Maximum likelihood estimation of the parameters of a statistical model involves maximizing the likelihood or, equivalently, the log likelihood with respect to the parameters. The parameter values at which the maximum occurs are the maximum likelihood estimates of the model parameters. The likelihood is a function of the parameters and of the data.
Let be the observations in a random sample, including the failures and censoring times (if the data are censored). Let be the probability density of failure time, be the reliability function, and be the cumulative distribution function, where is the vector of parameters to be estimated, . The probability density, reliability function, and CDF are determined by the specific distribution selected as a model for the data. The log likelihood is defined as






where
is the sum over failed units
is the sum over rightcensored units
is the sum over leftcensored units
is the sum over intervalcensored units
and is the interval in which the ith unit is interval censored. Only the sums appropriate to the type of censoring in the data are included when the preceding equation is used.
The RELIABILITY procedure maximizes the log likelihood with respect to the parameters by using a NewtonRaphson algorithm. The NewtonRaphson algorithm is a recursive method for computing the maximum of a function. On the rth iteration, the algorithm updates the parameter vector with

where is the Hessian (second derivative) matrix, and is the gradient (first derivative) vector of the loglikelihood function, both evaluated at the current value of the parameter vector. That is,

and

Iteration continues until the parameter estimates converge. The convergence criterion is



for all where c is the convergence criterion. The default value of c is , and it can be specified with the CONVERGE= option in the MODEL, PROBPLOT, RELATIONPLOT, and ANALYZE statements.
After convergence by the preceding criterion, the quantity

is computed. If then a warning is printed that the algorithm did not converge. is called the relative Hessian convergence criterion. The default value of d is 0.0001. You can specify other values for d with the CONVH= option. The relative Hessian criterion is useful in detecting the occasional case where no progress can be made in increasing the log likelihood, yet the gradient g is not zero.
A locationscale model has a CDF of the form

where is the location parameter, is the scale parameter, and G is a standardized form of the cumulative distribution function. The parameter vector is =( ). It is more convenient computationally to maximize log likelihoods that arise from locationscale models. If you specify a distribution from Table 16.57 that is not a locationscale model, it is transformed to a locationscale model by taking the natural (base e) logarithm of the response. If you specify the lognormal base 10 distribution, the logarithm (base 10) of the response is used. The Weibull, lognormal, and loglogistic distributions in Table 16.57 are not locationscale models. Table 16.58 shows the corresponding locationscale models that result from taking the logarithm of the response.
Maximum likelihood is the default method of estimating the location and scale parameters in the MODEL, PROBPLOT, RELATIONPLOT, and ANALYZE statements. If the Weibull distribution is specified, the logarithms of the responses are used to obtain maximum likelihood estimates ( ) of the location and scale parameters of the extreme value distribution. The maximum likelihood estimates (, ) of the Weibull scale and shape parameters are computed as and .
The parameters of the threeparameter Weibull distribution are estimated by maximizing the log likelihood function. The threshold parameter must be less than the minimum failure time , unless , in which case, can be equal to . The RELIABILITY procedure sets a default upper bound of for the threshold in the iterative estimation computations and a default lower bound of 0.0. You can set different bounds by specifying an INEST data set as described in the section INEST Data Set for the ThreeParameter Weibull.
If the shape parameter is less than one, then the density function in Table 16.57 has a singularity at , and the log likelihood is unbounded above as the threshold parameter approaches the minimum failure time . For any fixed , maximum likelihood estimates of the scale and shape parameters and exist. If in the iterative estimation procedure, the estimate of the threshold is set to the upper bound and maximum likelihood estimates of and are computed.
You can specify a SAS data set to set lower bounds, upper bounds, equality constraints, or initial values for estimating
the parameters of a threeparameter Weibull distribution by using the INEST= option in the ANALYZE or PROBPLOT statement.
The data set must contain a variable named _TYPE_
that specifies the action that you want to take in the iterative estimation process, and some combination of variables named
_SCALE_
, _SHAPE_
, and _THRESHOLD_
that represent the distribution parameters. If BY processing is used, the INEST= data set should also include the BY variables,
and there must be at least one observation for each BY group.
The possible values of _TYPE_
and corresponding actions are summarized in Table 16.66.
Table 16.66: _TYPE_
Variable Values
Value of 
Action 

