It14 Belotti PDF

Contribution to the community
Consistent estimation of the fixed-effects SF model

The sftfe command
Monte Carlo results
References
sftfe: A Stata command for fixed-effects

stochastic frontier models estimation
Federico Belotti?
Giuseppe Ilardi
? CEIS, University of Rome Tor Vergata
Bank of Italy
2014 Italian Stata Users Group Meeting

Milan. November 13, 2014
Belotti, Ilardi
Fixed-effects stochastic frontier estimation in Stata
Outline
The sftfe command
Monte Carlo results

The sftfe command
Monte Carlo results
References
The fixed-effects stochastic frontier (SF) model
yit
= i + xit + it ,
(1)
it
= vit uit ,
(2)
i = 1, . . . , n, t = 1, . . . , T ,
where, for each unit i and period t:

yit represents the output;
xit is a 1 k vector of exogenous inputs;
is a k 1 vector of technology parameters;
i is the unit fixed-effect;
vit is the idiosyncratic error;
uit the one-sided disturbance which represents inefficiency.
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
Distributional assumptions - homoskedastic model
vit
uit
IID N (0, 2 ),
2
IID Fu (, ),
(3)
i = 1, . . . , n, t = 1, . . . , T ,
(4)
vit and uit are independently distributed;

The inefficiency uit has distribution with support defined over
R+ , mean and variance parameter 2 (e.g., half-normal
( = 0), exponential ( = ) or truncated-normal);
vit is normally distributed with variance 2 .
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
Heterogeneity
Heterogeneity: can be observable or unobservable;
Model (1)-(2) adds i (unobservable) to shift the production
(cost) function;
Observable heterogeneity is reflected in measured variables;
Examples are:
1
2
3
Heteroskedastic inefficiency it = exp(zit );

Heteroskedastic noise it = exp(rit );
Heterogeneity in the inefficiency mean it = sit ;
It might be that zit = rit = sit .

Belotti, Ilardi

The sftfe command
Monte Carlo results
References
The Maximum Dummy Variable approach

Greene (2005) propose to estimate model (1)(4) by treating
the unit-specific intercepts as parameters to be estimated;
This approach has been implemented in the sfpanel
command (Belotti et al., 2013);
However, as Greenes simulations suggest, this approach leads
to inconsistent variance estimates, especially in short panels.
Since these parameters represent the key ingredients in the
post-estimation of inefficiencies, a solution to this issue is
crucial in the SF context.
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
Our contribution
The new command sftfe allows the estimation of the fixed-effects
SF models via three alternative estimators (Belotti and Ilardi, 2012;
Chen et al., 2014)1 ;
They exploit the first-difference data transformation to eliminate the
fixed-effects achieving consistency for both fixed-n and fixed-T
asymptotics;
sftfe allows to estimate models in which inefficiency follows a
first-order autoregressive process as well as to model inefficiencys
variance (eventually also the mean) as a function of exogenous
covariates.
1
Belotti and Ilardi (2012) has been revised including the extension of the Chen et al. (2014) approach to
heteroskedastic and dynamic inefficiency models. The updated version is available from
http://www.econometrics.it.
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
Eliminate the nuisance parameters

We get rid of the nuisance parameters through a first-difference
data transformation
yi
= Xi + i ,
(5)
= vi ui ,
(6)
where yi = (yi2 , . . . , yiT ) with yit = yit yit1 and Xi is

the T 1 k matrix of time-varying covariates with the t-th row
denoted by xit = (xit1 , . . . , xitk ), t = 2, . . . , T .
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
First-differenced model
Idiosyncratic error - vi
The normality assumption for vit implies that vi has a

T 1-variate normal distribution with covariance matrix
= 2 T 1 , where T 1 is
T 1
2 1 0 0
1 2 1 0
..
..
..
.
.
.
=
0 1
..
.
.
.
.
.
.
.
.
.
. 1
0
0 . . . 1 2
Belotti, Ilardi
(7)

The sftfe command
Monte Carlo results
References
First-differenced model (ctd)

Inefficiency - ui
The multivariate distribution of ui is generally unknown;

Nevertheless, given the independence assumption between
vi and ui , the marginal likelihood contribution Li can be
defined in general terms as
Li () =
Z
f (vi , ui ) dui =
f (vi )f (ui ) dui
Z
=
f (yi |ui )f (ui ) dui
(8)
where is the parameter vector to be estimated.

