the West African savanna started in earnest when two respective railways reached Bamako (the Soudan) in 1904 and Bouaké (northern Côte d’Ivoire) in 1912 (Bassett Reference Bassett2001, pp. 56–8; Roberts Reference Roberts1996, p. 80). Despite high expectations and a persistent politique cotonnière, the volume and quality of smallholder-grown cotton disappointed enormously and remained far behind Uganda’s (Figure 2). The French attempted to generate cotton exports in other ways. In the Niger River Valley in the Soudan, they established the “Office du Niger,” a highly capitalized and irrigated cotton-growing scheme (Roberts Reference Roberts1996, pp. 118–44, 223–48; van Beusekom 2002). In northern Côte d’Ivoire, they invited European planters, who already dominated coffee and cocoa production in the country’s southern forest regions (Bassett Reference Bassett2001, p. 77). However, these projects struggled to attract sufficient labor and settlers and incurred persistent losses, which made an influential strand of the French administration stick to the promise of rainfed cotton cultivation by African farmers (Roberts Reference Roberts1996, pp. 163–91, 149–82). After independence, this promise was finally delivered through a belated but impressive “peasant cotton revolution” (Bassett Reference Bassett2001). From 1961–1965 to 1995–1999, Benin, Burkina Faso, Côte d’Ivoire, Mali, and Togo significantly expanded their joint share of sub-Saharan cotton production from 2.8 to 35.4 percent, and in worldwide production from 0.2 to 2.8 percent (FAOSTAT). I will discuss the changing conditions that enabled this post-colonial take-off in a final section of the paper.

Figure 2 COTTON PRODUCTION IN UGANDA, CÔTE D’IVOIRE, AND MALI, 1900–2000

Note: Data for the Sudan before 1937 refer to export, not production.

Sources: Data post-1960 are from the FAOSTAT database, earlier data from Reference MitchellMitchell (1995, p. 244–45), and Côte d’Ivoire before 1948 from Bassett (Reference Bassett2001, p. 52), and the Soudan before 1948 from Roberts (Reference Roberts1996, pp. 268, 245). Whenever applicable, a 3:1 seed-cotton-to-cotton-lint conversion is used.

The Uganda and FWA cases not only present a compelling contrast of outcomes but also of rainfall seasonality patterns. Importantly, the two regions have similar agro-climatic conditions, which allows us to focus on the effect of rainfall seasonality. Annual rainfall quantities (1901–1960) suggest that environmental conditions in the two regions were overlapping rather than distinct: 1,425 mm in Korhogo (northern Côte d’Ivoire), 1,313 mm in Kampala (Buganda), and 1,373 mm in Lira (northeast Uganda), and a considerably lower annual total of 994 mm in Koutiala (Mali) (Matsuura and Willmott Reference Matsuura and Willmott2018). Aside from southern Uganda’s banana farmers, cropping patterns and agricultural practices in Uganda and FWA were also similar, revolving around annual crops such as grains, oil crops, and beans, supplemented by some cassava and yams (McMaster Reference McMaster1962; Parsons Reference Parsons1960a, pp. 14–67, 1960b, pp. 1–45; Tothill 1940; Jameson 1970). Importantly, Uganda did not outperform the other two countries in terms of cotton yield potential. Instead, the aggregate agro-climatological yield potential of cotton (under rainfed conditions) from Côte d’Ivoire and the Soudan jointly was over 3.5 times as large as Uganda’s (FAO/IIASA 2011). Where environmental conditions in Uganda and FWA clearly diverge is in their very different patterns of rainfall seasonality. As shown in Figure 3A, farmers in northern Côte d’Ivoire had to make do with a fairly short rainy season of seven months, during which they had to procure all of their food and cash crops. The rainy season in the Soudan was even shorter—five months. Ugandan farmers (Figure 3B) instead benefited from a bimodal rainfall distribution with 9 to 11 wet months, which enabled them to harvest two consecutive crops each year.