LB 
Lower bound 
UB 
Upper bound 
EQ 
Equality 
PARMS 
Initial value 
For example, you can use the INEST data set In
created by using the following SAS statements to specify bounds for data that contain the BY variable Group
with three BY groups: A, B, and D. The data set In
specifies a lower bound for the threshold parameter of –100 for groups A, B, and D, and an upper bound of 3 for the threshold
parameter for group D. Since the variables _Scale_
and _Shape_
are set to missing, no action is taken for them, and these variables could be omitted from the data set.
data In; input Group$1 _Type_$ 211 _Scale_ _Shape_ _Threshold_; datalines; A lb . . 100 B lb . . 100 D lb . . 100 D ub . . 3 ;
You can specify a regression model by using the MODEL statement. For example, if you want to relate the lifetimes of electronic parts in a test to Arrheniustransformed operating temperature, then an appropriate model might be

where , and is the centigrade temperature at which the ith unit is tested. Here, [ 1 ].
There are two types of explanatory variables: continuous variables and classification variables. Continuous variables represent physical quantities, such as temperature or voltage, and they must be numeric. Continuous explanatory variables are sometimes called covariates.
Classification variables identify classification levels and are declared in the CLASS statement. These are also referred to
as categorical, dummy, qualitative, discrete, or nominal variables. Classification variables can be either character or numeric. The values of classification variables are called
levels. For example, the classification variable Batch
could have levels ‘batch1’ and ‘batch2’ to identify items from two production batches. An indicator (01) variable is generated
for each level of a classification variable and is used as an explanatory variable. See Nelson (1990, p. 277) for an example that uses an indicator variable in the analysis of accelerated life test data. In a model, an explanatory
variable that is not declared in a CLASS statement is assumed to be continuous.
By default, all regression models automatically contain an intercept term; that is, the model is of the form

where does not have an explanatory variable multiplier. The intercept term can be excluded from the model by specifying INTERCEPT=0 as a MODEL statement option.
For numerical stability, continuous explanatory variables are centered and scaled internally to the procedure. This transforms the parameters in the original model to a new set of parameters. The parameter estimates and covariances are transformed back to the original scale before reporting, so that the parameters should be interpreted in terms of the originally specified model. Covariates that are indicator variables—that is, those specified in a CLASS statement—are not centered and scaled.
Initial values of the regression parameters used in the NewtonRaphson method are computed by ordinary least squares. The parameters and the scale parameter are jointly estimated by maximum likelihood, taking a logarithmic transformation of the responses, if necessary, to get a locationscale model.
The generalized gamma distribution is fit using log lifetime as the response variable. The regression parameters , the scale parameter , and the shape parameter are jointly estimated.
The Weibull distribution shape parameter estimate is computed as , where is the scale parameter from the corresponding extreme value distribution. The Weibull scale parameter is not computed by the procedure. Instead, the regression parameters and the shape are reported.
In a model with one to three continuous explanatory variables x, you can use the RELATION= option in the MODEL statement to specify a transformation that is applied to the variables before model fitting. Table 16.67 shows the available transformations.
Table 16.67: Variable Transformations
Relation 
Transformed variable 

ARRHENIUS (Nelson parameterization) 

ARRHENIUS2 (activation energy parameterization) 

POWER 

LINEAR 
x 
LOGISTIC 

In some situations, it is desirable for the scale parameter to change with the values of explanatory variables. For example, Meeker and Escobar (1998, section 17.5) present an analysis of accelerated life test data where the spread of the data is greater at lower levels of the stress. You can use the LOGSCALE statement to specify the scale parameter as a function of explanatory variables. You must also have a MODEL statement to specify the location parameter. Explanatory variables can be continuous variables, indicator variables specified in the CLASS statement, or any interaction combination. The variables can be the same as specified in the MODEL statement, or they can be different variables. Any transformation specified with the RELATION= MODEL statement option will be applied to the same variable appearing in the LOGSCALE statement. See the section Regression Model with Nonconstant Scale for an example of fitting a model with nonconstant scale parameter.
The form of the model for the scale parameter is

