sfa

Stochastic frontier analysis in R. A single, consistent interface to a wide range of cross-sectional, panel, latent-class, zero-inefficiency, two-tier, sample-selection, endogenous-regressor, copula and nonparametric stochastic frontier models, with a common formula syntax for modelling the variance of each error component and a common "sfareg" result object that works with the standard R modelling generics.

Beyond fitting, it provides the tools to choose among the fifteen cross-sectional inefficiency distributions rather than assume one, and to check whether the chosen specification is defensible -- see Model selection and diagnostics.

Written by David H. Bernstein, Christopher F. Parmeter and Alexander D. Stead.

Installation

The released version from CRAN:

install.packages("sfa")

The development version from GitHub:

# install.packages("remotes")
remotes::install_github("davidhbernstein/sfa")

The development version here is ahead of CRAN; see NEWS.md for what has changed, including one deliberate breaking change to psfm(model_name = "TFE").

Quick start

library(sfa)

## Simulate a cross-section with known true parameters
cs <- data_gen_cs(N = 1000, rand = 1, sig_u = 0.3, sig_v = 0.3,
                  cons = 0.5, beta1 = 0.5, beta2 = 0.5, a = 4, mu = 1)

## Normal-half-normal frontier
fit <- sfm(y_pcs ~ x1 + x2, model_name = "NHN", data = cs)

summary(fit)
coef(fit)              # lambda, sigma, (Intercept), x1, x2 -- see note below
logLik(fit)
head(fit$exp_u_hat)    # technical efficiency, E[exp(-u) | e]
head(fit$med_u_hat)    # median efficiency prediction (NHN only)

Which efficiency predictors come back depends on the model: exp_u_hat (Battese and Coelli 1988) is returned by most of sfm()'s models, and the Jondrow et al. (1982) point predictor u_hat = E[u | e] by NE, NTN, NU, NGE, NLN, NW, THT and tHN. See ?sfm for the full list.

A panel model, with a determinant of the inefficiency variance:

pd <- data_gen_p(t = 10, N = 100, rand = 100, sig_u = 1, sig_v = 0.3,
                 sig_r = 0.2, sig_h = 0.4, cons = 0.5, beta1 = 0.5, beta2 = 0.5)

fit_p <- psfm(y_tre_z ~ x1 + x2 | z_gtre, model_name = "TRE_Z",
              data = pd, individual = "name")

Which model should I use?

Start from what your data and question look like:

Situation	Use
A cross-section, the standard case	sfm() with "NHN" or "NE"; then check the distribution (next rows)
You are not sure which inefficiency distribution fits	spec_test_all() and moment_range() on OLS residuals before fitting; vuong() or TIC() to compare fits; sfma() to average over them
Residuals are skewed the "wrong" way, so sigma_u collapses to 0	skewness_test() and inefficiency_test() first; then esfm(), sfm(estimator = "acols" | "cmle"), or copsfm(vdist = "glogistic") -- see ?sfa-wrongskew
Inefficiency depends on firm characteristics	sfm(y ~ x | z, model_name = "NHN_Z") (or "NE_Z"); muhet with "NTN" for Battese and Coelli (1995)
The noise variance depends on covariates	vhet = ~ z_v in sfm()
Some firms are fully efficient	zsfm()
Firms use different technologies	lcsfm()
Deviations are two-sided (e.g. bargaining)	ttsfm()
The sample is selected	selsfm()
A regressor is endogenous	ivsfm()
Noise and inefficiency may be dependent	copsfm(), preferring copula = "frank" or "clayton"
You do not want to assume a functional form	npsfm()
A panel, separating firm effects from inefficiency	psfm() with "TRE" (random effects) or "TFE" (fixed effects, needs long T)
A panel with persistent and transient inefficiency	psfm() with "GTRE" / "GTRE_Z"
A panel, without a distribution for inefficiency	psfm() with , , , , or

