Frontier Functions: Stochastic Frontier Analysis (SFA) & Data Envelopment Analysis (DEA)

Sponsored by: The Martin School of Public Policy and Administration The Department of Economics The Research Office University of Kentucky

Production and cost functions

A researcher wishes to estimate a production function or a cost function. The object is to estimate not the average production or average cost, but the maximum possible production given a set of inputs or the minimum possible cost of a set of outputs. OLS regression estimates the mean of the dependent variable conditional on the explanatory variables; Quantile regression is based on a quantile (e.g. 10th,25th, median, 75th, 90th), not the maximum or minimum; The max or min cannot be detected directly and used to define the sample for selection bias analysis; Limited dependent variable models truncate the dependent variable into categories or limits but not the maximum or minimum.

Frontier functions: definition

None of those standard econometric models is the answer. The answer is frontier functions, econometric stochastic frontier analysis (SFA) or linear programming data envelopment analysis (DEA). Frontier functions estimate maxima or minima of a dependent variable given explanatory variables, usually to estimate production or cost functions. All frontier functions come from one paper, Aigner and Chu (1968).

Aigner and Chu (1968)

D.J. Aigner and S.F. Chu (AER 1968), On Estimating the Industry Production Function invented this area. A viable distinction between the average and frontier functions as predictors of capacityderives from a probability interpretation of alternative forecasts.the frontier we construct is truly a surface of maximum points. This became Stochastic Frontier Analysis, Stochastic = probability interpretation. Estimation, for primary metals production in state aggregates: one stage least squares and two stage least squares, quadratic programming (now rarely estimated), and linear programming, developed into Data Envelopment Analysis in Charnes, Cooper, and Rhodes (1978) and subsequent research.

Varian (1984)
Varian shows how to estimate and test for the Weak Axiom of Cost Minimization (WACM) and other microeconomic assumptions Varian suggests using either regression (SFA) or linear programming (DEA) The WACM applies to for-profit, not-for-profit, private, and public producers The only requirement is that minimum inputs are intended to be used to produce desired output, or maximum output is intended from inputs used Profit maximization is not required

SFA and DEA

Two large differences and another possible difference SFA has a stochastic frontier with a probability distribution DEA has a non-stochastic frontier SFA has one output, or an a priori weighted average of multiple outputs DEA often has more than one output, no a priori weights, but assumes input-output separability Both can have stochastic inefficiency, SFA always does, DEA sometimes does

One-sided disturbances
In frontier functions, the disturbance has a distribution all on one side of zero the maximum production must be greater than or equal to any value in the sample, the minimum cost must be less than or equal to any value in the sample. produced quantities are bounded by the maximum, with non-positive disturbances costs are bounded by the minimum, with nonnegative disturbances

MLE with a one-sided disturbance does not work well

MLE and the Cramr-Rao lower bound (minimum variance of an asymptotically unbiased estimator, usually the MLE) are questionable! Begin with a likelihood function L which shows the probability of the data x given the parameters , The parameters might be the mean and standard deviation or might just be mathematical parameters.

Setting up MLE
Limits are a function of parameters in a nonstochastic frontier function: production function (max), cost function (min) L is Likelihood, L* is log likelihood. L() is always a probability distribution, so it follows that it integrates to 1.0 over the range of the data, from lower bound A to upper bound Z. AZ L(x | )dx = 1. Take the derivative wrt :

MLE: problems
AZ[dL(x | )/d] dx + [dZ/d]L(Z)[dA/d]L(A) = 0 E(dL*/d) + [dZ/d]L(Z) [dA/d]L(A) = 0. The first derivatives of the log likelihood do not have mean 0 if those extra terms stay. Second derivatives add more unwanted derivatives if the limits are functions of the parameters. The negative inverse Hessian is not the variance of the MLE. This is not working at all.

MLE: possible repairs

Make the frontier stochastic and limits of production or cost not a function of the parameters, completely eliminating the problem. Make the probability distribution have pdf of 0 and derivatives of 0 at the limits, even though the limit itself is a function of the parameters: [dZ/d]L(Z) = 0 and [dA/d]L(A) = 0 The Gamma Distribution can do that (Greene (1980)).

The Gamma Distribution

The Gamma Distribution describes a nonnegative random variable with two parameters (shape) and (spread) If ~ (, ), E() = /, V() = /2 pdf() = exp(-)-1/(), with different shapes for > 0, in ranges: less than 1, 1, between 1 and 2, 2, greater than 2 A graph follows; > 2 is required for the pdf and its derivatives to be zero at the limits.