Figure 3 AVERAGE MONTHLY RAINFALL (IN MILLIMETERS), 1900–1960

Sources: Calculated based on closest grid-point in Matsuura and Willmott (Reference Matsuura and Willmott2018).

LABOR SEASONALITY AND COTTON OUTPUT:A SIMULATION

To quantify how seasonality affected the cotton-growing capacity of farmers in colonial Uganda versus FWA, I estimate and compare the maximum cotton output of a uniform, stylized farming unit under two different seasonality regimes. The farm maintains two adults and three children; cultivates cotton and grain; and is assumed to maximize cotton output facing three constraints: (1) a maximum daily labor input, (2) food self-sufficiency, and (3) a given seasonal distribution of agricultural labor inputs. In this simulation exercise, we hold all conditions constant, except for the intra-annual distribution of labor demands, which include clearing, planting, weeding, and harvesting operations. As the timing of these different operations was strictly circumscribed by the rainfall cycle, this allows us to isolate the effect of seasonality on cotton output.Footnote ³

As a first step in the simulation, I establish a feasible maximum labor availability. The vast majority of farm households in the cotton-growing regions of northeastern Uganda, Côte d’Ivoire, and the Soudan did not procure labor outside the household, aside from “work parties”— short-term village- or clan-based labor exchanges that alleviated some pressure during the peak season (Bassett Reference Bassett2001, pp. 130–2; Tosh Reference Tosh1978). While men and women traditionally had separate tasks (clearing and weeding, respectively), such distinctions largely faded during the colonial period, so that we can assume that households efficiently coordinated their farming duties, especially during the peak season (Boserup Reference Boserup1970, pp. 16–24; De Haas Reference De Haas2017b, p. 622). Women farmed at least as many hours as men during the year as a whole, but men could allocate their full labor time to farming during the peak season, while women had to also attend to reproductive duties, including child-rearing, food preparation, and water and firewood gathering (Bassett Reference Bassett2001, p. 134; Cleave Reference Cleave1974, p. 191; Vail Reference Vail1972, p. 37). Young children, even those not in school, generally performed little field labor (Cleave Reference Cleave1974, p. 191) but contributed by chasing away animals, tending livestock, and sorting the cotton post-harvest. Based on these considerations, I estimate that households were able to provide up to 1.6 adult male equivalents of agricultural field labor per day, every single day of the month during the peak season.

Second, I assume that the farm supported five consumers with an average nutritional requirement of 2,100 kilocalories (kcal) per day (De Haas Reference De Haas2017b) and produced an additional 25 percent surplus to hedge against partial harvest failures. The farming unit cultivated just enough food crops for self-sufficiency and used the remaining agricultural labor to maximize cotton output. This is consistent with a “safety first” farming strategy in a context of land abundance and thin food crop markets (Binswanger and McIntire 1987; De Janvry, Fafchamps, and Sadoulet 1991). In such conditions, an inverse relationship exists between the size of the food crop harvest and the price, which discourages specialization. Absent external demand, food crops cannot be sold profitably in years of localized abundance, leading farmers to diversify into other commercial activities (including cotton cultivation) to obtain non-subsistence commodities and fulfill taxes, school fees, and other cash obligations. In the absence of an external supply of food crops, farmers also avoid relying on non-edible crops, as food cannot be bought affordably and reliably in years of localized shortage (De Janvry, Fafchamps, and Sadoulet 1991, pp. 1401–3). Even in the postbellum southern United States, a context of much better food market functioning than FWA and Uganda, cotton-growing smallholders still pursued a “safety first” strategy. To explain this, Wright (Reference Wright1978, p. 64) points out that “using cotton as a means of meeting food requirements involved the combined risks of cotton yields, cotton prices, and corn prices. The man who grows his own corn need only worry about yields.” In a later section, I will discuss in more detail the potential substitution effects between cotton and food crop marketing.

