J Prod Anal (2008) 29:193199 DOI 10.

1007/s11123-007-0080-4

Farrell efciency under value and quantity data

re Robin Cross Rolf Fa

Published online: 23 January 2008 Springer Science+Business Media, LLC 2008

Abstract In 1957, Farrell introduced a nonparametric method to estimate technical efciency. His original illustration, however, utilized value-based, rather than quantity-based data (p. 279). This common practice raises the question of how value-based DEA models coincide with quantity-based models. It is well known that the two models coincide when rms face identical prices. In practice, however, prices vary across rms and the two models yield materially different results. We decompose the resulting difference into its technology and rm-related components and then use Farrells original data set to show that the expected difference is systematic and one-sided. Keywords Data envelopment analysis Farrell efciency measures Gamma distribution Monte Carlo JEL Classications C43 D24

1 Introduction In his classic paper, The Measurement of Productive Efciency, Farrell (1957) introduced a radial approach for estimating efciency, referred to today as data envelopment analysis or DEA. His input oriented model is related to the Mahler (1939) inequality, which states that the normalized cost function is less than or equal to the input distance function. This duality between a support function, the cost function, and the corresponding distance function, Shephards (1953) input distance function, is the basis for gauging and decomposing input efciency. DEA estimates have been shown to be consistent (Kneip et al. 1998) with a known asymptotic distribution (Gijbels et al. 1999). Farrell originally dened technical efciency in terms of input and output quantities (p. 254). However, it is common for data to be furnished in expenditure and revenue (value) terms, rather than physical input and output (quantity) terms, e.g., labor expenditure rather than labor hours. This is due, in part, to regulatory requirements, such as tax reporting rules, and in part to the difculty of determining quantity measures for certain inputs, for example capital. As a result, many DEA studies have utilized value-based data for one or more inputs or outputs. One important example is Farrell himself. Recent examples from a variety of research areas include Tsai et al. (2006), Ventelou and Bry (2006), and Cullinane et al. (2005). An alternative value-based DEA model for cost efciency was considered by Tone (2002) and extended to directional re and DEA by Fukuyama and Weber (2004) and Fa Grosskopf (2006). In this paper, we address two questions central to the use of value data in estimating technical efciency. First, how does the value-based DEA model relate to quantity-based DEA models? Second, if they do not coincide, then what

R. Cross (&) Department of Agricultural and Resource Economics, Oregon State University, 221B Ballard Hall, Corvallis, OR 97331-3601, USA e-mail: robin.cross@oregonstate.edu re R. Fa Department of Economics, Oregon State University, 312 Ballard Hall, Corvallis, OR 97331-3601, USA e-mail: rolf.fare@oregonstate.edu re R. Fa Department of Agricultural and Resource Economics, Oregon State University, 312 Ballard Hall, Corvallis, OR 97331-3601, USA

123

194

J Prod Anal (2008) 29:193199

exactly does the value-based DEA model measure and how do we interpret the difference? It is well known that the two models coincide when all re and Grosskopf rms face the same set of prices (Fa 1985). However, prices are known to vary across rms in practice (Carsten 2001; Engle and Rogers 1996; Rogoff et al. 2001), and when prices vary across rms, as we illustrate, the two models do not coincide. Farrell initiates the discussion of what value-based models measure when the prices are assumed, incorrectly, to be the same across all rms. He states, In this case, the (technical efciency) index will reect not only the technical efciency of the rm but also the extent of its adaptation to a set of factor prices different from those facing it, (p. 264). We formalize this claim and answer our second question by introducing a multiplicative decomposition of the difference between the two models into a technology and rm-related component. Finally, we illustrate the potential magnitude and direction of this difference. Re-considering Farrells original 1957 empirical data set, in which prices varied across the decision-making units (DMUs) studied, we nd both the technology-related and rm-related effects to be material, disrupting the relative efciency ranking of the DMUs. Employing Monte Carlo simulation, we show that the expected difference is one-sided and systematically increasing in the level of price variation. Condence intervals are provided.

