Practical Sensitivity Analysis

for Environmental Modeling

Thorsten Wagener
[email protected]
With Francesca Pianosi
[email protected]
 The morning session
will have four parts

[1] Introduction to S.A.

(Terminology, Use, etc.)
[2] Sensitivity to what?

[3] Visual, Screening and

Variance-based Methods
[4] Validation and
Robustness of Results

2
[1] INTRODUCTION TO
SENSITIVITY ANALYSIS

3
I will use the following terminology

Factor 1 Output 1

Factor 2 Output 2
Model
… …
Factor N Output M

Sensitivity Analysis
The study of how uncertainty in the output of a model
can be apportioned to different sources of uncertainty
in the model input (factors) (Saltelli et al., 2004).
4
What is the relationship to the material
of the previous days?
Output
Variance

Output Variance
Decomposition

So what kind of questions might we ask with SA?

Saltelli 2009 http://sensitivity-analysis.jrc.ec.europa.eu 5
Question 1: What factor contributes most
to the output uncertainty?

Example:
Flood forecasting

Todini 2004 Hydrological Processes 6

 Question 2: Which factors do not impact
my output?

Example:
Model calibration

“After much whittling down, this is my

most parsimonious model structure…”
7
Sampling needs increase exponentially
No. parameters Number of samples for consistent grid sampling

1 100
Many environmental
2 10,000 models, of course,
have many more
3 1,000,000 uncertain parameters
and inputs and will be
subject to the problem
4 100,000,000
of making enough
runs to characterize
5 10,000,000,000 the whole model
space.
6 1,000,000,000,000
8
Question 3: What value does a factor have
to take to achieve the desired model output?

For example: Under

what conditions do we
not reach a global
temperature increase of
2°C?

9
Question 4: How can we reduce the
output variance below a chosen threshold?

We might try to fix or reduce the uncertainty in the smallest

number of factors to achieve this objective

10
The 4 possible questions (objectives) in
summary:
Factors prioritization (FP) Assume that, in principle, the uncertain input factors can be ‘discovered’, i.e. determined or measured,
so as to find their true value. One legitimate question is then “which factor should one try to determine first in order to have
the largest expected reduction in the variance of the model output”? This defines the ‘factors prioritization’ setting. Saltelli and
Tarantola (2002) have shown that the variance-based main effect provides the answer to the Factor Prioritization setting.
Factors fixing (FF) Another aim of sensitivity analysis is to simplify models. If a model is used systematically in a Monte Carlo
framework, so that input uncertainties are systematically propagated into the output, it might be useful to ascertain which
input factors can be fixed, anywhere in their range of variation, without sensibly affecting a specific output of interest. This may
be useful for simplifying a model in a larger sense, because we may be able then to condense entire sections of our models if
all factors entering in a section are non-influential. Saltelli and Tarantola (2002) also showed that the variance-based total effect
provides the answer to the Factor Fixing setting. A null total effect is a sufficient condition for an input factor to be irrelevant,
and therefore to be fixed.
Factors Mapping (FM) In this case, the analyst is interested to as many information as possible, either global and local, i.e. which
values of an input factor (or of group of factors) are responsible for driving the model output in a given region? Which
conditions are able to drive a specified model behaviour? In this case, a full array of methods, from local ones, to Monte Carlo
Filtering, to model emulators, to variance-based and entropy-based methods can provide useful insights about model
properties.
Variance Cutting (VC) In other cases the objective of SA can be the reduction of the output variance to a lower threshold
(variance cutting setting) by simultaneously fixing the smallest number of input factors. This setting could be of use when SA is
part of a risk assessment study and e.g. when a regulatory authority was to find the width of the impact estimate distribution
too wide. Note that the variance cutting and factor prioritization settings may appear to be very similar, as they both aim at
reducing the output variance. However, in the case of factor prioritization the scope is to identify the most influent factors one
by one, while in the variance cutting setting the objective is to reduce the output variance down to a pre-established level by
fixing the smallest subset of factors at once.