Belotti, Ilardi

The sftfe command
Monte Carlo results
References
How to estimate the model: MMLE

Marginal Maximum Likelihood estimator (MMLE, Chen
et al., 2014);
1
The basic idea is to exploit the Closed Skew Normal class of

distributions (CSN, Gonzalez-Farias et al., 2004) that, thanks to its
closeness property under marginalization and linear transformations,
allows to derive a closed form expression for the marginal likelihood
function in equation (8);
Feasible only when inefficiency has truncated-normal (or

half-normal) distribution;
Extension to heteroskedastic (or dynamic) inefficiency is

cumbersome when T > 5 since the estimation requires the
approximation of T-variate Gaussian integrals (see Kumbhakar and
Tsionas, 2011; Chen et al., 2014).
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
How to estimate the model: MMSLE

Marginal Maximum Simulated Likelihood estimator (MMSLE,
Belotti and Ilardi, 2012);
1
The basic idea is that estimation can be accomplished via

simulation, treating the marginal likelihood function in equation (8)
as an expectation with respect to the random vector ui ;
Feasible when inefficiency has half-normal or exponential

distribution;
Extension to heteroskedastic inefficiency is feasible but constrained

(only time-invariant covariates can be used to model inefficiency
variability);
Extension to dynamic inefficiency not feasible.

Belotti, Ilardi

The sftfe command
Monte Carlo results
References
How to estimate the model: ...

The MMLE is cumbersome when the inefficiency (and/or the
idiosyncratic error) is allowed to be heteroskedastic and
T > 5;
The MMSLE imposes a restriction: the variance can only be
expressed as a function of time-invariant exogenous
explanatory variables.
Solution: Pairwise Difference estimator (PDE, Belotti and
Ilardi, 2012).
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
How to estimate the model: PDE

Pairwise Difference estimator (PDE, Belotti and Ilardi, 2012);
1
The basic idea is to exploit the closeness property of the

normal-exponential (or the normal-truncated normal via the
CSN framework) marginal likelihood function when T = 2 to
define a U-estimator based on all pairwise quasi likelihood
contributions;
Feasible and computationally efficient when inefficiency is

heteroskedastic and has half-normal, exponential or
truncated-normal distribution;
Extension to dynamic inefficiency is feasible and

straightforward when the latter has truncated-normal (or
half-normal) distribution.
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
The basic sftfe syntax is the following

sftfe depvar
indepvars

if

in

, options
Factor variables are allowed.
Options:
estimator(type) specifies the estimator to be used. May be
mmle, mmsle and pde. Default is pde.
cost specifies a cost frontier model; default is production frontier
model.
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
MMLEs specific options
distribution(distname) specifies the inefficiency

distribution. Can be hnormal or tnormal. Default is hnormal.
ghkdraws([#] , [type(string) antithetics]) governs the draws
used in Geweke-Hajivassiliou-Keane (GHK) simulation of
higher-dimensional cumulative multivariate normal
distributions. if # is omitted, the number of draws is set to
100. The type(string ) suboption specifies the type of
sequence in the simulation, can be halton, hammersley,
ghalton, random, with halton being the default; antithetics
requests antithetic draws; If this option is omitted, the
estimation is performed exploiting the result outlined in Kotz
et al. (2000) through Gauss-Hermite quadrature.
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
MMSLEs specific options

distribution. Can be exponential or hnormal. Default is
exponential.
usigma(varlist [, noconstant]) specifies that inefficiency is
heteroscedastic, with variance expressed as a function of
time-invariant covariates defined in varlist. Specifying
noconstant suppresses the intercept in this function.
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
MMSLEs specific options - 2
simtype(string ) specifies the method to generate random