x2

a b L(y) 0

x1

Fig. 1 Quantity-based model with distance from the origin to the efcient frontier 0b, distance from the origin to the rm 0a, and inputoriented technical efciency score 0b/0a, a simple ratio of the two distances

xv2 c

L v (y)

xv1

Fig. 2 Value-based model with technical efciency score 0d/0c

2 Decomposition of the value-based DEA model In this section, we consider the case in which prices vary across rms and the two models do not coincide. In this context, we ask the question: If the value-based technical efciency model does not coincide with the quantity-based model, then what does it measure? We answer this by decomposing the value-based model into three multiplicative components: the quantity-based technical efciency score, a frontier-related term, and a rm-specic term. Later, we will use this decomposition to illustrate the magnitude and direction of the difference resulting from prices that vary across rms. To illustrate the decomposition, Fig. 1 shows a simple quantity-based model, in which the location of the rm is labeled a and the efcient frontier b. In Fig. 2, the value-based practitioner observes the efcient expenditure frontier, labeled c, and the location of the rms expenditures, labeled d. Because prices vary across rms, there is no way that the practitioner can recover all of the quantity information using a single price index vector. As a result, two components of interest may differ between Figs. 1 and 2. First, the distances between the origin and the rms may differ. We will refer to this as the rm effect and measure it as 0a/0c. Second, the distances between the origin and the efcient frontiers may differ. We will call this the technology effect and measure it as 0d/0b. We then have the following identity, 0d =0c 0b=0a 0d =0b 0a=0c : 1

This identity states that the value-based technical efciency score (0d/0c) can be decomposed, multiplicatively, into the quantity-based technical efciency score (0b/0a), a technology effect (0d/0b), and a rm effect (0a/0c). We assign notation to this identity, TEv yvk0 ; xvk0 TEi yk0 ; xk0 ck0 dk0 ; 2 where the technology and rm effects for rm k0 are denoted ck0 and dk0 ; respectively.

3 Empirical illustration Next, we use Farrells original empirical example to show what may happen when prices vary across rms. In Farrells

123

J Prod Anal (2008) 29:193199

195 Table 1 Sample of Farrells original 1950 data set State 1. 2. 3. 4. 5. 6. Maine New Hampshire Vermont Massachusetts Rhode Island Connecticut Land 11.8 13.6 23.5 5.1 4.6 5.3 Labor 153.9 175.4 207.1 138.2 151.5 125.7 Materials 222.7 411.2 328.0 309.6 331.0 317.3 Capital 341.8 346.0 454.9 286.9 274.5 258.3

example, output prices varied rather than input prices as illustrated in the previous section. Under the constant returns to scale assumption, however, output price variation is exactly reciprocal to input price variation in the single output model, and Farrells example is well suited to our point. Farrell states that, due to scarcity of detailed price data and the variation in prices across rms, allocative efciency, ...is a measure that is both unstable and dubious of interpretation; its virtue lies in leaving technical efciency free of these faults, (p. 261). In his empirical illustration, Farrell utilizes value-based data, imparting to his technical efciency estimates those properties he had reserved for estimates of allocative efciency. Farrell collected much of his data on state-level agricultural output from a 1952 publication by the United States Department of Agriculture. Output, in this case, is reported in value terms. Located a few pages away, in the same source document, we nd agricultural commodity prices by states. Without exception, these prices vary across states. Specically, Farrell uses 1950 agricultural production data from the United States. He considers four inputs, including land and labor, measured in physical units, and materials and capital, measured in dollars. Farrell denotes these inputs as b, c, d, e, respectively, not to be confused with our labels in the previous section. Agricultural output, which he denotes a, includes cash receipts from farming plus the value of home consumption, measured in dollars. We use the following DEA model to recreate Farrells technical efciency scores: ^k0 minfkk0 : TEv 1; x
K X k 1 K X k 1

Table 2 Sample of 1950 cattle prices by state

State 1. Maine 3. Vermont 4. Massachusetts 5. Rhode Island 6. Connecticut

Cents per lb 23.3 23.4 32.6 39.1 24.9

2. New Hampshire 25.0

zk ! 1 ^k1 zk x . . . ^k0 1 kk0 x 3

K X k 1

^k4 zk x

^k0 4 kk0 x

zk ! 0; k 1; :::; 48g; where b ^1 ; x a c ^2 ; x a d ^3 ; x a e ^4 ; x a a 1 : a 4

Table 1 provides an example of the resulting data that appear in his paper (p. 279) for the rst six states. From the same source document, we see that average selling prices for cattle varied across states in 1950. An example of these average selling prices is given in Table 2 for the same six states. Milk and other crop and livestock commodity prices also vary. In 1950, cattle and milk represented 18.9% and