11
In general, there are just a few basic
steps in any sensitivity analysis
Sensitivity Analysis is the study of how uncertainty in the output
of a model can be apportioned to different sources of uncertainty
in the model input (factors) (Saltelli et al., 2004).

1. How do I sample the factor space?

2. What output or output-based error metric should I
calculate?
3. What sensitivity metric should I calculate? How do I
visualize the result?

12
[2] SENSITIVITY TO WHAT?

13
There are two major distinctions in how
we approach this question. [1]
[1] Analyze the model output directly

e.g. flash
flooding in
Africa

In many cases we will not have observations of the variable of

interest, especially in relation to extremes

14
We can directly estimate the sensitivity
to the simulated output
• This approach means
that we put all our
stock into our model!
• Works only if we are
rather confident in
the realism of our
model.
• For example,
integrated
assessment models

15
Or [2]. We can do sensitivity analysis in
either case, but with different objectives
[2] Analyze some error metric (objective function, likelihood etc.)

e.g. hurricane occurrence

If we do have observations, then we can test the model in

relation to the data, rather than just in itself

16
In this case we typically estimate some
type of (statistical) error metric

y ytsim(θ) ~ yttrue for all t = 1, T

ytobs

et = yobst – yt(q)
et yttrue

ytsim(θ)

time

E(q) = { e1(q), e2(q), e3(q), … ,en(q) } 0

17
A typical error metric (objective or cost or
likelihood function) is the Root Mean Squared Error

e.g.

Gupta et al. 2008 Hydrological Processes 18

The sensitivity analysis result is very
dependent on the metric chosen, e.g.
• Mass balance (e.g. bias)
• Dynamics (e.g. RMSE)
• Peaks over threshold
• Periods below threshold
• ...

19
Keep in mind that part of the sensitivity
analysis has to be a performance analysis
• The model sensitivities are more likely to be
meaningful if the model shows a good performance!
• Performance of your model (i.e. how well it
reproduces the data) is a large part of the
justification for trusting your model in the first place.
• TIP: Assess the performance for all the samples you
estimate during your sensitivity analysis (check the
histogram).

20
[3] VISUAL, SCREENING AND
VARIANCE-BASED METHODS

21
We can distinguish between local and
global approaches to sensitivity analysis
[1] Local methods analyze [2] Global methods
sensitivity around some attempt to analyze
(often optimal) point in variability across the full
the factor space factor space

22
Local methods require a good ‘baseline’ or
‘nominal point’

Local derivative of output Y

with respect to factor Xi at
fixed point x0
A priori or
optimized
estimate

We will later discuss how looking at multiple starting

points can help considerably in making the SA more
robust
23
The simplest strategy (and most widely
used) is the One-at-A-Time SA
• OAT is a local strategy. The term ‘local’ refers to the fact that all derivatives
are taken at a single point, known as ‘baseline’ or ‘nominal value’ point, in
the hyperspace of the input factors.
• Why modellers like OAT:
– the baseline vector is a safe starting point where the model properties are well
known;
– all OAT sensitivities are referred to the same starting point;
– moving one factor at a time means that whatever effect is observed on the
output, this is due solely to that factor;
– conversely, a non-zero effect implies influence, i.e. it never detects
uninfluential factors as relevant;
– the chances of the model crashing or otherwise giving unacceptable results is
minimized, as these are likely to increase with the distance from the baseline.
The model has more chances to crash when all its k factors are changed than
when just one is. A global SA is by definition a stress testing practice.

