Imperfect competition in agricultural markets: evidence from Ethiopia

Journal of Development Economics 76 (2005) 405 – 428 www.elsevier.com/locate/econbase

Imperfect competition in agricultural markets: evidence from Ethiopia Theresa Osborne Department of Economics, European University Institute, Via della Piazzola, 43, 50133 Firenze, Italy Received 1 February 2001; accepted 1 February 2004

Abstract Drawing upon unique transaction-level data from rural Ethiopia, this paper tests for general forms of imperfect competition among rural wholesale traders. These are key to the grain distribution system as they purchase from farmers and perform interregional trade. Tests show that traders in a typical source market engage in imperfectly competitive behavior in purchasing from farmers, driving down the price paid to farmers approximately 3%. In contrast, there is no conclusive evidence of imperfect competition among traders in the larger, more centrally located market studied. Thus, efficiency losses due to market structure are likely to be greatest in markets which also have poor road links and lesser volumes of marketed grain. D 2004 Elsevier B.V. All rights reserved. JEL classification: D43; L13; O12; Q13 Keywords: Imperfect competition; Seasonality; Agricultural marketing system; Africa

1. Introduction This paper examines the performance of agricultural output markets in Ethiopia, a country noted for its low agricultural productivity and high food insecurity. There is growing recognition among economists and policymakers of the importance of market performance in providing appropriate production and consumption incentives in such settings.1 Yet, little E-mail address: [email protected]. 1 The classic works on the relationship between markets and famine or food security are Ravallion (1987) and Sen (1981). 0304-3878/$ - see front matter D 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.jdeveco.2004.02.002

406

T. Osborne / Journal of Development Economics 76 (2005) 405–428

is known about even the most critical markets. Some Ethiopian observers allege, for instance, that grain traders bcheatQ farmers by paying too low a price for their grain. Traders and brokers themselves admit to discussing (and in the case of brokers, setting) prices (source: Interviews by author). If present, imperfect competition in these markets would have important efficiency and distributional effects. It could, moreover, raise the price paid by net grain consumers, among whom number the most vulnerable sectors of the population.2 Concerns about the distributional consequences of unchecked market forces, in addition to other economic and political agendas, have prompted many governments to intervene directly in output markets using such means as marketing boards, price controls, and strategic stock-holding. Ethiopia has been a particularly unfortunate example of this, having endured a period of severe state intervention in the grain trade. In 1979/1980, the former (Derg) government adopted a set of measures, collectively called the dquota systemT, which heavily taxed both farmers and wholesale traders, restricted trading licenses, and imposed severe penalties (including imprisonment and death) on violators. Abandoned in 1990, the quota system has been partly blamed for Ethiopia’s persistent problems of low agricultural productivity, food insecurity, and famine (see, e.g., Franzel et al., 1989, Lirenso, 1987).3 Throughout this period, however, a private grain market survived and some traders actually thrived. The abolition of the quota system in 1990–1991 led, as might be expected, to a general reduction in the risks and costs of trading. Entry resulted, and the efficiency of these markets has improved dramatically. Dercon (1993), for example, has shown that market integration within the country improved substantially following 1990. In addition, traders surveyed in 1994 unanimously reported that trading margins were much higher (and volumes lower) under the quota system, suggesting a marked improvement in efficiency (source: Traders’ Survey by author). This paper tests for deviations from perfect competition in interregional trade under the relatively free market conditions that prevail in post-1990 Ethiopia. There is a substantial body of research on imperfect competition in developed markets in goods ranging from fish (see, e.g., Graddy, 1995) to asset trading (e.g., Wang, 1999), automobiles and airline seats. Yet, despite considerable policy attention, there is little empirical research on conditions in developing country markets, where the effects would be arguably more severe (here, exceptions include, e.g., Aleem, 1990; Ellis, 1993; Usher, 1968).4 The primary reason for the general paucity of empirical studies, it seems, is the lack of suitable data. Thus, unique data were collected for the purpose of this paper on

2

Earlier studies of Ethiopian grain markets have found suggestive evidence of imperfect competition. For example, Pickett (1991) estimates the internal rate of return on trading to be 48% using data collected in 1957. This may indicate super-normal profits. 3 The country’s long period of civil war is also considered a part of the explanation, although the true causes are undoubtedly many. 4 Other studies of agricultural markets include Fafchamps and Minten (2002), which examines the role of social network capital for traders in Madagascar, and Gabre-Medhin (1999a, 1999b), who examines the role of brokers in reducing transactions costs in Ethiopia.

T. Osborne / Journal of Development Economics 76 (2005) 405–428

407

transactions by key agents in the marketing system.5 These are rural wholesale traders, who purchase farmers’ grains and ship them for sale to other regions—particularly to the main national market in Addis Ababa, where they may then be rerouted to other deficit areas. Ethiopia is poor and rural even by African standards, yet it possesses relatively high agro-climatic diversity, which makes regional trade in food grains extremely important. Thus, the link between farmer and wholesale trader is an important determinant of farmers’ income, marketed supply, and interregional grain flows. The aim of this paper is to test for imperfect competition in this link using detailed data on traders’ purchases. To do this, I estimate a general demand relations equation for the traders’ bid price as a function of expected net marginal revenue (ENMR; anticipated selling price less marginal costs) and quantities supplied. I deal with possible selectivity bias and simultaneity using three different estimators. The analysis is conducted for two markets, one relatively lowvolume market which is a day’s journey from the central grain market in Addis Ababa and one high-volume market about one hour from the capital. The results provide clear evidence of imperfect competition in the smaller and more remote market. In contrast, there is no conclusive evidence of imperfect competition in the higher volume, nearer market. This comparison across markets suggests that the problem of excessively high marketing costs may be even greater in the many smaller and more remote markets than either of those studied here. In the next section, I briefly describe the Ethiopian grain marketing system and the key actors. In Section 3, I describe the data and present the main patterns they exhibit. In Section 4, I motivate the estimation strategy, which is then presented in Section 5. In Section 6, I present the results, and Section 7 concludes.

2. Market background bWholesaleQ or bruralQ traders operate in fairly similar ways throughout Ethiopia’s traditional surplus areas. Traders in Yetmen, the first market analyzed, can thus be considered typical.6 Yetmen, a village of approximately 2500 inhabitants, is situated in the highlands along an all-weather (albeit dirt) road 248 km northwest of Addis Ababa. Traders each own or rent a storehouse, where they open on market days 3 days/week, awaiting the arrival of farmers selling grain (usually on foot, with the grain sometimes carried by donkeys). Traders will then weigh the grain using a large special-purpose scale, examine its quality, and make a cash offer, which farmers may then accept or reject. Some time after thus collecting these grains, the traders clean, package, and ship them by truck 5 The data were collected under a project financed by the Mellon Foundation with support from Princeton University. The data collection was designed and managed by the author, and data recorded with the help of Million Taffesse, and Makuanint Gatenet and Kebede. 6 My analysis herein is aided by a trader survey which I conducted with the help of a field assistant in several areas of the country. The survey included questions about various aspects of traders’ operations, in particular about the information environment, cost structures, price expectations, and credit and liquidity issues. This survey showed that a market’s exact roles depend fundamentally on the size of local consumer demand and the need for external sources of supply. Nonetheless, it demonstrated that traders’ business relationships and practices were very similar among surplus markets.