14.1% of the total value US agricultural production, respectively, together about one third of total output. The average price for cattle across the 48 continental states was 25.9 cents per pound, and the standard deviation was 6.1 cents. Milk prices averaged 4.38 dollars per hundredweight with a standard deviation of 1.13 dollars. Recall that if prices are identical across all DMUs, the two models coincide. To illustrate the impact of allowing a single set of prices vary, we relax the price assumption slightly by allowing cattle and milk prices vary across states, but continue to assume that they are identical across farms within each state. Under this weaker assumption, we wish to compare Farrells original scores to those estimated from data with milk and cattle price variation removed. To remain consistent with Farrells model specication in Eq. 3, we remove the price variation by multiplying the physical units of cattle and milk production for each state by a single national average price for cattle and milk. This retains Farrells use of revenue data, but without variation across states. We add these resulting cattle and milk revenues to the non-milk and cattle state revenues to obtain a single output value for each state. Finally, we divide the four inputs by this corrected output as in Eq. 4 to obtain a data set directly comparable to Farrells, but free from price variation in approximately one third of its output. Output aggregation in technical efciency models leads to a number of problems separate from those we consider re et al. 2004). For comparability, we in this paper (Fa maintain the same level of aggregation as Farrells original data set, removing only price variation.

123

196 Table 3 Resulting technical efciency estimates 1. 2. 3. 4. 5. 6.

J Prod Anal (2008) 29:193199

State Maine New Hampshire Vermont Massachusetts Rhode Island Connecticut

Farrells (TEv) 94 71 58 99 100 100

Corrected (TEi) 94 70 60 96 94 100

Tech. effect (c) 96 96 97 94 92 93

Firm effect (d) 104 106 101 109 116 107

Table 4 Summary statistics Mean Standard deviation

Farrells (TEv) 82.2 12.6

Corrected (TEi) 84.5 12.2

Tech. effect (c) 99.6 2.6

Firm effect (d) 98.8 5.9

Using the same model and normalization as given by Eqs. 3 and 4 above, we estimate the technical efciency scores TEi, the technology effect c, and the rm effect d. An example of these estimates, all multiplied by 100, is shown in Table 3 for the same six states, starting with the recreated Farrell value-based technical efciency scores TEv. Statistical rst moments and standard deviations for all 48 states are given in Table 4. As can be seen from Table 3, value-based technical efciency scores do not coincide with those of the partially corrected model. By correcting for price variation in just one third of total output and leaving all other expenditure and revenue terms unaltered, individual state efciency scores differed between the two models by 3.8% up or down, on average and by as much as 22% in the case of Wisconsin, a state heavily dependant on milk production. On average, Farrells value-based estimates are 2.7% lower than the corrected model, (82.284.5)/84.5, a characteristic that we will see again later. An important utility of DEA efciency analysis is its ability to assign a rank to each DMU relative to its peers. The level of price variation described above was sufcient to disrupt the relative ranking of states by 2.7 rankings, up or down, on average. In particular, Wisconsin dropped 17 rankings from the 11th most efcient state under the corrected model to the 28th under Farrells value-based model.

questions. Is the expected magnitude of the technology and rm effect commensurate with the level of price variation or coincidental? If commensurate, are the expected magnitudes linear in price variation? Does the variation in these terms behave similarly? Are these terms one-sided? If so, what is their expected direction? Here, we use Monte Carlo simulation to estimate the magnitude and direction of the expected technology effect c and the rm effect d over a range of price variation. We begin by introducing a model to illustrate the expected values of interest, along with the corresponding 95% condence intervals. Then, we verify that the differences found in Farrells empirical example lie within our 95% condence interval, given the level of price variability in the markets for cattle and milk in 1950. Finally, we map the expected magnitudes of the differences and condence intervals over a range of price variation, and nd that the expected technology effect introduces systematic and onesided bias to technical efciency scores.

4.1 Model and assumptions As in our empirical example, we allow cattle and milk prices to vary across states. Price vectors for milk and cattle in the 48 states, denoted pm 2 <48 ; where m = milk, cattle, are vectors of real-valued, non-negative IID random variables, drawn from their respective two-parameter Gamma distributions. While a number of distributions are used to model prices, the Gamma family is desirable because it is non-negative (Greene 1980, 1990) with exible higher moments. Symmetric distributions, such as the Normal and the Students t, are consistent with prices when truncated at zero, but lack exible higher moments. The Gamma family of distributions, expressed in terms of

4 Expected differences over a range of price variation The empirical illustration suggests that Farrells valuebased technical efciency scores were, on average, lower than quantity-based scores and the differences varies across states. These observations raise a number of important

123

J Prod Anal (2008) 29:193199

197
m 1X c0 m j 1 k ; j

random variable p with density function f, is given by Spanos (1999) as the set F satisfying, ( )     b1 p a1 p F f p; a; b exp ; a; b [ 0 b Ca b 5 where C[a] is the Gamma function (Abramowitz and Steguns 1964). Special cases include the exponential density when a equals 1 and the Chi-square density when b equals 2 (Wackerly et al. 2002). The mean and variance of the Gamma density are ab and ab2, respectively. We will estimate the shape parameters using a Maximum Likelihood Estimation (MLE) model by solving the usual maximization problem ( ) 48 X max lnf p; a; b ; 6
a;b k 1

