Testing Oligopolistic Behaviors: Conduct and Cost in Taiwan's Flour Market

ABSTRACT

Testing the validity of the new empirical industrial organization (NEIO) proposition is often made more complicated by a lack of reliable cost data. Nevertheless, the author takes advantage of a more sophisticated dataset to directly measure cost in Taiwan's flour market. In contrast to the previous findings, the evidence indicates that the NEIO technique performs reasonably well in estimating both conduct and cost parameters.

INTRODUCTION

The new empirical industrial organization (NEIO) studies suggest that economic marginal cost (MC) cannot be straightforwardly observed from accounting data; hence, the industry price-MC margin is unavailable.Researchers must therefore use price and quantity data to infer MC and conduct by focusing on the firms' equilibrium behavior (naturally through a first-order condition of optimization).1 The NEIO methodology has been tested by several empirical studies based on the detailed cost data of certain industries, and the evidence indicates that the actual cost and conduct parameters are statistically different from the NEIO estimates. For example, under the NEIO framework, Genesove and Mullin (1998) estimated industry MC and conduct using data for the American refined sugar industry. They then tested the NEIO proposition by comparing the true MC and conduct grounded in complete cost information with the NEIO estimates. As shown in their evidence, the NEIO estimates significantly understated the market power.2 Although they claimed that the difference was economically small, Kim and Knittel (2006) and Carruthers (2006) still indicated that their NEIO estimate significantly understated the market power. Clay and Troesken (2003) also performed exactly the same test for the American whisky industry. Unlike Genesove and Mullin, they found that the NEIO estimate overstated the market power. Kim and Knittel (2006) thus concluded that the NEIO technique did a poor job of estimating the market power.

Since Clay and Troesken's study is essentially based on Genesove and Mullin's model, I hereby refer to these two studies collectively as the G-C model. In general, the literature strongly supports the structural form of the MC in their model because they can directly observe the true costs. In this article I employ the same empirical specification as the G-C approach, except that the observations are obtained from an industry with a huge excess capacity. Specifically, I use a unique setup of Taiwan's flour cartel to verify the accuracy of the NEIO technique. The interesting part of the case is that Taiwan's flour industry has maintained an extremely low level of capacity utilization rate at around 40% for more than 20 years. I believe that this kind of excess capacity can provide some advantages for researchers to resolve the difficulties that might arise when attempting to test the validity of the NEIO method, particularly with the cost specification.3

Unlike the G-C findings, the evidence presented in this article indicates that the NEIO approach performs reasonably well in estimating the cost and conduct parameters. The NEIO estimates are not only close to the true values derived from the full cost information, but are also insensitive to the assumed demand specification. It thus appears to be successful in assessing market power in Taiwan's flour industry.

The remainder of this article is structured as follows. In Section 'HISTORICAL BACKGROUND AND NEIO SPECIFICATION', I describe the background of the case and in Section 'CHARACTERISTICS OF THE DATASET' explain how the characteristics of my dataset can be used to solve the empirical problem of the NEIO method. In Section 'DATA AND MARKET DEMAND', I estimate the market demand, and then compute the direct estimates of MC and conduct using Taiwan's flour market data in Section 'CONDUCT AND COSTS'. I estimate MC and conduct using the NEIO technique, and compare these indirect estimates with direct estimates that are grounded in detailed cost information in Section 'VALIDITY OF THE NEIO ESTIMATES'. Section 'CONCLUSION' concludes.

HISTORICAL BACKGROUND AND NEIO SPECIFICATION

Historical Background

In 2000 Taiwan's Fair Trade Commission (TFTC) brought a collusion charge against the Taiwan Flour Industry Association. In this case, millers had colluded in the market for years until they were found guilty of collusion. The TFTC alleged that the flour firms had restricted competition and that this constituted concerted behavior that was prohibited under Article 14 of the Fair Trade Law. There were two interesting features of this cartel case. First, the industry had exhibited an extremely low capacity utilization rate of about 40% for too long. Second, price wars never occurred among the cartel members. After the case was closed, the TFTC (2001) released a report based on its inquiry into pricing behavior in the flour market and provided detailed data on prices and costs that can be used to verify the accuracy of NEIO estimates.

G-C Conditions

It is difficult in practice to obtain suitable data to test the validity of the NEIO methodology. To conduct their empirical study, G-C may have had no choice but to select products with the following four characteristics (the G-C conditions, hereafter):

1. The targeted product has to be homogeneous such that output prices tend toward uniformity across firms.
2. The production technology has to be simple and a single input cost accounts for a significant portion of output price such that MC can be accurately described.
3. Firms utilize a common technology in which the main variable input is transformed at a fixed and generally accepted coefficient into the output.
4. Firms face the same and exogenously determined input price.

Accordingly, to fulfill the G-C conditions, the sample has to be a simply processed product for which input materials are mixed, heated, crushed, scratched, and so forth.

