HYDROLOGY

Lesson 4
Probability plot and method of moments

Stefania Tamea
 PROBABILITY PLOT
How to verify how good a probabilistic model (Gumbel, lognormal) is?
Use a PROBABILITY PLOT.

This is a graphical tool to assess how good a probabilistic model is (in terms of
«shape») in fitting the data sample, before estimating the parameters.
It is a special case of a Q-Q (quantile-quantile) plot where EMPIRICAL AND
THEORETICAL DISTRIBUTIONS ARE COMPARED.
[A quantile is the value associated to a given probability in a probability distrib.]

ui
If difference is only due to location and
scale, the plot represents a straight line
x(i )

If no allignment of points in the Probability Plot  no need to keep the

probabilistic model in the statistical inference procedure
 PROBABILITY PLOT PROCEDURE
PROCEDURE: F (x)
- Sort the sample x(i) 1
- Compute the plotting position, or empirical F ( xi )
distribution function F(xi )
- Build the reduced variate ui for the
distribution to be assessed 0
xi* xi
- Plot the points (xi, ui) x

PLOTTING POSITION is the non-exceedance probability associated to the value of

an element in the sample
 Cumulative relative frequency F ( x(i ) ) = i / n (problems with upper extreme)
 Weibull plotting position F ( x(i ) ) = i /( n + 1)

 Hazen plotting position F ( x(i ) ) = (i − 0.5) / n

 REDUCED VARIATE
REDUCED VARIATE: is the value corresponding to the same probability of xi if the
probabilistic model was correct!
It is derived from the quantile function, by isolating the parts containing the
random variable and the parameters from the rest.

EXPONENTIAL DISTRIBUTION

𝑃𝑃 𝑥𝑥 = 1 − 𝑒𝑒 −𝑥𝑥/𝜃𝜃  𝑥𝑥 = −𝜃𝜃 ∙ ln(1 − 𝑃𝑃)

𝑥𝑥
= −ln(1 − 𝑃𝑃) The reduced variate, for the
𝜃𝜃
exponential distribution, reads:

only elements dependent only elements dependent 𝑢𝑢𝑖𝑖 = −ln(1 − 𝐹𝐹(𝑥𝑥𝑖𝑖 ))

on the random variable 𝑥𝑥 on the probability 𝑃𝑃, to be
and on the parameters 𝜃𝜃 replaced by 𝐹𝐹(𝑥𝑥𝑖𝑖 )
 REDUCED VARIATE
GUMBEL DISTRIBUTION
𝑥𝑥−𝜃𝜃1
− 𝑥𝑥−𝜃𝜃1
−𝑒𝑒 𝜃𝜃2
𝑃𝑃 𝑥𝑥 = 𝑒𝑒  = −ln(− ln 𝑃𝑃 )
𝜃𝜃2

The reduced variate, for the Gumbel distribution, reads: 𝑢𝑢𝑖𝑖 = −ln(−ln(𝐹𝐹(𝑥𝑥𝑖𝑖 ))

LOGNORMAL DISTRIBUTION

First modify the variable taking the natural logarithm: 𝑦𝑦𝑖𝑖 = ln 𝑥𝑥𝑖𝑖
Then you can verify if the random variable 𝑦𝑦 has a normal (Gaussian) distribution, thus the
probability distribution reads
𝑦𝑦 − 𝜃𝜃1 The reduced variate, for the
𝑃𝑃 𝑦𝑦 = Φ lognormal distribution, reads: 𝑢𝑢𝑖𝑖 = Φ−1 (𝐹𝐹 𝑦𝑦𝑖𝑖 )
𝜃𝜃2

Standard Normal Distribution (𝜇𝜇 = 0, 𝜎𝜎 2 = 1) INVERSE of Standard Normal Distrib

 PROBABILITY PLOT
CONTINUOUS FUNCTIONS
𝑢𝑢𝑖𝑖 GUMBEL PP
Adding a line to the probability plots is equivalent
to estimate the parameter values 𝑥𝑥𝑖𝑖 − 𝜃𝜃1
𝑢𝑢𝑖𝑖 =
𝑥𝑥𝑖𝑖 − 𝜃𝜃1 𝜃𝜃2
𝑢𝑢𝑖𝑖 = −ln(−ln(𝐹𝐹 𝑥𝑥𝑖𝑖 ) =
𝜃𝜃2 𝑥𝑥𝑖𝑖

And vice-versa: knowing the parameter values, the continuous distribution can be
added to the plot (line)

0.99
0.9
Sometimes, the use of probability plots is based on
0.8
pre-printed charts (PROBABILITY GRAPHS) in which 0.5
the vertical axis is stretched to reproduce the 0.2

spacing dictated by the specific distribution 0.1

0.01
 EXAMPLE
𝑥𝑥𝑖𝑖 𝐹𝐹 𝑥𝑥𝑖𝑖 𝐹𝐹 𝑥𝑥𝑖𝑖 𝑢𝑢𝑖𝑖,𝑒𝑒𝑒𝑒𝑒𝑒 𝑢𝑢𝑖𝑖,𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺
Sample: Calculations:
17 1/8 0.125 0.134 -0.732
32 19
19 2/8 0.25 0.288 -0.327
21 24
21 3/8 0.375 0.470 0.019
44 28
17 24 4/8 0.5 0.693 0.367
28 5/8 0.625 0.981 0.755
n=7 32 6/8 0.75 1.386 1.246
41 7/8 0.875 2.079 2.013

