Evaluation of con®dence intervals for a steady-state
leaky aquifer model
S. Christensen
a,*& R.L. Cooley
b aDepartment of Earth Sciences, Aarhus University, Ny Munkegade b. 520, 8000 Aarhus C, Denmark
b
Water Resources Division, U.S. Geological Survey, Box 25046, Mail Stop 413, Denver Federal Center, Denver, Colorado 80225, USA
(Received 20 May 1997; revised 22 December 1997; accepted 20 November 1998)
The fact that dependent variables of groundwater models are generally nonlinear functions of model parameters is shown to be a potentially signi®cant factor in calculating accurate con®dence intervals for both model parameters and functions of the parameters, such as the values of dependent variables calculated by the model. The Lagrangian method of Vecchia and Cooley [Vecchia, A.V. & Cooley, R.L., Water Resources Research, 1987, 23(7), 1237±1250] was used to calculate nonlinear Schee-type con®dence intervals for the parameters and the simulated heads of a steady-state groundwater ¯ow model covering 450 km2 of a leaky aquifer. The nonlinear con®dence intervals are compared to corresponding linear intervals. As suggested by the signi®cant nonlinearity of the regression model, linear con®dence intervals are often not accurate. The commonly made as-sumption that widths of linear con®dence intervals always underestimate the actual (nonlinear) widths was not correct. Results show that nonlinear eects can cause the nonlinear intervals to be asymmetric and either larger or smaller than the linear approximations. Prior information on transmissivities helps reduce the size of the con®dence intervals, with the most notable eects occurring for the parameters on which there is prior information and for head values in parameter zones for which there is prior information on the parameters. Ó 1999 Elsevier
Science Ltd. All rights reserved
Key words: con®dence interval, nonlinearity, groundwater ¯ow, model, regres-sion.
1 INTRODUCTION
Estimates of parameters and dependent variables from a calibrated groundwater model are generally uncertain because the data used for calibration are uncertain and because the model never perfectly represents the system or exactly ®ts the data. Con®dence intervals on the es-timated (calibrated) model parameters and model de-pendent variables can be used to express the degree of uncertainty in these quantities, but calculation of con-®dence intervals is not straightforward because the so-lution of the groundwater ¯ow equation for hydraulic heads, and quantities such as ¯ows that are a function of hydraulic heads, is generally a nonlinear function of the
model parameters. Cooley and Vecchia13 derived a
general method for the calculation of simultaneous con®dence intervals for the output from a groundwater ¯ow model when the statistical distribution of model
parameters is known. Vecchia and Cooley23and Clarke8
independently derived a similar methodology that can be used to compute simultaneous con®dence intervals on both the estimated parameters and the output from a nonlinear regression model with normally distributed
residuals. Vecchia and Cooley23 showed that the same
algorithm can be used both to estimate the parameters by nonlinear regression and to calculate con®dence limits from which the desired simultaneous con®dence
intervals are obtained. Beven and Binley3 presented a
Bayesian method that is similar in intent to the method
of Vecchia and Cooley23, and Brookset al.4presented a
nonstatistical variant.
Ó1999 Elsevier Science Ltd Printed in Great Britain. All rights reserved 0309-1708/99/$ ± see front matter
PII: S 0 3 0 9 - 1 7 0 8 ( 9 8 ) 0 0 0 5 5 - 4
*
Corresponding author.
Con®dence intervals considered here are Schee-type
intervals22, which are simultaneous for all linearizable
functions of model parameters and so are the largest of all simultaneous intervals for these functions. Individual con®dence intervals and con®dence intervals that are simultaneous for ®nite numbers of functions, which are
smaller than Schee-type intervals, could also have been
calculated, but published theory (for example, Ref.16) is
not as complete as the theory for the Schee-type
in-tervals (for example, Ref.10). As will be shown,
conclu-sions obtained for the present study will apply to the other types of con®dence intervals.
For brevity we will term con®dence intervals calcu-lated using the nonlinear regression model as nonlinear con®dence intervals. The nonlinear regression model can be linearized using a ®rst order Taylor series ex-pansion. Con®dence intervals calculated using the line-arized model are the standard ones found in most
regression texts22and are termed here `linear con®dence
intervals'.