Third, I estimate the total adult-male-equivalent days of farm labor required per hectare of food and cotton crops. Millet was the most important grain in farmers’ diets in the savannas of FWA and Uganda. I follow De Haas Reference De Haas2017b, p. 609) in estimating a millet labor requirement of 99 adult-male-equivalent days per hectare. Also, following De Haas (Reference De Haas2017b, p. 615) for Uganda, and Benjaminson (Reference Benjaminson, Benjaminsen and Lund2001, p. 264) and Labouret (Reference Labouret1941, pp. 211–6) for FWA, I estimate a typical millet yield of 606 kg/hectare after accounting for waste, losses, and seed retention, and a caloric value of 3,417 kcal/kg. We can now establish that the stylized farming unit required 2.31 hectares of food crops to be self-sufficient, which equates to a labor input of 229 adult-male-equivalent days per year. I estimate total annual labor requirements for cotton at 198 adult-male-equivalent days per hectare (De Haas Reference De Haas2017b, p. 615).

Fourth, I establish the monthly distribution of labor requirements for cotton and food crops which—crucially and unlike the previous three steps—differed between the stylized farms in FWA and Uganda due to their different seasonality patterns. For Uganda, I obtain monthly labor distribution data from 1964–1965, based on 30 farms each in Aboke (Lango district) and Koro (Acholi district), and one experimental farm observed ca. 1970 in Teso district (Cleave Reference Cleave1974, pp. 87, 121; Vail Reference Vail1972, p. 104). For FWA, I rely on data from seven farms in Katiali (northern Côte d’Ivoire), collected by Bassett in 1981 and 1982 (Bassett Reference Bassett2001, p. 126).

The Ivorian data have several issues that need to be resolved up front. First, the observed intra-annual labor-input distribution pertains to a period where farmers had gained improved access to labor-saving innovations. Applying it to an earlier period, before these innovations were adopted, implies that the intervening labor-saving effects were smoothly distributed throughout the year. For most of the cotton-growing cycle, this is a realistic assumption, as animal traction reduced labor inputs early in the season, while the application of herbicides reduced the need for mid-season weeding (Bassett Reference Bassett2001, pp. 107–45).Footnote ⁴ However, harvest labor demands increased between c. 1960 and 1980, as yields increased fourfold for cotton (Bosc and Hanak Freud 1995, p. 282).

It is possible that this increased labor demand was partly alleviated by increased picking efficiency, which could result, for example, from the introduction of new varieties. Indeed, Olmstead and Rhode (2008) have shown that biological innovation increased cotton-picking efficiency on the southern U.S. slave plantations by a factor of four between 1801 and 1862. Unlike in the southern United States, however (Olmstead and Rhode 2008, p. 1142), picking was not “the key binding constraint on cotton production” in FWA, so this type of innovation was not urgent. To be on the conservative side, I still assume that Ivorian cotton-picking efficiency doubled between 1960 and 1980. Footnote ⁵ I have no reason to account for changes in food crop harvest efficiency. I assume that the reported yield increase from c. 600 to 1,000 kilogram per hectare (Benjaminson Reference Benjaminson, Benjaminsen and Lund2001, p. 264) between 1960 and 1980 came with a proportional increase of required labor inputs.

A final limitation of the Ivorian data is that rainfall in the year of data collection tapered off early, in September instead of November (Bassett Reference Bassett2001, p. 120). Importantly, however, we will find that the seasonal labor bottleneck occurred before the year’s rainy season’s premature end, which largely mitigates this concern.

Figure 4 shows the intra-annual distribution of agricultural labor requirements, expressing the monthly inputs as shares of the annual total labor requirements of cotton and food crops, respectively. As can be seen on Panel A, Ivorian labor demands for food crops and cotton peaked jointly in a single agricultural cycle. In their writings, French colonial administrators espoused a clear awareness that cotton and food crops directly competed for scarce labor inputs during this time of the year. According to Bassett (Reference Bassett2001, p. 126), Ivorian officials (1912) observed “that it would be difficult to expand cotton cultivation […] without improving labor productivity. Otherwise, there simply was not enough time in the agricultural calendar if farmers gave priority to food security.” An official in the Soudan (1924) remarked that “the increase in cotton production in this colony will have as its corollary the reduction in the volume of grains […]. The agricultural labor […] will be thus allocated differently, but will not vary” (Roberts Reference Roberts1996, p. 167).

