Journal of Productivity Analysis, 10, 289–304 (1998

c 1998 Kluwer Academic Publishers, Boston. Manufactured in The Netherlands.

Productivity Measurement from a Reference Technology: A Distance Function Approach
DAVID K. LAMBERT* Department of Applied Economics and Statistics, University of Nevada, Reno

Abstract An approach to productivity measurement based on distance function measurement is developed in this paper. A series of reference technologies are constructed by augmenting observed inputs and outputs such that the resulting effective netputs are consistent with the weak axiom of profit maximization. Geometric means of year-ahead and year-back distance function values serve as measures of multifactor productivity changes. Both primal measures from the inner bound and dual measures from the outer bound of the reference technologies are calculated. Resulting productivity measures are similar to those derived using index number approaches, but potentially provide greater information regarding the reference technologies.
Keywords: Technical change, Index numbers, Multifactor productivity, Agricultural production JEL classification: D24



In this paper we develop a measure of productivity growth based on distance function measurements of observed production bundles from constructed reference technologies. The procedure is based on both quantity based indexes discussed by Malmquist (1953) and Farrell (1957) and on distance functions calculated from a dual profit frontier (F¨ re a and Primont, 1995). However, the approach differs from the techniques developed by F¨ re a et al. (1989, 1994) due to its suitability for measuring productivity changes for time series, rather than panel, data. The approach is similar to the primal-dual approach developed in Chavas and Cox (1994). However, our approach is quite different in the development of a reference technology from which distance function values are derived. Whereas Chavas and Cox (1994) derive distance function measures from unadjusted prices and quantities, we first develop reference technologies by calculating effective netputs that are consistent with Varian’s weak axiom of profit maximization (WAPM). The technique is nonparametric, thus eliminating the need to assume a particular functional form for the production relationships. The procedure is ideally suited to estimating multi-output productivity.
* The author is an associate professor in the Department of Applied Economics and Statistics at the University of Nevada, Reno, NV, 89557-0105, USA, phone number (702) 784-1675. Financial assistance from the Nevada Experiment Station, project number NEV05145 is gratefully acknowledged.

The latter assumption can be dropped. although the use of flexible functional forms increases the approximation properties of the estimating equations to the underlying technology. Another approach to estimating technical change is growth accounting. can greatly exaggerate estimates of factor biases in technical change. Another problem with the econometric approach arises from the time series nature of the data. 1976. input-output separability. The Fisher index has been shown to better satisfy various axiomatic. nonparametric procedures have recently been developed to decompose changes in the Fisher productivity index into technical and allocative efficiency changes. Lambert and Shonkwiler (1995) have shown how standard translog representations of the dual cost function. The T¨ rnquist index. However. (1994) about the inability of index numbers to differentiate the influence of a technical change from shocks or other sources of production inefficiency between periods. but seldom is in applied work. As commonly employed. using index numbers as discrete approximations to measuring continuous changes in technology (Diewert. and econometric estimates of deterministic or stochastic trend may serve as measures of technical change. Chambers. exact for the translog (Diewert. The economic justification for the use a of index numbers establishes a consistency between an underlying production technology and a given index. Estimation generally requires that inputs and outputs be highly aggregated. criteria than alternative index number measures. is o commonly used to measure changes in output not accounted for by changes in inputs. 1988). problems may arise using the econometric approach.1 Finally. primal or dual approaches may be employed to represent the underlying production technology. such as that between the translog technology and the T¨ rnquist index. and shifts in the cost function over time (Ray and Mukherjee. Lovell.290 2. such as including a deterministic trend. Common approaches to representing technical change. may be invalid. 1994). 1980. .2 o The two most common measures used in applied analyses are the T¨ rnquist and the o Fisher productivity indexes. Consequently. 1983. 1957). 1992). as well as satisfying more general economic qualities desirable in index number construction (Diewert. Since true production relationships cannot be induced from a few observations. F¨ re. or at least that inefficiencies are constant proportionally over time. and Sickles. scale changes. This recent work has addressed a problem raised by F¨ re et al. Solow. Efficient production is also generally assumed. One approach is based on an econometric approximation of production technology. several approximation techniques must be used to estimate intertemporal shifts. 1988). or test. and Grosskopf. The time series properties of the data must be considered for proper identification of the state of technology. Hicks-neutral technical change. with time included as a deterministic proxy for the state of technology. and constant returns to scale (Antle and Capalbo. 1976). 1988). productivity measurements are still seen to be dependent upon model specification (Guilkey. In addition. 1996). Capalbo. the T¨ rnquist multifactor productivity index (MFP) index measures o technical change under the assumptions of competitive behavior. 1980. Background LAMBERT Productivity analysis assumes that production possibility sets change over time (Diewert. thus imposing assumptions of separability on the technology (Antle and Capalbo.