Synthetic case studies9,10,13,17,23show that: (i)
corre-sponding linear and nonlinear con®dence intervals are often oset or shifted relative to one another, and the
nonlinear intervals are generally larger10,17; (ii) the
variability in sizes of nonlinear intervals is generally larger than the variability in sizes of corresponding
lin-ear intervals10,17; (iii) the dierences between the sizes of
corresponding nonlinear and linear con®dence intervals
increase as the sizes of the intervals increase9; (iv) use of
prior information can have a signi®cant eect on the
sizes of con®dence intervals10. In the present study we
compute and analyze both linear and nonlinear con®-dence intervals using ®eld data to test the calculation procedures and the concepts derived from the synthetic case studies. Linear and nonlinear con®dence intervals are compared for two dierent sets of data used for calibration: (i) observed heads and (ii) observed heads and prior information on the parameters that is ob-tained from ®eld measurements.
2 BACKGROUND
2.1 Regression model and basis for con®dence intervals
The groundwater ¯ow model and accompanying infor-mation on the groundwater system are stated as a nonlinear regression model of the form
yf b e 1
where y[yi] is a vector of n observations of the
groundwater system (in this study hydraulic head and
selected model parameters); b[bj] is a vector ofp
un-known, true, model parameters;f(b)[fi(b)] is a vector
of n model-computed values corresponding to y; and
e[ei] is a vector ofn true errors. The true errors are
treated as random variables assumed to have zero means and be distributed normally as
eN 0;Wÿ1r2 2
where W[Wij] is an n´n known, positive-de®nite
weight matrix andr2 is an unknown scalar.
MatrixWÿ1r2 is the variance±covariance matrix for
e. In this study this matrix is assumed to be block
di-agonal with one block for hydraulic heads and the other for prior information on selected parameters.
Diagonal entries of Wÿ1r2 are variances and
o-diag-onal entries are covariances. Because all head obser-vations are assumed to be uncorrelated here, all o-diagonal entries are zero in the hydraulic head block. The prior information on parameters is assumed to be
correlated. Vectors y, f(b), and e are partitioned to
conform with the blocks in W. Weights are initially
estimated, and can be reestimated as the regression progresses if analysis of residuals indicates that they are
incorrect12. Thus, although the weights are assumed to
be known, it is apparent that they are approximate. To this extent, the computed con®dence intervals are also
approximate. The unknown scalar r2 is estimated by
the regression.
Schee-type con®dence intervals are to be obtained
on some function of parametersg(b). In this study the
functiong(b) can represent hydraulic head or a
param-eter, and thus can be the same as f(b). However, in
generalg(b) can also represent other nonlinear or linear
functions of parameters for which a con®dence interval
is desired. As indicated earlier, Schee-type intervals are
on all linearizable (see Appendix A) functions simulta-neously, and thus can be de®ned by the probability statement
Probg bL6g b6g bUfor all linearizableg b
1ÿa 3
where g(bL) and g(bU) are the lower and upper
con®-dence limits, respectively, and sets bL and bU are in
general dierent for each function,g(b). These limits are
computed as minimum and maximum values of g(b)
over the parameter con®dence region. Schee-type
in-tervals should be contrasted with the more common individual con®dence intervals that are on a single function. A good way of visualizing the dierence is to
note that, from eqn (3), the probability isa that there
will be any (1ÿa)´100% Schee-type interval that
does not contain the true value, whereas the probability
is a that the individual con®dence interval does not
contain the true value. Because of this dierence,
Schee-type intervals are larger than individual
inter-vals. Calculation of Schee-type con®dence intervals is
discussed in Ref.23 and in Appendix A.
2.2 Measures of model nonlinearity
Nonlinearity of the regression model results when the
model function f(b) is a nonlinear function of model
non-linearity, parameter eects nonlinearity and intrinsic
nonlinearity2. Parameter eects nonlinearity is the
component of the nonlinearity that can in theory be removed by a suitable transformation of model param-eters, whereas the intrinsic nonlinearity is the compo-nent of the nonlinearity that cannot be removed by any parameter transformation. The sum of the two compo-nents is the total nonlinearity, as measured by eqn (4) below, say, and, through some transformation of pa-rameters, it has a minimum achievable value equal to the intrinsic nonlinearity. As described later, nonlinearity
aects the Schee-type intervals.
Measures of nonlinearity have been developed by
Beale2, Linssen20, and Bates and Watts1, among others.
Linssen's measures20are easily computed and are useful
to gauge the degree of total and intrinsic nonlinearity of
the model. The square of Linssen's measure20 of total
nonlinearity can be stated as12
^
l is the linear model
ap-proximation offl, computed from
f0l ^fX b^ l
ÿ^b 5
s2 is the error variance, de®ned as
s2S ^b
nÿp 6
andmis the number of parameter setsbl used to
com-pute the measure. In eqn (5) X^ is the n ´p sensitivity
and Na12give a justi®cation of eqn (4) and recommend
usingm2p pointsbl at the maximum and minimum
linear Schee-type con®dence limits for the parameters.