Figure 4 MONTHLY LABOR REQUIREMENTS (SHARE OF TOTAL ANNUAL LABOR INPUTS PER CROP CATEGORY)

Sources: Monthly shares calculated on the basis of data from Bassett (Reference Bassett2001, p. 126), Cleave (Reference Cleave1974, pp. 87, 121), and Vail (Reference Vail1972, p. 104).

In contrast, Uganda’s food crops could be sown and harvested twice during the year, while cotton was typically relegated to the second rainy season (Figure 4, Panel B). Because labor demands for Uganda’s two seasons overlapped from May to August, farmers had to calibrate their cropping choices to accommodate for the parallel tasks of harvesting first-season crops and planting second-season crops. To allow for the harvest of first-season food crops, as well as to mitigate the impact of adverse weather events (such as a hailstorm or short drought, which interfered with the germination of seeds), cotton planting was “staggered” over four months (May to August). Early cotton was sown in newly opened fields, while late cotton was planted in freshly harvested fields (Tothill 1940, pp. 42–52).

As a fifth and final step, I multiply the monthly labor shares (Figure 4) with the annual adult-male-equivalent day requirements per hectare (noted above) to obtain the actual number of adult-male-equivalent days required for each month per hectare of cotton and food crops. By dividing the monthly inputs by the number of days of each month, I establish a daily labor requirement for each month. The hectarage of cotton that can be cultivated is therefore given by the month in which the combined labor requirement for 2.31 hectares of food crops (at self-sufficiency) and cotton (maximized number of hectares) first reaches the point where the 1.6 daily available labor units are fully engaged.Footnote ⁶ Because farmers had access to work parties, which may have enabled them to shift labor inputs during the peak-labor month, I slightly relax this single peak-month condition, instead assuming that labor inputs in the three most labor-constrained months could be smoothed (June, July, and August in both FWA and Uganda). Because there is only one growing season, the procedure of calculating the maximum potential cotton hectarage is straightforward in FWA. In the case of Uganda, an intermediary calculation is necessary to calibrate the optimal shares of food crops cultivated in the first and second seasons to maximize the cotton acreage. In the optimized simulation, Ugandan households cultivated 63 percent of their food crops in the first season and 37 percent in the second season.Footnote ⁷

Figure 5 visualizes the final result. The difference between Côte d’Ivoire and Uganda is large. While farmers in FWA (Panel A) were capable of cultivating at most 0.40 hectares of cotton while retaining food self-sufficiency, farmers in Uganda (Panel B) could cultivate close to three times as much: 1.17 hectares. It is important to recall that this rift in simulated cotton output arises solely from differences in the intra-annual distribution of labor inputs. This, in turn, is a result of contrasting rainfall patterns, which allowed for only a single growing cycle in FWA versus two growing cycles in Uganda. It is reassuring that the simulated results are very close to farm survey data from colonial Uganda (Table 2), as well as farm-level data from colonial FWA. A large Soudanese farming unit was surveyed in 1937 consisting of eight adults (four men, four women) and seven children cultivated 10.84 hectares, including 7.0 hectares of millet, 3.14 hectares of secondary food crops, and 0.7 hectares of cotton (Labouret Reference Labouret1941, pp. 211–6). This translates to 2.54 hectares of food crops and 0.18 hectares of cotton per adult pair (the equivalent of our stylized farm unit), which is within the simulated cotton production possibilities in FWA.

Figure 5 SIMULATION OF MONTHLY FOOD CROP AND COTTON LABOR INPUTS

Sources: Author’s calculations (see text). Rainfall information taken from Figure 4.

Table 2 UGANDAN HOUSEHOLD SIZES AND CROP ACREAGES: SIMULATION VERSUS ACTUAL FARM SURVEY DATA

Sources: De Haas (Reference De Haas2017b, Online Appendix, Table A2).