or political changes might affect year-to-year comparisons. quantifying a change in MFP between two periods. 1989. A Distance Function Measure of Productivity Change Identifying Changes in Technology Consider the production possibilities set S available at time t: t S = {(x.. and convex. the distance function productivity measures are compared with both T¨ rnquist and Fisher o productivity estimates. without recourse to a presumed functional form underlying the aggregation functions. Output distance functions have recently been exploited to measure changes in production possibilities sets between periods (Caves et al. However. or product-product relationships. 1995). the constructed reference technologies are available for analysis of input and output relationships should such information be desired. additional information cannot be derived concerning factor-factor. as well as substitution possibilities among inputs and outputs (Chavas and Cox. however. 1982. Consider a similarly defined production possibilities set available at period s. where exogenous factors such as weather. In addition. St is conditional upon the technology available at time t. Finally. if possible. the output set is bounded. This paper demonstrates how the nonparametric approach can provide estimates of MFP changes over time that are very similar to those resulting from index number approaches. One problem relates to identifying and. 3. Index number approaches provide no information on the nature of the production technologies generating the sample of netput vectors. business cycles. and positive levels of inputs . 1990). The next section discusses two of the problems associated with measuring shifts in production possibility sets over time. If production set St a conforms to a standard set of economic properties including weak disposability of inputs and outputs. The productivity measures may uncover rates of growth in outputs that are unexplained by corresponding changes in inputs. F¨ re et al. Ss . but Ss ⊂ St . y): x can produce y at time t} (1) where x ∈ n is a vector of inputs and y ∈ m is a corresponding vector of outputs. Progressive MFP change has occurred when s > t. A distance function approach to measuring productivity change is developed for measuring distance from a reference production technology in which all netput vectors are technically and allocatively efficient by construction. Changes in multifactor productivity (MFP) have occurred between periods t and s if St ⊂ Ss .PRODUCTIVITY MEASUREMENT FROM A REFERENCE TECHNOLOGY 291 Ray and Mukherjee’s work relies on linear programming techniques not dissimilar to those developed here. MFP change is regressive when t > s. closed. factor-product. can indicate the presence of technical change affecting the nature of production.. Trends over time. 1994). Procedures to identify changes in MFP are not straight-forward. Both measures will be local. The nonparametric approach has been shown to provide information on both input and output biases in MFP changes (Cox and Chavas.

A distance function value equal to 1 indicates the observed level of netput. o ms (xs . yt ) = 1/ Ds (xt . yt ) = [mt (xt+1 . ys ) = (sup(θ : (xs . yt ) o = 1/(sup(θ: (xt . xt . y) is on the frontier of St . mt (xs . yt+1 )/ Dt+1 (xt . ys . θ yt ) ∈ Ss ))−1 (6) (5) Caves et al. (1982) show how the geometric mean of m t and m s is a generalization of the T¨ rnquist productivity measure when the underlying production technology is translog. lies on the production surface of St . 1402). ys . yt ) = 1. mt (xs . xt . the Malmquist measure can also be derived relative to technology existing at period s. ys ) measures the proportional deflation in ys to conform with technology set St . although their approach obviates the need for presuming a structure on the underlying technology. yt ) o o Since Ds (xs . 1995. yt+1 . Values of Dt0 (·) less than 1 will lie within the boundary of St . (x. xt . yt ) = Ds (xs . yt+1 . (1982) assume that (xt . ys ) is assumed to equal 1. xt . yt ) = Dto (xs . p. Values of Dt0 (·) greater than 1 indicate that y cannot be produced given x and St . yt+1 . yt ) is the minimal output deflation factor required so that the deflated measure of output in period s.. p. so that Dto (xt . Therefore. yt )]1/2 o (7) . In order to avoid the choice of an arbitrary technology set (e.292 LAMBERT must be used to produce positive levels of output (F¨ re and Primont. ys . then the a technology can be completely described by the output distance function: Dto (x. yt )]1/2 = [Dto (xt+1 . (1994) similarly define their Malmquist measure of productivity change based on a the geometric means of the Malmquist measures from two adjacent time periods. (1982) define this measure ms (xs . yt ) = Dto (xs . ys . (2) The output distance function is the reciprocal of the maximum proportional expansion in output y given x and St . xt . y/θ ) ∈ St } = (sup{θ: (x.g. ys /mt . ys )/ Ds (xt . ys . No increase in the observed levels of outputs y is possible given x and the technology available at t. o F¨ re et al. ys )/ Dto (xt . We adopt F¨ re et al. Caves et al. xt . y) = inf{θ: (x. implying that a proportional increase in outputs could occur for the observed level of inputs. yt ) (3) The numerator Dto (xs . (1982. xt . m(xt+1 . yt )ms (xt+1 . θ would indicate the minimal shrinkage of y to be on the boundary of St . Caves et al. either t or s) as a benchmark for measuring productivity changes between periods.’s approach and define our productivity index as the geometric mean of a Malmquist measures calculated at two adjacent time periods. 27). yt ) can always be found on the boundary of St . θ y) ∈ St })−1 . Caves et al. θ ys ) ∈ St ))−1 (4) Following Caves et al. (1982) developed productivity indexes based on distance function calculations from a reference technology. xt . Their Malmquist based index for an observed netput bundle in period s relative to technology associated with period t is: mt (xs .