draws for the first-differenced inefficiency. Can be uniform, for
uniformly distributed random variates, or halton (the default)
for Halton sequences.
nsimulations(#) specifies the number of draws used in the
simulation. The default is 250.
base(#) specifies the number, preferably a prime, used as a
base for the generation of Halton sequences. The default is 5.
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
PDEs specific options

distribution. Can be exponential, hnormal or tnormal. Default
is hnormal.
dynamic specifies that inefficiency follows a first-order
autoregressive process. Only when distribution(distname)
is hnormal or tnormal.
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
PDEs specific options - 2
emean(varlist m [, noconstant]) may be used only with

distribution(tnormal). With this option, sftfe specifies
the inefficiency mean as a linear function of the covariates
defined in varlist m.
usigma(varlist u [, noconstant]) specifies that inefficiency is
heteroscedastic, with variance expressed as a function of
covariates defined in varlist u.
vsigma(varlist v [, noconstant]) specifies that idiosyncratic
error is heteroscedastic, with variance expressed as a function
of covariates defined in varlist v.
Specifying noconstant suppresses the constant in this function.

Belotti, Ilardi

The sftfe command
Monte Carlo results
References
Postestimation
predict
type
newvar
if

in

, statistic
where statistic includes:

xb, the default, calculates the linear prediction.
stdp calculates the standard error of the linear prediction.
u produces estimates of (technical or cost) inefficiency via
E(u|) using the Jondrow et al. (1982) estimator.
jlms produces estimates of (technical or cost) efficiency via
exp [E(u|)].
alpha produces estimates of fixed-effects.
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
Syntax examples
Homoskedastic normal-truncated normal model via MMLE:
sftfe y x1 x2, est(mmle) dist(tn)
Homoskedastic normal-exponential model via MMLE:

sftfe y x1 x2, est(mmsle) dist(exp) nsim(250) base(7)
Heteroskedastic normal-exponential model via PDE:

sftfe y x1 x2, est(pde) dist(exp) usigma(z1 z2)
Heteroskedastic and dynamic normal-half normal model via PDE:

sftfe y x1 x2, est(pde) dist(hn) dynamic usigma(z1 z2)
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
MMSLE vs MMLE - Data Generating Process

We consider the homoskedastic normal-half normal model
investigated by Chen et al. (2014), that is
yit
= i + xit + vit uit ,

2
vit
N (0, ),
uit
0,
(9)
(10)
i = 1, . . . , n, t = 1, . . . , T ,
(11)
where
the fixed-effect parameters 1 , ...,
n are drawn from a standard Gaussian
random variable; xit = 0.5i + 0.52 wit with wit N (0, 1);
For each experiment, we use the same i and xit in all replications, thus
only uit and vit are redrawn in each replication;
We set = 1,
= = 2, and consider different sample sizes
(n = 100, 250) and panel lengths (T = 5, 10);
The analysis is based on 250 replications for each experiment.

Belotti, Ilardi

The sftfe command
Monte Carlo results
References
Results: MMSLE vs MMLE (n = 100)
T =5
E(u|)
ru,u
MMSLE
Bias
MSE
-0.002
0.004
-0.025
0.028
-0.006
0.010
-0.036
0.324
0.707
( 0.644 )
T = 10
MMLE
Bias
MSE
-0.002
0.004
-0.050
0.061
4.5e-04 0.012
-0.055
0.348
0.707
( 0.644 )
Belotti, Ilardi
E(u|)
ru,u
MMSLE
Bias
MSE
-0.001
0.001
-0.051
0.010
0.036
0.003
-0.047
0.268
0.752
( 0.692 )
MMLE
Bias
MSE
-0.001
0.001
-0.007
0.009
0.004
0.003
-0.012
0.266
0.752
( 0.692 )