Based on the previous conditions, G-C can reasonably assume that MC depends on an exogenously determined factor price P_RAW and is constant over different output levels. Thus, industry's MC is exhibited as:

• (1)

in which c₀ represents all marginal costs beyond the cost of the main input and k is the fixed relation at which one unit of the main input is transformed into one unit of output. For example, to infer the MC of refined sugar, Genesove and Mullin (1998) used several different sources of evidence and presumed a value of c₀ = 26¢ (per hundred pounds of sugar) as the baseline specification. They also pointed out that the production of one kilogram of refined sugar needed 1.08 kilograms of raw sugar, and assumed that firms faced the same and exogenously determined raw sugar price P_RAW. Thus, MC = 0.26 + 1.08P_RAW was constant over a wide range of output up to capacity (see curve ABC in Fig. 1), and the direct estimate of industry's MC could therefore be calculated. The same kind of specification was also used by Clay-Troesken.

Figure 1. Constant Marginal Cost (MC) and Average Variable Cost (AVC).

NEIO Specification

The empirical specification of the NEIO model is mainly based on the following first-order condition in aggregate form:

• (2)

Here, P is price, Q is industry's aggregate output, and θ is the industry conduct parameter.4 By letting η be the elasticity of demand, one can rewrite Equation (2) as

• (3)

such that θ is the elasticity-adjusted Lerner index (L_η). With complete cost and demand information, researchers can directly measure θ and MC, and can compare these true values with the NEIO estimates obtained from the regression of Equation (2).

CHARACTERISTICS OF THE DATASET

Taiwan's flour cartel dataset has several advantages to satisfy the G-C conditions.

Excess Capacity vs. Full Capacity

The G-C specification (MC = AVC = c₀ + kP_RAW) gives rise to a case similar to that of a natural monopoly as plotted in Figure 1, in which the production is characterized by decreasing AC and hence the MC pricing rule causes firms to lose their AFC. Under pure strategy profile, Shapiro (1989) indicated that adding even a small fixed cost to the model of constant MC would lead to the non-existence of the competitive equilibrium, unless, for instance, there exists a very strong demand (or very small capacity) such that the market price is higher than the price that generates the unconstrained monopoly profit, P(Q_monopoly). In this situation, industry capacity (K) is less than Q_monopoly. Because P(K) is higher than P(Q_monopoly), the best the oligopolist can do is to name the same price P(K) and produce at full capacity. Thus, there is no excess capacity.5

Using this kind of capacity constraint to justify the NEIO methodology is quite unreasonable in empirical applications because firm conduct and industry conduct are viewed as unknown parameters to be estimated in the NEIO. If firms always produce at full capacity, then there will be no behavior equation to be estimated. As a result, the profit margin cannot be directly linked to analytical notions of firm and industry conduct. Thus, to obtain data to satisfy the G-C conditions, researchers must choose an industry that fulfills the following two requirements:

1. The industry has to be characterized by a substantial excess capacity.
2. The industry conduct should exhibit a certain degree of market power (such as monopoly or collusion), which allows firms to earn some revenues in excess of variable costs so as to cover AFC and earn profits.

To solve explicitly for the equilibrium, Genesove and Mullin (1998, 2006) introduced a dominant-fringe firm competition, in which the capacity of the dominant firm always suffices to meet market demand, but the fringe capacity does not. Thus, the dominant firm always carries enormous excess capacity. By contrast, the fringe firms exactly produce up to capacity. In this situation, the dominant firm sets the equilibrium price to maximize profits based on its residual demand, which is the difference between market demand and fringe capacity. Conversely, the fringe firms take price as given and produce to capacity regardless of the price above MC that the dominant firm sets. Because MC is constant until capacity, there is no output reaction by the fringe firms for a price exceeding MC. The equilibrium price must be higher than OA, and the fringe firms must produce to capacity; at the same time, the dominant firm always holds excess capacity.6

This study alternatively focuses on another line of game theoretic contributions, such as those discussed by Osborne and Pitchik (1986, 1987) and Davidson and Deneckere (1990), which emphasizes that the correlation between excess capacity and collusion is positive. Based on this line of reasoning, production capacity serves two roles in the implementation of cartels that identify Taiwan's flour cartel as a possible case to fulfill the requirements for G-C conditions. That is, the industry is characterized by excess capacity, and the industry conduct exhibits a certain degree of market power.

The Industry is Characterized By Excess Capacity

The flour cartel allocates market share according to a firm's capacity. The firm takes the collusive price as given and produces up to its quota allocation. This arrangement encourages firms to add capacity in an effort to increase their market shares because large capacity gives colluders a stronger bargaining position in their negotiations regarding quotas. It inevitably leads to a larger capacity than actual output and hence “all” firms carry excess capacity.7 As the TFTC report pointed out, Taiwan's flour industry has maintained an extremely low capacity utilization rate of around 40% for more than 20 years. Although there were 32 flour firms in the industry, the report contained only the capacity of 25 firms (see Table 1 for the details). The TFTC still indicated that excess capacity was prevalent throughout the industry. Accordingly, the TFTC data could match the G-C conditions and could allow the technology of MC = c₀ + kP_RAW to be applied in this case.