From eqn (4) it can be seen that Linssen's measure is a scaled length of the discrepancy between the correct
model valuesfl and the linearized valuesfol.
The minimum value ofM^b for any transformation of
parameters is the intrinsic nonlinearity and is computed
using sets of transformed parameters d0l for which the
numerator of eqn (4) is a minimum. Linssen20 shows
that each set can be found by solving the set of linear equations
^
XTWX^d0lX^TW flÿ^f 8
Thus, the square of Linssen's measure of intrinsic non-linearity is
Note thatflin eqn (9) is computed using parameter sets
bl.
Degrees of total nonlinearity have been classi®ed by
Beale2 and Cooley and Na12. For example, Beale2
classi®es a model as disastrously nonlinear if M^b>
1=Fa p;nÿp and eectively linear if M^b<0:01=
Fa p;nÿp, where Fa(p,nÿp) is the upper a point of
theFdistribution withpandn ÿp degrees of freedom.
Cooley and Na12 added an intermediate class
^
Mb<0:09=Fa p;nÿp, for which linear Schee-type
intervals for parameters were found to approximate the nonlinear ones fairly well in test cases. Large values of
^
Mb indicate that linear con®dence intervals may not be
accurate. However, small values may not always
indi-cate that linear intervals will be accurate becauseM^b is
an average value that may not measure the relevant degree of nonlinearity as it aects a particular interval, and because of the possible in¯uence of nonlinearity of
the function g(b).
As shown in Appendix A,M^/ can be used to correct
a Schee-type interval for intrinsic nonlinearity. The
resulting interval is theoretically conservative to the
degree of approximation used in its derivation, if M^/
were the true value and not an estimate2. In practice, the
correction should be regarded as approximate and the resulting intervals as probably conservative.
3 FIELD CASE
The test case is a steady-state groundwater ¯ow model
of a 450 km2 leaky aquifer in Quaternary deposits of
glacial till and ¯uvioglacial sand and gravel. The aquifer is located in the western part of the Danish island Zee-land within the catchment of the stream Tude aa (Fig. 1). The aquifer system is outlined in Fig. 2. The aquifer is overlain by 10±80 m of till. In high-lying areas, the aquifer is recharged by downward leakage from a phreatic aquifer in the more permeable upper zone of the till, whereas in low-lying stream valleys and areas near the coast, the aquifer is discharged by upward leakage. More information about the hydrogeology is
given by Christensen5.
The model was originally calibrated by trial and
er-ror5 and subsequently by nonlinear regression7 using
MODFLOWP18. Choice of model parameters was
based on a detailed description of the hydrogeology5.
parameters in Ref.7are nearly the same, which indicates redundancy in the parameterization.
Deletion of redundant parameters improves the conditioning for the calculation of nonlinear con®dence intervals, so the number of model parameters was
re-duced to a total of 11: the log10-transmissivity of the
nine zones shown in Fig. 3 Y1 Y9, and the log10
-vertical hydraulic conductivity of the semi-con®ning till
layer within zones 1±5 (Z1±5) and zones 6±9 (Z6±9).
Each aquifer zone is characterized by speci®c geo-logical conditions. In zone 1 the aquifer is thin, and well logs show that it is absent in some places. The expected eective transmissivity of this zone is therefore small, which is also indicated by the large head gradients in the northern part of zone 1 (Fig. 3(b)). In the southern part of the zone, the head gradients are smaller because the groundwater ¯uxes are reduced by scattered ground-water withdrawals. The thickness and the transmissivity of the aquifer varies signi®cantly within zones 2, 6 and 7. The aquifer is rather thick and continuous within zones 3±5, with observed transmissivities of the order of
10ÿ3±10ÿ2 m2/s. Zones 8 and 9 represent an extremely
heterogeneous area having local sand layers (zone 9) surrounded by till (zone 8). Fig. 3(b) shows that the head gradients are very large in this area. (Note that the observed head ®eld is not shown, but it compares well with the simulated head ®eld in Fig. 3(b).)
The simulated leakage is a function of the vertical hydraulic conductivity, the thickness of the semi-con-®ning till layer, and of the head dierence between the water table in the upper part of the till and the head in the aquifer. The hydraulic conductivity was estimated by calibration as mentioned, whereas the thickness of the semi-con®ning layer was interpolated from well logs and thus was not a model parameter. The water table in the till was ®xed at an elevation 3 m below the ground level, whereas the head in the aquifer was a dependent vari-able computed by the model.