408

T. Osborne / Journal of Development Economics 76 (2005) 405–428

for distant sale. For traders in Yetmen, as for many source markets, the destination is the central grain market (ehil beranda) in Addis Ababa. There, they are typically resold to institutions, other traders, or to consumers. Casual observation suggests that there is potential for noncompetitive behavior among rural wholesale traders in purchasing farmers’ grain. Although significant entry has occurred since liberalization (1990), there are barriers to entry in the form of capital market imperfections coupled with high fixed costs. There is also some evidence of asymmetric information between traders and farmers regarding the bfinalQ sale price in Addis Ababa, which could disadvantage farmers in an imperfectly competitive setting.7 Information flows between traders locally appear to be relatively free, as they meet and discuss prices regularly (source: Traders Survey). Moreover, any attempts at collusion can be facilitated by the frequent interaction of traders and the ease of observing each others’ stores. A deviation which attracted more sellers of grain than usual could be quickly detected and bpunishmentQ swiftly follow. Furthermore, if traders face capacity constraints, the benefit of deviating might be small (see, e.g., Brock and Scheinkman, 1985). 2.1. Debre Zeit market Debre Zeit is a large town with a relatively active market approximately 80 km from Addis Ababa, with road links to demand centers to the north and south and rail links to the east. The Debre Zeit market transacts more volume and has more traders than Yetmen, and wholesale traders also sell retail (small amounts) in the local market. Some Debre Zeit traders also buy grains in bulk on nonmarket days through agents, in addition to directly from farmers on market days. Purchase prices do not vary within the day or across traders in the sampled transactions, in contrast to those observed in Yetmen. Debre Zeit’s relative proximity to Addis Ababa and its greater number of traders may make imperfect competition less likely there. 2.2. The Addis Ababa mercato, or ehil beranda An estimated 53,000 tons is traded in the Addis Ababa market annually, on the order of a third of nationally marketed grains. Ehil beranda brokers coordinate all wholesale trade there by acting as sale agents for rural wholesale traders, and greatly economize on traders’ transactions cost of selling (Gabre-Medhin, 1999a, and Traders’ survey). First, they are an important conduit for information on market conditions. In addition, they make it unnecessary for sellers to travel to the market themselves. Finally, they reduce wholesale traders’ risk by storing unsold grains at minimal cost and providing funds in advance of sale at zero interest (keeping a fixed fee of 1.5 birr/quintal for their services). While brokers may coordinate to obtain a higher selling price for wholesale traders, this would be 7 Grain prices vary substantially by quality, color, and point of origin, and there is no quality standardization to facilitate price comparisons. Thus, of 10 farmers in Yetmen tested by the author for their knowledge of the current Addis Ababa price of Gojjam white teff, only 1 could provide an answer, and this was approximately 5 birr too low (relative to a price of approximately 200 birr). In addition, Yetmen traders quote total payments for a given purchase. This must make it more difficult for farmers to know what unit price they are being offered.

T. Osborne / Journal of Development Economics 76 (2005) 405–428

409

of little value to traders if these gains were competed away in their source markets, such as those studied here.

3. The data Ten wholesale traders were randomly selected for detailed observation in Yetmen and Debre Zeit. For four pre-established hours of the day, two in the morning and two in the afternoon, on major market days (3 days per week) throughout one year (February 1994– February 1995), one trader at a time was observed on a rotating basis. Enumerators recorded the price, quantity, and type of grain on all purchases in each of these hourly intervals. Direct observation was necessary because traders do not keep written records and could not give accurate answers to how much business they transacted recently or at what prices. To capture the total volume and timing of traders’ sales, enumerators recorded all shipments by date and trader out of these markets. In Yetmen, because the price of trucking varies daily, the enumerator also noted the trucking price per quintal for each shipment. In Debre Zeit, the trucking price is fairly constant and was therefore not recorded for each shipment. For data on traders’ selling prices, I used Addis Ababa wholesale price data obtained from the government’s grain trading agency, the Ethiopian Grain Trade Enterprise (EGTE), which conducts price surveys thrice weekly in the ehil beranda. These data are considered fairly reliable, as the EGTE has an active role in the market and an interest in accuracy. Important quality differences are accounted for, as data are collected separately for bwhiteQ, bredQ, or bmixedQ teff, sorghum, barley, and wheat varieties, and by main source regions.8 The difference in price was substantial between, for example bGojjam red teffQ and bAd’a magna (white) teffQ, the most valuable type. I therefore use the appropriate color and source region to match up sale prices for each transaction. Using these data, contemporaneous observed net trading margins, denoted m ˜ t , can be calculated as P t
p t
c˜t for all observations t, where the local purchase price by traders is denoted as p t , the selling price in Addis Ababa as P t , and c˜ t represents observable marginal costs. Observed marginal costs include labor to clean and load grains, sacks, brokerage fees, and trucking charges, which represent the largest share of c˜ . Examination of m ˜ , as well as of quantities transacted, reveals a pronounced seasonal pattern. Fig. 1 shows the proportion of total observed purchases and sales occurring in each week over the year. Purchases are higher at the beginning of the period, which is the postharvest period, then drop in the lean season and rise again as the following harvest comes in. This seasonality is much more pronounced in Yetmen than in Debre Zeit, where farmers are on average wealthier. As shown in Fig. 2, average weekly m ˜ are highest in the postharvest season (February–March), lower in the lean season, and then rise again in the following harvest season (December–February). Again, the pattern is more pronounced in Yetmen.9 8

Teff is a high-value grain grown almost uniquely in the Horn of Africa. There are some periods in which margins are negative between Debre Zeit and Addis Ababa. This is unlikely to mean that grain flows reversed, but both markets could have become sources for deficit areas to the south. Alternatively, because the data are averages over a few sellers, the sale price reported in Addis could be systematically lower due to the high quality of Debre Zeit area grains. 9

410

T. Osborne / Journal of Development Economics 76 (2005) 405–428

Fig. 1.

These patterns are consistent with the anecdotal evidence, both in Ethiopia and elsewhere, that farmers who are the most liquidity-constrained at the time of the harvest tend to sell right away at a blowQ price. Lirenso (1987), for example, presents evidence of early selling by poorer Ethiopian farmers. Ellis (1993) also finds that early sales by farmers in Indonesia are attributable to the need for cash. In a small follow-up survey for this paper, all farmers surveyed in Yetmen (N=10), and all but one in Debre Zeit (N=10) claimed that they could not hold their grain and wait for higher prices, as they bneededQ the cash. In addition, the peak in m ˜ at the start of the new harvest season shown (week 41, Fig. 2) coincides with the repayment date on government-sponsored fertilizer loans, suggesting that traders can at least occasionally set lower p (higher m ˜ ) when grain is supplied inelastically. Thus, farmers who sell most of their output early are likely to be those with less cash savings, those who have seasonal loan repayments due, or those for whom the cost of financing grain storage is prohibitively high. Those farmers, such as many around

T. Osborne / Journal of Development Economics 76 (2005) 405–428

411

Fig. 2.

Debre Zeit, who have other liquid assets or alternative income streams are better positioned to hold out for a higher price.

4. Theoretical motivation To motivate the estimation strategy, I first examine a simple theoretical model of trader behavior under the null of perfect marginal pricing. Traders observe information known in each period t, I t , and form expectations about future selling prices and costs. On this basis,

412

T. Osborne / Journal of Development Economics 76 (2005) 405–428

they decide whether to store grain speculatively. Defining net marginal revenue (NMR) as selling price less full marginal cost, c, or P
c, expected NMR (ENMR) is the trader’s expectation of the highest obtainable NMR in the future. Under perfect competition, traders would observe the rule: buy if today’s purchase price is less than or equal to ENMR. This condition is written as:10   l pt V Et max ðPtþk
ctþk ðqt ÞÞjIt ð1Þ k¼0

where c includes all observed and unobserved marginal costs (including a risk premium) and may be varying in quantity, q. Traders are assumed here to be sale price-takers. Bertrand style (perfect) competition for grain would result in a one-for-one pass through of all changes in ENMR to farmers, as below:   l ð2Þ pt ¼ Et max ðPtþk
ctþk ðqt ÞÞjIt : k¼0

Note, given Eq. (2), that the pattern of expected future price changes would not explain the seasonality in observed margins (m ˜ ) under perfect competition. In the early harvest season, market participants generally expected substantial price increases, and this is indeed what occurred.11 The price of teff rose approximately 50% and the price of maize doubled in the few months just following the harvest.12 Thus, if anything, consideration of future prices would have increased p (and lowered m ˜ ) in the harvest season vis-a´-vis contemporaneous NMR. Deviations from Eq. (2) may result for a couple of reasons. First, in the presence of declining average costs, traders may not reach minimum efficient scale, so that restrained competition may be necessary to cover average costs. Second, liquidity constraints may inhibit competition (creating super normal profits), particularly if constraints bind for several traders at once. Finally, traders may act to restrict competition, either tacitly or overtly, with the aim of capturing supernormal profits. Each of these is considered a form of imperfect competition. However, conditioning on the market’s structure (which itself may not be optimal) not all represent inefficient behavior. Thus, any deviation from marginal pricing would have to be interpreted carefully. To motivate departures from Eq. (2) arising from noncompetitive behavior, consider a trader’s static optimum offer price:   Bpt l pt ¼ Et max ðPtþk
ctþk ðqt ÞÞjIt
qt s : ð3Þ k¼0 Bqt Here, Bp t /Bqst is the slope of the supply curve faced by the trader. This would equal the slope of market supply under monopsony and zero under perfect competition. For N 10

This means that it could sometimes be profitable for traders to buy at a price greater than current NMR, but not if they are also simultaneously selling grain. 11 Source: Interviews of market participants (traders and government officials) in early harvest season, and follow-up survey of Traders. 12 The issue of whether overall storage in Ethiopia is efficient and possible reasons why (or why not) are explored in Osborne (2004).