Table 1. Firms' Utilization Rates (1994:1–1999:1, Quarterly Averages)

Firm	Utilization rate (%)	Firm	Utilization rate (%)	Firm	Utilization rate (%)
Note The figures are the quarterly averages between 1994:1 and 1999:1 for each firm. The Taiwan Fair Trade Commission (TFTC) data do not reveal the names of the firms to protect their privacy.
Source: TFTC (2001).
1	25.53	10	30.29	18	71.36
2	29.68	11	21.85	19	50.00
3	26.72	12	35.23	20	42.33
4	53.83	13	42.20	21	48.61
5	29.03	14	34.70	22	53.23
6	21.50	15	85.73	23	28.62
7	49.97	16	52.33	24	73.50
8	20.77	17	12.50	25	33.42
9	34.88

In summary, the main difference between Taiwan's flour industry and the American sugar industry is that the former is characterized by excess capacity for “all” firms, whereas in the latter case excess capacity exists only in the dominant firm.8 Because this study follows exactly the same procedure as for the G-C, this difference might lead to the conclusion why the empirical results presented later are different from those of the G-C. In other words, I hope to provide some explanations about why NEIO works well in the Taiwan flour industry, but does not work well in a similar industry studied earlier.

The Industry Conduct Exhibits a Certain Degree of Market Power

Millers with enormous excess capacity can punish deviators more harshly, so that cheating is unlikely to happen and the industry conduct exhibits a certain degree of market power. For example, if cheating is observed by the cartel, then all firms could produce at full capacity and dump a large amount of product on the market. This predictably will cause prices to collapse, and hence excess capacity becomes a credible threat to the enforcement of collusion. Thus, the collusive output is sustained as the equilibrium in every period.9

Dynamic Setting Vs. Static Setting

The second advantage of the TFTC dataset stems from the Corts critique. Corts (1999) illustrated that the conduct parameter estimator may be biased in a dynamic oligopoly. The intuition is straightforward. Equation (2) assumes that θ is constant across different periods. Nevertheless, in a dynamic game, θ depends on the incentive compatibility condition associated with collusion. If the incentive compatibility condition is correlated with demand shocks, then the NEIO estimator might yield a biased estimate of the mean conduct parameter.

Corts auspiciously indicated that there still exist some possibilities for to accurately measure the market power. For example, if the colluding regime is sustained in every period, then one can safely claim that is unbiased in measuring market power. However, in practice, I (Ma, 2005) used the argument of Bresnahan (1989) to emphasize that most cartel cases experience periods of falling apart and restructuring such that the data reveal both the periods of cooperation and the periods of price wars. Therefore, collusion is unlikely to be sustained as the equilibrium in every period, which means that the static NEIO models might be subject to the potential risk of triggering Corts' objection.10 Empirically, the researchers should allow θ to be time-varying. Nevertheless, modeling θ to vary would quickly exhaust the degrees of freedom available.

Fortunately, this study investigates a cartel that is unlikely to occur in which Friedman's trigger strategy equilibrium is sustained in every period. I (Ma, 2005) used TFTC data to examine the stability of the collusive equilibrium in Taiwan's flour market. I calculated the payoff streams following a deviation or adherence for each firm. The evidence showed that the specified punishment path was credible and could sustain the collusive allocation in every period. Thus, a price war never occurred among the cartel members.11 Because industry conduct (θ) was constant over the sample period, the of NEIO becomes an unbiased estimator used to measure the average collusiveness of conduct. Accordingly, this study does not have to distinguish whether there are two distinct regimes (cooperation or collusion) and hence can take advantage of the data in all periods.

Exogeneity vs. Endogeneity

For the sake of data convenience, most empirical studies on the NEIO model focused on very big markets (e.g., the United States), which are large enough to influence world prices. In so doing, P_RAW might be endogenous. For example, in the study of Genesove and Mullin (1998). it is possible that demand shocks in the U.S. refined sugar (output) market could affect world prices and create excess demand or supply in the U.S. raw sugar (input) market, which then affects P_RAW. Thus, they have to worry that the input price (explanatory variable) might be correlated with the error term in the output price equation, and use a valid instrument to address endogeneity. However, Taiwan is a small open economy that consumes less than 1% of global wheat production and trades wheat at given world prices. This makes the change in P_RAW an exogenous variable that is unrelated to the price of output.