A no-¯ux condition was used along the entire model boundary. In some areas, the no-¯ux condition is due to geologic boundaries, whereas groundwater divides
de-®ne the boundary condition in other areas5.
Ground-water is withdrawn from several major and minor well ®elds throughout the model area. The total withdrawal
is about 8 million m3 per year, and around 40% of the
total withdrawal is withdrawn from the four major well ®elds shown in Fig. 1. The withdrawal corresponds to approximately 45% of the ground water leaking to the aquifer, whereas 50% is leaking upward through the till
in the low stream valleys, and 5% is leaking to the sea5.
4 PARAMETER ESTIMATION AND MODEL NONLINEARITY
The model parameters were estimated by nonlinear
re-gression using a modi®ed code of Cooley and Na12that
uses a combined modi®ed Gauss±Newton and
quasi-Newton method for the optimization11. The model
Fig. 1.Location of aquifer.
Fig. 2. Schematic representation of the aquifer system along the pro®le line in Fig. 1. The arrows give the ¯ow directions in
the aquifer and in the overlying till.
Fig. 3.Ground-water model: (a) zonation and head data (sis estimated standard deviation); (b) simulated head and location
parameters were estimated from two dierent sets of data: (a) synchronously measured head in 100 wells (Fig. 3); and (b) these same heads supplemented with
prior information on four zonal log10-transmissivities
(Y4,Y5,Y6 andY7).
The hydraulic heads in wells were observed with an accuracy of a few centimeters. However, due to pro-cesses such as small-scale variability in hydrogeology, which are unknown and/or are not represented in the model, the groundwater model can only simulate the large scale variations (here termed ``the drift'') of the
true head ®eld7. The weights of the head observations
were therefore estimated as the inverses of the variances
between a polynomial approximation of the drift6and
the measured heads. The estimated variances vary
be-tween 1 and 4.4 m2(Fig. 3(a)), whereas the covariances
are assumed to be zero.
Prior information on zonal log-transmissivities was
obtained by Christensen7in the following way.
Local-scale transmissivity was estimated by analyzing short-duration pumping tests in 90 wells. These results were used to estimate the transmissivity semi-variogram, and to estimate by kriging the local-scale transmissivity at regularly spaced grid points within each of the four zones. The zonal estimates and the corresponding
co-variance matrix were obtained by combining the kriged
estimates as described by Journel and Huijbregts19(c.f.
pp. 320±324). The inverse covariance matrix was used as the weight matrix of the prior information.
The estimated parameters from the nonlinear re-gression are shown in Fig. 4. The estimated trans-missivities lie within the range of transtrans-missivities determined by analysis of numerous pumping tests
within the various aquifer zones5,7, and the estimated
hydraulic conductivities of the con®ning layer are within
the range of values (10ÿ9±10ÿ8m/s) found by analyzing a
smaller number of pumping tests5. The parameter
esti-mates also compare well to those of previous model
studies in the area5,7.
Residual analyses12 show that, in both cases (a) and
(b), the weighted residuals,W1=2e(whereeis the residual
vector which is an estimate of e), have nearly constant
variance; they appear to be unbiased and normally dis-tributed. Thus, there is no reason to suspect that the weight matrix used is not a good estimate of the true covariance matrix for the calibration data (head and
prior information). The error variances2increases from
1.06 in case (a) to 1.12 in case (b). This may indicate a weak incompatibility between the ¯ow model and the prior information.
In case (a) the total nonlinearity,M^bis 0.17, and in case (b)M^b is 0.12. These values lie between 1/F0:05(p,n ÿp)
0.52 and 0.09/F0:05(p,nÿp)0.05, which means that
there is signi®cant nonlinearity in both cases. The large values suggest that linear con®dence intervals may not be
accurate. The intrinsic nonlinearity,M^/, is 0.06 in case
(a) and 0.05 in case (b). Both values are probably not signi®cant, but the con®dence region was corrected to improve the approximation of the con®dence intervals.
We used the correction method of Beale2and Linssen20,
which is summarized in Appendix A. Finally note that the prior information slightly reduces the total as well as the intrinsic nonlinearity of the regression model.
5 CALCULATION OF CONFIDENCE INTERVALS
In this study, linear and nonlinear 95% Schee-type
con®dence intervals were calculated on the estimated parameters and hydraulic head. The linear intervals
were calculated by the standard method22, and the
nonlinear intervals were calculated by the likelihood method described in Appendix A. The latter intervals were calculated using a modi®ed Gauss±Newton
algo-rithm, as suggested by Vecchia and Cooley23.