T. Osborne / Journal of Development Economics 76 (2005) 405–428

413

traders, 1bNbl, Eq. (3) represents the static Cournot outcome. Thus, a general equation capturing traders’ pricing behavior can be expressed in the following Demand Relations (DR) curve for trader i at time t:  
l pit ¼ l þ Et max ðPtþk
ctþk ÞjIt þ /q4 vit ; pit ; pjt ; k¼0

ð4Þ

where the variable q* is quantity purchased in a given observation (an hour of trader activity) as a function of exogenous supply factors v it , other traders’ prices, p jt , and the current bid price, which is in turn affected by ENMR. Eq. (4) nests perfect competition (/=0), Cournot behavior (/=Bp/Bq s ), monopolistic competition, and other more collusive types of behavior. More generally, the more negative the value of / (V0), the greater the degree of competitive restraint as a response to quantity supplied. With negative / (and fixed l), farmers would tend to receive a lower share of ENMR in the harvest season, when q is high, as observed in the data. Eq. (4) also nests brule of thumbQ supernormal margin setting, wherein traders fix margins (or prices) independently of q by agreeing on a lower l, although this will be difficult to identify separately from unobserved average costs (see Corts, 1999). In addition, l and / may vary over time due to the varying sustainability of imperfectly competitive behavior. Indeed, one possible explanation of the seasonality observed in net margins is the presence of imperfect competition among traders which is relaxed or breaks down in the lean season.13

5. Estimation strategy To test for deviations from marginal pricing, I specify the estimating equation based on Eq. (4) as shown:
pit ¼ l þ a Pi;tþk
c˜ i;tþk þ /qit þ bXit þ git ;

ð5Þ

where full ENMR is now replaced by observable ENMR (OENMR). That is, c is replaced by observable marginal costs and ENMR is replaced by P
c˜ at the time of sale (t+k). I will not be able to control directly for unobserved marginal costs, but as I will argue below, given the technology used in trading this is unlikely to lead to spurious findings. X is defined as a vector of exogenous shifters of farmers’ supply, traders’ cost, and other traders’ demand factors and g is a composite error term, to be defined more precisely below. Note that in Eq. (4), a (which appears in Eq. (5)) is implicitly equal to 1. Thus, an estimate of a=1 would mean that shifts in OENMR are well captured in the estimation and we could interpret /ˆ as the slope of the DR curve.14 13

It is not problematic in a dynamic framework to establish the sustainability of collusion under seasonal shifts in farmers’ supply. (See Rotemberg and Saloner, 1986; Green and Porter, 1984; Slade, 1985.) 14 This is true as long as the DR curve is linear in q. I experiment with nonlinear functional forms, and the linear form performs best in estimation.

414

T. Osborne / Journal of Development Economics 76 (2005) 405–428

One approach often taken in the literature to distinguish between forms of imperfect competition is to first estimate the (average) slope of the traders’ supply curve. This estimate is then compared to the estimate of /ˆ from an equation analogous to Eq. (5), to distinguish the influence of bmarket powerQ alone (i.e., Cournot behavior) from that of bconductQ, which involves traders’ departure from Cournot behavior.15 This approach, called the bConduct parameter methodQ, can present various problems of interpretation, as pointed out by Corts (1999), and proved infeasible here given the limitations of the data.16 This paper, moreover, is primarily concerned with testing for general forms of imperfect competition. 5.1. Unobserved marginal costs A key difficulty in drawing inferences through estimation of Eq. (5) lies in the unobservability of some elements of marginal cost, which I denote as c u (uc
c˜ ). Observed marginal costs, c˜ , may of course be rising or falling in q; however, these will be captured in OENMR and thus will not lead to a biased estimate of /. However, if unobserved marginal costs of capital, risk, and traders’ time are rising in q, then the OENMR–c u curve would be downward sloping. Thus, one would obtain /ˆ b0 even under perfect marginal pricing, and one could not identify deviations from perfect competition. If, as is more likely, however, c u is declining in q, then /ˆ will be upward biased (OENMR–c u will be upward sloping). In fact, imperfectly competitive behavior aimed solely at recapturing falling c u would still lead to the finding of /ˆ =0: The OENMR–c u curve would be horizontal. Given the simple trading technology used one can essentially rule out the possibility of rising unobserved marginal costs. Consider some of the various elements of c u in turn. First is the potential cost of physical storage space, which can lead to rising c u ( q). However, in the output range studied here, physical storage constraints rarely if ever bind (source: Trader survey), and additional costs to expand storage space are not actually incurred. This leaves as the main unobserved costs those of capital, risk, and the trader’s time. Consider the cost of capital. Given supply conditions, and assuming a fixed shadow cost of funds, r, it is difficult to see how the cost of capital per unit traded could be rising in q. As the time it takes to turn over invested capital increases, so does its opportunity cost. In markets, like Yetmen, where grain flows are relatively slow, it can take days or weeks for traders to assemble a full truckload. Traders may wish to hold grain beyond that point, but this would be for speculative purposes rather than technological necessity. If traders buy larger volumes, however, turnover can be rapid 15

Still, under less than full monopsony, / comprises total competitive restraint (and does not equal

Bp Bc s
Bq S
Bq .) Thus, even if Bp/Bq could be identified, it would still be somewhat problematic to separate the u

effects of any unobserved marginal costs from market power per se. 16 Attempts to estimate the slope of the supply curve using weekly data produced negative coefficients, most probably due to the frequency of supply curve shifts which cannot be explained by the observable variables. Thus, there is a great deal of endogeneity between p and q. However, under the estimation method used, one can control for this endogeneity without separately identifying supply.

T. Osborne / Journal of Development Economics 76 (2005) 405–428

415

and the working capital cost per unit thereby lowered. For this reason, the unobserved cost of capital should be falling in q.17 Moreover, the marginal cost of capital curve should shift higher in the lean season when supplies dwindle and prices (or amount invested per unit, p) are higher. One must also consider, however, whether r adjusts either seasonally or simultaneously with quantity supplied. In principle, traders could borrow from the Commercial Bank of Ethiopia at a fixed interest rate (12%, although collateral requirements were more onerous). Moreover, informal credit contracts in Ethiopia as reported by traders and farmers are surprisingly invariable. Traders lend to each other short term at zero interest and obtain zero-interest loans from their brokers (source: Traders’ Surveys), so that the marginal cost of capital can be very low regardless of season. One form of informal loan contract even suggests that the opportunity cost of capital may be higher in ˆthe lean season.18 Finally, a seasonal shift in the shadow interest rate can be absorbed in l s , where this parameter is allowed to vary by season s. Risk is likely to play a role in traders’ decisions and may also create a deviation from observed marginal (OENMR) pricing (under true marginal pricing). However, risk will not be strictly positively related to quantity transacted. A trader’s risk would increase the greater is his total investment (i.e., total stock), and the higher is the purchase price ( p). While stockholdings temporarily increase more in the short run when q is high, the trader’s total risk can also be more quickly unwound at those times by facilitating a quicker sale of all current stock. In addition, prices tend to be highest when volumes are lowest, at least partially offsetting any direct effect of q. Finally, as discussed in Section 2, traders are able to mitigate risk somewhat through the services brokers provide. These factors would make a systematic positive effect of quantity transacted on assumed risk unlikely. Lastly, consider the unobserved cost of the trader’s time per unit traded.19 At a fixed shadow wage (w), this will be constant given supply conditions, but will shift down when supply ( q) is higher. Greater supplies reduce the time it takes to collect a certain amount of grain and, thus, the trader’s labor cost (at a constant shadow wage). Therefore, this cost will generally be falling in observed q as well.20 Given the above considerations, I maintain the assumption of constant c u in estimating Eq. (5). Given that these costs could be declining in q, the estimate of / obtained may actually understate the true deviation from marginal pricing. 17 However, the cost curve should be lumpy. The first quintal purchased in a load will have to be stored longer than the later quintals, and working capital invested in it is turned over more slowly, whereas for the last quintal, capital can be recouped right away. Under perfect competition, this lumpiness should not affect the market price, unless all traders are at similar phases of stock accumulation. 18 Moneylender loans to farmers (to which traders are usually not a party) are observed in many parts of the country only in the lean season. These loans are repayable only after the following harvest, with implied annualized rates of interest of 200% (or the option of repayment with 100 kg of teff). 19 Gabre-Medhin (1999a) estimates the shadow daily wage of her sample of traders and the shadow cost of capital to be 40 birr/day and 15% on average, respectively. Unfortunately, estimating these prices requires the assumption of perfectly competitive behavior. 20 On the margin, traders’ disutility of work (and w) may be higher when supplies are greatest on a per hour basis, since they work more hours. However, this would not translate to higher costs on a per unit traded basis.