. Confirming the existence of a production set does not provide information on the curvature of the production frontier. pi ). i = j.e. . In other words. If we have a set of K observations (xi . capital stock and employment.3 Consequently. (9) and proves that SO rationalizes the data (xi . .. .e. for all i. Varian (1984) has shown that there exists a closed. K. (1994) a measured gross domestic output for 17 OECD countries resulting from two factors. K}. referred to their Malmquist measure as a theoretical index since the reference technologies could not be represented without reliance upon. Varian has developed procedures to identify an outer bound on set S if one assumes observed netput bundles result from optimizing behavior such as profit maximization. Varian proves that SI forms the tightest lower boundary of all other production sets S that rationalize the data (i. yi ) and envelopes all production sets S that rationalize the data (i. this characterization defines a lower bound on the frontier of set S. establishing the existence of a production set that rationalizes the K observations involves checking the following K(K −1) inequalities: pi yi − ri xi ≥ pi y j − ri x j . the tightest inner bound of production set S can be defined: SI = com− {(xi . However. Varian (1984) also characterized the outer bound on the true production possibilities set: SO = {(x. for example. . Caves et al. . SI ⊂ S for all sets S that rationalize the data). . yi ) for all i = 1. . . K}. A production possibilities set is known to exist if observed production relationships are consistent with the weak axiom of profit maximization (WAPM) (Varian. . K. However. K. i = 1. . . F¨ re et al. . . If the data does satisfy WAPM. Price and Weyman-Jones (1996) examined efficiency and total productivity gains in the UK gas industry before and after the 1986 privatization of the industry. This raises the second problem in identifying shifts in the production possibilities set. convex. yi ) and corresponding prices (ri . All of these studies employed panel data on physical quantities. y) for each set of prices they faced. . for all i. consistency with WAPM requires pi yi − ri xi ≥ pi y j − ri x j .PRODUCTIVITY MEASUREMENT FROM A REFERENCE TECHNOLOGY 293 The Malmquist productivity index will naturally be conditional upon the location of the frontier of set S. Most studies employing a distance function approach to productivity measurement define changes in MFP relative to the inner bound representation of technology. . .’s (1982) original discussion of the Malmquist productivity index. a characterize technology as the convex hull of a set of production observations available at a given time period. an assumption of a translog technology. j = 1. . F¨ re and Grosskopf (1995) used a similar Malmquist index a in measuring differences in MFP for the agricultural sectors of nine EU countries and the United States. . Bureau.. (8) where com− {} is the negative monotonic hull of S. Employing Varian’s terminology. How should the production possibilities set S be represented in the calculation of the distance functions underlying the Malmquist productivity measures? F¨ re et al. Availability of panel data circumvents the problems mentioned in Caves et al. lower and upper bounds can be placed on production set S. negative monotonic production set that rationalizes the data. decision makers made the profit maximizing choices of (x. S ⊂ SO). j = 1. . for all S that rationalize the data. y): pi y − ri x ≤ pi yi − ri xi for all i = 1. . 1984).

convex. technical change may affect the quality of both inputs and outputs. . T . negative monotonic production technology that rationalizes the effective inputs and outputs. . a We turn next to a procedure to estimate productivity changes based upon a distance function approach when panel data are not available. Under an assumption of profit maximization. Distance function values were then calculated from these annual characterizations of technology. negative monotonic production set that rationalizes the observations. . . y∗ ): x∗ can produce y∗ for all i = 1. 1991). . 1992). Consider two production sets. Augmentation factors are sought that would result in pi y∗ − ri x∗ ≥ pi y∗ − ri x∗ . . measure by using panel data and the a inner bound characterization of technology to represent each year’s production possibilities frontier. yt+1 ) ⊂ St .’s productivity index. In this case. neutral technical change affecting aggregate output levels (Harvey and Marshall. failure of (xt+1 . . Augmentation factors are presumed to exist that explain qualitative differences between inputs and outputs in different periods. Creation of a Reference Technology Creation of a reference technology using time series data is more problematic since only one observation per period is available. (x∗ . for the observed production bundles to represent different technologies (see appendix 1). i. . page 584). 1990). though not necessary. convex. technology has changed between the two periods since St = St+1 . Binswanger. Consider a series of T observed production bundles (xi . j = 1. y∗ ). b). T . Testing for WAPM provides information on the existence of a closed. However. The resulting augmented inputs and outputs are termed effective netputs (Chavas and Cox. T } i i i i (11) The proof of (11) follows directly from Varian’s theorem 3 (1984. Technical change is generally attributed to differential improvements in the ability of factors to produce output (Lambert and Shonkwiler. . . we can ascertain whether technical change has affected the production set by imposing minimal behavioral assumptions. Values of a and b that satisfy WAPM indicate the existence of a closed. . yt+1 ) and (xt . yt ) to satisfy WAPM is sufficient. Technical change has occurred that render what would be profit maximizing choices using year t + 1 technology unachievable in year t. 1974). where: y∗ = y(y. . S∗ = {(x∗ . and x∗ = x(x. a). yi ). . or technological progress that has disproportionate benefits in the production of different outputs (Chavas and Cox. i = 1. Under this hypothesis. Call the i i vectors of effective quantities. (10) The vector valued functions y and x are one-to-one correspondences between each input and i i j j output.294 LAMBERT F¨ re et al. St and St+1 . It is often found in empirical work that WAPM is violated when comparing earlier and later time periods. (1989. where (xt+1 . 1995. forming the basis for F¨ re et al. 1994) operationalized the Caves et al. The augmentation hypothesis central to all of these studies has proven to be useful for modeling technical change.