FOOD YIELDS AND COTTON PLANTING IN UGANDA:A PANEL ANALYSIS

Rainfall seasonality also impacted farmers’ cotton cultivation through a second mechanism. When facing short rainy seasons, farmers had to plant, tend, and harvest all food crops and cotton simultaneously. This created a risk assessment problem; in the Soudan, as noted by Roberts (Reference Roberts1980, p. 53), “where a farmer has devoted a significant portion of his energy to the cultivation of nonedible cash crops, a poor harvest may result in a reduced capacity to survive.” Not being able to anticipate fluctuations in harvest outcomes, farmers had to hedge against the possibility of partial harvest failure. This is why we assumed, in the simulation above, that farmers structurally planned for a “normal surplus” of 25 percent more food crops than their subsistence needs. This implies an inefficient allocation of resources: overinvestment in subsistence crops to reduce risk at the expense of cotton planting and cash income.

What we have not yet considered is that a longer growing season (and especially two growing cycles per year) enabled farmers to avoid growing a surplus, as they could assess the yields of their first season food crops before deciding whether to grow additional food crops in the second season to achieve self-sufficiency or to instead invest in cotton to augment cash income. If Ugandan farmers indeed made an informed decision to adjust their resource investment in cotton planting based on their mid-year food position, we should find that they reduced their second-season cotton acreage after a bad food crop yield in the preceding season, and vice versa. To test this hypothesis, I analyze a panel consisting of 10 districts in colonial Uganda over 38 years (1925–1962). The key variables are the rainfall during the first four months of the year, which proxies for the first-season food-crop harvest (independent) and cotton acreages planted during the second season (dependent).Footnote ⁸

Because cotton was crucial to Uganda’s economy, and output depended on farmers’ annual planting decisions, the colonial administration devised a system to monitor acreage. Local African chiefs were required to count cotton “gardens” in their administrative area. These counts were transformed into an acreage estimate using a standardized conversion based on a sample of representative fields and accumulated at the district level to be reported in the Annual Report of the Department of Agriculture.Footnote ⁹ Chiefs inflated acreages to meet performance expectations, and measuring practices were periodically adjusted.Footnote ¹⁰ However, the acreage statistics emerged from systematic and bottom-up data collection. After the inclusion of district and year controls, we do not expect non-random bias.

Unfortunately, colonial-era data on food crop acreages, let alone yields, was of very poor quality. Instead, we can use rainfall, measured monthly at numerous locations using precipitation gauges, as a workable proxy for harvest outcomes. In tropical conditions, rainfall during the growing season was the prime determinant of yield outcomes. A voluminous body of econometric studies has demonstrated that annual or seasonal rainfall variability explains a variety of economic and social outcomes in Uganda (Asiimwe and Mpuga 2007; Björkman-Nyqvist Reference Björkman-Nyqvist2013; Agamile, Dimova, and Golan 2021) and in economies with high dependence on rainfed agriculture more generally (Carleton and Hsiang 2016; Dell, Jones, and Olken 2014). The effects of rainfall variability have been found to be most pronounced during extreme events in both directions (droughts or floods), but smaller deviations from the expected rainfall pattern also adversely affect output.

To proxy first-season harvest outcomes, I consider the total rainfall during the first four months of the year. While the millet growing cycle lasted from January to August (Figure 5), the rainfall during the first four months of the year (January to April) was most crucial for yields because it was during these months that the newly planted seeds germinated and transformed moisture and nutrients into biomass. Another reason for considering these four months only is that rainfall conditions from May onwards had a direct bearing on cotton-planting decisions, as almost all cotton planting took place from May to August, which interferes with our identification of the food crop harvest effect. I express rainfall deviation during the first four months of the year in z-scores and transform to absolute (non-negative) values to capture the expectation that deviation from the long-run mean has an adverse linear impact on harvest outcomes:

$$Absolute\;RainfallDeviatio{n_{i,t}} = \left| {\left( {{x_{i,t}} - \;{{\bar x}_i}} \right)/{\sigma _i}} \right|$$

where $${\bar x_i}$$ is the long-term mean (1925–1962) of each district, $${x_{i,t}}$$ is the annual observation in time t for district i, $${\sigma _i}$$ and is the standard deviation of each panel (that is, for every i).