416

T. Osborne / Journal of Development Economics 76 (2005) 405–428

5.2. Estimators There are two important econometric issues involved in estimating Eq. (5). First is the potential simultaneity both between p and OENMR and between p and q. Second is the fact that price ( p) is observed only when a positive transaction takes place. That is, letting q* represent latent quantity and q observed quantity, one only observes q* for which q*N0. This is a particular issue in Yetmen, whereas q=0 observations are very rare in Debre Zeit. I address the potential simultaneity and censoring issues by using three different estimation methods. 5.2.1. Method I In the first method (I), I assume that farmers are not price-sensitive in the short run, and that one can properly treat hourly supply as fixed (inelastic), given information on recent prices. Thus, I treat any correlation between q and p as purely the trader’s response to changing supply and market conditions. Here, I also treat the lack of an observation due to q*V0 as a randomly distributed occurrence, conditioning on all observable variables. Under this assumption, there is no need to correct for censoring. However, there is still likely to be simultaneity between OENMR and q. Factors unobservable to the econometrician, such as weather events, input market conditions, and crop forecasts, will affect both farmers’ supply to the market ( q) and the expected final price (E[ P]). Under Method I, I therefore use two stage least squares (2SLS). The first step is a reduced form equation for the OENMR to be received on grain purchased by trader i in period t and sold in t+k, also denoted P i,t+k
c˜ i,t+k . This can be written as follows: Pi;tþk
c˜ i;tþk ¼ p þ XIit þ nit :

ð6Þ

All observable variables known to traders at time t, denoted I it ={X it , Z it } are included in the regression, and X represents the vector of coefficients on I it . This imposes the orthogonality condition that errors in expectation are uncorrelated with available information, E[In]=0, so that predicted P i,t+k
c i,t+k is an unbiased estimate of OENMR. X and Z can be exogenous shifters of supply, OENMR, or both, as identification does not depend on separate estimation of supply. The Z (detailed below) are exogenous variables (again, shifting one or both of supply or OENMR) which should not affect traders’ bid price ( p) once one accounts for costs, trader’s expected future NMR, and any potential simultaneity and selectivity. After estimating Eq. (6), I augment the main equation (Eq. (5)) by including in it the ˆ estimated error n it from Eq. (6). Thus, the term g from Eq. (5) in this case can be written g=kn it +v it , where k is a parameter capturing the correlation between n and g. v it is a (mean-zero) disturbance term which is uncorrelated with the expectation ˆ error, n it . Including n it (along with P i,t+k
c˜ i,t+k ) in the estimation produces identical estimates to the usual instrumental variables procedure, but has the advantage that it provides a direct test of whether the variable in question is endogenous—i.e., whether the coefficient on the residual is statistically significant and its sign appropriate.

T. Osborne / Journal of Development Economics 76 (2005) 405–428

417

For this and subsequent estimation methods, the validity of the exclusion restrictions (Z) used is examined using a TR2 statistic for the regression of the final residual, vˆ on the X and Z variables. This is the Sargan test. However, since for Yetmen, I estimate the first step equations using the entire sample and the second step on the subsample for which qN0, the Sargan test is not valid. Unfortunately, there is no equivalent test available for these methods. Nonetheless, the TR2 statistic provides an indication of the orthogonality of the errors to I it . In the case of Debre Zeit, the samples are the same in all steps, and the TR2 statistic is the Sargan value. 5.2.2. Method II Under Method II, I relax one of the restrictive assumptions maintained under Method I. In particular, if farmers’ supply is not fixed in the short run, Method I will produce biased estimates. Thus, under Method II, there is an additional (reduced form) first stage regression to account for the possible endogeneity between p and q, as shown: qit ¼ 1 þ cXit þ #Zit þ eit ;

ð7Þ

where X and Z are as in Eq. (6). The estimated residual from Eq. (7) is included in the second step along with that from Eq. (6), and g equals k 1n+k 2e+v, where v is a random error uncorrelated with n and e The k’ s capture the correlation between the first step errors and the errors in Eq. (5). 5.2.3. Method III Under Method III, both simultaneity and censoring are addressed using an estimator proposed by Vella (1993). The first stage equation for q becomes: qit ¼ 1 þ cXit þ #Zit þ eit ; qit ¼ qit4ðqit4 N0Þ

ð8Þ

where e it is assumed to be a normally distributed random disturbance with mean zero. The generalized residual from this equation is then included in the second step (Eq. (5)), in addition to the residual from Eq. (6).21 Thus, g it in Eq. (5) is defined here as g=k 1n+k 3e+v, where (again) v is a random error term, here uncorrelated with n and e. Under Method III, therefore, estimation entails the following steps: I first estimate Eq. (8) using Tobit maximum likelihood and Eq. (6) using linear regression, both over the entire sample. Finally, I estimate Eq. (5) using OLS over the subsample for which q*N0 and include the generalized residual from Eq. (8) and the estimated residual from Eq. (6) in the regression (see Vella, 1993 for more details). 21 The generalized residual is the nonlinear counterpart to the residual in linear models. In nonlinear models where the dependent variable is censored, the residual cannot be defined as the difference between the actual and predicted values. Accordingly, the generalized residual is defined as the object which has many of the properties of a residual. In particular, it has zero mean and is uncorrelated with the explanatory variables. Vella (1993) shows that in many models involving sample selection or censored endogenous regressors, the appropriate control function, which accounts for the simultaneity and selection bias, is the generalized residual. Note that given the definition of the generalized residual for the tobit model, combined with the fact I only use the uncensored observations in the second step, the generalized residual is simply the difference between the actual and predicted values.

418

T. Osborne / Journal of Development Economics 76 (2005) 405–428

5.3. Details of estimation The DR curves are estimated separately for Yetmen and Debre Zeit, given the distance between them and their differing market structures. Since traders were randomly sampled and represent a substantial fraction of the total—approximately one-half of all wholesale traders in Yetmen and one-fourth in Debre Zeit, I assume that their behavior is on the whole representative of the entire (local) source market. I estimate the model for Debre Zeit using all observed transactions (full sample), including bulk purchases (see Section 2), and using only direct purchases from farmers in the usual observation hours (subsample). To define quantity q it for each observation, I use hourly purchased amounts for trader i observed in hourly interval t. I combine all grain purchases in one measure of q it to avoid the difficulty of considering market conditions separately for each grain. For p it , I calculate a weighted average purchase price for that trader-hour. In Yetmen, since the vast majority (75% by weight) of transactions recorded were for the white teff variety, I use the weighted average hourly price of white teff as a proxy for the combined offer price p. This is equivalent to assuming perfect correlation between the price of white teff and of mixed teff and chickpeas, the other two grains traded in the sample. Correlations in weekly mean price with that of white teff are 0.99 and 0.96, respectively, so this simplification should not affect results.22 In Debre Zeit, since there are many more grains transacted, I constructed a weighted average price using the share of total hourly purchases represented by each grain as the weight. Again, since these prices are highly correlated, this should not affect results. Next, OENMR ( P i,t+k
c˜ i,t+k ) had to be calculated as the weighted average sale price of the grains purchased by trader i in hour t less observed marginal costs at the time of presumed sale (t+k). This was complicated by the nonobservability of the exact date on which each unit of grain was sold (i.e., k). However, in many cases, I observe the timing of the trader’s subsequent shipment, and I infer this to be the approximate sale date. For some observations, however, the date of next shipment was not observed, particularly towards the end of the sample period. Alternatively, as in the case of Debre Zeit, shipments may have been missed by the enumerator due to the greater size and activity of the marketplace. Thus, the date of shipment was estimated for Debre Zeit for all observations without a clear subsequent ship date. In particular, I predicted the shipping date using a linear regression of shipdate on purchase date, the current price of Gojjam (another region’s) white teff, a quadratic in month, and trader effects. I then set P t+k equal to the next available sale price (sampled thrice weekly) for those predicted dates. (In Debre Zeit, since observed marginal costs are constant, for all current, expected, and lagged NMR, I simply use observed selling price P.) For Yetmen, there were relatively few observations with no clear ship date, and for these, the last NMR obtained was used. For all Methods, the X variables for Yetmen are: (1) controls for (monthly meannormalized) rainfall and rainfall squared from nearby weather stations (RainStation#), as well as national production-weighted rainfall (rainf, rainf 2 ), and all these rainfall variables interacted with a dummy for whether or not the month’s rainfall is important for future 22