Gamma Distribution: shapes

Gamma distributions (lambda =1)
1 0
0

.2

.4

.6

.8

2 x alpha=1.0 alpha=2.0

3 alpha=1.5 alpha=2.5

Ok, so a Gamma Distribution?

No, not really. The parameters are restricted mathematically. That really annoys researchers. Some other distribution? No, no other onesided distribution has the required properties at the limits. This is why no one has just one disturbance .

Composite disturbances
The disturbance has two parts Stochastic frontier (v), unlimited range as usual. The limits of the production or cost function are at infinity, not a function of the parameters Inefficiency (u), one sided, non-positive for production, non-negative for cost Finally, yj = xj + uj + vj , that is, j = uj + vj So there are two disturbance terms to keep the parameters from affecting the limits

Panel data: Fixed effects

Panel data researchers would like to include fixed or random effects in everything, so why not frontier models? Greene (2005) addresses this in detail. Fixed effects have special problems in non-linear models, but they can work Random effects are offered by Stata. Now there are three disturbance terms! yjt = xjt + j + ujt + vjt

Fixed effects in non-linear models

Fixed effects have well known advantages in linear models but in non-linear models they: are inconsistent (too small sample for each fixed effect), cannot be differenced out (differences of non-linear models are still non-linear), spread their inconsistency to other coefficients (assuming correlation with other explanatory variables, which is the motivation for fixed rather than random effects).

Wait, maybe fixed effects are ok

With few units and many observations, fixed effects work because the sample size for each fixed effect might be large enough. Greene (2005) points this out. Stata refuses to enter fixed effects in the model. The user can enter fixed effects. Random effects, normally distributed, are offered by Stata. As always, they must be assumed to be uncorrelated with explanatory variables. The independence assumption cannot be tested by Stata, and there is no Hausman test, but Estimate fixed effects by direct inclusion and regress the fixed effects on explanatory variables to test the independence required for consistent random effects.

Stata: all MLE, all the time

Stata offers MLE with composite disturbances. The one-sided distribution is half-normal, truncated normal, or exponential (restricted Gamma) frontier dependent explanatory, d(hn) or d(tn) or d(e) In Stata, u is one-sided inefficiency and v is the twosided stochastic frontier. Stata uses notation from Greene (1990) in which = ratio of standard deviations u/v, so that = 0 means there is no inefficiency. Fixed effects sneaked in by the user under frontier, or random effects by Stata (normally distributed). xtfrontier dependent explanatory, re i(group_id) For minimization, use the option , cost

Stata: heteroscedasticity
Stata offers a lot of heteroscedasticity: either u or v can be heteroscedastic, or both. Heteroscedastic u (one-sided error, inefficiency) Heteroscedastic v (two-sided error, random variation) The same explanatory variables, or different variables, can appear in the frontier and in the heteroscedasticity. frontier, uhet(var_name) vhet(var_name)

Stata estimates the inefficiency

Stata estimates the technical efficiency, the percentage of estimated frontier output attained or the extra percentage spent beyond frontier cost predict var_name, te As usual, many other options exist using predict. Successful Stata estimation is illustrated at this point.

Is MLE necessary?
If you always use Statas options, yes! If not, no! Not-MLE (1) Corrected OLS Not-MLE (2) Fixed effects in panels Not-MLE (3) Gamma-distributed inefficiency Note: the Gamma distribution or any other distribution of inefficiency is unrestricted if MLE is not used; only MLE has a range problem

Not MLE(1) Corrected OLS

Estimate OLSthats all, just OLS yj = xj + j Estimate residuals ej and interpret them as inefficiency Assuming production, inefficiency<0, most efficient = max(e1, e2, en) = emax Inefficiency of unit j = emax - ej Substitute min and ej - emin for a cost function

Not MLE (2) Fixed effects as inefficiency

Schmidt and Sickles (1984) but not in Statafixed effects required! Given panel data and fixed effects, assume that inefficiency is the fixed effect Estimate yjt = xjt + j + vjt by xtreg predict the fixed effects j and define the most efficient (production) max Inefficiency = max - j Min and reverse sign for cost functions

Not MLE (3) Gamma-distributed inefficiency

Greene (1990), not in Stata Not a panel, yj = xj + (uj + vj) by reg and predict the residuals ej = uj + vj Adjust residuals to one side of 0 by the max or min; the constant absorbs emax/min Assume v ~ (0, v2) and u ~ a Gamma distribution and estimate v2, , and E(e) isnt useful, fixed to 0 by OLS but

Not MLE (3) Gamma-distributed inefficiency

V(e) = v2 + /2 Skewness(e) = 2/3 Kurtosis(e) = V2(e) + 6/4 Three equations in three unknowns: V(v), two parameters of the distribution of u Standard errors by delta method or GMM But the range of the data is a function of the parameters? No problem, not MLE!