where is the intercept term. The intercept term can be excluded from the model by specifying INTERCEPT=0 as a LOGSCALE statement option.
The parameters are estimated by maximum likelihood jointly with all the other parameters in the model.
The location and scale parameters are estimated by maximizing the likelihood function by numerical methods, as described previously. An alternative parameterization that is likely to have better numerical properties for heavy censoring is , where and is the pth quantile of the standardized distribution. See Meeker and Escobar (1998, p. 90) and Doganaksoy and Schmee (1993) for more details on alternate parameterizations.
By default, RELIABILITY estimates a value of from the data that will improve the numerical properties of the estimation. You can also specify values of p from which the value of will be computed with the PSTABLE= option in the ANALYZE, PROBPLOT, RELATIONPLOT, or MODEL statement. Note that a value of p = 0.632 for the Weibull and extreme value and p = 0.5 for all other distributions will give and the parameterization will then be the usual locationscale parameterization.
All estimates and related statistics are reported in terms of the location and scale parameters . If you specify the ITPRINT option in the ANALYZE, PROBPLOT, or RELATIONPLOT statement, a table showing the values of p, , , and the last evaluation of the gradient and Hessian for these parameters is produced.
An estimate of the covariance matrix of the maximum likelihood estimators (MLEs) of the parameters is given by the inverse of the negative of the matrix of second derivatives of the log likelihood, evaluated at the final parameter estimates:

The negative of the matrix of second derivatives is called the observed Fisher information matrix. The diagonal term is an estimate of the variance of . Estimates of standard errors of the MLEs are provided by

An estimator of the correlation matrix is

The covariance matrix for the Weibull distribution parameter estimators is computed by a firstorder approximation from the covariance matrix of the estimators of the corresponding extreme value parameters as









For the regression model, the variance of the Weibull shape parameter estimator is computed from the variance of the estimator of the extreme value scale parameter as shown previously. The covariance of the regression parameter estimator and the Weibull shape parameter estimator is computed in terms of the covariance between and as

Table 16.68 shows the method of computation of approximate twosided confidence limits for distribution parameters. The default value of confidence is . Other values of confidence are specified using the CONFIDENCE= option. In Table 16.68, represents the percentile of the standard normal distribution, and and are the MLEs of the location and scale parameters for the normal, extreme value, and logistic distributions. For the lognormal, Weibull, and loglogistic distributions, and represent the MLEs of the corresponding location and scale parameters of the locationscale distribution that results when the logarithm of the lifetime is used as the response. For the Weibull distribution, and are the location and scale parameters of the extreme value distribution for the logarithm of the lifetime. denotes the standard error of the MLE of , computed as the square root of the appropriate diagonal element of the inverse of the Fisher information matrix.
Table 16.68: Confidence Limit Computation
Parameters 


Distribution 
Location or 
Scale 
Shape 
Threshold 

Normal 






Lognormal 






Lognormal 



(base 10) 



Extreme value 






Weibull 






Exponential 






Logistic 






Loglogistic 






Generalized 



Gamma 



Threeparameter 



Weibull 



Approximate confidence limits for the regression parameter are given by


The maximum likelihood estimate of the percentile for the extreme value, normal, and logistic distributions is given by

where , G is the standardized CDF shown in Table 16.69, and are the maximum likelihood estimates of the location and scale parameters of the distribution. The maximum likelihood estimate of the percentile for the Weibull, lognormal, and loglogistic distributions is given by

where , and G is the standardized CDF of the locationscale model corresponding to the logarithm of the response. For the lognormal (base 10) distribution,

The maximum likelihood estimate of the percentile for the threeparameter Weibull distribution is computed by

where , and G is the standardized CDF of extreme value distribution.
Table 16.69: Standardized Cumulative Distribution Functions
LocationScale 
LocationScale 


Distribution 
Distribution 
CDF 
Weibull 
Extreme value 

Lognormal 
Normal 

Loglogistic 
Logistic 

The variance of the MLE of the percentile for the normal, extreme value, or logistic distribution is

Twosided approximate confidence limits for are






where represents the percentile of the standard normal distribution.
The limits for the lognormal, Weibull, or loglogistic distributions are






where refers to the percentile of the corresponding locationscale distribution (normal, extreme value, or logistic) for the logarithm of the lifetime. For the lognormal (base 10) distribution,






Approximate limits for the threeparameter Weibull distribution are computed as






where refers to the percentile of the standard extreme value distribution.
For the extreme value, normal, and logistic distributions shown in Table 16.69, the maximum likelihood estimate of the reliability function is given by

The MLE of the CDF is .
Let . The approximate variance of u is

Twosided approximate confidence intervals for are computed as


where


and represents the percentile of the standard normal distribution. The corresponding limits for the CDF are