In addition, the volume of chick peas is insignificant and sporadic as compared to teff, and the selling price series is much less complete, so it would not be possible to estimate the equations separately.

T. Osborne / Journal of Development Economics 76 (2005) 405–428

419

production.23 These will affect both supply and demand, since rainfall provides information about future production to both farmers and traders, as well as final consumers. Rainfall bnewsQ may also affect traders’ risk or anticipated future liquidity position and may thus need to be included in Eq. (5). In addition, X includes: (2) Trader effects, which could capture differential fixed (or average) costs and/or appetite for risk (affecting ENMR); (3) An indicator variable for whether or not the trader had shipped grain to market in the past 10 days (shiplast; affecting ENMR) as a crude control for the trader’s current liquidity position; and (4) Dummy variables for the day of the week, hour of the day, and whether it was a religious holiday (relighol; potentially affecting both ENMR and supply). The Z variables used for Yetmen are various lags of observed NMR ( P
c˜ )—lagged by one market day ( P t
1
c˜ t
1) and by five market days ( P t
5
c˜ t
5)—and a quadratic term in month (month and month 2 ). The lagged P
c˜ terms capture any delayed supply or demand response to information contained in P
c˜ (or to trends in it). The quadratic in month is used to capture seasonality in both supply (see Fig. 1) and future price expectations. Once the effects of these Z variables on OENMR and quantity are accounted for in the first stage(s), they are assumed to have no direct effect on the price the trader pays. In particular, under perfect competition, the quadratic in month should not be correlated with v; nor should it be under imperfect competition where btrueQ l and / do not vary seasonally (as a quadratic in month). A maintained assumption of the test is therefore that neither unobserved costs nor the degree of competitive restraint follow this quadratic form over the period. To control for seasonality, however, I will also re-estimate the relationships allowing /ˆ and lˆ to vary by season. X variables for Debre Zeit include the same rainfall variables as described for Yetmen, with the addition of (nearby) Addis Ababa rainfall, and indicators for trader, day and hour. Since traders shipped grain fairly constantly, shiplast was omitted. However, a dummy variable for whether the sale date was estimated was included. The Z variables for the full sample were the current selling price in Addis Ababa, P t , and the Addis Ababa price lagged by 5 days ( P t
5). I used current P t rather than P t
1, as given the proximity of Debre Zeit to Addis Ababa and traders’ close contact with brokers, it is likely that these traders know the current selling price. Once prices are set in Addis Ababa in the morning, they do not normally adjust during the day. For the nonbulk sales subsample, I also included as a Z variable a quadratic in month to capture seasonal effects. These were not, however, good instruments over the full sample, as in this case they were correlated with vˆ.24 23

These months were selected using OLS regression on a longer time series for 1969–1994 (see Osborne, 2004). For Yetmen, these data were from the nearest weather stations of Bahir Dar and Debre Markos as proxies for local and regional rainfall. Each month’s rainfall was normalized by its long-run mean (over 1961–1994). In addition, I included national (mean production-weighted) rainfall. 24 Before proceeding to estimation, it is also important to check that the price series used do not contain unit roots, or that if they do, they are sufficiently cointegrated that one can perform inference using regressions in levels. For the Yetmen data, one cannot reject that both p and P
c have unit roots. However, an augmented Dickey–Fuller test shows that the series are cointegrated, and thus standard inference can be done. For Debre Zeit, neither price series exhibited a unit root. Serially correlated errors are an issue for estimation, however, indicating that the estimated covariance matrix will be incorrect. Unfortunately, there was no specification which eliminated this problem. I calculated the Newey West robust standard errors on the second step (without adjustment), and the variables of interest remained significant, albeit with higher standard errors.

420

T. Osborne / Journal of Development Economics 76 (2005) 405–428

6. Results 6.1. Yetmen Table 1 shows the coefficients of interest for the first steps for Yetmen. At least one of the Z variables and/or the quadratic in month is significant in each first step. In particular, P t
5
c˜ t
5 is negative and statistically significant at the 5% level in the Tobit equation for q (and almost significant at the 10% level in the linear equation). Higher P t
5 (abstracting from c˜ t
5) may predict lower supply in t, due to unobserved information shocks, or traders may expect downward trends in P (i.e., higher P t
5, given P t
1) to persist, making purchasing less attractive. Month and month 2 are (jointly) significant in the OENMR equation, and P t
1
c˜ t
1 is positive and significant as expected. In addition, the liquidity measure shiplast is positive and statistically significant (or nearly so) in the q equations and negative and significant in Eq. (5). This suggests that greater liquidity permits greater purchasing but coincides with a lower expected NMR. Rainfall variables were jointly significant in the OENMR equation, but individual coefficients are not of particular interest for this analysis. Similarly, some time and trader dummies (not reported) were individually significant in each first stage equation. Table 2 shows the second step results for Yetmen using the three methods, plus OLS for comparison. The coefficient on OENMR is close to and not significantly different from 1

Table 1 First step estimates, Yetmen (T-statistics in parentheses) q P t
1
c˜ t
1 P t
5
c˜ t
5 Month Month 2 Shiplast Relighol (National)Rainfall Rainfall 2 Rainfall*Important month (Imp) Rainfall 2 *Imp Rain Station 1 2 /*Imp Rain Station 2/*Imp Rain Station 2 2 /*Imp T LR v 2 R2

OENMR

Linear

Tobit

Linear


0.142 (
0.124)
1.63 (
1.43) 111.59 (1.05)
12.6 (
1.30) 51.3 (2.18)
39.56 (
2.04) 3535.92 (1.56)
1913.51 (
1.56)
5929.92 (
0.855)

0.343 (0.143)
4.89 (
2.18) 49.85 (0.239)
8.28 (
0.44) 73.22 (1.63)
81.43 (
2.02) 3123.23 (0.711)
1572.36 (
0.657)
3732.173 (
0.269)

0.17 (2.48) 0.03 (0.45) 8.52 (1.34)
0.267 (
0.46)
2.91 (
2.08) 1.98 (1.71) 37.39 (0.278)
14.36 (
0.196) 214.25 (0.518)

3339.09 (0.986) 366.81/
457.54 (1.60)/(
1.47)
10540.47/9077.57 (
1.42)/(1.03) 7656.66/
6954.68 (1.45)/(
1.22) 682

2350.99 (0.349) 420.07/
444.58 (0.935)/(
0.719)
8907.72/6858.05 (
0.619)/(0.392) 6636.68/
5827.8 (0.646)/(
0.522) 682 295.34 –


142.51 (
0.705)
7.75/28.92 (
0.568)/1.55) 47.92/
165.20 (0.108)/(
0.313)
60.24/99.11 (
0.191)/(0.291) 682

.3834

.7129

Trader, day, and hour effects included in each equation. Some rainfall variables ommitted due to collinearity.

