Beruflich Dokumente
Kultur Dokumente
Observation-Parameter Statistics
1.
2.
3.
4.
5.
6.
7.
8.
9.
An alternative:
The relative
parameter coefficients
of variation can also
be plotted.
- They equal the
parameter standard
deviation divided by
the parameter value.
Composite scaled sensitivities
-Not fit-independent,
from the final steady-state model
but all are multiplied
by the same value of
the standard error of
the regression so
relative values are
(Book, Fig. 7-5a, p. 146)useful.
V(b)=s2(XT X) -1
b is a vector of parameter values
s2 is he calculated error variance (measure of model fit)
X is the matrix of sensitivities of the simulated equivalents
to the observations, calculated at b
is the weight matrix
V(b)=s2(XT X)-1
Five versions
a.
b.
c.
d.
e.
Var (1)
Cov (2,1)
Cov (3,1)
Cov(1,2)
Var (2)
Cov (3,2)
Cov(1,3)
Cov (2,3)
Var (3)
(Book, p. 126)
The variance indicates the range over which a parameter value could
extend without affecting model fit too adversely.
The parameter correlation coefficients indicate whether coordinated
changes in the parameter values could produce the same simulated
values and, therefore, the same model fit.
b2
~Var(b2)
minimum
~Var(b1)
b1
b2
minimum
b1
Based on V(b)
Important to understand the role that observations
play in the regression
Estimates can be largely affected by very few
observations
Two characteristics are important
Leverage depends only on the type, location,
and time of the observation
Influence depends on the observed value as well
hi x i X X x i
T
where
hi is leverage of ith parameter
x i is a vector composed of the scaled sensitivities of the ith
parameter or the transpose of the ith row of the X matrix
10
130
19
19
120
8
110
100
90
Exclude obs 18
Observations excluding 18
80
3
70
All obs
All observations
Observations excluding 19
Exclude obs 19
60
High leverage
High influence
18
18
50
0
10
20
30
40
0
50
10
15
20
25
30
Exercises
Parameter correlation coefficients (from the _pcc file) are used primarily
to assess parameter uniqueness
Computed from parameter variance-covariance matrix values
cor bi , b j
cov bi , b j
var bi
var b j
Correlation coefficients far from 1.0 may mean either there is a lack
of correlation or the correlations are not accurate enough.
R=
R N2 =
(6.11a)
m
e 0 m T e o m T
T
cov j , k
pcc j , k
1/ 2
1/ 2
var j var k
(6.18)
(7.5)
First, perform a regression run without flow observation and prior information:
In MFI2K, save the ex5.2c dataset as dataset ex7.1e and use the ex7.1e dataset in
this exercise
To omit the flow observation, click on Deactivate under the Options menu in MFI,
and deactivate RVOB
In Observations>Single-Head Observations, make the name of the output file
ex7.1e._os
From the UCODE menu, remove the prior information equations
Perform regression
Do part 1 of the Problem on p. 150 of Hill and Tiedeman
Run UCODE_2005 with SensitivityAnalysis=yes to obtain a parameter correlation
matrix
Next, include the flow observation and prior information, and start the
regression from different initial values:
bj t n,1.0 sb
2
where
(Book, p. 138)
of containing the true parameter value (when the model is sufficiently linear with
respect to the parameters)
a.
b.
c.
there is a 95%
chance that the
true parameter
values vall within
the calculated
ranges.
Parameter
Value
Situation
1: Parameter estimate and most of confidence intervals lie within reasonable range
2: Parameter estimate and confidence intervals lie outside reasonable range
3: Parameter estimate lies outside the range but part of confidence intervals lies
within the range
Range of
Reasonable
Values
Parameter
Value
1
Situation 3: Modeler needs to think about (1) what additional data could provide
more information towards estimating the parameter, and (2) improvements to the
conceptual model of the system.
2
Situation
Situations 2 and 3: Adding prior does not address the fundamental problem. There
is enough information for the regression to converge to a particular parameter value.
The modeler needs to think about what is causing the unreasonable estimates.
DO EXERCISE 7.1g (p. 153) Compare estimated values with reasonable ranges
Confidence intervals,
starting, final, and true
values, compared to
reasonable ranges
(Book, Figure 7-7,
p.153)
GW_Chart can
make graphs like
these.
_pc
1.
2.
3.
4.
5.
6.
7.
8.
9.