24
Output Plots
Scatter Plots
Regional Sensitivity Analysis Plots

[3.1] GRAPHICAL METHODS

25
A pitch for looking at your results!

• Graphical methods provide representations of

sensitivity in the form of graphs, charts, or surfaces
• Graphical methods are typically not the full solution,
but rather in support of a more detailed analysis
• Often the results produced for another (not visual)
method can easily be visualized as well (hence little
extra work)

26
Scatter plots (factor vs output-metric)
12

12
10

10
RMSE (m3/s)

RMSE (m3/s)
8

8
6

6
4

4
2 4 6 8 10 0.0 0.2 0.4 0.6 0.8 1.0

N K

27
Scatter plots (factor vs factor)

1.0
o o
o o o o
o o
o o o o
o o
o
o o
o o
o o
o o
o o
o o
o o
o o o
o
o o o
o o
o o o
o o
o o o
o o
o
o
o o
o o o o
o o o
o o
o o o
o
o o
o o
o o
o o
o o
o o
o o
o o
o o
RMSE = 13.1 o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o o
o
o
o
o
o
o
o
o
o
o
0.8 o
o o
o o o o
o o
o o o o
o
o o o o
o o
o o
o o
o o o
o o
o o o
o o o
o o
o o o
o o
o o
o
o
o o
o o
o o
o o
o o
o o
o o o o
o o o
o o
o o
o o
o o
o o
o o
o o
o
o
o
o o o
o o o
o o o
o o
o o o
o
o o
o o o
o o o
o o o o
o o
o
0.6

o
o o o o o o o o
o o o
o o o
o o o
o o
o o
o o
o o
o o
o
o
o o
o o
o o o
o o
o o
o o
o o
o o
o
o
o o
o o o
o o o
o o
o o
o o
o o
o RMSE = 3.37
K

o o
o o
o o
o o
o o o
o o
o o
o o
o
o o o o o o o o o
o o
o
o
o
o o
o o
o o
o o o
o o
o o o
o o
0.4

o o o
o o o
o o o
o o
o o
o o
o
o
o o o o
o
o o
o
o o
o o
o o
o o o
o
o
o o
o o
o o
o o o
o o
o o
o o
o o
o
o
o o o o
o o
o o
o o
o o
o o o
o
o o
o o
o o
o o o
o o o o o
o
o o o o o
o
o o o
o o o
o
o o o o
o o o o
o o o
o o
0.2

o o
o o
o o
o o o o o o o
o
o
o o
o o
o o
o o
o o
o o
o o o o
o
o o
o o o
o o
o o
o o o
o o
o o
o
o o
o o
o o
o o
o o
o o
o o
o o o
o
o o
o o
o o
o o
o o o
o o o
o o
o
o
o o
o o o o o
o o
o o
o o o
o
o o o
o o
o o
o o
o o o o
o o
o
0.0

2 4 6 8 10

28
Method of Morris

[3.2] SCREENING METHODS

29
What if we have a very large number of
parameters?
e.g. large scale
groundwater
pollution model

We’d like to reduce the number of parameters in a first step, so that we can
then assess the key parameters more thoroughly
30
In such cases we can applied what is
called a ‘screening method’

Screening methods are preliminary

numerical experiments whose
purpose is to isolate the most
important factors from amongst a
large number that may affect a
particular model response. By using
screening methods, factors can be
ranked in order of their importance.
However, the percentage of the
output variation that each factor is
accounting for cannot be quantified.

[Pappenberger and Vandenberghe, 2007. Sensitivity analysis, guidance 31

manual. EU Harmoni-CA Project.]
We have earlier discussed OAT, which
can easily be improved
• A good OAT would be one where, after having moved
of one step in one direction, say along X1, one would
straightway move of another step along X2 and so on
till all factors up to Xk have been moved of one step
each.
• This is the basis for elementary effects methods, a
popular method of this type is the Method of Morris

32
A popular strategy to implement this is
the Method of Morris
• Derives measures of global sensitivity from a set of local
derivatives, or elementary effects (EE)
• Each factor xi is perturbed along a grid of size Δi to
create a trajectory through the factor space, where f(x)
is the baseline
• Each trajectory yields one estimate of the elementary
effect for each factor, i.e. , the ratio of the change in
model output to the change in that parameter