Aggregate Industry Data vs. Individual Firm Data

To test the applicability of the NEIO method, one needs detailed cost information on an individual firm level. This can be seen using the following argument. If firms are heterogeneous in terms of their cost functions, then one should expect the low-cost firms to have large market shares and they are particularly important in the determination of industry MC and conduct. This effect will eventually result in a lower industry MC. Thus, Bresnahan (1989, p. 1044, Equation (17′)) indicated that, instead of Equation (2), the crucial equation should be:

• (2)

in which w_i is firm i's market share, is the share-weighted average collusiveness of conduct, and . Based on Equation (2), researchers have to interpret the NEIO estimate of MC (or θ) as the weighted average notion of the cost (or conduct). They should disaggregate the cost data at the individual firm level and compute Average[MC], so that they can compare this weighted actual cost with the NEIO estimate. Conversely, assuming that MC is the same across firms will give rise to the risk of obtaining biased inferences on the validity of the NEIO. Nevertheless, in practice, because it is difficult to trace back the data over the past two previous centuries, G-C took the approach that the industries they studied had such small variation in MC across firms that any such differences could be safely ignored.12 Thus, they could accurately compare these directly assigned measures of MC and θ with the NEIO estimates obtained from Equation (2).

On the other hand, Taiwan's flour cartel case contains two advantages that allow researchers to avoid being faced with the measurement problems. First, the TFTC dataset contains detailed cost information at the individual firm level that allows one to calculate the weighted average value of c₀. Second, the input cost (P_RAW) is the same across Taiwanese millers. Because Taiwan does not produce any wheat, all of the raw materials have to be imported from abroad. To stabilize firms' wheat costs, Taiwan's Ministry of Economic Affairs (MOEA) set up a Wheat Stabilization Fund (WSF) and posted “wheat standard prices” in the 1980s. If a miller's wheat import cost was higher (or lower) than the posted standard price, then the WSF disbursed (or collected) the amount of the deviation. This made P_RAW nearly the same across millers. Thus, I can safely claim that my c₀ is rooted in the weighted average notion, my P_RAW is identical across firms, and the TFTC data can be used to compute a true measure of MC = c₀ + kP_RAW.

DATA AND MARKET DEMAND

Based on the previous argument, the TFTC data seems to provide a realistic empirical setting in which researchers can test the validity of the NEIO technique with less data constraints. I now describe the sources and frequency of the data.

Data Source and Frequency of Data

The cost data source is based on the inquiries conducted by the Taiwan Fair Trade Commission (TFTC, 2001). The dataset covers 5 years and 3 months. The aggregate data on the flour price and industry output are obtained from the Yearbook of Industrial Production Statistics published by the MOEA. Although the TFTC and MOEA data are monthly, I still follow the earlier empirical literature in aggregating up to the quarterly level. This is because the short-run elasticity is generally smaller than the long-run one. The former one estimates a significantly higher monopoly price than would be suitable for a forward-looking monopolist.

The data extend from 1994:1 through 1999:1 such that there are 21 observations. The main drawback of the TFTC dataset is that it relates only to 5 years and 1 quarter. Nevertheless, the number of observations is still greater than the 16 observations of Clay-Troesken.

Production Technology and Marginal Cost

Flour is a homogeneous product. It is shipped to grocers who in turn pack the product and ship it to final users without any identification of the brands. Price therefore tends toward uniformity, and millers compete in terms of the quantity in the market.

The production of flour is a common and simple process. The wheat is transformed at a fixed and generally accepted coefficient into the flour. The TFTC indicated that the production of one kilogram of flour needs 1.37 kilograms of wheat on average. This coefficient has remained constant over the decades. In addition, the value-added of the flour is not high. Estimates of the TFTC show that the wheat cost comprised 69% of flour price between 1994 and 1998. Because wheat is the main variable input, it is transformed into flour at a fixed coefficient, its price is exogenously determined and is identical across firms, and flour production is operating within the range of huge excess capacity, such that the cost structure can be expressed as:

• (3)

Because the TFTC report provides detailed information of c₀ at the individual firm level, one can compute the weighted average of New Taiwan (NT) dollars.13 Note that all prices and costs are deflated by the GNP deflator to provide comparability.

On the other hand, in the production process, firms could also sell some by-products such as wheat bran and wheat middlings. Thus, c₀ is more than offset by the by-product revenue (r), which is worth NT 2.80 dollars on average ( ). As in the case of whisky in Clay and Troesken (2003), I hereby take 0 as the best estimate of c₀. Therefore, the following discussion uses as the baseline specification.

Market Demand

Following the G-C approach, the demand for flour is estimated by fitting four functional forms—linear, quadratic, log-linear, and exponential. The general form can be exhibited as:

• (4)

where P is the price of flour; Q is industry output in kilograms; β measures the size of market demand; γ is an index of concavity; and if γ is positive, then α is the maximum willingness to pay. Given this, the forms of the demand function include the linear (γ = 1), quadratic (γ = 2), log-linear (α = 0, γ<0), and exponential (α, γ→∞, and γ/α is a constant) forms.