. θ ≥ 0 We can similarly derive the distance function of (xt . where the t subscript t indicates the year of normalization for constructing the reference production set. . . . t = s. the observed netputs for year t and effective netputs for other years will form the vertices of the hull of S∗ . .PRODUCTIVITY MEASUREMENT FROM A REFERENCE TECHNOLOGY 295 By normalizing the augments for a particular year. yt )1/2 (15) Calculation of Netput Augmentation Parameters Construction of a WAPM consistent set of effective inputs and outputs requires finding solution values of a and b to pt [y(yt . bs )] ≥ 0 for all t. yt ) from the inner bound SI∗ by t+1 reversing the order of t and t + 1 in problem (12). yt ) on the surface of SO∗ . . . This deflation factor φ can be found in a similar fashion to the distance t function calculations under SI: (Do t (xt+1 . the inner bound measure of productivity change between t and t + 1 is then: mSI = [Do t (xt+1 . . at ) − y(ys . t The distance of a particular observation (xt+1 . Values of the input and output augments result from finding feasible . yt )]1/2 t SI∗ SI∗ (13) The output distance function with respect to the outer bound will measure the minimum proportional change in output necessary for situating a candidate observation. yt+1 )/ Do t+1 (xt . s = 1. The outer bound measure of productivity changes can then be derived: mSO = [Do t (xt+1 . T . yi ) for i = 1. yt+1 )/ Do t SO∗ SO∗ t+1 (xt . T (14) where (x∗i .. (xt . yt+1 )) = Max θ subject to θyt+1 ≤ T i=1 T i SI∗ (12) T i=1 λi y∗ i λi x∗ ≤ xt+1 λi ≤ 1 i=1 λi . From (7). yt+1 ) from the inner bound SI∗ under an t assumption of nonincreasing returns to scale can be found by solution of the following linear programming problem (F¨ re et al. . T can be calculated relative to both the inner and outer bounds on the frontier of S∗ . . as )] − rt [x(xt . . 1994): a Do t (xt+1 . yt+1 ))−1 = Max φ subject to pi (φyt+1 ) − ri xt+1 ≤ pi y∗i − ri x∗i t t SO∗ for all i = 1. bt ) − x(xs . The distance function value for observed bundles (xi . pi ) t t t are corresponding input and output prices. say t. y∗i ) is the i th effective netput relative to reference technology set S∗ and (ri .