For each district, I use observations from the meteorological station for which most data points are available. Footnote ¹¹ I merge two adjacent small cotton-growing districts, Mubende and Toro, because (i) consistent rainfall observations for Mubende are lacking; (ii) unlike other districts, these two districts were effectively treated as a single cotton zone by the colonial authorities; and (iii) for some years their acreage statistics were reported jointly. Although some earlier rainfall and acreage statistics are available, I take 1925 as the starting year because (1) we have almost complete data for all districts by this time; (2) acreage expansions before 1925 were substantial and haphazard and spurred by government cotton campaigns; (3) by 1925 cotton was widely diffused, and acreages (and data quality) had stabilized across districts; and (4) starting in 1925 excludes the sharp acreage fluctuations related to a currency realignment that took place in the early 1920s. The last year considered coincides with the last planting season under colonial rule (which ended on 9 October 1962).

To assess farmers’ investment in cotton, I estimate the following regression model:

$$Ln{(CottonAcres)_{i,t}} = {{\rm{\beta }}_0} + {\beta _1}\;AbsoluteRa\inf allDeviation{}_{i,t} + {\nu _i} + {\mu _t} + {\delta _{i,t}} + {\varepsilon _{i,t}}$$

where $$Ln{(CottonAcres)_{i,t}}$$ denotes the natural logarithm of the cotton acreage planted per district iand year t. District and year fixed effects are denoted $${\nu _i}$$ and $${\mu _t}$$ respectively, while $${\delta _{i,t}}$$ captures unobservable district characteristics $$({\nu _i})$$ interacted with a linear time trend (t) to account for district-specific time trends. The coefficient of interest, $${\beta _1}$$ , is the estimated effect of a one standard deviation change (either positive or negative) in rainfall on the log cotton acreage. A negative sign, $${\beta _1} < 0$$ , indicates that, on average, rainfall deviations from January to April are associated with a lower second-season cotton acreage.

The results are reported in Table 3. Column (1) shows the results without district-specific time trends, which are added in Column (2). In Column (3), I add the rainfall from May to August (the cotton-planting season) and September to December (the cotton-growing season) as controls. In Column (4), I remove the years 1939 to 1945, as war conditions affected both cotton cultivation and data-reporting quality. Footnote ¹² The results are significant and stable in each of these specifications, with a one standard deviation change in first-season rainfall reducing the cotton acreage by approximately 6.0 percent. Over the 38 years, cotton acreages potentially followed district-specific non-linear time trends or had structural breaks (e.g., unobserved changes in acreage measuring practices related to the preferences of a specific administrator or agricultural officer). To ascertain that the results are not driven by any trends and breaks in the time series that may correlate with rainfall, I replace the log cotton acreage (a level variable) with the log cotton acreage minus the log cotton acreage of the previous year in Column (5), and relative to the median of the previous three years in Column (6) (both change variables). The results hold up to this change of dependent variable.

Table 3 EFFECT OF FIRST-SEASON ABSOLUTE RAINFALL DEVIATION ON SECOND-SEASON COTTON ACREAGE, 1925–1962

Notes: Robust standard errors in parentheses.

*** p<0.01, ** p<0.05, * p<0.1

Source: Cotton acreages from Uganda Protectorate, Annual Reports of Agriculture for the Year [1926 to 1962]. Rainfall, see Online Appendix A.3.3.