Although the demand for flour did not exhibit significant seasonality, it did expand over time. I thus allow demand to evolve as per capita income continues to rise in the path of economic growth. For the sake of the identification concern, the specification allows either α or β depending on income, but not both. The four functional forms conditioned on α = α₀ + α₁y are as follows:14, 15

• (5)

Endogeneity

Because the price of flour in the demand function is endogenous, following the G-C approach, these demand functions are estimated using nonlinear instrumental variables (NLIVs). Given that wheat is the main component of cost, the NLIVs use international wheat prices as the instrumental variable to obtain consistent estimates.16 This instrumental variable takes advantage of the fact that Taiwan is a small open economy that trades wheat at given world prices. Therefore, the wheat price becomes an exogenous variable that is unrelated to the error term in the demand equation.

Demand for Flour

Table 2 presents the demand estimates corresponding to each alternative specification.17 In all cases the standard errors are corrected for conditional heteroskedasticity and serial correlation.18 The empirical results show that the estimates are reasonable predictions from economic theory. Increases in flour prices result in lower sales, and increases in per capita income result in higher sales.19

Table 2. Demand for Flour

	(1) Linear (γ = 1)	(2) Quadratic (γ = 2)	(3) Log-Linear (α = 0)	(4) Exponential (α,γ→∞)
Note All regressions were estimated using the NLIVs. Standard errors are in parentheses. Significance levels are denoted by the symbol *, for which **denotes significance at the 1% level, and *denotes significance at the 10% level.
α0	35.034** (11.843)	70.356** (29.218)
α1	0.00014* (0.00008)	0.00014* (0.00008)
β	5,037.292* (2585.669)	35.840 (36.160)	−13.078** (0.443)	2,66,560.623** (49,464.460)
γ			−0.398* (0.181)
γ/α				−0.034* (0.016)
Degrees of freedom	18	18	19	19
Adjusted R^2	0.42	0.50	0.15	0.14

CONDUCT AND COSTS

Generalized Pricing Rules

In this section I discuss the direct measures of price, cost, and conduct (θ) using demand estimates from Table 2 and cost information from the TFTC. Given Equations (2) and (4), the pricing rule depends on cost and conduct:

• (5)

By substituting the mean value of and α = α₀ + α₁y into Equation (5), row 1 of Table 3 presents the generalized pricing rules implied by different specifications of the demand function. The pricing rule varies substantially across specifications, unless θ is small. When θ is small, for example θ = 0 (perfect competition) in row 2, the flour price equals marginal cost in all specifications (). Alternatively, row 3 shows the implied monopolized price for θ = 1, in which the price ranges from −3.39 to 34.08 depending on the specification. Two pieces of evidence indicate that the monopoly-pricing rule fails to explain our case. First, most hypothetical monopoly prices are much higher than the observed prices listed in row 7. Second, if the market demand is specified as log-linear form, then θ must be lower than 0.4 so as to conform to the existing cost data. (i.e., true MC = 5.08.) Otherwise, the flour price will become negative.

Table 3. Pricing Rules

	(1) Linear	(2) Quadratic	(3) Log-linear	(4) Exponential
Note The figures in parentheses are standard errors.
(1) P(MC,θ)				29θ + MC
(2) P^Competition (θ = 0, MC = 5.08)	5.08	5.08	5.08	5.08
(3) P^Monopoly (θ = 1, MC = 5.08)	26.04	31.05	−3.39	34.08
(4) P^Cournot (θ = 0.03, MC = 5.08)	6.30	6.23	5.49	5.95
(5) P^Colluison (θ = 0.18, MC = 5.08)	11.47	11.51	9.24	10.30
(6) η at full sample mean	0.32	0.33	0.40	0.35
(7) Observed price (P)	11.63 (0.72)	11.63 (0.72)	11.63 (0.72)	11.63 (0.72)
(8) Observed Lerner Index	0.56 (0.08)	0.56 (0.08)	0.56 (0.08)	0.56 (0.08)
(9) Observed adjusted Lerner Index	0.18 (0.03)	0.18 (0.03)	0.22 (0.03)	0.20 (0.03)

When θ = 1/32 = 0.03, a 32-firm Cournot noncooperative oligopoly, the price ranges from 5.49 to 6.30, which is lower than the observed prices. This indicates that millers did exercise market coordination to a certain extent and the collusive price should lie somewhere between the perfect monopoly and Cournot levels.20

Elasticity-Adjusted Lerner Index (θ)

Because the TFTC report provides complete price and cost information, one can calculate the actual price-cost margin () in row 8. To do this, multiply this margin by demand elasticity η in row 6 to determine the direct measure of industry conduct or the elasticity-adjusted Lerner index ) in row 9.21 Here, the demand elasticity is calculated using the full sample mean. The evidence indicates that the direct measure of θ ranges from 0.18 to 0.22 according to the different demand functions. These actual values of θ are above the Cournot level (θ^Cournot = 0.03), but below the monopolistic one (θ^M = 1). I hereby use θ = 0.18 in the linear case as the baseline specification and substitute it in Equation (5). Row 5 then shows that the simulated price ranges from 9.24 to 11.51, which is quite near to the observed price in row 7.