33
The computational cost is r(k+1) model runs,
where k is the no. of parameters and r=4 to 10

• We repeat this for N trajectories in the parameter

space to avoid the dependence on a baseline
• We can then estimate the mean elementary effects μ
(first order effects) and their standard deviation σ
(interactions)
• Campalongo et al. (2007, EM&S) suggested
calculating the absolute values of the mean

[Herman et al. 2013, HESS] 34

The result provides and indication of
factor importance and interactions

[Chu-Agor et al. 2011, EM&S] 35

Advantages and Disadvantages
Screening procedures do not give any quantitative information about the
sensitivity, so they are very useful as a first screening when the number of
parameters is too high to perform a quantitative analysis. The low
computational cost is one of the main advantages of the screening
methods.

Key References
Saltelli, A., Ratto, M., Andres, T., Campolongo, F., Cariboni, J., Gatelli, D.,
Saisana, M., and Tarantola, S., 2007. Global sensitivity analysis. Gauging
the worth of scientific models. John Wiley & Sons. In press.
Morris, M. D., 1991 Factorial sampling plans for preliminary computational
experiments, Technometrics, 33, 161–174
Campolongo, F., Cariboni, J. and Saltelli, A., 2007. An effective screening
design for sensitivity analysis of large models. Environmental Modelling
and Software 22: 1509-1518.
36
Sobol’s Method

[3.3] VARIANCE-BASED
METHODS

37
Variance-based methods quantify sensitivity by
decomposing the variance of model outputs into
factors related components
In particular, the variance is decomposed into main effects (or
first-order effects) and interaction effects. The main effect of a
parameter quantifies the portion of the variance of the model
output which is explained by that parameter, by allowing all
other parameters to be varied at the same time.
The total effect of a parameter measures the residual variance of
the model output that remains by removing the portion
explained by all other parameters, i.e. quantifies the variance
(i.e. the uncertainty) in the model output that would be left by
fixing any other factor to its ‘true’, albeit unknown, value.

[Pappenberger and Vandenberghe, 2007]

38
The steps in this type of approach are:
Model evaluation
Sampling Compute sensitivity
against sampled
of parameter space indices
parameter sets
Latin-hypercube sampling main effects
Sobol’s sequences interaction effects
... total effect

39
A variance-based approach is called FAST
(and extended versions of it)
FAST (Fourier Amplitude Sensitivity Test) is a methodology which
allows to estimate the entire set of main effect sensitivities by
Fourier transformation (Koda et al., 1979; McRae et al., 1982),
using a single sample of size N.
Extensions of the FAST method are described in Saltelli et al.
(1999) and Tarantola et al. (2006). In classic FAST only the main
effect terms are computed. Extended FAST allows the
computation of higher order terms, in particular it allows to
compute the entire set of main and total effects, at the cost of
kN model runs.
FAST decomposes the output variance V(Y) by means of spectral
analysis: Where Vi is the amount of variance explained by
factor Xi and K is the residual.
[Pappenberger and Vandenberghe, 2007] 40
Sobol’ is becoming a very popular strategy
in environmental modeling
The Sobol’ method is a Monte Carlo procedure that allows to
compute any term of the variance decomposition, each at the
cost of N model runs (Sobol’, 1993).
Following Saltelli (2002), the cost of estimating the entire set of
main and total effects is of (2+k)N model evaluations, which
roughly halves the computational cost with respect to the
original Sobol’ algorithm.