The nine entry points

Function	Fits	Estimators
sfm()	Cross-sectional frontiers	15
psfm()	Panel frontiers	21
lcsfm()	Latent-class frontiers	3
zsfm()	Zero-inefficiency frontiers	2
ttsfm()	Two-tier frontiers	3
selsfm()	Sample-selection frontiers	2
ivsfm()	Frontiers with endogenous regressors	3
copsfm()	Dependence between the error components, and skewed noise	6 copula families (15 with rotations) x 2 u x 3 v
npsfm()	Nonparametric frontiers	5

All but npsfm() return an object of class "sfareg". npsfm() returns "npsfareg" instead — a kernel-estimated frontier has no parameter vector with standard errors, so coef(), vcov() and logLik() would have nothing to return.

selsfm() and ivsfm() do not take their equations through the | pipes: selsfm() takes selection and frontier as separate formulas, and ivsfm() takes formula, endogenous and instruments. Both reject a | segment, because a pipe already means "variance determinant" everywhere else and reusing it would give one character two meanings.

sfm() — cross-sectional

model_name	Distribution of u
NHN, NHN_Z	half-normal (_Z: with variance determinants)
NE, NE_Z	exponential (_Z: with variance determinants)
NTN	truncated normal
NR	Rayleigh
NU	uniform
NG	gamma
NNAK	Nakagami
NGE	generalized exponential
NLN	lognormal
NW	Weibull
tHN	half-normal, with Student-t noise
THT	half t, with Student-t noise
TSL	truncated skew-Laplace
NGB2	generalized beta of the second kind — nests NHN, NE, NG and NW as limits, and the half-t exactly
NB	binomial (Carree 2002) — estimator = "cols" only; the one u here that can be skewed either way

sfm() also offers estimator = "cols" — corrected OLS (Olson, Schmidt and Waldman 1980), closed-form and deterministic, for NHN, NE, NG and NB — and robust divergence-based alternatives to MLE via robust = "mlqe" | "psi" | "mdpd" for NHN.

The tuning parameter of those robust criteria does not have to be guessed: hscore_select() chooses it by minimising the Hyvarinen score, calibrate_c() gives a fixed weight-matching alternative, and density_weights() shows what the estimator did to each observation. influence_sfa() reports the influence function of any fit — which observations move it, and whether the specification lets any single one of them move it without bound.

fit <- sfm(y ~ x1 + x2, data = d, model_name = "NHN")
sel <- hscore_select(fit, method = "mlqe")          # data-driven
calibrate_c(sigma_v = 0.3, sigma_u = 0.6)           # or fixed
sfm(y ~ x1 + x2, data = d, robust = "mlqe", c_mlqe = sel$c)

Any of NHN, NE and NTN can additionally take covariates in more than one error component, as named formulas rather than further pipe segments:

sfm(y ~ x1 + x2 | z_u, vhet = ~ z_v, model_name = "NHN_Z")   # heteroskedastic v
sfm(y ~ x1 + x2, muhet = ~ z_mu, model_name = "NTN")         # Battese–Coelli (1995)

vhet drives the noise scale, uhet the inefficiency scale (the same thing the | z segment does), and muhet the pre-truncation mean.

psfm() — panel

model_name	Estimator
TRE, TRE_Z	true random effects (Greene 2005)
GTRE, GTRE_Z	generalized true random effects, four-component
GTRE_FML	GTRE by full maximum likelihood
GTRE_SEQ1, GTRE_SEQ2	sequential/moment-based GTRE
TFE	true fixed effects (Greene 2005)
TFE_WMLE	within MLE (Chen, Schmidt and Wang 2014)
FD	first differences
SSFE	Schmidt and Sickles (1984) fixed effects (within)
SSRE, SSCRE	Schmidt–Sickles random effects, and correlated random effects (Mundlak 1978)
CSS	Cornwell, Schmidt and Sickles (1990), firm-specific quadratic in time
LS	Lee and Schmidt (1993), one common temporal pattern scaled per firm
KSS	Kneip, Sickles and Song (2012), data-driven temporal basis
PL80	Pitt and Lee (1980), time-invariant
PL80_MVTN	Pitt and Lee (1981) Model III: inefficiency varies over time and is correlated within a firm, through a multivariate normal truncated to the negative orthant. Balanced panels only
BC92	Battese and Coelli (1992) time decay
K1990, K1990modified	Kumbhakar (1990) time patterns

The last five are one family: each writes the firm effect as alpha_it = sum_r theta_ir * g_r(t) and reads inefficiency off it as distance from the best firm, assuming no distribution for inefficiency at all. They differ only in how much of that structure is assumed rather than estimated — SSFE fixes L = 1 with a constant basis, LS frees the basis, CSS fixes L = 3 to {1, t, t^2}, and KSS estimates both. KSS needs a balanced panel; the others do not.