The relatively high values of the direct measure of θ suggest more market power than one would expect from an industry composed of 32 firms. The evidence also suggests that millers did have a certain degree of market power. On the other hand, the evidence also rejects the fully collusive regime (θ = 1). The likely explanations include the following two grounds. First, when excess capacity exists, perfect collusion can be sustained only if the cost of capacity is zero such that firms can carry considerable capacity to support the monopoly price in the unconstrained equilibria.22 Nevertheless, the capital cost cannot be zero. Therefore, excess capacity can only support a price between the Cournot level and the monopoly level, but not the monopoly price.23 Second, industry pricing is constrained by threats of foreign imports.24

Variation in Elasticity-Adjusted Lerner Index (θ)

The variation in the price-cost margin is not significant across periods. Under a linear specification, Table 4 indicates that both L_η and L were quite stable in the sample period, except for a slight decrease between 1995:3 and 1996:2. The mean value of L (L_η) is 0.56 (0.18) and the standard deviation is 0.08 (0.03). However, column 3 shows that the Herfindahl index barely changed and there were no new entrants or price wars. Thus, the market structure has remained stable over the sample period. The variation in L_η was unlikely to be due to the changes in market power.

Table 4. Lerner Indices by Quarter (Linear Specification)

	(1) Lerner index (L)	(2) Adjusted Lerner index (Lη = θ)	(3) Herfindahl index
1994:1	0.63	0.20	441.93
1994:2	0.64	0.20	417.38
1994:3	0.62	0.20	449.95
1994:4	0.58	0.19	468.59
1995:1	0.62	0.20	492.52
1995:2	0.59	0.19	474.31
1995:3	0.45	0.14	439.79
1995:4	0.41	0.13	419.52
1996:1	0.47	0.15	484.28
1996:2	0.37	0.12	497.51
1996:3	0.53	0.17	525.78
1996:4	0.55	0.18	500.72
1997:1	0.53	0.17	517.46
1997:2	0.50	0.16	513.39
1997:3	0.56	0.18	530.90
1997:4	0.57	0.18	520.73
1998:1	0.61	0.19	469.22
1998:2	0.63	0.20	442.12
1998:3	0.67	0.21	476.67
1998:4	0.63	0.20	458.49
1999:1	0.67	0.21	431.74
Average	0.56	0.18	474.91
Standard deviation	0.08	0.03	35.56

VALIDITY OF THE NEIO ESTIMATES

Validity of the NEIO Technique

Following exactly the same approach as that of G-C, I substitute MC = c₀ + kP_RAW into Equation (5) and yield the following pricing rule:

• (6)

In what follows, Equation (6) will be estimated by using nonlinear least squares under two information structures: (a) both c₀ and k are unknown to the researchers; and (b) k is known to be 1.37, but c₀ is unknown. Table 5 reports the nonlinear estimates of c₀, k, and θ together with direct estimates of these parameters.25 demand parameter estimates are taken from Table 2. Note that α_t = α₀ + α₁y_t takes different values for different period observations and hence one can achieve identification from changes in α. The covariance matrix is corrected for conditional heteroskedasticity and serial correlation.26

Table 5. Nonlinear Estimates of Pricing Rule Parameters

	Linear		Quadratic		Exponential		Direct
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
Note The figures in brackets are standard errors. Wk denotes the Wald statistic, which is based on the distance of from 1.37. Wθ denotes the Wald statistic, which is based on the distance of from 0.18. The 90% χ^2(1) critical value is 4.61 and the 95% critical value is 5.99. df = degrees of freedom. ***Indicates that the estimates are significant at the 1% level; **at the 10% level; and *at the 20% level.
c0	1.273 (5.915)	0.235 (5.175)	−0.691 (4.487)	−0.742 (3.848)	0.012*** (0.001)	0.012*** (0.001)	0
k	1.301*** (0.201)		1.364*** (0.170)		1.251*** (0.216)		1.37
θ	0.196* (0.152)	0.214* (0.144)	0.234** (0.108)	0.234** (0.014)	0.215** (0.132)	0.193* (0.143)	0.18
Wk	0.12		0.002		0.30
Wθ	0.01	0.06	0.25	0.27	0.07	0.01
df	18	19	18	19	18	19

The results show that the NEIO technique performs well in estimating the cost parameters. This can be seen from two pieces of evidence. First, is always close to 1.37. This is evidenced by the Wald statistics (W_θ) reported in the third row from the bottom. These statistics show that one cannot reject the null hypothesis of k = 1.37 even at the significance level of 10%. Second, the hypothesis of c₀ = 0 cannot be rejected in both the linear and quadratic cases. Although c₀ = 0 is rejected in the exponential case, the estimated value of 0.012 is quite small and close to zero. In brief, the NEIO estimates of the cost parameters are consistent with the direct measures.

