Laub Chi-Square Data Fitting

How Bad is Good?
A Critical Look at the Fitting of

Reflectivity Models using the Reduced
Chi-Square Statistic
Charlie Laub and Tonya L. Kuhl

Department of Chemical Engineering and Materials Science
University of California, Davis
Abstract:
The origins of the chi-square statistic that is typically employed
as a goodness of fit metric for iterative fitting of reflectivity data
will be reviewed. The basis for the rule of thumb that states that
a "good fit" is achieved with a reduced chi-square value of 1.0
will be derived and it will be shown that, even for modestly
larger errors, the "fit" is not significant in a statistical sense.
We propose a new rule of thumb that states that the reduced

chi-square error must be less than 1.5 to achieve a minimum
level of significance, e.g. at the 0.1 percent level.
The Chi-Square Statistic
⎧1 2⎫
χ = ∑ ⎨ 2 [y i − f ( xi )] ⎬
2
0< χ <∞ 2
N ⎩ σi ⎭
N = number of data points
σ2i = “variance”, related to the measurement error for yi
y = independent variable, x = dependent variable
f = assumed relationship between x and y.
yi = “observed mean”
f(xi) = “predicted mean”
• Used as a “figure of merit” when fitting a function, f, to data, {xi, yi} to

evaluate the “goodness of fit”
Reflectivity is not a direct technique
• Most reflectivity experiments capture only intensity information
• Moving from inverse space to real space is required in order to obtain
information about the sample (e.g. the scattering length density profile)
• This is commonly performed by iteratively fitting a model to the data
• Iterative fitting attempts to minimize the chi-square error between the
data set and the model-generated reflectivity
• The model encompasses the SLD profile and mathematical formalism
employed to calculate the reflectivity
• When fitting reflectivity data, it is assumed that when the difference
between the data and the model approaches zero, the scattering length
profile accurately represents the distribution of material inside the sample
region (within the limits of non-uniqueness due to lack of phase info)
Statistical Concepts Relevant to Data Fitting
• The “Parent Distribution” – to be determined by taking data
• has a mean, variance and form that are unknown!
• Collecting data can be thought of as “taking samples” from the “Parent
Distribution” that defines the relationship between the independent and
dependent variables
• The parent distribution is a “Probability Surface”
• The fitting function, f, describes the assumed functional relationship
between the independent and dependent variables. f predicts the mean for
each data point, µi.
• The resulting difference between the predicted mean f(xi) and the
observed mean, yi, is the deviation, ∆yi.
• It is typical to assume that the deviations from the mean will follow some
distribution, e.g. normal, poisson, etc.
The Origins of the Chi-Square Statistic
• If the deviations from the mean follow Gaussian statistics, the probability
of making any one observation is given by:
⎡ ⎛ y − f (x ) ⎞ 2 ⎤
PG (xi ) = ⎟⎟ ⎥ where x − µ → yi − f (xi )
1
exp ⎢− 12 ⎜⎜ i i
σ 2π ⎢⎣ ⎝ σi ⎠ ⎥⎦
• The total probability of obtaining a set of N measurements, {xi,yi}, is

equal to the product of the probabilities for each data point:
⎧ 1 ⎫ ⎪ ⎧ ⎡ N
⎛ y − f ( x ) ⎞
2
⎤ ⎫⎪
P{x , y} = ∏ PG = ⎨∏ ⎬ ∗ ⎨exp ⎢− 2 ∑ ⎜⎜ ⎟⎟ ⎥ ⎬
1 i i
N ⎩ N σ 2π ⎭ ⎪⎩ ⎢⎣ i =1 ⎝ σi ⎠ ⎥⎦ ⎪⎭
• Maximizing the probability is equivalent to minimizing the sum in the
exponential term of P{x,y}, specifically the sum of the deviations, ∆y.
• The chi-square statistic is defined by this sum:
⎛ yi − f ( xi ) ⎞
2
N
χ 2
≡ ∑ ⎜⎜ ⎟⎟
i =1 ⎝ σi ⎠
What is the reduced chi-square error (χ2/ν ) and
why should it be equal to 1.0 for a good fit?
• The method of least squares is built on the hypothesis that the optimum
description of a set of data is one which minimizes the weighted sum of squares
of deviations, ∆y, between the data, yi, and the fitting function f.
• The sum of squares of deviations is characterized by the “estimated variance of

the fit”, s2, which is an estimate of the variance of the parent distribution, σ2.
• The ratio of s2/σ2 can be estimated by χ2/ν, where ν = N – p – 1, N is the
number of observations and p is the number of fitting parameters. χ2/ν is called
the reduced chi-square statistic.
• If the fitting function accurately predicts the means of the parent distribution,
then the estimated variance, s2, should agree well with the variance of the parent
distribution, σ2, and their ratio should be close to one.
• This explains the origin of the rule of thumb for chi-square fitting that states that
a “good fit” is achieved when the reduced chi-square equals one.
Assigning Significance to the reduced chi-
square statistic
• The reduced chi-square statistic is just a 2-parameter distribution!