GTRE_SEQ1, GTRE_SEQ2, SSFE, SSRE, SSCRE, CSS, LS and KSS are not maximum likelihood, so logLik() (and hence AIC()/BIC()) returns NA for them.

psfm_bootstrap() provides a parametric bootstrap for GTRE-family fits, parallelised over cores.

zsfm() — zero inefficiency

ZISF and ZISF_Z: a mixture of a fully efficient regime and an inefficient frontier regime, with the regime probability optionally parameterised by covariates (ZISF_Z).

lcsfm() — latent class

LCM and LCM_Z: the latent class frontier (Greene 2005; Orea and Kumbhakar 2004) — n_class unobserved technologies, each with its own frontier and its own two scales, mixed by a multinomial logit whose covariates are optional (LCM_Z). Returns posterior class probabilities and posterior-weighted efficiency alongside the class-conditional predictions.

ttsfm() — two tier

TTNE (normal–exponential–exponential), TTHN (normal–half-normal–half-normal), and TTNLS (nonlinear least squares, no distributional assumption beyond the means of the two one-sided components).

selsfm() — sample selection

Two estimators, for two different reasons a selected sample misleads. Both take their two equations as separate arguments rather than through pipes:

model_name	Estimator
"greene" (default)	Greene (2010), two-step: probit on the full sample, then simulated ML on the selected subsample
"kts"	Kumbhakar, Tsionas and Sipilainen (2009), single-step ML

selsfm(selection = participate ~ z1 + z2,
       frontier  = y ~ x1 + x2, data = d)                      # Greene

selsfm(selection = adopt ~ z1 + z2,
       frontier  = y ~ x1 + x2, data = d, model_name = "kts")  # KTS

They are not variants of one model. In Greene the units in the sample are there for reasons correlated with the noise, so estimating on the selected sample alone is biased; the correction is a selectivity term, and because the selection equation is a probit on an observed binary outcome it can be estimated first and the frontier second.

In KTS there are two technologies, both observed, each with its own frontier and its own pair of scales, and the choice between them depends on inefficiency itself. That forces a single step: the choice equation cannot be a probit when the u entering it is unobserved, so there is no first stage to run. Use "greene" when one regime is unobserved and selection is on the noise; use "kts" when both regimes are observed and the switch is on efficiency.

ivsfm() — endogenous regressors

Amsler, Prokhorov and Schmidt (2016): one or more regressors correlated with the statistical noise. Three estimators of the same model, not three models — IVLIML (full-information maximum likelihood), IVCF (two-step control function) and C2SLS (corrected 2SLS):

ivsfm(y ~ x1 + x2, endogenous = ~ x2, instruments = ~ w1 + w2,
      data = d, model_name = "IVLIML")

With uhet the model becomes that of Amsler, Prokhorov and Schmidt (2017), in which the environmental variables entering the inefficiency scale may themselves be endogenous.

endogeneity_test() asks whether the correction was needed at all — a Wald test of H0: rho = 0, i.e. that the noise is uncorrelated with the reduced-form errors and a plain sfm() fit would already have been consistent.

fit <- ivsfm(y ~ x1 + x2, endogenous = ~ x2, instruments = ~ w1 + w2,
             data = d, model_name = "IVLIML")
endogeneity_test(fit)

copsfm() — dependence between v and u

Drops the independence assumption between the noise and inefficiency components, coupling them with a copula and integrating the resulting density by Gauss–Legendre quadrature (n_nodes). It also drops the assumption that the noise is symmetric: vdist offers a generalized logistic carrying its own skewness parameter (Bonanno and Domma 2022), which is where a positive residual skew can go other than into sigma_u = 0.