41
[Pappenberger and Vandenberghe, 2007]
Sobol’ attributes the the variance in the
model output as follows …

42
The first-order and total sensitivity
indices are defines as …

43
Interpretation of the sensitivity indices
The main (or first-order) effect (Si) measures the contribution to the output variance
from varying the i-the factor alone (but averaged over variations in other factors)
(i) the higher the value of Si, the higher the influence of the i-th factor on the output
(i) if Si = 0, then the i-th parameter has no direct influence on the output (but it might still have
some in interaction with other parameters!)
(iii) the sum of all Si is always lower or equal to 1. If it is equal to 1, then there are no interactions
between the parameters (“additive” model)

The total effect (STi) measures the total contribution to the output variance of the i-
th factor, including its direct effect and interactions with other factors
(i) STi must be higher or equal to Si. If it is equal, then the parameter has no interactions with the
other parameters
(ii) if STi = 0, the i-th parameter has no influence (neither direct or indirect) on the model output
(ii) the sum of all STi is always higher or equal to 1. If it is equal to 1, then there are no interactions
between the parameters
[Pappenberger and Vandenberghe, 2007] 44
+&-

Advantages
Extremely robust, they work with any type of discontinuous (even randomised)
mapping between input factors and the output. Sobol’ estimator is unbiased. They
do not rely on any hypothesis about the smoothness of the mapping. The only key
assumption is that variance (i.e. the second moment) is an adequate measure for
quantifying the uncertainty of the model output.
Computing main effects and total effects for each factor, while still being far from a full
factors mapping, gives a fairly instructive description of the system. Moreover, they
provide unambiguous and clear answers to well specified sensitivity settings
(prioritisation and fixing).
Disadvantages
The computational cost is relatively high, which implies that these methods cannot be
applied to computationally expensive models. They do not provide any mapping,
i.e. they decompose the output uncertainty but they do not provide information
about, e.g., the input factors responsible for producing Y values in specified
regions, such as extreme high/low or any behavioural classification.
[Pappenberger and Vandenberghe, 2007] 45
Sobol’ sequences of quasi-random points
sample spaces very evenly

256 points from a pseudorandom number source (left); compared with the
first 256 points from the 2,3 Sobol sequence (right). The Sobol sequence
covers the space more evenly (red=1,..,10, blue=11,..,100, green=101,..,256)
46
[3.4] WHICH METHOD
SHOULD I SELECT?

47
There is not single best strategy for all
problems!

[Frey and Patil 2002. Risk Analysis] 48

 Various techniques available and their use as a function of
computational cost of the model and dimensionality of the
input space

AD means “automated
differentiation”.
[Cariboni et al., 2007. Ecological Modellling] 49
MCF
“Monte Carlo
Filtering”

FF
“Factor
Fixing”

FP
“Factor
Prioritization”

VC
“Variance Cutting”

FM
“Factor Mapping”

[Cariboni et al., 2007.

Ecological Modellling]

50
[4] VALIDATION AND
ROBUSTNESS ANALYSIS

51
Andres (1997) suggested a simple
strategy to verify that all important
factors have been identified

Vary all sensitive factors

Vary all insensitive

Optimal: Perfect
factors

Optimal: No
correlation correlation

Vary all factors Vary all factors

Andres 1997 J. Statistical Computing and Simulation 52

Example: Verification results for 4
different S.A. strategies

Tang et al. 2007 Hydrology and Earth System Sciences 53

And finally, we can use bootstrapping to
assess the robustness of the results

Blatman and Sudret 2010. Reliability Engineering & System Savety 54

[5] A BRIEF OUTLOOK
REGARDING DEVELOPMENTS

56
[a] Better visualization of complex spaces
and interactions
Total Order First Order Interactions
Total effect = Identifiability + Interactions

A As
A A
As
As A
D As Dα
α

Outside Inside, white space Inside, connection

Color corresponds to parameters: D A As α

Kelleher et al. In Press. Water Resources Research 57
[c] Understanding space-time variability
of controls in environmental systems

Van Werkhoven et al. 2008. Geophysical Research Letters 58

Having the capability of being
wrong gracefully should be a
convincing argument for
practicing … modelers’ routinely
to estimate the uncertainty
associated with their
predictions!
(Beven, 2000, Wiley)
59