In Column (7), I take the baseline specification with log cotton acreage but exclude outliers of cotton acreage growth or decline relative to the previous year (over two standard deviations). In Column (8), I exclude outliers with extreme rainfall (over two standard deviations). Reassuringly, the effect is not driven by extreme spikes of the dependent and independent variables. Finally, to establish whether there was a heterogeneous impact of excessively dry versus wet conditions, Column (9) shows the result using shock dummies, taking 1.25 standard deviations as the cut-off point. Both types of shocks are, on average, associated with a smaller cotton acreage, but only significantly so for the positive (excessive rainfall) shocks. Column (10) replicates this shock-specification using log cotton growth as the dependent variable (as in Column (5)).Footnote ¹³ The coefficient for positive shocks remains similar, and negative shocks also have a significant effect in this specification. All results are robust to correcting standard errors for the small number of clusters using the “wild bootstrap” procedure (Cameron, Gelbach, and Miller 2008).Footnote ¹⁴ A range of falsification exercises with rainfall lags, leads, annual rainfall, and linear (non-absolute) rainfall effects does not yield significant associations, as we should expect.Footnote ¹⁵

REPERCUSSIONS OF SEASONALITY IN FWA VERSUS UGANDA

We have seen that it is plausible that differences in seasonality set apart the cotton-growing capacity of farmers in FWA and Uganda; however, to come to a closer understanding of the actual impact of seasonality on divergent cotton outcomes, we must consider a range of alternative explanations and contextual dynamics. I draw from the literature to identify the most important such explanations and dynamics and discuss how they interacted with the constraints that seasonality imposed on agriculture.

OVERCOMING SEASONALITY IN FRENCH WEST AFRICA

Strikingly, FWA realized a major cotton take-off in the post-colonial period (Bassett Reference Bassett2001), which shows that the constraining role of agricultural seasonality was not immutable. I identify two contextual factors that explain why seasonality was a major constraint in the colonial period and not afterward. First, specialization was inhibited by poorly developed markets for food, the causes of which I have already discussed. With rapidly rising urbanization and economic diversification after independence, food marketing became more lucrative. Second, agricultural output was constrained by the absence of yield-enhancing and labor-saving technological breakthroughs during the colonial period, which were achieved after independence.

During the colonial period, agricultural innovations such as new crop varieties (Arnold Reference Arnold and Jameson1970, pp. 155–64; Dawe Reference Dawe1993, pp. 149–59; Roberts Reference Roberts1996, pp. 224, 253–4) or the plow (Tosh Reference Tosh1978, p. 435; Roberts Reference Roberts1996, pp. 147, 176–7; Vail Reference Vail1972, p. 71) were haphazardly developed and had marginal impacts on labor productivity at best. In light of these persistent constraints, the cotton take-off in post-colonial FWA (Figure 2) is truly remarkable. How were farmers suddenly able to overcome the seasonality bottleneck that had prevented cotton adoption and frustrated colonial officials for over a half-century? The answer to this question largely resides in persistent research and extension efforts by the French former colonizers in collaboration with post-colonial governments (Lele, van de Walle, and Gbetibouo 1989). Even though they were unable to effectuate their ambitions, some colonial officials had understood early on that only by transforming labor productivity through higher yields and more efficient farming practices could cotton achieve the desired success (Roberts Reference Roberts1996, p. 168). From the 1960s onwards, in a context of a global “Green Revolution,” the renewed concerted efforts to generate export cotton finally produced a well-rounded package of new technologies and inputs that increased crop yields and reduced labor requirements per hectare. That farmers in FWA proved willing and able to adopt these technologies testifies to their suitability in a context of seasonality labor–constrained savanna agriculture (Bassett Reference Bassett2001; Benjaminson Reference Benjaminson, Benjaminsen and Lund2001; Bosc and Hanak Freud 1995; Lele, van de Walle, and Gbetibouo 1989).

As a result of the introduction of improved varieties and the application of mineral fertilizers, grain yields in Côte d’Ivoire and Mali rose from an initial 600–800 kg/hectare to over 1,000 kg/hectare for millet and sorghum, and over 1,700 kg/hectare for maize by the early 1980s (Benjaminson Reference Benjaminson, Benjaminsen and Lund2001, p. 264), before stagnating at this higher level. Seed cotton yield gains over the same period were even more impressive, rising approximately fourfold from c. 300kg/hectare to 1,200 kg/hectare (Bassett Reference Bassett2001, p. 186; Fok et al. 2000, p. 14).Footnote ²⁹ Various innovations were adopted to a greater or lesser extent across FWA, substantially increasing labor productivity (Bosc and Hanak Freud 1995).