copsfm(y ~ x1 + x2, data = d, copula = "frank")

# skewed noise and an exponential inefficiency, with and without dependence
copsfm(y ~ x1 + x2, data = d, copula = "independent",
       udist = "exponential", vdist = "glogistic")
copsfm(y ~ x1 + x2, data = d, copula = "fgm",
       udist = "exponential", vdist = "glogistic")

argument	choices
udist	"hnormal", "exponential"
vdist	"normal", "logistic", "glogistic" (the last two are the generalized logistic, "logistic" fixing its shape at 1)
copula	the fifteen families below, plus "independent", which estimates no dependence parameter at all

copula	parameter	independence at	dependence it can express
"gaussian"	rho in (−1, 1)	0	both signs, no tail dependence
"fgm"	theta in [−1, 1]	0	both signs, but weak — Spearman rho = theta/3
"frank"	theta real	0	both signs, full range, no tail dependence
"clayton"	theta > 0	0	positive, lower-tail dependence
"gumbel"	theta >= 1	1	positive, upper-tail dependence
"joe"	theta >= 1	1	positive, heavier upper tail than Gumbel

Clayton, Gumbel and Joe carry only positive dependence. Nothing rules out a negative association between noise and inefficiency, so each also has 90 and 270 rotations that reverse the sign ("clayton270"), and 180, the survival copula, which preserves it.

Which of these actually work, measured rather than assumed. Every density here is verified three ways — it equals the second mixed partial of its own CDF, it integrates to 1 over the unit square, and it is exactly 1 at the independence parameter. That establishes the densities are right. It does not establish that the dependence parameter is recoverable, and for most of them it is not. Fitting each family to 25 samples generated from itself at n = 400:

family	recovers its own theta?	collapsed to the independence bound
frank	yes (5.43 against a truth of 5)	0%
clayton	yes (2.24 against 2)	0%
gaussian, fgm	yes	0%
gumbel	no	36%
joe	no	40%
clayton270	no	56%
gumbel90	no	60%

On data generated from a Gumbel copula with Spearman rho = 0.685 at n = 2000, every family — including the true one — returns the independence boundary, and their log-likelihoods differ by less than 0.04. The likelihood is flat in the dependence parameter. This is a property of the model, not of the code.

So: prefer "frank" or "clayton", and treat the others as exploratory. copsfm() warns when you select a family that did not recover, quoting its measured collapse rate; the rotations that were never measured warn that they were not, rather than implying either outcome. More generally, the dependence parameter is estimated imprecisely even when it is recoverable — on a Gaussian design at n = 600 its sampling standard deviation is 0.385 against a truth of 0.5 — so a comparison across families is descriptive rather than evidence for a dependence structure.

npsfm() — nonparametric

Estimates the frontier by kernel regression instead of assuming it linear.

method	Estimator
FLW	Fan, Li and Weersink (1996). Kernel regression for E[y|x], then the scale parameters from the residuals. Also supports dist = "exp", "gamma", "unif"
SVKZ	Simar, Van Keilegom and Zelenyuk (2017). Local method of moments; sigma_u(x) and sigma_v(x) vary with the covariates
PSZ (alias KPST)	Park, Simar and Zelenyuk. Local maximum likelihood
MY	Martins-Filho and Yao. Iterative local likelihood
SZ	Simar and Zelenyuk (2011). DEA monotonization of a prior smooth fit

f <- npsfm(y ~ x1 + x2, data = d, method = "FLW", dist = "hn")
head(fitted(f))       # the estimated frontier
head(f$exp_u_hat)     # technical efficiency

PSZ and MY run one numerical optimization per observation — for MY, per observation per iteration — so expect them to be one to two orders of magnitude slower than FLW. npsfm() takes a single-part formula and rejects a | z segment: its heteroskedasticity is nonparametric in the covariates themselves.

