The purpose of a SSA is to gain an understanding of the impact of uncertainties associated with the input parameter of your design. This goal is achieved using a variety of statistical measures and postprocessing tools.
Convention: Set of data .
Mean is a measure of average for a set of observations. The mean of a set of
observations is defined as follows:
(70) |
Standard deviation is a measure of dispersion from the mean for a set of
observations. The standard deviation of a set of observations is defined as follows:
(71) |
Sigma level is calculated as the inverse cumulative distribution function of a
standard Gaussian distribution at a given percentile. Sigma level is used in conjunction
with standard deviation to measure data dispersion from the mean. For example, for a
pair of quantile and sigma level
, it means that value
is about
standard deviations away from the sample mean.
Skewness is a measure of degree of asymmetry around the mean for a set of
observations. The observations are symmetric if distribution of the observations looks
the same to the left and right of the mean. Negative skewness indicates the distribution
of the observations being left-skewed. Positive skewness indicates the distribution of
the observations being right-skewed. The skewness of a set of observations is defined as follows:
(72) |
Kurtosis is a measure of relative peakedness/flatness of distribution for a set of
observations. It is generally a relative comparison with the normal distribution.
Negative kurtosis indicates a relatively flat distribution of the observations compared
to the normal distribution, while positive kurtosis indicates a relatively peaked
distribution of the observations. As such, the kurtosis of a set of observations is defined with calibration to the normal distribution as
follows:
(73) |
where and
represent mean and standard deviation, respectively.
Table 1: Different Types of Kurtosis
![]() |
![]() |
![]() | |||
| κ < 0 | κ = 0 | κ > 0 | |||
| |||||
Entropy is a concept first introduced in classical thermodynamics. It is a measure of disorder in a system. The concept of entropy was later extended and recognized by Claude Shannon in the domain of information/communication as a measure of uncertainty over the true content of a message. In statistics, the entropy is further reformulated as a function of mass density, as follows:
(74) |
where represents mass density of a parameter. In the context of
statistics/probability, entropy becomes a measure of complexity and predictability of a
parameter. A more complex, and less predictable, parameter carries higher entropy, and
vice versa. For example, a parameter characterized with uniform distribution has higher
entropy than that with truncated triangular distribution (within the same bounds).
Hence, the parameter characterized with uniform distribution is less predictable
compared to the triangular
distribution.
(75) |
The probability mass is used in a normalized fashion, such that not only the shape,
but the range of variability, or the distribution is accounted for. This is shown in
Equation 75, where is the relative frequency of the parameter falling into a certain
interval, and
is the width of the interval. As a result, Shannon entropy can have a
negative value. Following are some comparisons of the Shannon entropy, where S2 is
smaller than S1.

Taguchi Signal-to-Noise Ratios
Three signal-to-noise ratios are provided in the statistics of each output in your SSA. These ratios are as follows:
Nominal is Best
Smaller is Better
Larger is Better
Signal-to-noise (S/N) ratios are measures used to optimize control parameters to achieve a robust design. These measures were first proposed by Dr. Taguchi of Nippon Telephone and Telegraph Company, Japan, to reduce design noises in manufacturing processes. These design noises are normally expressed in statistical terms such as mean and standard deviation (or variance). In computer aided engineering (CAE), these ratios have been widely used to achieve a robust design in computer simulations. For a design to be robust, the simulations are carried out with an objective to minimize the variances. The minimum variance of designs/simulations can be done with or without targeting a certain mean. In design exploration, minimum variance targeted at a certain mean (called Nominal is Best) is provided, and is given as follows:
(76) |
where and
represent mean and standard deviation, respectively.
Nominal is Best is a measure used for characterizing design parameters such as model dimension in a tolerance design, in which a specific dimension is required, with an acceptable standard deviation.
In some designs, however, the objectives are to seek a minimum or a maximum possible at the price of any variance.
For the cases of minimum possible (Smaller is Better), which is a measure used for characterizing output parameters such as model deformation, the S/N is expressed as follows:
(77) |
For the cases of maximum possible (Larger is Better), which is a measure used for characterizing output parameters such as material yield, the S/N is formulated as follows:
(78) |
These three S/N ratios are mutually exclusive. Only one of the ratios can be optimized for any given parameter. For a better design, these ratios should be maximized.