The calculation of each con®dence limit (the upper or lower bound of the con®dence interval) is a separate optimization problem in which the bound is obtained as an extreme value on the edge of the parameter
con®-dence region23. To guard against obtaining a local
ex-treme, the calculation of each limit was carried out with nine dierent sets of starting values for parameters. The starting values were obtained by adding or subtracting up to two linearized standard deviations to or from the values estimated by nonlinear regression. This proce-dure was followed for the calculation of nonlinear in-tervals for the eleven parameters, as well as for the hydraulic head at 20 grid points homogeneously dis-tributed over the model area (Fig. 3).
The variation of parameter starting-values did not produce signi®cant dierences in the con®dence limits for the parameters. Also, for all limits but one, the calculations converged in 5±14 iterations. In the case where no prior information was used, the calculation of
the lower bound value ofZ1±5converged more slowly, in
17±25 iterations. The convergence criterion was 0.001 for the relative parameter change from one iteration to the next.
The variation of parameter starting values also did not produce signi®cant dierences in the con®dence limits for the head at the grid points. For all limits but one, convergence was achieved in 5±19 iterations. For the upper limit at point 13, however, convergence was not achieved for any of the starting parameters. In this case convergence was achieved by manually adjusting
the value of one parameter (Y1) and letting the
algo-rithm estimate the other 10 parameter values.
For con®dence intervals on selected parameters and heads, we not only searched for extreme values (i.e. con®dence limits) on the edge of the parameter con®-dence region, but also searched the inside of the region. In all of these cases, parameter sets giving the con®dence limits were found on the edge of the region. The results thus strongly indicate that the con®dence intervals are calculated correctly.
Based on the above results, we conclude that the al-gorithm is a fairly robust and ecient method for
cal-culating nonlinear con®dence intervals on both
parameters and heads for the present ®eld case. We therefore calculated the head con®dence interval at 485 grid points evenly spaced throughout the model area. At 400 of these points, the calculations converged in less than 25 iterations. These 400 points were used to con-tour characteristics of the con®dence intervals. The calculation of nonlinear intervals was made on a 50 MHz 80486 PC. The CPU time used was about 5 min per parameter interval, and about 6 min per head in-terval.
6 RESULTS
6.1 Con®dence intervals on parameters
Fig. 4 shows the calculated con®dence intervals for the
parameters. For the log10-vertical conductivities, Z, the
sizes of the intervals are about half an order of magni-tude, and the linear and nonlinear intervals are quite
similar. For the log10-transmissivities,Y, the sizes of the
intervals vary from one to more than three orders of magnitude. The widest intervals are for: (i) the zones
with a small Y-estimate and/or small head gradients
throughout the zone (Y1,Y2,Y4), or (ii) a zone without
head observations (Y8). A comparison of linear and
nonlinear con®dence intervals supports conclusions from previous synthetic case studies. Nonlinear intervals are generally larger than the linear intervals, except for
Y3, and the two types of intervals are often oset relative
to one another10,17. The dierences among nonlinear
intervals are greater than the dierences among
corre-sponding linear intervals10,17. Absolute dierences
be-tween the sizes of nonlinear and linear intervals tend to
increase as the sizes of intervals increase9, but dierences
relative to the sizes of the corresponding linear intervals are more nearly constant. Addition of prior information signi®cantly reduces the sizes of the con®dence intervals for the parameters with prior information, and the lin-ear and nonlinlin-ear intervals are similar for these
pa-rameters. Intervals for parameters without prior
information are nearly unaected by adding prior in-formation on other parameters.
Fig. 4 shows that the linear con®dence interval onY3
construct the linear con®dence interval for g(b)
corre-sponds to the Cramer±Rao bound21, and therefore,
should be the lower bound. However, the linearized variance only corresponds to the Cramer±Rao bound if it is calculated at the unknown true set of parameters;
i.e., for^bb. Because^bin general will dier fromb, the
linearized variance actually used will generally not cor-respond to the Cramer±Rao bound, and linear con®-dence intervals may therefore be larger than their nonlinear counterparts.