The sftfe command
Monte Carlo results
References
T =5
E(u|)
ru,u
MMSLE
Bias
MSE
-0.002
0.004
-0.025
0.028
-0.006
0.010
-0.036
0.324
0.707
( 0.644 )
T = 10
MMLE
Bias
MSE
-0.002
0.004
-0.050
0.061
4.5e-04 0.012
-0.055
0.348
0.707
( 0.644 )
Belotti, Ilardi
E(u|)
ru,u
MMSLE
Bias
MSE
-0.001
0.001
-0.051
0.010
0.036
0.003
-0.047
0.268
0.752
( 0.692 )
MMLE
Bias
MSE
-0.001
0.001
-0.007
0.009
0.004
0.003
-0.012
0.266
0.752
( 0.692 )

The sftfe command
Monte Carlo results
References
T =5
E(u|)
ru,u
MMSLE
Bias
MSE
-0.002
0.004
-0.025
0.028
-0.006
0.010
-0.036
0.324
0.707
( 0.644 )
T = 10
MMLE
Bias
MSE
-0.002
0.004
-0.050
0.061
4.5e-04 0.012
-0.055
0.348
0.707
( 0.644 )
Belotti, Ilardi
E(u|)
ru,u
MMSLE
Bias
MSE
-0.001
0.001
-0.051
0.010
0.036
0.003
-0.047
0.268
0.752
( 0.692 )
MMLE
Bias
MSE
-0.001
0.001
-0.007
0.009
0.004
0.003
-0.012
0.266
0.752
( 0.692 )

The sftfe command
Monte Carlo results
References
T =5
E(u|)
ru,u
MMSLE
Bias
MSE
-0.002
0.004
-0.025
0.028
-0.006
0.010
-0.036
0.324
0.707
( 0.644 )
T = 10
MMLE
Bias
MSE
-0.002
0.004
-0.050
0.061
4.5e-04 0.012
-0.055
0.348
0.707
( 0.644 )
Belotti, Ilardi
E(u|)
ru,u
MMSLE
Bias
MSE
-0.001
0.001
-0.051
0.010
0.036
0.003
-0.047
0.268
0.752
( 0.692 )
MMLE
Bias
MSE
-0.001
0.001
-0.007
0.009
0.004
0.003
-0.012
0.266
0.752
( 0.692 )

The sftfe command
Monte Carlo results
References
T =5
E(u|)
ru,u
MMSLE
Bias
MSE
-0.002
0.004
-0.025
0.028
-0.006
0.010
-0.036
0.324
0.707
( 0.644 )
T = 10
MMLE
Bias
MSE
-0.002
0.004
-0.050
0.061
4.5e-04 0.012
-0.055
0.348
0.707
( 0.644 )
E(u|)
ru,u
MMSLE
Bias
MSE
-0.001
0.001
-0.051
0.010
0.036
0.003
-0.047
0.268
0.752
( 0.692 )
MMLE
Bias
MSE
-0.001
0.001
-0.007
0.009
0.004
0.003
-0.012
0.266
0.752
( 0.692 )
The bias may be reduced by increasing the number of draws
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
T =5
E(u|)
ru,u
MMSLE
Bias
MSE
-0.002
0.004
-0.025
0.028
-0.006
0.010
-0.036
0.324
0.707
( 0.644 )
T = 10
MMLE
Bias
MSE
-0.002
0.004
-0.050
0.061
4.5e-04 0.012
-0.055
0.348
0.707
( 0.644 )
Belotti, Ilardi
E(u|)
ru,u
MMSLE
Bias
MSE
-0.001
0.001
-0.051
0.010
0.036
0.003
-0.047
0.268
0.752
( 0.692 )
MMLE
Bias
MSE
-0.001
0.001
-0.007
0.009
0.004
0.003
-0.012
0.266
0.752
( 0.692 )