The minimum and maximum values of a set of observations are:
(79) |
(80) |
Note: The minimum and maximum values strongly depend on the number of samples. If you generate a new sample set with more samples, then chances are that the minimum value is lower in the larger sample set. Likewise, the maximum value of the larger sample set is most likely higher than for the original sample set. Hence, the minimum and maximum values should not be interpreted as absolute physical bounds.
Statistical Sensitivity Measures
The sensitivity charts displayed under Six Sigma Analysis are global sensitivities based on statistical measures. For more information, see see Single Parameter Sensitivities.
Generally, the impact of an input parameter on an output parameter is driven by:
The amount by which the output parameter varies across the variation range of an input parameter.
The variation range of an input parameter. Typically, the wider the variation range, the larger the impact of the input parameter.
The statistical sensitivities displayed under Six Sigma Analysis are based on the Spearman-Rank Order Correlation coefficients that take both those aspects into account at the same time.
Basing sensitivities on correlation coefficients follows the concept that the more strongly an output parameter is correlated with a particular input parameter, the more sensitive it is with respect to changes of that input parameter.
Spearman Rank-Order Correlation Coefficient
The Spearman rank-order correlation coefficient is:
(81) |
where:
|
|
|
|
|
|
Significance of Correlation Coefficients
Because the sample size is finite, the correlation coefficient
is a random variable. Hence, the correlation coefficient between two
random variables
and
usually yields a small, but nonzero value, even if
and
are not correlated at all in reality. In this case, the correlation
coefficient would be insignificant. Therefore, you need to find out if a correlation
coefficient is significant or not. To determine the significance of the correlation
coefficient, assume the hypothesis that the correlation between
and
is not significant at all, meaning that they are not correlated, and
(null hypothesis). In this case the variable:
(82) |
is approximately distributed like the student's -distribution with
degrees of freedom. The cumulative distribution function student's
-distribution is:
(83) |
where:
|
|
There is no closed-form solution available for Equation 83. See Abramowitz and Stegun (Pocketbook of Mathematical Functions, abridged version of the Handbook of Mathematical Functions, Harry Deutsch, 1984) for more details.
The larger the correlation coefficient , the less likely it is that the null hypothesis is true. Also, the
larger the correlation coefficient
, the larger is the value of t from Equation 82
and consequently also the probability
is increased. Therefore, if the probability that the null hypothesis
is true is given by
. If
exceeds a certain significance level, for example 2.5%, you can assume
that the null hypothesis is true. However, if
is below the significance level, then it can be assumed that the null
hypothesis is not true and that consequently the correlation coefficient
is significant. This limit can be changed in Design Exploration Options.
Cumulative Distribution Function
The cumulative distribution function of sampled data is also called the empirical
distribution function. To determine the cumulative distribution function of sampled
data, you need to order the sample values in ascending order. Let be the sampled value of the random variable
having a rank of
, which makes
smallest out of all
sampled values. The cumulative distribution function
that corresponds to
is the probability that the random variable
has values below or equal to
. Because you have only a limited amount of samples, the estimate for
this probability is itself a random variable. According to Kececioglu (Reliability Engineering Handbook, Vol. 1, 1991,
Prentice-Hall, Inc.), the cumulative distribution function
associated with
is:
(84) |
Equation 84 must be solved numerically.
Probability and Inverse Probability
The cumulative distribution function of sampled data can only be given at the
individual sampled values using Equation 84. Hence, the evaluation of
the probability that the random variable is less than or equal to an arbitrary value x
requires an interpolation between the available data points.
If is, for example, is between
and
, the probability that the random variable
is less or equal to
is:
(85) |
The cumulative distribution function of sampled data can only be given at the
individual sampled values using Equation 84. Hence, the evaluation of
the inverse cumulative distribution function for any arbitrary probability value
requires an interpolation between the available data points.
The evaluation of the inverse of the empirical distribution function is most
important in the tails of the distribution, where the slope of the empirical
distribution function is flat. In this case, a direct interpolation between the points
of the empirical distribution function similar to Equation 85 can
lead to inaccurate results. Therefore, the inverse standard normal distribution function
is applied for all probabilities involved in the interpolation. If
is the requested probability for which you are looking for the inverse
cumulative distribution function value, and
is between
and
, the inverse cumulative distribution function value can be calculated
using:
(86) |