Using accounting data in cartel damage calculations: blessing or menace?

Abstract

Standard methods for calculating cartel-damage rely on data of prices charged and quantity sold. Such data may not easily be available. In this paper, it is shown that accounting data can be used for computing a lower bound for cartel-damage. Previous literature indicates that economic profits can hardly be inferred from accounting data. Therefore, it is shown under which econometrically testable assumptions on accounting costs a meaningful lower bound for cartel-damage can consistently be estimated when using accounting data. However, the aggregation-level and the publication-frequency of accounting data pose a challenge to the estimation of cartel-damage. A further challenge is to appropriately reflect the strength respectively effectiveness of the collusive agreement in the specification of any such estimation.

Appendix

1.1 A dataset for the vitamins cartel

In this section the above method for quantifying the cartel-induced damage is applied to the vitamins cartel. Connor (2001, p. 277) considers the vitamins cartel to be “the first, the biggest, most elaborate, most complete, longest lasting, and most influential of the cartel pandemics of the 1990s. [… Moreover, the vitamins cartel] extracted historically unprecedented monopoly overcharges from customers. These overcharges later appeared on the profit statements of the vitamin manufacturers.” First, I briefly describe the vitamins cartel and the data used. Second, I show how such data would be used to calculate a lower bound for damages. In this context, I demonstrate some of the difficulties associated with such calculations as described in Sect. 3 above.

Between 1989 and 1999 several European, Japanese, and North-American producers of vitamins colluded in prices.¹² Among these firms were BASF, Roche, Daiichi, and Takeda that agreed to jointly raise prices of the vitamins shown in Table 1. This table provides some basic information on the structure of these cartels. The supposed ringleader (i.e. Hofmann-La Roche in five cases) is highlighted in bold. The intensity of the collusive agreement varied over both the affected vitamins and over time, as the prices were raised gradually rather than bouncing up immediately after the formation of the cartel. After the discovery of the cartel, prices dropped by about 50% within 2 months leaving prices 30–40% below the pre-cartel level. These price movements are not uniform across all cartelized vitamins. For example, prices for liquid vitamin A almost tripled during the collusive phase while those for vitamins B1, B6, folic acid, C, and B2 were raised less strongly (Connor 2001, p. 322).

Table 1 The vitamins cartel

Yearly firm-level data of the above firms’ revenues and costs of goods sold is collected from the ThomsonReuters Datastream database for the period 1985–2007.¹³ These variables are measured in thousands of each national currency, i.e. Euro for BASF, Swiss Francs for Roche, and Yen for Daiichi and Takeda. Figure 2 provides an overview on the development of these four firms’ revenue and costs of goods sold as well as the difference and ratio of these two variables.

Fig. 2

Revenues and costs

Fig. 1

Revenue effects

1.2 Quantifying cartel damages in the vitamins cartel

In Sect. 2.1, I show that a lower bound for cartel-damages can in principle be found as the regression-coefficient of a cartel-dummy in a regression of the difference between a firm’s revenue minus its costs of goods sold on possibly explanatory variables. Such a regression is performed for BASF, Roche, Daiichi, and Takeda using the above publicly available data. The aim of this exercise is to show some of the difficulties that are outlined in Sect. 3 and may arise when performing regressions of this type.

A first difficulty consists in the definition of the cartel-dummy. Note that only one cartel-dummy can be used for BASF’s and Roche’s participation in the cartels of vitamins B5 and beta-carotene as both cartels occurred at the same time. As is standard, all dummies take a value of zero when a cartel is inactive and a value of one when the cartel is in force. Such a dummy does not reflect variations in the strength of a collusive agreement that lead to the price movements as described above. As a second difficulty, one should note that the use of yearly data keeps time-series short (23 respectively 21 observations). Therefore, one must be parsimonious in specifying the set of further explanatory variables X. These variables and the estimation methodology are presented in turn.

I start with describing the regression methodology for BASF and present the regressions for the other firms further below. For BASF the set of regressors X comprises an IFRS-dummy that covers the effects of BASF switching from German-GAAP to IFRS in 2004. Since the US-market for vitamins is quite important, the US-$/Euro exchange rate is also included in X. A further variable is incoming orders (IO _t ) in the chemical industry. This value index takes a base value of 100 in year 2000 and is provided on a monthly basis by the German Statistical Office (Destatis). A third difficulty arises because this value index applies to the reunified Germany so that the time series only starts in 1991. The missing values for the period 1985–1990 are backcasted on basis of a regression of IO _t on IOw85 _t , which is a monthly value index for incoming orders in the West German chemical industry with base year 1985 (IOw85 _t ) and observations ending in 1994.

A fourth difficulty may arise because of statistical problems or an inadequate specification of the regression. In this context, estimating the damage-regression (9) by standard OLS¹⁴ does not provide informative results. None of the coefficients is statistically significant except for the one associated with incoming orders (p-value 0.01%). The Durbin-Watson statistic indicates positive, first-order autocorrelation in the error term which implies misspecification of the model. An Augmented Dickey-Fuller test indicates that both, the dependent variable and incoming orders, are non-stationary, which potentially leads to a spurious correlation problem. Such a problem can, for example, be overcome by estimating an error correction model or by performing a regression in first differences. In this specific case, it turns out that estimating an error correction model does not provide additional information in comparison to a regression in first differences (see Eq. (16)). Therefore, the latter is used to draw further inferences.

$$ \Updelta \pi_{t} = \alpha + {\varvec{\upbeta}} \cdot {\mathbf{\Updelta D}}_{{{\mathbf{Kj,t}}}} + \gamma \cdot e_{EURUSD} + \delta_{1} \cdot \Updelta IO_{t} + \varepsilon_{t} $$

(16)

The adjusted R² of this regression amounts to 53% with a, γ, and δ ₁ being significant at the 1%-level and taking the expected signs (γ < 0 and δ ₁  > 0). The β-coefficients for the cartels in vitamins B4, C, and B2 are insignificant, which is reasonable since these vitamins were subject to competition. The β-coefficients for the cartels in vitamins A, E, beta-carotene and B5 are significant at the 1%- respectively 10%-level. These β’s are negative which is counterintuitive. Therefore, in light of the above difficulties a lower bound for cartel damages cannot be quantified. A similar picture can be seen in the regressions for Roche.

A fifth difficulty is imposed by the existence of structural breaks. In case of Takeda it is found that π kinks upwards in 1999 such that a simple linear trend rises too strongly prior to 1999 and too softly afterwards. As a consequence, a second linear trend is included starting in 1999. Since for the Japanese chemical industry no data of orders or sales was available the only additional variable is the YEN/USD exchange rate. Therefore, a sixth difficulty arises because of the lack of data. As a consequence, the estimated coefficients of the cartel-dummies do not provide good estimates of the damage caused by Takeda because the high value of R² is mainly due to the inclusion of the trend terms and the constant. A similar result applies to the regression for Daiichi.

Table 2 reports summary statistics for the above regressions, i.e. the adjusted coefficient of determination, the p-values of F-tests examining the hypotheses that (a) all coefficients or (b) all cartel-coefficients jointly equal zero, the p-value of the Jarque–Bera test for normally distributed error terms, and the Durbin-Watson statistics. The latter’s values indicate misspecification for at least BASF and Daiichi. However, the effects of autocorrelation must be traded off with parameter-parsimony that is required by the limited number of observation per series.

Table 2 Summary statistics