T. Osborne / Journal of Development Economics 76 (2005) 405–428

421

in Methods II and III. Thus, shifts in OENMR are well captured in these Methods, and under constant c u /ˆ should accurately estimate the average slope of the demand relations curve (Eq. (4)). The negative and statistically significant coefficient on the residual from Eq. (6), denoted pcres, indicates that an overestimate (underestimate) of ENMR given the information present would tend to push up (down) the trader’s purchase price, as would be expected. As shown, /ˆ is negative and significant for all methods, and the perfect marginal pricing hypothesis is rejected. When either a Tobit or linear projection is used to account for possible endogeneity and/or selectivity of q, /ˆ is of much greater magnitude than under the exogenous q assumption. A comparison across Methods II and III, moreover, shows that a failure to control for selectivity would induce substantial downward bias in /ˆ, and Method III is the prefered approach. That is, the observations for which q=0 tend to occur when traders would need to increase p to attract a positive supply, but it may not be profitable to do so in those times because margins are already thin. Indeed, traders are much more likely to close their stores (and stop purchasing) in the lean season, when m ˜ are thin, and rarely do so during the harvest season. The coefficient on the error from the respective q equations (Eq. (7) or Eq. (8)), denoted qres, is typically statistically significant at the 10% level, indicating that, without the correction, the endogeneity of p and q would bias /ˆ . The positive coefficient indicates that unobservables which increase q are positively correlated with unobservables which increase p, as would be the case if supply responds positively to price. Table 2 Second step estimates, Yetmen, various models (Robust, corrected T-statistics in parentheses) q (in quintals) qres OENMR pcres Shiplast Relighol Rainfall/*Imp Rainfall 2 /*Imp Rain Station 1/*Imp Rain Station 1 2 /*Imp Rain Station 2/*Imp Rain Station 2 2 /*Imp T R2 DW Stat TR2(Z)

OLS

Method I

Method II

Method III


0.58 (
3.08) – 0.037 (1.06) –
1.67 (
1.55) 0.535 (0.535) 207.41/
1431.14 (8.82)/(
5.81)
117.80/686.78 (
8.62)/6.30) 460.69/– (19.52)/–


0.54 (
2.81) – 1.32 (2.82)
1.29 (
2.75) 2.17 (0.98)
2.19 (
1.25)
39.34/613.76 (
1.14)/(4.90)
82.30/23.85 (
0.626)/(0.496) –/–


8.56 (
1.97) 8.12 (1.86) 1.06 (4.23)
1.03 (
4.07) 5.77 (2.33)
4.72 (
2.41) 98.53/
835.05 (2.28)/(
2.42)
55.11/439.51 (
2.47)/(2.82) –/–


3.43 (
1.93) 2.92 (1.64) 0.90 (1.59)
0.87 (
1.53) 3.22 (1.45)
3.63 (
1.92) 122.74/
998.13 (1.56)/(
3.31))
66.19/508.68 (
1.61)/3.77) 90.03/– (0.586)/–


435.49/151.93 (
19.43)/(14.50)
322.26/1354.3 (
7.11)/(5.44) 338.96/
744.05 (10.75)/(
6.54) 285 .9509 – –


158.78/1.75.02 (
1.76)/(3.71) 117.58/
487.33 (1.22)/(
3.11)
1220.32/59.77 (
4.32)/0.402) 273 .9492 0.687 53.843


27.73/9.68 (
11.89)/(1.47)
151.48/702.90 (
1.99)/(2.06) 94.37/
313.33 (1.83)/(
2.03) 273 .9511 0.699 0.0819


104.66/34.48 (
0.776)/(0.69)
226.54/900.94 (
2.16)/(3.00) 172.34/
440.58 (1.61)/(
2.79) 273 .9506 0.665 3.72

See Table 1 legend. T-statistics are corrected for the first stage(s).

422

T. Osborne / Journal of Development Economics 76 (2005) 405–428

The estimate of / of
3.4 says that for each additional quintal of grain traded in a given hour, 3.43 birr/quintal less is offered to farmers (with a price range of from approximately 120 to 240 birr/quintal.) The deviation produced by /ˆ q at mean q (over all observations for which q*N0) of 5 birr/quintal represents an average of 2.7% of the price received. In the harvest season, when the poorer farmers tend to market their grain, the average effect of estimated restraint (/ˆ q) on p rises to 3.8%. In addition, the coefficient on shiplast was positive and significant, suggesting that greater liquidity leads traders to pay a somewhat higher price (albeit given /ˆ b0), possibly as high as 3 birr/quintal on average. In each regression, some of the trader, day, hour effects and rainfall variables were individually significant. Note that the second stage R 2 statistics for all estimations of Eq. (5), including by OLS, are very high: the included regressors (particularly OENMR and rainfall) explain a great deal of the variability in the data.25 Moreover, the TR2 statistics (Table 2), if interpreted as v 2 statistics as in a Sargan test, would not reject the null of E[Zv]=0 (except under Method I). 6.1.1. Seasonal effects I now allow /ˆ and lˆ to vary between the harvest and lean seasons (where lean season is defined as from June–October 1994).26 In the case of lˆ , this may also capture seasonal unobserved cost shifts, so inference regarding either cost or conduct shifts will be somewhat inconclusive. The main results are summarized in Table 3. For both Methods II and III, the magnitude of the point estimate for / was slightly higher for the harvest season. However, using a Wald test on the second step, one cannot reject that these coefficients are the same across seasons. At the same time, the lean season dummy (lean) was large, positive and statistically significant in the second step (and not in either of the first steps). This suggests either a seasonal increase in costs in the harvest season or a shift in brule-of-thumb conductQ which raises profits in that season. I have argued above that even with seasonally varying r and w, it is unlikely that unobserved costs are higher in the harvest season, particularly in Yetmen where the volume of grain flows to traders is much more seasonal. Moreover, the coefficient on lean is extremely high (98 birr/quintal). Thus, this finding is suggestive of successful collusion in the harvest season which partially or fully breaks down in the lean season, when grain supplies are scarce. At the same time, given that /ˆ b0 even when controlling for season, restrained competition is identified within season (though q/ˆ ) as well. Finally, the coefficient on shiplast remains positive and statistically significant here, so this variable does not appear to be proxying for seasonality in the main estimations (Table 2), but rather for liquidity. 6.1.2. Robustness tests Due to concerns about unobserved capital costs, I repeat the estimation including two different proxies for unobserved capital costs. These are not included in the main

25 With 31 included varibles and 273 observations in the main specification, there are sufficient degrees of freedom to interpret the R 2 in the usual way. 26 Here, I do not allow the endogeneity (pcres and qres) to enter differently by season, however.

T. Osborne / Journal of Development Economics 76 (2005) 405–428

423

Table 3 Test of seasonal effects, Yetmen and Debre Zeit (Robust, corrected T-statistics in parentheses) Yetmen

q h /q ql qres OENMR pcres Lean Leanbulk Shiplast T R2 p-value Lean=Leanbulk a

Debre Zeit

Method II

Method III

Method I


8.76 (
2.02)
8.17 (
1.91) 8.24 (1.90)
0.740 (
1.28) 0.774 (1.33) 83.78 (2.38) – 4.95 (1.84) 273 .9516 –


3.44 (
1.94)
2.83 (
1.59) 2.86 (1.63) 0.913 (1.64)
0.882 (
1.58) 98 (3.12) – 6.08 (2.40) 273 .9509 –

0.02 (0.36) – – 1.00 (29.74) – 12.91 (3.23) 4.20 (2.35) – 710 .9449 0.14

X and Z variables as for nonseasonal specifications. T-statistics corrected for first stage(s). a Test done after regular IV estimation.

specifications reported in Table 2 due to endogeneity concerns. In one case (using both Methods II and III), I included as an additional Z variable the average time it took until the next shipment of grain on the part of all other traders observed on the same day (aveship). In the second case, I included as an X variable the actual time to next shipment (timeship). The coefficients of interest are reported in Table 4. In each case, the new variable was both positive and significant in the first stage equation for q, indicating that more grain is transacted at times when traders will wait to sell. Moreover, the coefficients on the timing variables were positive in the OENMR equations as well, indicating that, on average, waiting longer led to higher NMR. For both tests, for Methods II and III, the coefficient /ˆ remains negative and significant. Thus, it appears unlikely that unobserved capital costs related to the speed of turnover (given fixed interest rates) have induced a spurious finding of /ˆ b0.