Kernel regression comes from np, and SZ's DEA step solves one linear program per unit with lpSolve. Both are in Suggests, not Imports, so they are only required if you actually call npsfm():

install.packages(c("np", "lpSolve"))

Formula syntax

Variance determinants are supplied in extra pipe-delimited segments:

y ~ x1 + x2 | z | zp

• the first segment is the frontier,
• the second parameterises the variance of the first one-sided component (sigma_u),
• the third parameterises a second one-sided component where the model has one (sigma_w in two-tier models, sigma_h in GTRE_Z).

Omitted segments default to 1, i.e. homoskedastic.

The link function differs by model family. sfm()'s NHN_Z/NE_Z and ttsfm()'s TTNE/TTHN use sigma = exp(z'delta), while psfm()'s GTRE_Z/TRE_Z use sigma = sqrt(exp(z'delta)) — that is, delta parameterises the variance rather than the standard deviation. Check which convention applies before interpreting a coefficient on z.

Model selection and diagnostics

The package offers seventeen cross-sectional inefficiency distributions. These are the tools for choosing among them, and for asking whether the choice is defensible at all.

Function	Question it answers
TIC(), vuong()	Which of two non-nested specifications fits better, without assuming either is correct
spec_test(), spec_test_all()	Is this pair of noise/inefficiency distributions defensible, from OLS residuals alone
moment_range()	Can this pair produce the residuals' skewness and kurtosis at all — the range check that precedes the test
sfma()	What if the data does not identify one — average over distributions instead of choosing
lcsfm_homogeneity()	Does a latent-class fit beat a single technology
gof_test()	Is the assumed inefficiency distribution right, holding normal noise fixed
cw_test()	The same question without ever forming the composed density, from OLS residuals
skewness_test(), inefficiency_test()	Is there evidence of inefficiency at all (the wrong-skew problem)
uhet_test()	Does inefficiency depend on firm characteristics (validly, from a two-step fit)
endogeneity_test()	Was the endogeneity correction needed — is rho different from zero
esfm(), symmetry_test()	Fit a frontier when the residual skewness has the wrong sign
sfm(estimator = "acols"), sfm(estimator = "cmle")	Two more answers to the same problem: corrected OLS without the third moment, and maximum likelihood under moment constraints. See ?sfa-wrongskew for the whole suite
skewness_decomp()	Where the composed error's skewness comes from — inefficiency, noise, or the dependence between them
tt_boundary_report()	Is a two-tier fit's collapsed sigma_v the maximum, or did the optimizer stop short
influence_sfa()	Which observations move the fit, and can any single one move it without bound

## The score-based diagnostics differentiate the fitted likelihood, so the fit
## has to have kept it: pass keep_objective = TRUE. TIC(), vuong() and
## influence_sfa() all need this; the others do not.
fit_hn <- sfm(y ~ x1 + x2, data = d, model_name = "NHN", keep_objective = TRUE)
fit_e  <- sfm(y ~ x1 + x2, data = d, model_name = "NE",  keep_objective = TRUE)
vuong(fit_hn, fit_e)                            # neither assumed correct
influence_sfa(fit_hn)                           # who is driving this fit

spec_test_all(residuals(lm(y ~ x1 + x2, d)))    # before fitting anything
moment_range(residuals(lm(y ~ x1 + x2, d)))    # which pairs are even possible
sfma(y ~ x1 + x2, data = d, models = c("NHN", "NE", "NTN"))

## Is the assumed inefficiency distribution itself defensible? Hold the
## normality of the noise fixed and it implies a distribution for the composed
## error, so testing that is testing the assumption on u.
gof_test(fit_hn, data = d, B = 199)             # KS and Pearson chi-square

## Is there any inefficiency to speak of? The one-sided LR test is the one to
## quote: the Wald ratio and the naive LR test both have the wrong size here,
## because the null sits on a boundary.
inefficiency_test(fit_hn)

spec_test(), gof_test(), lcsfm_homogeneity() and sfma() default to a bootstrap null rather than the published asymptotic one. Their help pages give the measured size distortions behind that choice: in two cases the asymptotic limit is badly mis-sized for the models this package fits, because it is stated for a restricted specification the package does not impose.