As to the estimation of θ, the results are barely satisfactory. Although θ is slightly overestimated, the Wald statistics (W_θ) reported in the second row from the bottom show that one cannot reject the null hypothesis of θ = 0.18 at a significance level of 5%. In economic terms, the difference is not significantly large. This argument is further supported by row 5 of Table 3, in which the implied price for θ = 0.18 ranges from 9.24 to 11.51 and is quite near to the actual price (11.63). Finally, columns 2, 4, and 6 indicate that the partial cost information does not improve the estimates of θ in the linear as well as the quadratic cases.

In conclusion, the regimes for the perfectly competitive (θ = 0), noncooperative Cournot oligopoly (θ = 0.03), and monopoly (θ = 1) pricings are all rejected under various demand specifications. The collusive profits should lie somewhere between the perfectly collusive and Cournot levels.27 The intuition can be developed by an examination of the firms' behavior. In considering the case of joint profit maximization, the lower-cost firms should produce to capacity and let the higher-cost firms sell up to the residual demand. Evidently, in our case, this proposition is inconsistent with the fact that all millers produced at levels that were substantially below capacity. Therefore, one can reject the solution of perfect collusion, even if one has no information about whether the cartel used side payments to increase the colluding gains and to distribute the joint profits.

Estimating Cost Under the Assumed Regime of Conduct

The literature now and then estimates unknown cost parameters under the assumption of a particular regime of conduct—that is, by restricting θ to equal a specific value. Here, I estimate three models: perfectly competitive (θ = 0), Cournot noncooperative (θ = 0.03), and monopolistic (θ = 1). To maintain brevity, Table 6 reports only the results for the linear specification.28 As is expected, columns 1, 3, and 5 indicate that all models do a very poor job of estimating c₀ and k. The monopolistic assumption leads to a substantial underestimate of c₀. On the other hand, the competitive and Cournot assumptions lead to an over-estimated c₀. Because the direct measure of θ (0.18) is relatively lower than 1, it is natural for the discrepancy to be greater under the monopoly regime.

Table 6. Cost Estimates for Different Regimes

	Perfect competition		Cournot		Monopoly		Direct
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
Note Demand estimates are taken from the linear specification in Table 2. The figures in parentheses are standard errors. df = degrees of freedom. ***Indicates that the estimates are significant at the 1% level; **indicates significance at the 10% level; and *indicates significance at the 20% level.
c0	8.660*** (0.959)	7.919*** (0.117)	7.527*** (0.974)	6.842*** (0.119)	−29.094*** (2.518)	−28.005*** (0.261)	0
K	1.236*** (0.167)		1.246*** (0.169)		1.568*** (0.383)		1.37
df	19	20	19	20	19	20

Finally, columns 2, 4, and 6 indicate that the partial cost information (i.e., k is set to be 1.37) does not improve the results. An even worse situation is the scenario where all estimated coefficients being highly significantly different from 0 might lead researchers to arrive at an incorrect conclusion regarding industry cost.

Columns 1, 3, and 5 indicate that the impact of the assumed conduct on k is mild. For example, in the linear case, it can be seen from the Taylor series expansion of the pricing rule (6) around θ = 0 that:

• (9)

Because the mean value of the maximum willingness to pay is much larger than , the intercept involving c₀ in the pricing rule is more sensitive to the specification of θ than the slope involving k.

CONCLUSION

Using data from Taiwan's flour industry, this study conducts tests on the NEIO methodology similar to those conducted by the G-C model. In contrast to their findings, the evidence shows that the NEIO technique appears to perform reasonably well in estimating market power and cost parameters. The estimated conduct and cost parameters are close to the actual values derived from full cost information. The result is also insensitive to different demand specifications. This opposite conclusion is reached mainly because of the differences in the dataset source. It is hoped that by using a more subtle data structure can be prevented from deviating from their pursuit of the true relationship between market power and profits.

However, this study only concludes that the NEIO technique performs well in Taiwan's flour industry. The reason that it performs well is that I have the detailed cost data to compare the actual cost with the NEIO estimate. Then the question is, if one does not have the detailed cost data, one will still not know whether the NEIO techniques work well or not. Admittedly, it is difficult to extend our results to other industries because we do not have the detailed cost data. However, the evidence obtained still suggests that the NEIO technique is largely validated, with proven feasibility, although policy makers should be cautious when relying on NEIO techniques to diagnose industry performance. Furthermore, the result also underlines the importance of data collection requirements for antitrust law enforcement agencies.