6.2 Con®dence intervals on hydraulic head
Fig. 5 shows the calculated con®dence intervals for head at the 20 grid points, and Fig. 6 shows the contoured size of the nonlinear con®dence intervals throughout the model area. The size of the nonlinear intervals generally ranges from 2±10 m within the largest part of the area. The size, however, is as large as 40 m in an area within and east of zones 8 and 9, where head data are lacking. Head gradients in this area are signi®cant (toward two well ®elds in zone 9), which results in a large sensitivity
of head to the local parameters (Y7andY8). Because the
estimates of these parameters are uncertain due to the mentioned lack of data, large con®dence intervals result. Some reduction in uncertainty of parameter estimates is expected if head data are collected immediately south and east of zone 8. Similar conditions explain the large con®dence intervals around a minor well ®eld in zone 1. These results show that hydraulic head can vary over a large range when parameter values vary over their
con®dence region. Thus, when Schee-type intervals are
appropriate, as when gauging the uncertainty of a number of dierent dependent variables and (or) pa-rameters simultaneously, large uncertainties can result.
Comparison of the linear and nonlinear intervals in Fig. 5 shows that the two types of intervals are often oset from one another. Many of the nonlinear intervals are skewed, and the skewness tends to increase with the size of the interval.
Fig. 7 shows the dierences between the sizes of nonlinear and linear con®dence intervals. Note that the sizes of the nonlinear con®dence intervals in general are larger than the sizes of the linear intervals in the southern part of the model area, whereas the relation-ship is opposite in the northern part. As argued above,
this indicates that the linearized variance used to com-pute the linear con®dence interval does not in general correspond to the Cramer±Rao bound.
Figs. 5 and 6 show that prior information on Y
re-duces the width and skewness of the con®dence intervals somewhat in the zones with the prior information, and corresponding linear and nonlinear intervals are more alike in these zones. In contrast to this, the con®dence intervals are nearly unaected in zones without prior information. Fig. 8 shows the reduction in size of the con®dence intervals due to the use of prior information. It can be seen that the prior information reduces the sizes of the intervals by less than 0.5 m in most of the area, but in minor areas the reduction is of the order of 1 m. Within and east of zones 8 and 9, the reduction is, however, up to 16 m. The head gradient in this area is
large, and it is sensitive to Y7. Prior information
sig-ni®cantly reduces the uncertainty of this parameter and thus also reduces the head con®dence interval.
7 SUMMARY AND CONCLUSIONS
Nonlinear regression was successfully used to estimate the parameters of a steady-state groundwater ¯ow model of a leaky aquifer. The residuals are normally distributed and nonlinear con®dence intervals can therefore be calculated by the method of Vecchia and
Cooley23. Linssen's measure20 shows that total
nonlin-earity of the regression model is signi®cant, even when prior information on some transmissivities is added to the head data. To a lesser extent, this is also true for the Fig. 6.Size of nonlinear con®dence intervals on head [m].
intrinsic nonlinearity. We therefore corrected the size of the parameter con®dence region used to calculate the
nonlinear Schee-type intervals as suggested by Beale2
and Linssen20.
The method of Vecchia and Cooley23 was, for the
present case, a fairly robust and ecient algorithm for the calculation of nonlinear con®dence intervals on both the hydraulic head and the model parameters. Calcu-lated con®dence limits did not depend on starting pa-rameter values, which indicates that all con®dence limits are probably unique. One of the main assumptions of
Vecchia and Cooley's23 Lagrangian method, that the
limits of the con®dence interval can be calculated from parameter sets on the edge of the parameter con®dence region, is ful®lled for all the calculated con®dence in-tervals.
As suggested by the signi®cant model nonlinearity, linear con®dence intervals are often not accurate. Nonlinear eects can cause the nonlinear intervals to be oset from, and either larger or smaller than, the linear approximations. The nonlinear head intervals are largest in areas of the model with large head gradients, whereas linear intervals are largest in areas with small gradients. Thus, widths of linear con®dence intervals do not al-ways underestimate the actual (nonlinear) widths so that the linearized variance used to compute a linear con®-dence interval does not in general correspond to the Cramer±Rao bound. There is only correspondence if the linearized variance is calculated at the unknown true set of parameters.
The general agreement of regression estimates of parameters with and without prior information suggests that the parameter and head data are compatible. The prior information helps reduce and stabilize the con®-dence intervals, although the most notable eects occur for the parameters on which there is prior information and for head values in zones for parameters having prior
information. Use of prior information reduced the total nonlinearity of the model only slightly.
The present case study compares linear and corre-sponding nonlinear con®dence intervals and establishes geometric characteristics of nonlinear con®dence inter-vals. Thus the general conclusions should be valid not
only for Schee-type intervals, but also for individual
and other types of con®dence intervals as well.
ACKNOWLEDGEMENTS
This work was partly ®nanced by the Danish Natural Science Research Council via a research fellowship for Steen Christensen.