• A more useful distribution is obtained if we integrate it from x= χ2/ν to
x=infinity and tabulate the result as a function of χ2/ν and ν. This new
distribution (here called Q(χ2/ν,ν)) describes the probability that χ2/ν for
a set of deviations obtained by randomly sampling N observations from a
normal distribution would exceed the value for χ2/ν that was obtained by
the deviations obtained by fitting f to our N data points, {xi,yi}.
• Q(χ2/ν,ν) is a function that can be calculated numerically and has been
tabulated (consult appendix C-4 of Bevington, for example).
• Q(χ2/ν,ν) is a probability distribution, therefore: 0 < Q < 1.
How to statistically evaluate the goodness-of-fit.
• It’s simple!
• If Q is approximately 0.5, the deviations between f and {y} agree with
what you would expect statistically, meaning that the variance in the
data is approximated well by the variance of the fit because, at Q=0.5,
the reduced chi square value is near 1 which means that s2 and σ2 are
approximately equal.
• If Q is “small” then the fit is “poor” because a set of random samples

from the parent distribution have a higher probability of giving rise to
a reduced chi-square value that is equal to or less than the value
obtained by fitting of f to y. E.g. if Q=0.01, there is a 99% probability
that random deviations explain the deviations in the data better than the
predictor of the mean, f.
• If Q is very close to 1 because χ2/ν is very near zero, then most likely
the estimate of the uncertainties in the data, {σi}, is too large.
The probability that the reduced chi-square value obtained by randomly
Q: sampling N observations from a Gaussian distribution is larger than the
reduced chi-square value obtained via fitting a function to a data set
having ν degrees of freedom (ν=100 is typical).
ν 0.9 0.5 0.1 0.01 0.001 Probability

level for the
10 0.487 0.934 1.599 2.321 2.959
hypothesis that
20 0.622 0.967 1.421 1.878 2.266 the fit describes
30 0.687 0.978 1.342 1.696 1.990 the deviations
better than a
40 0.726 0.983 1.295 1.592 1.835
random sample
50 0.754 0.987 1.263 1.523 1.733
60 0.774 0.989 1.240 1.473 1.660
the reduced chi-
70 0.790 0.990 1.222 1.435 1.605
square value
80 0.803 0.992 1.207 1.404 1.560 that was ob-
90 0.814 0.993 1.195 1.379 1.525 tained by fitting
f to {x,y}.
100 0.824 0.993 1.185 1.358 1.494
140 0.850 0.995 1.156 1.299 1.410
200 0.874 0.997 1.130 1.247 1.338
Insight Gained by Using the Reduced
Chi-Square Statistic
• The reduced chi-square statistic simultaneously measures both:
• The deviations between the data and the mean of the parent distribution
that occur because there are less than an infinite number of
samples/observations.
• The discrepancy between the mean of the parent distribution and the
mean predicted by the fitting function or “model”, f.
• If the model is chosen poorly such that f cannot describe relationship
correctly, the values for the chi-squared statistic will be high. For example:
• Fitting a smooth SLD profile with too few discrete layers
• Using an approximate model (e.g. Born approximation, etc.)
• To be significant at a relatively low probability level of 0.001, the reduced
chi-square error must be less than about 1.5.

Laub Chi-Square Data Fitting

Uploaded by

Copyright:

Available Formats

Laub Chi-Square Data Fitting

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

Laub Chi-Square Data Fitting

Uploaded by

Copyright:

Available Formats

How Bad is Good?

A Critical Look at the Fitting of

Charlie Laub and Tonya L. Kuhl

We propose a new rule of thumb that states that the reduced

• Used as a “figure of merit” when fitting a function, f, to data, {xi, yi} to

• The total probability of obtaining a set of N measurements, {xi,yi}, is

• The sum of squares of deviations is characterized by the “estimated variance of

• The reduced chi-square statistic is just a 2-parameter distribution!

• If Q is “small” then the fit is “poor” because a set of random samples

ν 0.9 0.5 0.1 0.01 0.001 Probability

You might also like