Working with results

"sfareg" objects support the usual generics:

coef(fit); vcov(fit); logLik(fit); nobs(fit); AIC(fit); BIC(fit)
fitted(fit); residuals(fit); predict(fit, newdata = ...)
print(fit); summary(fit)

fit$out is the source of truth — a p x 3 matrix, one row per parameter, with columns par, st_err and t-val. Index it as fit$out[, "par"], never fit$out["par", ]. Its row names vary by model: several report the lambda = sigma_u/sigma_v, sigma = sqrt(sigma_u^2 + sigma_v^2) reparameterisation rather than the raw scale parameters, so read the names rather than assuming a position.

npsfm() fits are the exception. They carry no out matrix and no standard errors, so only fitted(), residuals(), nobs(), print() and summary() apply; read the frontier, its gradients and the scale estimates off the returned object ($frontier, $frontier.grad, $sigma.u, $sigma.v).

Troubleshooting

?"sfa-troubleshooting" is the long version of this list, with the message each state actually prints and what to change in which order.

• sigma_u is essentially zero. This reads two opposite ways and the residual skew decides which, so check fit$wrong_skew before concluding anything. With the skew the wrong way (wrong_skew = TRUE) the boundary is the maximum likelihood estimate -- OLS with no inefficiency -- and refitting will not move it; this is usually a feature of the sample, not a bug, and skewness_test() says whether the skew is significantly wrong at all. See the wrong-skew row of Which model should I use?. With the skew the right way the data say inefficiency is present, so the boundary is unlikely to be the maximum and the fit is suspect: refit from other starting values via start_val and check fit$opt$convergence. fit$sigma_u_at_bound and fit$residual_m3 record the state either way, and print() spells out which reading applies.
• Standard errors are NA. A parameter sits on its bound, or the Hessian cannot be inverted there, so the usual standard errors are not valid. They are reported as NA on purpose. sfa_diagnostics(fit) reports how the optimizer finished and whether the Hessian is positive definite.
• Different starting values give different fits. Some likelihoods have several local maxima. Compare logLik() across fits, or seed one fit with another's estimates: sfm(..., model_name = "NG", start_from = fit_ne). To make a fit reproducible run to run, set rand.psoptim (and, for the simulated-ML panel models, rand.gtre).
• TIC(), vuong() or influence_sfa() says the likelihood was not kept. Refit with keep_objective = TRUE.
• fitted(), residuals(), predict() or efficiency(logDepVar = FALSE) fails for a fit made inside a function. They rebuild the data from the stored call. Pass the data explicitly as newdata.
• The dependent variable is in levels, not logs. Use efficiency(fit, logDepVar = FALSE). For a cost frontier, fit with inefdec = FALSE; efficiency is then minimum over actual cost.
• An unbalanced panel. psfm(model_name = "GTRE") falls back from to with a warning. and need a balanced panel.

Simulating data

data_gen_cs() and data_gen_p() generate cross-sectional and panel data with known true parameters. Each returns a data frame with one response column per model family (y_pcs, y_pcs_z, y_pcs_r, y_tre_z, ...), so a given model_name is matched to the column generated under its own assumptions. These generators are how the package's estimators are checked against known truth, including npsfm()'s — NPSFM_FLW and NPSFM_SVKZ are registered in the root-n convergence framework and both pass.

Included data

Data set	Description
USUtilities	Panel of US investor-owned fossil-fuel steam electric utilities, 1986-1999
FinnishElec	Cross-section of Finnish electricity distribution firms, averaged over a four-year regulatory period
Indian	Panel of 14 paddy farmers in Aurepalle, India, 1975-76 to 1984-85
panel89	Cross-section of US commercial banks, 1989 (Kumbhakar, Parmeter and Tsionas 2013)

Citation

citation("sfa")

Bernstein, D. H., Parmeter, C. F., and Stead, A. D. (2026). Stochastic Frontier Analysis: The sfa Package. Working Paper.

License

GPL (>= 2). See LICENSE.md.