1988). energy. fertilizer. and a miscellaneous category. .5 This formulation calculates linear deviations in the observed netput quantities. b−.t . Ten inputs were identified: hired labor. 4. then an infinite number of feasible solutions exist. subject to the t=1 effective netputs satisfying WAPM. feeds.6 The vector of augmentation factors (a+. at ) = yt · at and x∗ = xt · bt . a translog form of both the input and output aggregator functions was presumed. There is a o striking similarity between the two index number estimates of MFP growth and the outer bound measure. T − 1. 1990. p. pesticides. If a feasible solution exists for the system. Agriculture The quantity and price indexes compiled by Capalbo and Vo (1988) have been used extensively in analyzing U.S. 1981). . self-employed labor. The objective function of the linear programming model is to find augmentation factors that minimize the absolute deviation of proportional changes between observed m n + − + − and effective netputs: min T ( i=1 yit (ait + ait ) + i=1 xit (bit + bit )). t t so that y∗ = y(yt . their approach did not involve first deriving effective netputs to represent a common reference technology across all observations.t and bt = 1 − b+. We use the scaling hypothesis in augmenting observed netputs (Pollak and Wales. The close correspondence between the Fisher and T¨ rnquist indexes arises from the o aggregation procedures used to develop the quantity and price indices in Capalbo and Vo (1988).t + b−. Even though the data used in the estimation were disaggregated into 6 outputs and 10 inputs. Thus. to each of the T (T − 1) rows results in a system of linear equalities. Lambert and Shonkwiler. 1988. Chavas and . field crops. Results from the calculation of the inner and outer bound representations of technology are compared with the T¨ rnquist and Fisher productivity indexes in figure 1 and tables 1 and . . structures. . other capital equipment. The T¨ rnquist index is exact for the translog o . 1995).296 LAMBERT solutions to T (T − 1) linear .S. vegetables. Their use is adopted here to permit comparisons with other productivity estimates derived from the same data set. It appears that first deriving a reference technology for year t for t = 2. Addition of a nonnegative slack variable. Further output aggregation is inappropriate for this data set given the inconsistency with Varian’s nonparametric weak separability test (Chavas and Cox. Six major output categories were considered: small grains. and animal products. Reference technology sets are constructed for each of the T−1 years in the dataset. some aggregation was still necessary to derive the original 16 general netput categories. Multioutput productivity measures are thus ideally suited to avoid aggregation bias. An Analysis of MFP in U. . ) is fixed to be a 2(m + n) × 1 vector of 1’s where the subscript to refers to the reference year for the constructed technology. fruits. This similarity is much closer than the dual model productivity estimates reported in Chavas and Cox (1994).t + a−. 125)). may yield productivity estimates closer to index number estimates than the Chavas and Cox approach. coarse grains. agricultural productivity (Capalbo.4 The problem is to append an objective function to choose the set of netput augments that satisfy a criterion imposed by the researcher. a−. However. where at = 1 − a+. land. δts . The T¨ rnquist approximation of the Divisia indexing procedure was used for o aggregation (Capalbo and Vo (1988.

such as the T¨ rnquist and the Fisher. have also o been noted. Diewert (1978) observed that superlative indexes (such as the Fisher and the T¨ rnquist) will be o similar numerically if the underlying technology is translog. the input based Malmquist productivity measure is equal to the Fisher productivity measure. including the translog. technology.PRODUCTIVITY MEASUREMENT FROM A REFERENCE TECHNOLOGY 297 Figure 1. Numerical similarities between distance function based productivity indexes. The Fisher index is superlative for a general class of arbitrary second order approximations to a flexible functional approximation. F¨ re and Grosskopf (1992) showed how. under constant returns to scale and a profit maximization. such as the Malmquist. Comparison of inner and outer bound nonparametric productivity measures with the Fisher and T¨ rnquist o index number estimates. and Divisia approximations. A parallel argument can be derived for the output based Malmquist .

1024 1.0590 1. US agricultural production.2180 1.0724 1.1293 1.5971 1.3278 1.2451 1. 1949–1983.2261 1.6286 1. However.4226 1.6407 1.1481 1.2292 1.2923 1.5998 1.2413 1.1672 1.2720 1.4005 1.1902 1.2590 1.1264 1..5250 1.3070 1. the input and output based Malmquist measures are identical under constant returns to scale (Caves et al.2613 1.1698 1.2452 1.0830 1.2177 1.4595 1.1544 1.1382 1.9933 1.4230 1.0042 1.1072 1.1615 1.4035 1.9769 1.2480 1.3428 1.0675 1.0487 1. Consequently.0862 1.4205 1.0900 1. .1288 1.1544 1.0042 1.0724 1.2047 1.2156 1.4304 1.1216 1.2431 1.1383 1.0159 1.3585 1.1796 1.2246 1. in which no scale assumptions are imposed. Small discrepancies between the outer bound productivity measure and the Fisher index arise from the estimation procedure for the outer bound.0848 1.1902 1.4570 1.1481 1.1610 1.0487 1.0849 1.2293 1.2012 1.0159 1.4871 Fisher 1 0.4600 1.1584 1.4604 T¨ rnquist o 1 0.5618 1.2479 1.6149 1.4948 1.0918 1.2459 1.9835 1.1383 1.1540 1.1070 1.2101 1.4734 1.1265 1.4468 1. 1982).3195 1.4059 1.4883 1.1677 1.1471 1.2555 1.6304 1.4220 1.1510 1.1481 1.6122 1.6433 1.298 Table 1.1472 1. Comparison of the output based inner and outer bound productivity indexes Fisher and T¨ rnquist index o number estimates.4974 1. the F¨ re and a Grosskopf results will hold for an output based Malmquist measure as well under constant returns to scale and profit maximization.1961 1.3058 1.1529 1.2925 1.4296 1.4063 1.4911 1.1235 1.4631 LAMBERT measure estimated here.4040 1.1198 1.4867 1.0826 1.9769 1.0830 1.5114 1.3072 1.0579 1.3752 1.2157 1.0919 1.2612 1.0558 1.2060 1.2046 1.1085 1.5021 1.1699 1.2579 1.4047 1 0. Inner Bound Outer Bound 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1 0.2178 1.3775 1.3280 1.