Failure of well-specified MLE: parameters

Failure to converge; estimation continues indefinitely through many iterations with no sign of stropping. Repeated non-concave loglikelihood means the log likelihood is not maximized, maximum likelihood fails; backed up means the loglikelihood decreases. Estimation fails to start, initial values not feasible. OLS starting values imply negative infinite log likelihood. Apparent estimates but the SD of inefficiency (u) is small or the ratio of u to the SD of the stochastic frontier (v), =u/v, is small, e.g. .01; sometimes goes as close to zero as Stata can make it, e.g. 0.00001.

Failure of well-specified MLE: distributions

The truncated normal distribution of inefficiency has an extra parameter, the mean of the normal truncated at 0, which often fails in estimation. The exponential distribution slopes down from 0 smoothly, which leads to initial values not feasible if inefficiency is not strongly skewed right. The stochastic frontier can disappear from the model, leaving one-sided inefficiency that violates the MLE range rule (range not a function of the parameters). The half-normal is the most often successful, the most common in the literature, and the default in Stata. Unsuccessful Stata estimation is illustrated at this point.

Data Envelopment Analysis (DEA)

Envelop the m inputs and n outputs in m+n space, i.e. a graph with points, with hyperplanes, i.e. lines/planes/etc. Linear programming Constant returns to scale (CRS) = CCR for Charnes, Cooper, and Rhodes (1978), Variable returns to scale (VRS) = BCC for Banker, Charnes, and Cooper (1984). Aigner and Chu (1968) did it first and also did quadratic programming

DEA assumptions
DMU = decision making unit, business, bank, farm, not-for-profit, government, university, etc. All actual observed inputs and outputs of any DMUs are feasible for all DMUs All linear combinations of observed inputs and outputs are feasible. Free disposal of inputs and outputs. The production function or cost function is piecewise linear, implying linear or nondifferentiable functions everywhere.

DEA efficiency without prices

Output-oriented technical efficiency is producing the greatest possible output in the sense of a linear function of a set of outputs given the value of a linear function of inputs. No prices are involved. Efficiency = output that could be produced from inputs used, if >100%, inefficient. Input-oriented technical efficiency is producing a given set of outputs with the smallest linear function of inputs. No prices are involved. Efficiency = percentage of actual inputs used that would be needed, if <100%, inefficient. Constant returns to scale: output and input orientation are the same. Variable returns to scale: output and input orientation are different.

DEA efficiency with prices

Allocative efficiency is minimizing the cost of the linear combination of the outputs produced, using input prices. Profit maximization: maximizing the value of outputs minus the value of inputs, using both output and input prices Scale efficiency is operating at the scale of operation maximizing the ratio of the linear sum of outputs to the linear sum of inputs. An economically efficient business is technically and scale efficient.

DEA: tiny example constant returns to scale

DMU x y y/x efficiency supereffic 1 1 6 6.00 1.0000 1.0909 2 2 8 4.00 0.6667 3 2 11 5.50 0.9167 4 3 9 3.00 0.5000 5 3 13 4.33 0.7222 6 5 15 3.00 0.5000 DMU#1 has the highest y/x and others are inefficient according to their ratios of y/x DMU#1 could drop to 5.5 and still be efficient (see DMU#3), DMU#1s superefficiency is 6.00/5.50 = 1.0909

DEA: tiny example variable returns to scale

DMU x y y/x efficiency supereffic 1 1 6 6.00 1.0000 2.0000 2 2 8 4.00 0.7000 3 2 11 5.50 1.0000 1.2143 4 3 9 3.00 0.5333 5 3 13 4.33 1.0000 1.1667 6 5 15 3.00 1.0000 big DMUs#1,3,5,6 define the frontier DMU#2 is inefficient relative to 0.6 X #1 + 0.4 X #3 DMU#4 is inefficient relative to 0.4 X #1 + 0.6 X #3 DMU#1 could use twice the input and still be efficient

DEA graph
Output-oriented DEA: input x, output y
30 0 10 20

2 x crs y

3 vrs

DEA: standard setup

N decision making units (DMU). Assume a linear function of n inputs produces m outputs. There is no economic production function or cost function in basic DEA. Assume the linear function of the inputs is minimized given the linear function of the outputs, Equivalent: the linear function of the outputs is maximized given the linear function of the inputs. Call inputs x and outputs y as in regression. Call the coefficients on inputs b and the coefficients on outputs c. These are shadow prices in economics.