We can re-run the same simulation introduced earlier for FWA but now based on the higher labor productivity in the post-colonial era. I take into account that food crop production shifted from millet towards maize, which has a similar caloric value per kg but higher yields (Bassett Reference Bassett2001, pp. 127–9; Bosc and Hanak Freud 1995, p. 290). I conservatively estimate that food crop yields doubled and that labor demands per hectare in both food crop and cotton cultivation halved. Under these new conditions, farmers were able to cultivate 6.5 hectares of cotton, more than a sixfold increase compared to the colonial era. Cotton yields per hectare also increased fourfold. As a consequence, smallholders’ cotton production possibilities had risen over 26-fold. Footnote ³⁰ This extension of agricultural production possibilities is reflected in the massive expansion of cotton output that was achieved over this period (Figure 2), while a further part of the increased production possibilities was allocated to surplus grain production (Bingen Reference Bingen1998 p. 271; Benjaminson Reference Benjaminson, Benjaminsen and Lund2001, p. 264).

CONCLUSION

Colonial efforts to spur agricultural exports from Africa, and cotton, in particular, produced unanticipated and uneven results. To understand such heterogeneous outcomes, we need to look beyond the dynamics of colonial coercion and investment, and instead consider how resource constraints informed African farmers’ adoption decisions and often thwarted colonial designs. I have shown that African savanna farmers facing short rainy seasons were unable, at the prevailing levels of technology and market access, to combine a substantial involvement in cotton production with sustained food security. In the context of short rainy seasons, farmers did not have enough labor at their disposal in the agricultural peak season and could not assess their food security before committing resources to an inedible cash crop. Colonial states, despite persistent attempts to generate cotton output, were unable to mitigate this bottleneck. As a result, cotton was far more widely grown in regions with comparatively long rainy seasons, most notably Uganda. Only after independence did the introduction of labor-saving technologies in the former French territories of West Africa enable widespread and large-scale cotton adoption among its savanna farmers.

Some of the dynamics studied for cotton will also be relevant to other savanna cash crops in Africa. Still, savanna farmers in northern Nigeria, Senegal, and the Gambia proved able to produce large quantities of groundnuts for export. Although these cases can only be addressed briefly here, it is important to note that their success relied on very specific conditions, which were not (yet) present in other parts of colonial Africa: large-scale rice imports in the Gambia and Senegal (Swindell and Jeng 2006; van Beusekom 2002) and agricultural intensification and fine-grained rural trade networks in northern Nigeria, linked to the region’s historically high population densities and pre-colonial processes of state formation (Hill Reference Hill1977; Hogendorn Reference Hogendorn1978). Emerging conditions in these areas, then, foreshadow a much wider breaking of resource bottlenecks in savanna agriculture, which colonial states, despite their efforts to export cotton from Africa, proved unable to effectuate, but which eventually occurred after independence in FWA and beyond.

Acknowledging the central role of seasonality in patterns of cash-crop diffusion in colonial Africa also has wider repercussions for how we evaluate the impact and legacies of colonial rule. Dynamics of colonial coercion and local resistance were certainly important features of cotton imperialism in Africa, as emphatically shown for the cases of the Belgian Congo (Likaka Reference Likaka1997) and Mozambique (Isaacman Reference Isaacman1996), among others. However, as I have shown, farmers in Uganda adopted cotton to a greater extent, not because they were subjected to more coercion than their West African counterparts, but because they benefited from more favorable rainfall seasonality. In a context of poorly developed markets and limited technological progress, the absence of stringent seasonal resource constraints proved decisive. Colonial interventions were at least partially endogenous to farmers’ responses: in the case of Uganda and FWA, more output spurred investment and lessened coercion. The implication is that processes of cash crop adoption as well as colonial coercion cannot be understood without giving full consideration to resource constraints in general, and agricultural seasonality in particular.