04% 1. The composite netput and output vectors in the inner bound problem may. the constraint t λt ≤ 1 was dropped).97% The inner bound measure of MFP closely follows the other three estimates.22% 1. Second. This appears to be the case in figure 2. The inner bound generally lies below the outer bound. the inner bound does exceed the outer bound MFP cumulative growth values in several years. However. The overall average annual rate of growth in MFP was 1.09% 0. differences between the input and output-based inner bound productivity measures were affected by the scale assumptions imposed in solution of the problems. in which the outer bound measures are illustrated under the two approaches. .67 percent using the input-based outer bound estimate as compared to the 1. The productivity measure resulting from the input-based distance functions exceeds the output-based measures in all years. under decreasing returns to scale.17% 1. For example.34% 1. unknown production possibility set.13% 1.99% 1. This discrepancy indicates departures from constant returns to scale in U. However.PRODUCTIVITY MEASUREMENT FROM A REFERENCE TECHNOLOGY 299 Table 2. If the direction were the same. however.5 percent reduction. standardization of all three measures to equal 1 in the first year of analysis can mask differences between the inner and outer bounds. Average annual growth rates in multifactor productivity (MFP) for selected years. yt ) interior to the reference technology. a 2. differ from (xt . o The inner bound estimates of MFP derived from the input and output-based approaches are nearly identical in all years.35% 1. (1982) show how. composite netput sets constructed on the frontier of each year’s reference technology set may not be the same for the inner and outer bound distance functions. yt ) to the frontier may be different for the two measures.24% 1. The constant returns to scale assumption of the T¨ rnquist index7 are thus somewhat questionable. if the inner bound is instead normalized to 0. yt ) under examination.29% 1.12% 1. consistent with the concepts of SI and SO bounding some true. Consequently.14% 1.e.06% 1.12% 1. Thus. the distance to the inner bound should be shorter than the distance to the outer bound for production sets (xt . This apparent discrepancy arises from two sources. infeasibilities in solving the input-based linear programming model required solving this problem under an assumption of constant returns to scale (i. The reverse will hold for production sets outside of the reference technology.11% 0. It is well known that the Malmquist productivity index will differ between input and output based distance functions. the input based productivity measure will be greater than the output based measure. the inner bound falls below the outer bound in all years. yt ) depending upon the vector of weights λ resulting from solution of the problem.975 in 1949. First. Inner Bound Outer Bound Fisher Index T¨ rnquist Index o 1950-1983 1950-1963 1964-1973 1974-1983 1.S.. agricultural production.22 percent rate of growth from the output-based measure. the direction of movement from the actual year’s netput vector (xt . Caves et al. Examination of the two linear programming models (12) and (14) show that the outer bound model results in a proportional change in all outputs and no change in the level of inputs for each netput vector (xt .

Under a behavioral assumption of profit maximization. Both the inner and outer bound measures closely track Divisia approximations of productivity changes. The procedure requires data on netput quantities and prices. the productivity measure derives from distance function measurements from a reference technology constructed from effective netputs. The Fisher measure has especially been shown to satisfy both axiomatic as well as . Conclusions The Malmquist based productivity measure developed in this paper is appropriate to calculate multifactor productivity for time series data. US agriculture. 1949–1983. The question remains concerning the rationale for determining the inner and outer bound productivity measures developed in this paper.300 LAMBERT Figure 2. 5. Comparison of output and input based distance function nonparametric estimates of MFP.

The inner and outer bound estimates do impose consistency with Varian’s weak axiom of profit maximization. A less restrictive model could be developed in which distance functions are calculated from cost or revenue function frontiers. reference technologies could be constructed subject to Varian’s (1984) weak axiom of cost minimization. no functional specifications need to be asserted for the nonparametric approach. 1985). ys ) and (xt . Both the index number and the linear programming productivity measures require the same data for their estimation. and negative monotonicity. yt ) to satisfy WAPM is sufficient for the observed production bundles to represent different technologies. and distance functions of observed netput bundles calculated relative to the inner and outer bounds on the cost function. Further. The nonparametric approach makes no requirement of the underlying technology other than convexity. . For example. Charnes et al. for example. (1994). S∗ . This technology can be completely characterized by projections t t represented by the distance functions upon the surfaces of SI∗ and SO∗ . Chavas and Cox t t (1995) have exploited this feature in deriving rudimentary information regarding netput quantity responses to price changes. be used to approximate changes over time in supply and demand responses to price. The reference technology is further t bounded by SI∗ and SO∗ . The derivation of T − 1 reference production sets could thus. Why then should the slightly greater mathematical complexity of the nonparametric methods be employed? First. Recall that the procedure establishes a reference technology set formed of effective netputs for each year. 1992). the nonparametric approach provides more information about the technology than the index number approaches.. projections onto the reference technology frontiers could measure changes in marginal products over time (see. Appendix 1 Under an assumption of profit maximization and efficient production. Similarly. The distance function approach is thus seen to provide productivity measures unencumbered from functional form or scale specifications. The nonparametric frontier model requires no functional specification for input and output aggregator functions. The T¨ rnquist is superlative for a translog technology. Any bias introduced by aggregation could thus be eliminated. directly observed input and output quantities and prices o could be used in the nonparametric approach. Such an assertion may be inappropriate in certain settings.PRODUCTIVITY MEASUREMENT FROM A REFERENCE TECHNOLOGY 301 economic justifications for its adoption (Diewert. The Fisher is more general and is o superlative for any arbitrary second order technology. This advantage of the nonparametric approach could be exploited in both the current time series approach as well as in the panel data models developed in F¨ re a et al. for example. Although the data used in this application were aggregated using a T¨ rnquist aggregator function. closedness. failure of (xs . Finally. changes in production characteristics over time can be derived from the construction of the reference technologies to provide further insight into the factors underlying estimated changes in MFP over time. These are much more general characteristics of the technology than those required for the index number measures.