APPENDIX A
It is shown here that if the intrinsic nonlinearity is
small, then nearly exact Schee-type con®dence intervals
for the function g(b) can be computed by ®nding
max-imum and minmax-imum values of g(b) over the standard
likelihood con®dence region forb. It is also shown that,
if the intrinsic nonlinearity is signi®cant, then the size of the con®dence region can be corrected, and the resulting con®dence intervals are conservative.
Assume ®rst that intrinsic nonlinearity is negligible.
Then there exists some transformation, a, of original
model parameters, b, that nearly linearizes the
regres-sion model, as measured by eqn (4)22(c.f. pp. 133±136).
In this case, model function f(b) can be written as
f(b)g(a), whereg(a) is approximately a linear function
ofa. In addition the sum of squared errors functionS(b)
can be written as S(a) using g(a) in place of f(b), and
S(b)S(a) 22 (c.f. pp. 195). Because the transformed
model is nearly linear, the assumption of normally dis-tributed errors (2) leads to
S b ÿS ^b=p S ^b= nÿp
S / ÿS ^a=p
S ^a= nÿp F p;nÿp
A:1
where /is the true set of transformed parameters
cor-responding to the true setbof original parameters, ^ais
the set of regression estimates of/corresponding to the
set of regression estimates^bofb, andF(p,nÿp) signi®es
the F distribution with p and n ÿp degrees of
free-dom14. Based on eqn (A.1), a (1 ÿa)´100% likelihood
con®dence region forb is given by
S b ÿS ^b6ps2Fa p;nÿp da2 A:2
whereFa(p,n ÿp) is the uppera point ofF14.
As demonstrated subsequently, a Schee-type
con®-dence interval is given by
min
b g b; maxb g b
A:3
subject to the constraint Fig. 8.Reduction of size of nonlinear con®dence interval on
S b ÿS ^b6da2 A:4
If there are no stationary points of g(b) (where
og=ob0) within the con®dence region given by
eqn (A.4), or if any stationary point produces an ext-remum less extreme than one on the boundary of eqn (A.4), then eqn (A.4) can be replaced by the
equality constraint S b ÿS ^b d2
a. In this case the
limits given by eqn (A.3) can be computed by ®nding extreme values of
L b;k0 g b k0hda2ÿS b S ^bi A:5
with respect tob and Lagrange multiplierk023. An
iter-ative solution for extreme values of eqn (A.5) is given by
Vecchia and Cooley23.
It is straightforward to show, using a proof
analo-gous to that used by Cooley10 for a Bayesian Sche
e-type interval, that eqn (A.5) gives a Schee-type
con®-dence interval for all functionsg(b) that are linearizable
using Taylor series. However, this can also be
demon-strated graphically for p2 by referring to Fig. 9. To
interpret the ®gure, which is a plot in parameter space,
note that g(b)constant describes a curve. Also note
that g(b1), g(b2), g(bL) and g(bU) are constants. The
con®dence region boundaryS b ÿS ^b da2is a closed
surface (here a closed curve) comprising all possible
values forS(b) on the boundary. Now ifb actually lies
outside of the con®dence region forb, such asb1, then
for at least one function of the formg(b)g(b1), such as
the one shown, g(b1) lies outside of its con®dence
in-terval given by (g(bL), g(bU)). Thus, the entire curve
g(b)g(b1) lies outside of the region bounded by
g(b)g(bL) andg(b)g(bU). However, ifbactually lies
within the con®dence region, such as b2, then all
line-arizable functions of the form g(b)g(b2) must lie
within their con®dence intervals because they must pass
through b2. Therefore, all linearizable functions are
contained within their con®dence intervals
simulta-neously with the same relative frequency, 1ÿa, thatbis
contained within its con®dence region. Although it cannot be graphically portrayed, this argument
gener-alizes for any p. We also argue that, because of the
physical nature of groundwater models, only lineariz-able functions are of interest and thus need be con-sidered.
If the intrinsic nonlinearity is signi®cant (which can
be measured by eqn (9)), then f(b) cannot be nearly
linearized and eqn (A.1) may not be nearly correct. As a
result, con®dence region eqn (A.2) may not contain b
with probability 1ÿa. Beale2 gave an approximate
correction to obtain a conservative (1ÿa)´ 100%
con®dence region. With the correction,da2 is given by
da21 n
where N/ is a theoretical measure of intrinsic
nonlin-earity obtained by letting setsblbe normally distributed,
and then lettingm®1in (9)2. (Actually, Beale2
made
an approximation that Guttman and Meeter15found to
yield signi®cant error and Linssen20 corrected. The
corrected version is used here.) In this study, we follow
Linssen's20suggestion and replaceN/byM^/as given by
eqn (9). The signi®cance of the intrinsic nonlinearity is
simply measured by the magnitude of the correcteddaas
compared to the uncorrected value.