The sftfe command
Monte Carlo results
References
T =5
E(u|)
ru,u
MMSLE
Bias
MSE
0.001
0.001
0.002
0.011
-0.011
0.004
-0.016
0.304
0.711
( 0.651 )
T = 10
MMLE
Bias
MSE
0.001
0.001
0.001
0.012
-0.011
0.005
-0.017
0.305
0.711
( 0.651 )
Belotti, Ilardi
E(u|)
ru,u
MMSLE
Bias
MSE
0.001
3.7e-04
-0.025
0.004
0.016
0.001
-0.026
0.261
0.752
( 0.691 )
MMLE
Bias
MSE
8.8e-04 3.6e-04
-0.004
0.004
-2.9e-04
0.001
-0.009
0.261
0.752
( 0.691 )

The sftfe command
Monte Carlo results
References
Dynamic PDE - Data Generating Process

We specify the following heteroskedastic normal-half normal model
with AR(1) inefficiencies
yi
= i T + xi + vi ui ,
(12)
vi
(13)
ui
NT (0, It ),
NT+
2 1
0, (1 )
i ,
i = 1, . . . , n,
(14)
where
i = {its }t,s=1,...,T with its = it is |ts| and it = exp(0 + zit 1 );
1 , ..., n and zit aredrawn from a standard Gaussian random variable
while xit = 0.5i + 0.52 wit with wit N (0, 1);
We set P
= 0.5,
PT = 0.5, 0 = 0.5 and 1 = 1 (this implies
n
= 1
i=1
t=1 it 2).
nT
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
Dynamic PDE - Data Generating Process

The simulation of the inefficiency vector ui is performed using
the MCMC approach outlined in Geweke (1991), which uses a
Gibbs algorithm for sampling from an arbitrary multivariate
truncated normal distribution;
We consider two different values for the parameter
( = 0.3, 0.7), different sample sizes (n = 100, 250) and panel
lengths (T = 5, 10);
The analysis is based on 250 replications for each experiment.
Belotti, Ilardi

The sftfe command
Monte Carlo results
References
Dynamic PDE ( = 0.3, n = 100)
T = 10
T =5
0
1
E(u|)
ru,u
Bias
-0.002
-0.061
-0.013
-0.002
0.071
0.028
0.952
( 0.781 )
MSE
0.001
0.051
0.011
0.001
0.039
0.249
Belotti, Ilardi
0
1
E(u|)
ru,u
Bias
-7.8e-04
-0.009
-0.004
5.4e-04
0.034
0.032
0.970
( 0.806 )
MSE
5.5e-04
0.019
0.005
6.0e-04
0.024
0.173

The sftfe command
Monte Carlo results
References
Dynamic PDE ( = 0.3, n = 250)
T = 10
T =5
0
1
E(u|)
ru,u
Bias
-0.001
-0.013
-0.009
-0.001
0.018
0.018
0.957
( 0.784 )
MSE
4.3e-04
0.016
0.004
6.2e-04
0.023
0.233
Belotti, Ilardi
0
1
E(u|)
ru,u
Bias
-6.9e-04
0.009
-0.010
0.002
0.036
0.037
0.971
( 0.810 )
MSE
2.1e-04
0.007
0.002
2.5e-04
0.011
0.170

The sftfe command
Monte Carlo results
References
Dynamic PDE ( = 0.7, n = 100)
T = 10
T =5
0
1
E(u|)
ru,u
Bias
-0.003
-0.220
-0.012
-6.3e-04
-0.008
-0.123
0.955
( 0.800 )
MSE
0.001
0.104
0.007
0.002
0.010
0.515
Belotti, Ilardi
0
1
E(u|)
ru,u
Bias
-9.7e-04
-0.088
-0.010
0.002
-0.019
-0.050
0.974
( 0.838 )
MSE
7.2e-04
0.030
0.003
7.1e-04
0.004
0.346