As an alternative, you can request that twosided likelihood ratio confidence limits for the reliability function and the CDF be computed by specifying the LRCLSURV option in the ANALYZE statement or the LRCLSURV option in the PROBPLOT statement.
Limits for the Weibull, lognormal, and loglogistic reliability function are the same as those for the corresponding extreme value, normal, or logistic reliability , where . Limits for the threeparameter Weibull use and the extreme value CDF.
You can create a table containing estimates of the reliability function, the CDF, and confidence limits computed as described in this section with the SURVTIME= option in the ANALYZE statement or with the SURVTIME= option in the PROBPLOT statement. You can plot confidence limits for the CDF on probability plots created with the PROBPLOT statement with the PINTERVALS=CDF option in the PROBPLOT statement. PINTERVALS=CDF is the default option for parametric confidence limits on probability plots.
In addition to estimating the parameters of the distributions in Table 16.57, you can estimate parameters, compute confidence limits, compute predicted values and prediction limits, and compute chisquare tests for differences in groups for the binomial and Poisson distributions by using the ANALYZE statement. Specify either BINOMIAL or POISSON in the DISTRIBUTION statement to use one of these distributions. The ANALYZE statement options available for the binomial and Poisson distributions are given in Table 16.5. See the section Analysis of Binomial Data for an example of an analysis of binomial data.
If r is the number of successes and n is the number of trials in a binomial experiment, then the maximum likelihood estimator of the probability p in the binomial distribution in Table 16.59 is computed as

Twosided confidence limits for p are computed as in Johnson, Kotz, and Kemp (1992, p. 130):

with and and

with and , where is the percentile of the F distribution with degrees of freedom in the numerator and degrees of freedom in the denominator.
You can compute a sample size required to estimate p within a specified tolerance w with probability . Nelson (1982, p. 206) gives the following formula for the approximate sample size:

where is the percentile of the standard normal distribution. The formula is based on the normal approximation for the distribution of . Nelson recommends using this formula if and . The value of used for computing confidence limits is used in the sample size computation. The default value of confidence is . Other values of confidence are specified using the CONFIDENCE= option. You specify a tolerance of number with the TOLERANCE(number) option.
The predicted number of successes X in a future sample of size m, based on the previous estimate of p, is computed as

Twosided approximate prediction limits are computed as in Nelson (1982, p. 208). The prediction limits are the solutions and of


where is the % percentile of the F distribution with degrees of freedom in the numerator and degrees of freedom in the denominator. You request predicted values and prediction limits for a future sample of size number with the PREDICT(number) option.
You can test groups of binomial data for equality of their binomial probability by using the ANALYZE statement. You specify the K groups to be compared with a group variable having K levels.
Nelson (1982, p. 450) discusses a chisquare test statistic for comparing K binomial proportions for equality. Suppose there are successes in trials for . The grouped estimate of the binomial probability is

The chisquare test statistic for testing the hypothesis against for some i and j is

The statistic Q has an asymptotic chisquare distribution with K – 1 degrees of freedom. The RELIABILITY procedure computes the contribution of each group to Q, the value of Q, and the pvalue for Q based on the limiting chisquare distribution with K – 1 degrees of freedom. If you specify the PREDICT option, predicted values and prediction limits are computed for each group, as well as for the pooled group. The pvalue is defined as , where is the chisquare CDF with K – 1 degrees of freedom, and Q is the observed value. A test of the hypothesis of equal binomial probabilities among the groups with significance level is
: do not reject the equality hypothesis
: reject the equality hypothesis
You can use the ANALYZE statement to model data by using the Poisson distribution. The data consist of a count Y of occurrences in a “length” of observation T. Observation T is typically an exposure time, but it can have other units, such as distance. The ANALYZE statement enables you to compute the rate of occurrences, confidence limits, and prediction limits.
An estimate of the rate is computed as

Twosided confidence limits for are computed as in Nelson (1982, p. 201):


where is the percentile of the chisquare distribution with degrees of freedom.
You can compute a length T required to estimate within a specified tolerance w with probability . Nelson (1982, p. 202) provides the following approximate formula:

where is the percentile of the standard normal distribution. The formula is based on the normal approximation for and is more accurate for larger values of . Nelson recommends using the formula when . The value of used for computing confidence limits is also used in the length computation. The default value of confidence is . Other values of confidence are specified using the CONFIDENCE= option. You specify a tolerance of number with the TOLERANCE(number) option.
The predicted future number of occurrences in a length S is