^ ck 0

Monte Carlo estimates are consistent, asymptotically normal, and accurate to an arbitrary level, depending on m (Campbell et al. 1997). Condence intervals can be readily calculated, as Probck0 2 ^ ck 0 z ; ^ ck0 z 95%; z 1:96rck0 ; j p : m 8 Equation 8 states that the true term ck0 lies within the given interval with 95% probability. Because the true standard deviation of the draws rck0 ;j is not known, we estimate it using the same Monte Carlo simulation method: m 2 1X ^ ck 0 ; j ck 0 ; j ^ ck 0 ; 9 r m j 1 where ^ ck0 is given by Eq. 7. Solving Eq. 8 for m, we determine that 200,000 draws are required to obtain an accuracy level z of 0.0001% at the 95% condence level.

where f is the Gamma density dened in Eq. 5. MLE parameter estimates are consistent and asymptotically normal (Greene 2003) under the null hypothesis, that is, when the true density g, is from the family of densities specied in (5). When the density is mispecied, i.e., when g 62 F, as may often be the case, solutions to (6) may be quasi-maximum likelihood estimates (QMLE). QMLE parameter estimates are consistent under certain restrictive conditions, but tend to be less efcient in nite samples than when density specication is correct (White 1982). In either case, Farrells sample size of K = 48 is quite small for inference purposes in this context. We use Monte Carlo simulation to calculate the expected value of c and d, the technology and rm effects, respectively. Monte Carlo is a useful tool for estimating expected values of functions of random variables when analytical solutions are not possible. In our case, the random variables of interest are cattle and milk prices. We will estimate the expected value of the technology-related and rm-related terms generated by these random prices. For instance, to estimate the expected value of technology-related term ck0 for rm k0 , we begin by generating 96 price observations, one for cattle and one for milk for each of the 48 states. Each price observation is a pseudorandom draw from the Gamma density given in (5) with parameter values a and b obtained by solving Eq. 6. Multiplying each states physical production levels of cattle and milk by these prices, we are able to reconstruct one value-based data set. We then run the DEA model in (3). With these efciency estimates and our corrected efciency estimates from the previous section, we calculate the mean technology-related term using Eq. 2. We repeat this process m times and average the results. Adapting the notation of Campbell et al. (1997), our model is,

4.2 MLE results MLE parameter estimates, normalized using the scalar average price, converged normally in the Matlab programming environment which uses the estimation procedure detailed in Hahn and Shapiro (1994). Parameters for normalized cattle prices were 18.7 and 0.054 for a and b, respectively, and 14.8 and 0.068 for milk. Parameter estimates were signicantly different from zero, though inference with respect to such parameters and with so small a sample size is dubious.

4.3 Monte Carlo results Our rst task is to verify that the value-based scores from the Farrell data TEv yvk0 ; xvk0 ; ck0 ; and dk0 ; listed in Table 1, lie within our 95% condence intervals. We nd that they do, as shown in Fig. 3 below. The technology-related term decreases at an increasing rate and ranges from a bias of zero (c = 1) when prices do not vary, to a 5% downward bias (c = 0.95) when the coefcient of price variation reaches 0.5. This downward sloping line suggests that price volatility leads to systematic downward bias in value-based technical efciency scores. Downward bias is not surprising, since we would expect volatility to disburse observations, thereby pushing the efcient input isoquant toward the origin. The width of the 95% condence interval increases from zero, when

123

198 Fig. 3 Expected technology and rm effects, (a) and (b), respectively, standard deviations, (c) and (d), with Farrell example results x and 95% condence intervals (dashed) over a range of price variation measured by the coefcient of variation
1.05 1.05

J Prod Anal (2008) 29:193199

(a)
Tech. Effect (Gamma) Firm Effect (Delta) 1.00 1.00

(b)

0.95

0.95

0.90

0.1

0.2

0.3

0.4

0.5

0.90

0.1

0.2

0.3

0.4

0.5

0.20

0.20

(c)
Std. Dev. of Tech. Effect Std. Dev. of Firm Effect 0.15 0.15

(d)