1. 1

   Here, the conduct is proxied by the elasticity-adjusted Lerner index. See Bresnahan (1989, p. 1012).

2. 2

   They found that the direct measure of conduct fell outside the 95% confidence interval of the NEIO estimate.

3. 3

   See Section 'CHARACTERISTICS OF THE DATASET' for details.

4. 4

   This equation nests several features of the oligopolistic models. For a monopoly or perfect collusion, θ = 1; for perfect competition, θ = 0; and for symmetric Cournot competition, θ = 1/N. Here, N is the number of firms.

5. 5

   On the other hand, if market demand is neither very large nor very small, then industry capacity () is larger than demand, but the individual firm's capacity (k_i) is less than market demand at a price equal to MC. In this situation, Edgeworth (1925) showed that there is no equilibrium with the firms charging equal prices above MC, for firms would have excess capacity and an incentive to slightly undercut each other. He suggested that the market would fail to settle down and that price would vary cyclically between high and low values. Thus, there is no equilibrium in terms of a pure price strategy, but a mixed strategy equilibrium does exist under quite general conditions. However, if the G-C conditions require that the analysis must relate to only a pure strategy and that output price must tend toward uniformity across firms, then Edgeworth essentially showed that no such equilibrium exists. Thus, in order to obtain data to satisfy the G-C conditions, researchers must choose an industry with a very large excess capacity. See Osborne and Pitchik (1986, 1987).

6. 6

   Genesove and Mullin (1998) indicated that the dominant firm (ASRC) had retained substantial excess capacity even in the depths of the Arbuckle-Doscher price war.

7. 7

   See Ma (2008).

8. 8

   At the same time, the fringe firms still produce to capacity.

9. 9

   See Ma (2005).

10. 10

    Genesove and Mullin (1998, p. 357, p. 369) also did not overlook the Corts' critique, and discussed its econometric implications in the context of the sugar industry. However, because they were interested in conducting a validation study, they wanted to see how the final predictions compared, even if they knew one of the assumptions underpinning the methodology.

11. 11

    Because it is quite easy to produce flour, one might wonder why there are not any threats of entry that might weaken the cartel. This is because a successful quota system instituted by the cartel effectively ruled out these threats completely. Because Taiwan produces no wheat, millers have to import their materials from abroad. The TFTC report indicated that firms used a 50,000-ton vessel for each voyage so as to minimize transportation costs. However, this capacity was much larger than the material needs of a single firm. Therefore, the individual firms had to procure and to ship wheat jointly under the supervision of the cartel. Evidently, this import scheme allowed the cartel to block entry by not allowing new entrants to join the procurement group. Since 1990, there has only been one entrant (the Global Flour Company) that joined the industry in 1998, and it was a joint venture of several incumbent flour firms in southern Taiwan. Thus, the collusive behavior of the incumbents was not influenced by the threat of new entry for years (Ma, 2005, p. 111).

12. 12

    For example, Genesove and Mullin (1998, p. 373) cited a businessman's testimony that the larger plants might have a 3- to 5-cent cost advantage over smaller plants. Clay and Troesken (2003, p. 155) indicated that existing estimates of AC per gallon net of tax for 1891–1895 ranged from 12.71–15.05 cents. However, from the empirical point of view, these differences are trivial in economic terms.

13. 13

    The G-C model did not consider labor cost as a component of MC because employment remains the same regardless of output.

14. 14

    Note that y denotes per capita income.

15. 15

    Here, ε_t is the error term and t refers to time.

16. 16

    The data source is the price of U.S. No. 1 hard red winter wheat in the San Francisco Bay area.

17. 17

    Another specification that permits yields similar results. For brevity and later convenience, Table 2 reports only the results for .

18. 18

    See White (1980).

19. 19

    Recall that α₀ is the constant, and α₁ is the coefficient for income.

20. 20

    Davidson and Deneckere (1990) indicated that Friedman's trigger strategy allows a continuum of equilibrium solutions. Thus, any price between the Cournot price and the monopoly price can be sustained as the equilibrium.

21. 21

    The G-C model used the notations L_η and θ interchangeably to denote industry conduct.

22. 22

    See Davidson and Deneckere (1990).

23. 23

    Davidson and Deneckere referred to these equilibria as constrained semi-collusive equilibria.

24. 24

    The tariff rate for the flour was 20%.

25. 25

    Under , the log-linear specification (α = 0) is unavailable in Table 5 for the sake of the identification problem.

26. 26

    See White (1980).

27. 27

    When excess capacity exists, Davidson and Deneckere (1990) emphasized that any price between the Cournot level and the monopoly level can be sustained in equilibrium.

28. 28

    The alternate demand specifications yield similar results.

Biography

• Tay-Cheng Ma is a professor at the Department of Economics, Chinese Culture University. He received his Ph.D. at Johns Hopkins University in 1998. His current research interest is in the field of industrial economics.