DEA: linear programming, input oriented (production)

Consider DMU t, 1tN, N total producers to study, with m outputs and n inputs. DEA estimates each DMUs efficiency by itself, not relative to one estimated frontier. Each DMU t has an individual input and output function. Max, over c.t and b.t, i=1mcityit/ j=1nbjtxjt s.t. bjt 0 and i=1mcityip/j=1nbjtxjp 1, all DMUs p. Linear fractional programming is difficult, maximizing the ratio of two linear functions; restate to maximize the numerator minus the denominator, which is a linear program.

DEA: avoiding linear fractional programming

Max i=1mcityit - j=1nbjtxjt s.t. bjt 0, all j, and j=1nbjtxjt = 1, a normalization of total cost, and i=1mcityip - j=1nbjtxjp 0, all DMUs p. Note on math: given real z, functions f(z), g(z) all >0; substituting max f(z)-g(z) for max f(z)/g(z) implies that f(z) and g(z) are near 1.0, so that ln(f(z)) and ln(g(z)) are approximately linear. Setting total costs = 1.0 is a normalization but setting total output (1) near 1.0 is an assumption that inefficiency is not too large. Linearizing overstates large inefficiencies. No standard errors, no statistical tests.

DEA including prices

Basic DEA has no economic production or cost function, but see Ray (2004, Chapter 9), linear programming, with additional constraints. Add constraints to the production or cost (linear) function using the market prices. Maximize output given inputs but add the linear constraint on inputs that cost adds up to a total variable cost budget. An explicit production function can be added as a constraint. DEA for profit maximization explicitly maximizes the total revenue from outputs minus the variable cost of inputs as a linear function.

DEA: what management consultants do

Rank DMUs by efficiency Benchmark to efficient units Estimate superefficiency Use the coefficients to suggest alterations in resource allocations. Assumption: the production function that applies to a particular DMU (farm, hospital, or university, e.g.) can be expanded or contracted linearly. DMUs with unusual combinations of inputs can appear efficient but be very difficult to emulate.

DEA: attempted standard errors

Interpretation as MLE on efficiency: estimate a probability distribution of estimated efficiencies. The frontier is still non-stochastic; the probability distribution is descriptive and post-estimation; this is not MLE. Chance-constrained linear programming: add a disturbance (maybe Gamma) to the non-stochastic frontier. The frontier is still non-stochastic; not MLE. Bootstrap variances: random sampling variation in estimated efficiencies does not represent behavior of DMUs or the observed frontier in the actual data. No method provides econometric standard errors, the reason many econometricians just say no.

Comparing DEA and SFA

Comparing SFA to DEA has not been done very much Some work on hospitals The correlation of efficiency estimates is not very high: 0.13-0.63 in hospitals, apparently similar elsewhere DEA focuses on individual DMUs, while SFA focuses on estimating the frontier.

Research on frontier functions SFA and DEA results

What systematic factors are associated with failure of SFA models: topic (banks, farms, hospitals, states, etc.), distribution (exponential, half-normal, truncated normal, gamma), explanatory variables, sample size, etc.? What systematic factors are associated with SFA and DEA results being similar or different?

Research on frontier functions: methodology

No theoretical reason to avoid the Gamma distribution, so use it in research and compare results. Apply SFA based on moments and compare with MLE. Quadratic programming (minimizing the sum of squared inefficiency terms) in DEA was difficult decades ago, but today? The method of Wolfe (1959) can be used. Aigner and Chu (1968) estimated quadratic programming and had apparently different estimates (with no standard errors) of capital-output elasticity and technology-output elasticity. Fractional linear programming also might be feasible in DEA given modern computing resources.

Go estimate frontier functions

Economics and policy are often concerned with efficiency of banks, farms, governments, private and public agencies, for-profit and not-for-profit producers. The weak axiom of cost minimization is reasonable; profit maximization is not required. Statas frontier and xtfrontier are available and Statas restrictions can be evaded. DEA is used by management consultants, estimated by general and specific linear programming packages. Comparative or methodological research is possible.