Exponential PP Gumbel PP
2.500 2.000
2.000 1.500
1.500 1.000
1.000 0.500
0.000
0.500
-0.500 16 26 36
0.000
16 26 36 -1.000
 STATISTICAL INFERENCE – STEP 3
STEP 3: ADAPT THE PROBABILISTIC MODEL TO THE DATA optimizing the fit
 Method of moments
 Method of L-moments
 Method of Maximum Likelihood (not suitable for small samples)

METHOD OF MOMENTS: equate the sample and distribution moments, with as

many equations as the number of parameters to estimate
xsup
1 n
µ=x µ= ∫ x ⋅ p( x) ⋅ dx x = ∑ xi
xinf n i =1
xsup
1 n
σ 2 = s2 σ2 = ∫ − µ ⋅ p ( x) ⋅ dx s = ∑ ( xi − x ) 2
2 2
( x )
xinf n i =1
xsup
1 n
M r = mr Mr = ∫ − µ ⋅ p ( x) ⋅ dx mr = ∑ ( xi − x ) r
r
( x )
xinf n i =1
 METHOD OF MOMENTS
x = µ (θ1 , θ 2 , θ 3 ,..)
Moments of the distributions are defined as
s 2 = σ 2 (θ1 , θ 2 , θ 3 ,..)
functions of the parameters
mr = M r (θ1 , θ 2 , θ 3 ,..)

ASSIGNMENT: GUMBEL and LOGNORMAL: 2 equations; GEV: 3 equations

ESE 1: Exponential µ =θ → x =θ

µ = θ1 + γ E ⋅ θ 2 = x  π
  θ = s ⋅
 2 π 2 →
2 2
ESE 2: Gumbel σ = θ 2
= s 6

 6
2 θ1 = x − γ E ⋅ θ 2

 µ (θ1 , θ 2 , θ 3 ) = x
ESE 3: GEV  2
 σ (θ1 , θ 2 , θ 3 ) = s
2

CA (θ ) = CA Numerical or graphical

 dist 3 solution (Excel functions)
sample
 METHOD OF MOMENTS

Distribution summary (for the

ASSIGNMENT)

PROS/CONS of method of moments

PRO: easy, robust
CONS: problems for sample variability
(n<30 with θ =2 and n<100 with θ=3),
biased (mean of estimates is different than parameter),
inefficient (no smallest variance among all possible estimators)
 EFFECT OF OUTLIERS
Q
(1) (2) Points (1) and (2):
more distant from the mean,
Q more important in the analysis of extremes
t but measured with lots of uncertainty

1 n
mr = ∑ (Qi − Q ) r
n i =1

Great uncertainty (extreme events may challenge instruments)

 Most uncertain points are most relevant in the definition of sample
characteristics  Uncertainty disturbs parameter estimates
Risk of inappropriate estimates
Need for a different method to estimate parameters: L-Moments
 METHOD OF L-MOMENTS
Principle is the same: equate the sample and distribution characteristics
xsup
Br = ∫ ⋅ ⋅ p ( x ) ⋅ dx
r
x P ( x ) Probability weighted moments
xinf

The r power is applied to the probability

( x − x ) r → x ⋅ P( x) r
instead of to the distance from the mean
Br is linear with respect to x

L-moments are linear combination of moments Br, thus are linear in x as well
L1 = B0 = µ Equal to the mean (location parameter)
Analog of the variance (amplitude around the mean,
L2 = 2 B1 − B0 measure of dispersion)
L3 = 6 B2 − 6 B1 + B0 Analog of the 3° order moment (linked to the simmetry),
L3=0 simmetric, L3>0 asimmetr positive

The method equates the L-moments (and NOT the probability-weighted

moments B) because L are more related to the distribution
 L-MOMENTS OF SAMPLE
Sample equivalent of the probability weighted moments
1 n (i − 1)(i − 2)...(i − r )
br = ∑ x(i ) x(i) are sample elements sorted
n i =1 ( n − 1)(n − 2)...(n − r ) in ascending order
1 n
b0 = ∑ x(i ) = x Equal to the sample mean (sorting desn’t matter)
n i =1
1 n (i − 1) (i-1)/(n-1) is a weight  Weighted mean
b1 = ∑ x(i )
n i =1 ( n − 1) Firts weight is 0 then values increase for larger elements
1 n (i − 1)(i − 2)
b2 = ∑ x(i ) First r weights are 0 then quadratic with i
n i =1 ( n − 1)(n − 2)

Relation between probability weighted moments l1 = b0 = x

l2 = 2b1 − b0
and L-moments is equal to those of the distribution
l3 = 6b2 − 6b1 + b0