The same arguments as originally used to justify the
use of eqn (A.5) to compute Schee-type intervals can
be used to incorporate the correction eqn (A.6). Because
the con®dence region is conservative, so are the Sche
e-type intervals.
REFERENCES
1. Bates, D.M. & Watts, D.G., Relative curvature measures of nonlinearity. Journal of the Royal Statistical Society, Series B, 1980,42(1), 1±25,.
2. Beale, E.M.L., Con®dence regions in non-linear estima-tion.Journal of the Royal Statistical Society, series B, 1960,
22(1), 41±76.
3. Beven, K. & Binley, A., The future of distributed models: Model calibration and uncertainty prediction. Hydrolog-ical Processes, 1992,6,279±298.
4. Brooks, R. J., Lerner, D. N. & Tobias, A. M., Determin-ing the range of predictions of a groundwater model which arises from alternative calibrations. Water Resources Research, 1994,30(11), 2993±3000.
5. Christensen, S., Hydrological model for the Tude aa catchment.Nordic Hydrology, 1994,25(3), 145±166. 6. Christensen, S., Calibration data accuracy in a
regional-scale groundwater model.Internal research report, Dept. of Earth Sciences, Aarhus University, Aarhus, 1995, 29 pp. 7. Christensen, S., On the strategy of estimating
regional-scale transmissivity ®elds.Ground Water, 1997,35(1), 131± 139.
Fig. 9. Relations between a con®dence interval (g(bL),
g(bU)) and possible true values g(b1) and g(b2) for g(b). S b ÿS ^b d2
8. Clarke, G.P.Y., Approximate con®dence limits for a parameter function in nonlinear regression.Journal of the American Statistical Association, 1987,82(397), 221±230. 9. Cooley, R.L., Exact Schee-type con®dence intervals for
output from groundwater ¯ow models, 1, Use of hydro-geologic information. Water Resources Research, 1993,
29(1), 17±33.
10. Cooley, R.L., Exact Schee-type con®dence intervals for output from groundwater ¯ow models, 2, Combined use of hydrogeologic information and calibration data. Water Resources Research, 1993,29(1), 35±50.
11. Cooley, R.L. & Hill, M.C., A comparison of three Newton-like nonlinear least-squares methods for estimat-ing parameters of ground-water ¯ow models. Computa-tional Methods in Water Resources IX, Vol.1,(ed. by T.F. Russell, R.E. Ewing, C.A. Brebbia, W.G. Gray & G.F. Pinder), 1992, 379±386.
12. Cooley, R.L. & Na, R.L., Regression modelling of ground-water ¯ow.U.S. Geological Survey Techniques of Water-Resources Investigations, 1990, book 3, ch.B4,232 pp.
13. Cooley, R.L. & Vecchia, A.V., Calculation of nonlinear con®dence and prediction intervals for ground-water ¯ow models.Water Resources Bulletin, 1987,23(4), 581±599. 14. Draper, N. & Smith, H.,Applied regression analysis. 2nd
ed., John Wiley, New York, 1981, 709 pp.
15. Guttman, I. & Meeter, D.A., On Beale's measures of non-linearity.Technometrics, 1965,7(4), 623±637.
16. Hamilton, D., & Wiens, D., Correction factors for F ratios in nonlinear regression. Biometrica, 1987, 74(2), 423±425.
17. Hill, M.C., Analysis of accuracy of approximate, simulta-neous, nonlinear con®dence intervals on hydraulic heads in analytical and numerical test cases. Water Resources Research, 1989,25(21), 177±190.
18. Hill, M.C., A computer program (MODFLOWP) for estimating parameters of a transient, three-dimensional ground-water ¯ow model using nonlinear regression.U.S. Geological Survey, Open-File Report 91-484, 1992, 358 pp. 19. Journel, A.G. & Huijbregts, C., Mining geostatistics.
Academic Press, New York, 1978, 600 pp.
20. Linssen, H.N., Nonlinearity measures: A case study.
Statistica Neerlandica, 1975,29,93±99.
21. Mood, A.M., Graybill, F.A. & Boes, D.C.,Introduction to the theory of statistics. 3rd ed., McGraw-Hill, New York, 1974, 564 pp.
22. Seber, G.A.F. & Wild, C.J., Nonlinear regression. John Wiley, New York, 1989, 768 pp.