Table 4 Second step estimates, Yetmen, with time variables (Robust, corrected T-statistics in parentheses) Method II

q qres OENMR pcres Shiplast Relighol Timeship T R2

Method III

Timeship X

Aveship Z

Timeship X

Aveship Z


4.41 (
2.07) 4.0 (1.85) 0.66 (1.49)
0.64 (
1.43) 2.30 (1.18)
2.11 (
1.44)
0.086 (
3.36) 259 .9543


5.82 (
2.99) 5.4 (2.77) 0.503 (1.69)
0.475 (
1.59) 2.26 (1.11)
2.57 (
1.55) – 259 .9538


2.51 (
1.85) 2.1 (1.52) 0.73 (1.68)
0.715 (
1.63) 1.88 (0.907)
2.39 (
1.49)
0.087 (
3.15) 259 .9543


3.12 (
2.94) 2.63 (2.53) 0.494 (1.83)
0.47 (
1.73) 1.23 (0.723)
2.52 (
1.62) – 259 .9532

Rainfall variables and trader, day, and hour dummies included in each. T-statistics correct for first stage(s).

424

T. Osborne / Journal of Development Economics 76 (2005) 405–428

6.2. Debre Zeit estimates Table 5 reports the first step estimates for Debre Zeit for the total sample of purchases and for the subsample of market-day purchases directly from farmers. As shown, both the Z variables P t and P t
5 are negative and statistically significant in the q equations, indicating that these capture information which is negatively correlated with current market supply. P t is highly positively correlated with OENMR, as expected, and P t
5 has a slight positive and statistically significant effect for the full sample. In addition, the quadratic in month is statistically significant for the subsample estimation. This probably proxies for seasonality in expected price movements (given P t ). The rainfall variables are jointly significant in the OENMR equation (but not in the q equation) for both the full and subsamples. Table 6 reports the second step results. The coefficient on OENMR is close to (and not statistically different from) 1 for all estimations, indicating that estimated shifts in traders’ DR curves were approximately one-for-one with OENMR. /ˆ is not significantly different from zero for either the full or the subsample estimation. The coefficient on qres is not significant under Method II, perhaps because of the lack of short run variation in p which could be correlated with (or affect q). Thus, Method I is the appropriate approach to use. Table 5 First step estimates, Debre Zeit (T-statistics in parentheses) q

Pt P t
5 Month Month 2 Rainfall /*Imp Rainfall 2 /*Imp Rain Station 1/*Imp Rain Station 1 2 /*Imp Rain Station 2 Rain Station 2 2 /*Imp Rain Addis /*Imp Rain Addis 2 /*Imp Esttime T R2 p-value, month terms

OENMR

Full sample

Subsample

Full sample

Subsample


3.54 (
2.03)
3.37 (
2.88) – – 1158/– (1.17)/–


6.83 (
3.05)
2.82 (
2.04)
118.04 (
0.29) 10.2 (0.33) 5315.60/– (0.809)/–

0.86 (31.47) 0.044 (2.40) – – 39.96/– (2.58)/–

0.926 (32.21) 0.012 (0.70)
4.20 (
0.80) 0.21 (0.53) 48.60/– (0.576)/–


671.83/30.40 (
1.14)/(0.146) 907.27/– (0.70)/–


2894.68/128.06 (
0.813)/(0.245) –/–


17.32/
6.49 (
1.87)/(
2.00)
14.79/– (
0.776)/–


22.02/
0.24 (
0.482)/(
0.03) –/–


932.97/388.34 (
0.716)/(0.557) –/– 153.42/
155.86 (0.37)/(
0.292)
293.68/– (
0.363)/–


319.16/694.02 (
0.655)/(0.852) –/–
938.12/507.42 (
0.893)/(0.675)
2148.4/– (
0.666)/–

12.78/
8.72 (0.670)/(
0.854) –/–
18.58/8.36 (
3.06)/(1.07)
53.02/– (
4.48)/–

168.23/30.60 (0.274)/(0.201) 1787.02 (1.655) 710 .4058 –

1330.46/
302.51 (0.635)/(
0.622) 176.80 (1.064) 452 .3862 0.89

28.27/
6.17 (3.15)/(
2.78)
1.34 (
0.859) 710 .9435 –

8.81/
7.29 (1.41)/(
0.697) –/–
5.10/10.84 (
0.378)/(1.12)
28.80/19.71 (
0.70)/(0.733) 19.71/
4.29 (0.733)/(
0.688)
4.16 (
1.95) 452 .9256 0.035

Day, trader, and hour dummies included.

Table 6 Second step estimates, Debre Zeit (Robust corrected T-statistics in parentheses) OLS

RainStation Station 1/*Imp RainStation Station 1 2 /*Imp RainStation Station 2/*Imp RainStation Station 2 2 /*Imp Rain Addis/*Imp Rain Addis 2 /*Imp T R2 Dw Stat TR2

Method II

Subsample

Full sample

Subsample

Full sample

Subsample


0.03 (
0.504) – 0.785 (28.67) – 1.09 (0.509) 110.19/– (
6.47)/–
55.61/27.07 (
6.47)/(9.98) 94.33/– (5.17)/–


0.04 (
0.789) – 0.827 (25.59) – 3.57 (1.13) 118.74/– (6.23)/–
60.89/23.4 (
5.61)/(6.83) 58.88/– (2.68)/–

0.01 (0.32) – 1.00 (29.72)
0.85 (
17.10)
0.93 (
0.42) 81.62/– (4.78)/–
41.97/23.61 (
4.38)/(8.21) 67.78/– (3.64)/–


0.00 (
0.026) – 0.976 (26.88)
0.754 (
11.05) 2.48 (0.79) 96.45/– (4.88)/–
50.06/20.92 (
4.50)/(6.25) 38.49/– (1.82)/–

0.49 (0.67)
0.48 (
0.65) 1.04 (15.59)
0.89 (
11.53)
1.81 (
0.675) 73.74/– (
0.977)/–
37.77/23.60 (
3.17)/(7.97) 68.08/– (3.52)/–


0.37 (
0.55) 0.37 (0.55) 0.937 (11.78)
0.715 (
7.11) 3.00 (0.95) 109.58/– (3.53)/–
57.02/20.94 (
3.37)/(6.21) 36.75/– (1.71)/–


89.17/20.65 (
4.92)/(2.31) –/–


59.07/12.67 (
2.71)/(1.17) –/–


64.37/15.75 (
3.51)/(1.72) –/–


40.15/8.02 (
1.92)/(0.77) –/–


64.26/15.48 (
3.37)/(1.61) 12.61/– (1.90)/–


38.97/9.17 (
1.85)/(0.844) –/–

16.08/
25.81 (2.80)/(
3.63)
51.65/– (
4.39)/– 27.54/
8.11 (3.07)/(
3.22) 710 .9096 1.33 –

6.91/
17.47 (.957)/(
1.98)
43.17/– (
2.93)/–
43.16/– (
2.94)/–

11.64/
24.28 (1.89)/(
3.30)
37.44/– (
3.01)/– 18.96/
7.32 (2.03)/(
2.78) 710 .9448 0.812 0.781

4.50/
16.79 (0.639)/(
1.99)
30.91/– (
2.10)/– 11.41/
3.58 (1.05)/(
1.16) 452 .9435 0.633 2.31

–/
25.36 –/(
3.21)

1.62/
14.85 (0.184)/(
1.59)
36.05/– (
2.07)/– 14.66/
4.31 (1.18)/(
1.29) 452 .9435 0.635 1.834

452 .9119 0.874 –


35.00/– (
2.63)/– 17.48/
7.43 (1.77)/(
2.73) 710 .9449 0.81 –

T. Osborne / Journal of Development Economics 76 (2005) 405–428

q qres OENMR pcres Esttime Rainfall/*Imp Rainfall 2 /*Imp

Method I

Full sample

All equations include trader, day, and hour effects.