1992). or WAPM is not satisfied.C. This also raises questions about the CRTS assumptions underlying the original aggregation functions used by Capalbo and Vo (1988) in compiling the dataset. D. Consider an alternative bundle (x j . 5. (1988).). there will be an infinite number of solutions to the equations (Noble and Daniel. M. M.. Capalbo.C. where the first term equals the number of positive and negative augmentation factors for all netputs and the second term equals the number of slacks for the T (T − 1) inequality constraints. Antle and S. Therefore. Varian refers to a production set consistent with WAPM as p-rationalizing the data. However. then S rationalizes the data. J. (1995). C. 4. “A Comparison of Econometric Models of U. S. page 92). y j ) such that pi y j − ri x j > pi yi − ri xi . Antle and S.: Resources for the Future. Agricultural Productivity: Measurment and Explanation. The number of unknowns under the scaling hypothesis is 2(n + m)(T − 1) + T (T − 1). however. The first year is lost in the procedure due to taking the geometric mean of year-ahead and year-back distance function measures. Bureau.302 LAMBERT Proof of sufficiency. S = S j . Capalbo (eds. J.” In J. 1996). and S. “The Measurement of Technical Change Biases with Many Factors of Production. y j ) cannot rationalize the data. and S. y) ∈ S. and solution of the augments result from assigning arbitrary weights to a linear combination of the input and output augments. Ax = b.S. if pi yi − ri xi ≥ pi y − ri x for all (x. Specifically. Since we are only concerned with profit maximizing behavior in this paper. (1974). M. such as the translating hypothesis where y∗ = t yt + at and x∗ = xt + bt .: Resources for the Future. the test approach to index numbers does not posit an underlying technology. 2. 3.” Journal of Agricultural Economics 46. (1988). H. Grosskopf.). The number of variables is thus greater than the rank of A. Agricultural Productivity and Aggregate Technology. which can be no greater than T (T −1). the set S j that includes (x j . as one reviewer has noted. The proof follows directly from Varian’s definition (1984. Notes 1. M. However. if there is one feasible solution. page 584) of a production set S that rationalizes a set of observed data. “A Comparison of Three Nonparametric Measures of Productivity a Growth in European and United States Agriculture. 309–326. Agricultural Productivity: Measurment and Explanation. M. F¨ re. Consequently. Although other augmentation procedures have been proposed. Consider the system of linear equalities. M. 7. Capalbo. always referring to production sets that p-rationalize the data.” In J. By Varian’s definition. 964–76. 6. M. where the p stands for profit. so that (x j . This is to differentiate the production set from those that c-rationalize the data with respect to cost minimization. t References Antle. Washington. Binswanger.. but instead proposes certain properties for price and quantity indexes that should be satisfied (Diewert. then the rank of the augmented matrix [A : b] is equal to the rank of A. Washington. P. We prefer the scaling hypothesis in which the proportional differences between observed and effective netputs are minimized. Capalbo (eds. . If a solution exists to the system. We are. rather than the arbitrarily weighted linear deviations from observed netputs resulting from the translating hypothesis. we do not emphasize the distinction. y j ) ∈ S.” American Economic Review 64. 1977. D. recent contributions have permitted varying levels of technical and allocative inefficiencies over time (Kumbhakar. “An Introduction to Recent Developments in Production Theory and Productivity Measurement. R.