The sftfe command
Monte Carlo results
References
Dynamic PDE ( = 0.7, n = 250)
T = 10
T =5
0
1
E(u|)
ru,u
Bias
-0.001
-0.180
-0.012
-0.002
-0.011
-0.106
0.959
( 0.801 )
MSE
5.2e-04
0.054
0.003
7.1e-04
0.004
0.496
Belotti, Ilardi
0
1
E(u|)
ru,u
Bias
-3.9e-04
-0.073
-0.013
0.002
-0.019
-0.039
0.975
( 0.842 )
MSE
2.5e-04
0.012
0.001
3.0e-04
0.002
0.336

The sftfe command
Monte Carlo results
References
References
Belotti, F., Daidone, S., Ilardi, G., and Atella, V. (2013). Stochastic frontier analysis using Stata. Stata Journal,
13(4):718758.
Belotti, F. and Ilardi, G. (2012). CEIS Research Paper 231, Tor Vergata University, CEIS.
Chen, Y., Wang, H., and Schmidt, P. (2014). Consistent estimation of the fixed effects stochastic frontier model.
Journal of Econometrics, 181(2):6576.
Geweke, J. (1991). Efficient simulation from the multivariate normal and student-t distributions subject to linear
constraints and the evaluation of constraint probabilities. In Keramidas, E. M., editor, Computing Science and
Statistics: Proceedings of the 23rd Symposium on the Interface, pages 571578. Interface Foundation of North
America, Inc.
Gonzalez-Farias, G., Dominguez-Molina, J., and Gupta, A. (2004). Additive properties of skew normal random
vectors. Journal of Statistical Planning and Inference, 126:521534.
Greene, W. (2005). Reconsidering heterogeneity in panel data estimators of the stochastic frontier model. Journal
of Econometrics, 126:269303.
Jondrow, J., Lovell, C., Materov, I., and Schmidt, P. (1982). On the estimation of technical efficiency in the
stochastic production function model. Journal of Econometrics, 19:233238.
Kotz, S., Balakrishnan, N., and Johnson, N. L. (2000). Continuous Multivariate Distributions, Volume 1, Models
and Applications, 2nd Edition. John Wiley & Sons.
Kumbhakar, S. C. and Tsionas, E. G. (2011). Some recent developments in efficiency measurement in stochastic
frontier models. Journal of Probability and Statistics, 2011.
Belotti, Ilardi

It14 Belotti PDF

Uploaded by

Copyright:

Available Formats

It14 Belotti PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

It14 Belotti PDF

Uploaded by

Copyright:

Available Formats

Contribution to the community

Consistent estimation of the fixed-effects SF model

sftfe: A Stata command for fixed-effects

? CEIS, University of Rome Tor Vergata

2014 Italian Stata Users Group Meeting

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

Consistent estimation of the fixed-effects SF model

The sftfe command

Monte Carlo results

Contribution to the community

The fixed-effects stochastic frontier (SF) model

where, for each unit i and period t:

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

Distributional assumptions - homoskedastic model

vit and uit are independently distributed;

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

Heteroskedastic inefficiency it = exp(zit );

It might be that zit = rit = sit .

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

The Maximum Dummy Variable approach

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

Eliminate the nuisance parameters

where yi = (yi2 , . . . , yiT ) with yit = yit yit1 and Xi is

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

The normality assumption for vit implies that vi has a

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

First-differenced model (ctd)

The multivariate distribution of ui is generally unknown;

f (vi )f (ui ) dui

f (yi |ui )f (ui ) dui

where is the parameter vector to be estimated.

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

How to estimate the model: MMLE

The basic idea is to exploit the Closed Skew Normal class of

Feasible only when inefficiency has truncated-normal (or

Extension to heteroskedastic (or dynamic) inefficiency is

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

How to estimate the model: MMSLE

The basic idea is that estimation can be accomplished via

Feasible when inefficiency has half-normal or exponential

Extension to heteroskedastic inefficiency is feasible but constrained

Extension to dynamic inefficiency not feasible.

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

How to estimate the model: ...

Fixed-effects stochastic frontier estimation in Stata

Contribution to the community

How to estimate the model: PDE