425

426

T. Osborne / Journal of Development Economics 76 (2005) 405–428

Pcres, however, was negative and significant in all specifications. Again, this indicates that unobservables which are positively correlated with OENMR, such as unanticipated changes in NMR, are negatively related to the price the trader pays (all else equal), as expected. In this market, the only possible suggestion of a deviation from marginal pricing is that l is significantly higher in the lean season. In the last column of Table 3, I show the results where X includes a dummy variable for the lean season (lean), and a dummy variable for the lean season interacted with a dummy variable for a bulk sale observation (Leanbulk). These variables are both positive and significant and the point estimate for lean is 8 birr/ quintal higher than that for Leanbulk. This difference in differences could indicate the presence of brule-of-thumbQ margin setting with regard to direct purchases from farmers which is more competitive in the lean season.27 Given the relative fixity of p in this market, it seems more likely that any imperfect competition would operate through rule-ofthumb margin setting, rather than through more continuous reactions to supply. Nonetheless, a Wald test does not reject that the coefficients on Lean and Leanbulk are equal at the 5% or 10% level (following standard instrumental variables). In addition, the magnitude of the lean coefficient is much smaller than that for Yetmen (13 versus 98). Since it is not possible to separate the potential influence of unobserved costs from conduct through inference on lˆ s , one cannot conclude that thinner lean season margins are due to changes in imperfectly competitive behavior in this market.28 6.3. Interpretation of results The above results provide conclusive evidence of a departure from marginal (ENMR) pricing in the Yetmen market. First, liquidity appears to affect traders’ incentives to compete more strenuously, even when controlling for season. In addition, traders appear to respond to greater supply by pricing less aggressively, as is predicted in standard models of uncompetitive behavior. However, as noted above, such a deviation may not signify inefficiency in the form of supernormal profits. This may be due to economies of scale (everywhere falling total marginal costs, c) which necessitate a deviation from marginal cost pricing to break even. Recall that if market power is used only to recapture falling c u (as a component of falling c), the estimate of / due to competitive restraint should be zero. However, further competitive restraint—that is, a /b0—may be needed to capture fixed costs under this scenario. In this case, average fixed costs—and observed margins (m ˜ )—should be higher in the lean season, when volume is lower.29 Yet, we observe higher margins in the harvest season. Thus, the economies of scale hypothesis alone would not explain the greater competitive restraint exhibited in that season. Finally, note that the estimated magnitude of the deviation from marginal pricing is high relative to estimated fixed costs. The average reduction in the price received by 27

This may also mean that the cost of assembling grain from individual farmers changes dramatically more over the seasons than the cost of bulk collection. This seems somewhat unlikely, since approximately the same labor input is required to clean and bag the grain, and labor piece rates do not vary dramatically. 28 This finding is therefore subject to the Corts (1999) critique. 29 This is true if traders consider the decision whether or not to operate (and incur fixed costs) within the year, but the amount of fixed costs incurred does not change within the year.

T. Osborne / Journal of Development Economics 76 (2005) 405–428

427

farmers (at mean q) due to / is three times average fixed costs, as estimated in Osborne (1997) using an interest rate of 25% (5 birr/quintal versus 1.8 birr/quintal).30 The total deviation from marginal pricing is therefore likely to be at least partly attributable to supernormal profits.

7. Conclusion While the abolition of the quota system in Ethiopia did much to increase competition and reduce the risks and costs of wholesale grain trading, free markets in remote rural areas still appear to fall short of efficiency in ways that are likely to have negative welfare and distributional effects. Trading margins in the markets studied, especially Yetmen, exhibit a seasonal pattern which is difficult to explain within a perfectly competitive framework, even with the possibility of storage and scale economies. Empirical tests show that traders in Yetmen in particular restrain competition in the bid price of grain when the supply offered is higher. And while this shows up partly as a seasonal shift in margins, this restraint is also identified through intraseasonal shifts in quantity (when controlling for season). The slope of the DR curve alone reduces the producer price to farmers of approximately 2.5%, and to those who must sell early at a lower price by at least 4%, representing a significant reduction in farmers’ income. In Debre Zeit, a larger market with better communication links, which also serves richer farmers, there is no conclusive evidence of imperfect competition in trading. The contrast in findings from these two markets suggests that inefficiencies resulting from market structure and imperfect competition may be even worse in the many smaller and more remote markets than Yetmen. These are also areas where transport and marketing costs generally tend to be higher and farmers typically poorer, and thus the impact of inefficiency arguably more deleterious.

Acknowledgements Field research used in this paper was primarily funded by the Mellon Foundation, with supplementary funding from the Center of International Studies, Princeton University. Many people assisted me in undertaking this project, especially Tim Besley, Angus Deaton, and Harvey Rosen. Institutional support was provided by the Institute for Development Research (IDR), Addis Ababa University (AAU), the Economics Department at AAU, and the Centre for the Study of African Economies at Oxford University. I thank the Ethiopian Grain Trade Enterprise for data and background assistance. Frank Vella, Adrian Pagan, and two anonymous referees provided valuable comments. Special thanks goes to my research assistants, Kebede, Makuanint Gatenet, and especially Million Taffesse. Any errors are mine. 30 These range by trader from 0.50 to 3.8 birr/quintal for those eight traders who are not simply engaged in low-volume speculative storage. The value of traders’ stores was estimated using the cost of a recent entrant’s store and using a declining average cost of capacity schedule.

428

T. Osborne / Journal of Development Economics 76 (2005) 405–428

References Aleem, I., 1990. Imperfect information, screening, and the costs of informal lending: a study of a rural credit market in Pakistan. World Bank Economic Review 4 (3), 329 – 349 (September). Brock, W.A., Scheinkman, J.A., 1985. Price setting supergames with capacity constraints. Review of Economic Studies 52 (3), 371 – 382 (July). Corts, K.S., 1999. Conduct parameters and the measurement of market power. Journal of Econometrics 88 (2), 227 – 250 (February). Dercon, Stephan, 1993. The consequences of liberalisation and peace for food markets in Ethiopia, preliminary draft, unpublished. Ellis, Frank, 1993. Rice marketing in Indonesia: methodology and results of a research study. Bulletin of Indonesian Economic Studies 29 (1), 104 – 123. (April). Fafchamps, M., Minten, B., 2002. Returns to social network capital among traders. Oxford Economic Papers 54 (2), 173 – 206 (April). Franzel, Steven, Colburn, Forrest, Getahun, Degu, 1995. Grain marketing regulation: impact on peasant production in Ethiopia. Food Policy, November. Gabre-Medhin, Eleni, 1999a. Transaction costs and market institutions: grain brokers in Ethiopia, MSSD Discussion Paper No. 31, International Food Policy Research Institute. Gabre-Medhin, Eleni, 1999b. Of Markets and Middlemen: The Role of Brokers in Ethiopia, MSSD Discussion Paper No. 39. International Food Policy Research Institute. Graddy, K., 1995. Testing for imperfect competition at the Fulton fish market. Rand Journal of Economics 26. Green, E.J., Porter, R.H., 1984. Noncooperative collusion under imperfect price information. Econometrica 52 (1), 87 – 100 (January). Lirenso, Alemayehu, 1987. Grain marketing and pricing in Ethiopia: a study of the impact of grain quota and fixed grain prices on grain producers, Research Report No. 28, Institute of Development Research, Addis Ababa University. Osborne, Theresa, 1997. The Efficiency of Grain Markets in Ethiopia. PhD thesis, Princeton University. Osborne, T., 2004. Market news in commodity price theory: application to Ethiopian grain market. Review of Economic Studies 71 (1), 133 – 164 (January). Pickett, James, 1991. Economic Development in Ethiopia: Agriculture, the Market and the State. OECD, Paris. Ravallion, Martin, 1987. Markets and Famines. Oxford [Oxfordshire]: Clarendon Press; New York : Oxford University Press. Rotemberg, J.J., Saloner, G., 1986. A supergame-theoretic model of price wars during booms. American Economic Review 76 (3), 390 – 407 (June). Sen, Amartya, 1981. Poverty and Famines: An Essay on Entitlement and Deprivation. Clarendon Press, Oxford. Slade, M.E., 1985. Price Wars in Price-Setting Supergames, Department of Economics Working Paper No. 85-35, UBC, Vancouver, BC. Usher, Dan, 1968. The Thai Rice Trade. In: Silcock, T.H. (Ed.), Thailand: Social and Economic Studies in Development, re-published in Welfare Economics and Public Finance. Edward Elgar Publishing, Brookfield, VA. Vella, Frank, 1993. A simple estimator for simultaneous models with censored endogenous regressors. International Economic Review 34. Wang, Elizabeth, 1999. Competition Among Exchanges: Does Multiple Listing Affect Trading Costs in Options Markets?, unpublished manuscript.