Golany. “Index Numbers and Indifference Curves. M.. R. Chavas. 260–267. W. Zhang. Guilkey. Output. Chavas. F¨ re. (1992) “A Nonparametric Analysis of the Influence of Research on Agricultural Productivity. Rolf. and R. W.: Resources for the Future. 473–487. 583–591. Agricultural Productivity: Measurment and Explanation. M.-P. Sten.” Journal of Productivity Analysis 3.. S. Malmquist.. Chavas. B. 883–900. “The Economic Theory of Index Numbers and the Measurement of Input. “Decomposition of the Fisher Ideal Index of Productivity: A Non-Parametric Dual Analysis of US Airlines Data.S. “A Nonparametric Analysis of Agricultural Technology. Quantity Indexes. Multi-Output Production and Duality: Theory and Applications. . Lovell. S. L. C. 209–42. (1996). (1995). 2nd ed. “A Non-Parametric Analysis of Productivity: The Case of U. S. F¨ re.” American Journal of Agricultural Economics 74. R. and T. “Malmquist Productivity Indexes and Fisher Ideal Indexes. Cox. Technical Change and the Demand for Energy in the UK Economy. (1976). 120. (1978). “Foundations of Data Envelopment Analysis for Pareto-Koopmans Efficient Empirical Production Functions. and Productivity. “A Review of the Evidence on Agricultural Productivity and Aggregate Technology. Harvey. R. Diewert.” American Journal of Agricultural Economics 77.). Pollak. Seiford (eds.” Applied Economics 23.” Journal of Productivity Analysis 5. Y. Methodology and Applications.-P. D. “Capital and the Theory of Productivity Measurement. “Productivity Developments in Swedish Hospitals: A a Malmquist Output Index Approach. “Fisher Ideal Output. 1-21. Grosskopf.. Diewert. Norris. (1980).-P. “A Nonparametric Analysis of Productivity: The Case of US Agriculture. and T. W.. J. Erwin. 303–310. Marshall.” Journal of the Royal Statisitical Society Series A.” American Journal of Agricultural Economics 77. Cox. Rolf. (1977). L. W.” Journal of Econometrics 30. Applied Linear Algebra. E. Boston: a Kluwer Academic Publishers. T. Input. and Productivity Indexes Revisited. W. Noble. Rolf. T Vo. J. W. K. F¨ re.” In J. (1982). Antle and S. 1996 Farrell. (1996). M.” Southern Economic Journal 63. F¨ re. Cox. (1988). 15–45. and J.” Bulletin of Economic Research 46. “Superlative Index Numbers and Consistency in Aggregation. M.” Journal of Econometrics 4. Technical Progress. C. Subhash C.” The Economic Journal 106.” Trabajos de Estatistica 4.C. 158–160. A.” European Review of Agricultural Economics 17. and D. (1985). W. Christensen.-P. Shonkwiler.. J. Boston: Kluwere Academic Publishers. Grosskopf.” In A.. M. F¨ re. Daniel. and Kankana Mukherjee. Cox. “The Measurement of Productive Efficiency.-P. (1994). 91–107. (1995).” Econometrica 49. “Efficiency. (1994) “A Primal-Dual Approach to Nonparametric Productivity Analysis: The Case of U. C. 591–616. B. 80–92. “Inter-Fuel Substitution. and Z. “Exact and Superlative Index Numbers.. and T. 29–39. Lambert. Cooper. (1989). D. Chavas.. Englewood Cliffs. Diewert. and T. A. and J. Chambers. L. Charnes. “Factor Bias under Stochastic Technical Change. Price. Diewert..S. J. and S.. W.. 450–464. Capalbo (eds. Weyman-Jones. and Productivity Indexes: A a Synthesis. 66–83. Primont. Cox.” Econometrica 50. Cooper.. Seiford. “Productivity Growth. Sickles. A. Ray. Wales. and T. 578–90. (1988).” American Economic Review 84.). Lindgren and P. R.” Applied Economics 28. Roos. (1990). Chavas. Charnes. 1659–1678. L.” American Economic Review 80. “Malmquist Indices of Productivity Change in the UK Gas Industry before and after Privatization. W. Erwin. (1953). “A Parametric Approach to Efficiency Measurement Using a Flexible Profit Function. S. and P.” The Economic a Journal 102.” International Economic Review 24. (1957). and W. B. and J. “Demographic Variables in Demand Analysis. J. (1981).. Grosskopf. (1992).. 211–248.” American Journal of Agricultural Economics 70.PRODUCTIVITY MEASUREMENT FROM A REFERENCE TECHNOLOGY 303 Capalbo. and Japanese Manufacturing. Grosskopf. and a Efficiency Change in Industrialized Countries. and T. (1991). (1992). Diewert. (1995) “On Nonparametric Supply Response Analysis.” Econometrica 46. 449–464. A. and L. Kumbhakar. C. (1983). Chavas. Cox. (1994). Erwin. and T. M. Data Envelopment Analysis: Theory. David K. A. (1990). L. Lewin and I. (1996). 253–82. Agriculture. Subal C. General. J. W. Caves. 1393–1414. Rolf.-P. NJ: Prentice-Hall. J.. and S. and T. 1533–51. Inc. 359–373.” American Economic Review 70. “A Comparison of the Performance of Three Flexible Functional Forms. D. Erwin. L. W. L. 1077–86. Washington..

R. Hal R. Varian.” Econometrica 52. 312–320.304 LAMBERT Solow. 579–97. . (1957).” Review of Economics and Statistics 39. (1984). “The Nonparametric Approach to Production Analysis. W. “Technical Change and the Aggregate Production Function.