0.10

0.10

0.05

0.05

0.00

0.1

0.2

0.3

0.4

0.5

0.00

0.1

0.2

0.3

0.4

0.5

prices do not vary, to 14% at a coefcient of price variation of 0.50. At this higher level of variation then, technical efciency scores may range from 1% overstated to 13% understated. As we saw from the Farrell empirical example, a 2.7% downward bias was sufcient to cause considerable disruption in the relative efciency rankings of states. It is also worth noting that we are considering the impact of variation in only one third of total output value. Clearly, price variation in other inputs and outputs could materially alter the magnitude of the bias demonstrated here. The expected rm effect appears to be invariant to price volatility, that is, delta is unity over all levels of price variation. The width of the 95% condence interval, however, increases from zero when prices do not vary, to 8% when the coefcient of price variation reaches 0.5. Again, average value-based technical efciency scores may be understated or overstated by as much as 4% at this level of price variation. This result is again not surprising, since we would expect price variation to disburse observations, but not necessarily in a particular direction. 5 Conclusion We have shown that when prices vary across rms, valuebased DEA models do not coincide with quantity-based DEA models. To better understand the value-based DEA score, we decomposed the resulting difference multiplicatively into a technology effect and rm effect. To illustrate the possible magnitude and direction of these effects we reexamined Farrells original empirical data set, in which

prices varied across rms. Correcting for a portion of the price variation, we found material differences between Farrells original value-based technical efciency scores and the corrected quantity-based scores. Relative rankings between states were also disrupted. Finally, mapping expected technology and rm effects over a range of price volatility levels, we found the bias resulting from the technology effect to be systematic and one-sided.

References
Abramowitz M, Stegun IA (eds) (1964) Handbook of mathematical functions. applied mathematics series 55. United States Department of Commerce, National Bureau of Standards, Washington, DC Campbell JY, Lo AW, MacKinlay AC 1997 The econometrics of nancial markets. Princeton University Press, New Jersey Carsten H-P (2001) A general refutation of the law of one price as empirical hypothesis. Jahrbucher f. Nationalokonomie u. Statistik 221(no. 1):4561 Cullinane K, Song DW, Wang T (2005) The application of mathematical programming approaches to estimating container port production efciency. J Prod Anal 24:7392 Engel C, Rogers JH (1996) How wide is the border? Am Econ Rev 86(5):11121125 re R, Grosskopf S (1985) A nonparametric cost approach to scale Fa efciency. Scand J Econ 87:594604 re R, Grosskopf S, Zelenyuk V (2004) Aggregation bias and its Fa bounds in measuring technical efciency. Appl Econ Lett 11:657660 re R, Grosskopf S (2006) Resolving a strange case of efciency. Fa J Oper Res Soc 57:13661368 Farrell MJ (1957) The measurement of productive efciency. J R Stat Soc Ser A 120(III):253281 Fukuyama H, Weber WL (2004) Economic inefciency measurement of input spending when decision-making units face different input prices. J Oper Res Soc 55:11021110

123

J Prod Anal (2008) 29:193199 Gijbels I, Mammen E, Park B, Simar L (1999) On estimation of monotone and concave Frontier functions. J Am Stat Assoc, Theory Meth 94(445):220228 Greene WH (2003) Econometric analysis, 5th edn. Prentice Hall, Saddle River, NJ Greene WH (1990) A gamma-distributed stochastic Frontier model. J Econ 46:141163 Greene WH (1980) Maximum likelihood estimation of econometric Frontier functions. J Econ 13:2756 Hahn GJ, Shapiro SS (1994) Statistical models in engineering. Wiley, New York Kneip A, Park B, Simar L (1998) A note on the convergence of nonparametric DEA estimators for production efciency scores. Econ Theory 14:783793 Mahler, K (1939) Ein Ubertragungsprinzip fur Konvexe Korper, vol. 64. Casopij pro Pestdvani

199 Rogoff, K, Froot, KA, Kim, M (2001) The law of one price over 700 years Shephard RW (1953) Cost and production functions. Princeton University Press, Princeton Spanos A (1999) Probability thoery and statistical inference, 1st edn. Cambridge University Press, Cambridge Tone K (2002) A strange case of the cost and allocative efciencies in DEA. J Oper Res Soc 53:12251231 Tsai HC, Chen CM, Tzeng GH (2006) The comparative productivity efciency for global telecoms. Int J Prod Econ 103:509526 Ventelou B, Bry X (2006) The role of public spending in economic growth: envelopment methods. J Policy Model 28:4003413 Wackerly DD, Mendenhall N, Sheaffer RL (2002) Mathematical statistics with applications, 6th edn. Duxbury, Pacic Grove White H (1982) Maximum likelihood estimation of misspecied models. Econometrica 50(1):125

123