Twosided approximate prediction limits are computed as in Nelson (1982, p. 203). The prediction limits are the solutions and of


where is the percentile of the F distribution with degrees of freedom in the numerator and degrees of freedom in the denominator. You request predicted values and prediction limits for a future exposure number with the PREDICT(number) option.
You can compute a chisquare test statistic for comparing K Poisson rates for equality. You specify the K groups to be compared with a group variable having K levels.
See Nelson (1982, p. 444) for more information. Suppose that there are Poisson counts in lengths for and that the are independent. The grouped estimate of the Poisson rate is

The chisquare test statistic for testing the hypothesis against for some i and j is

The statistic Q has an asymptotic chisquare distribution with K – 1 degrees of freedom. The RELIABILITY procedure computes the contribution of each group to Q, the value of Q, and the pvalue for Q based on the limiting chisquare distribution with K – 1 degrees of freedom. If you specify the PREDICT option, predicted values and prediction limits are computed for each group, as well as for the pooled group. The pvalue is defined as , where is the chisquare CDF with K – 1 degrees of freedom and Q is the observed value. A test of the hypothesis of equal Poisson rates among the groups with significance level is
: accept the equality hypothesis
: reject the equality hypothesis
Fitting to the probability plot by least squares is an alternative to maximum likelihood estimation of the parameters of a life distribution. Only the failure times are used. A least squares fit is computed using points , where and are the plotting positions as defined in the section Probability Plotting. The are either the lifetimes for the normal, extreme value, or logistic distributions or the log lifetimes for the lognormal, Weibull, or loglogistic distributions. The ANALYZE, PROBPLOT, or RELATIONPLOT statement option FITTYPE=LSXY specifies the as the dependent variable (’ycoordinate’) and the as the independent variable (’xcoordinate’). You can optionally reverse the quantities used as dependent and independent variables by specifying the FITTYPE=LSYX option.
Weibayes estimation is a method of performing a Weibull analysis when there are few or no failures. The FITTYPE=WEIBAYES option requests this method. The method of Nelson (1985) is used to compute a onesided confidence interval for the Weibull scale parameter when the Weibull shape parameter is specified. See Abernethy (2006) for more discussion and examples. The Weibull shape parameter is assumed to be known and is specified to the procedure with the SHAPE=number option. Let be the failure and censoring times, and let be the number of failures in the data. If there are no failures , a lower confidence limit for the Weibull scale parameter is computed as

The default value of confidence is . Other values of confidence are specified using the CONFIDENCE= option.
If , the MLE of is given by

and a lower confidence limit for the Weibull scale parameter is computed as

where is the percentile of a chisquare distribution with degrees of freedom. The procedure uses the specified value of and the computed value of to compute distribution percentiles and the reliability function.
In many applications, units can experience multiple causes of failure, or failure modes. For example, in the section Weibull Probability Plot for Two Combined Failure Modes, insulation specimens can experience either early failures due to manufacturing defects or degradation failures due to aging. The FMODE statement is used to analyze this type of data. See the section FMODE Statement for the syntax of the FMODE statement. This section describes the analysis of data when units experience multiple failure modes.
The assumptions used in the analysis are
a cause, or mode, can be identified for each failure
failure modes follow a seriessystem model; i.e., a unit fails when a failure due to one of the modes occurs
each failure mode has the specified lifetime distribution with different parameters
failure modes act statistically independently
Suppose there are m failure modes, with lifetime distribution functions .
If you wish to estimate the lifetime distribution of a failure mode, say mode i, acting alone, specify the KEEP keyword in the FMODE statement. The failures from all other modes are treated as rightcensored observations, and the lifetime distribution is estimated by one of the methods described in other sections, such as maximum likelihood. This lifetime distribution is interpreted as the distribution if the specified failure mode is acting alone, with all other modes eliminated. You can also specify more than one mode to KEEP, but the assumption is that all the specified modes have the same distribution.
If you specify the ELIMINATE keyword, failures due to the specified modes are treated as right censored. The resulting distribution estimate is the failure distribution if the specified modes are eliminated.
If you specify the COMBINE keyword, the failure distribution when all the modes specified in the FMODE statement modes act is estimated. The failure distribution , from each individual mode is first estimated by treating all failures from other modes as right censored. The estimated failure distributions are then combined to get an estimate of the lifetime distribution when all modes act,

Pointwise approximate asymptotic normal confidence limits for can be obtained by the delta method. See Meeker and Escobar (1998, appendix B.2). The delta method variance of is, assuming independence of failure modes,

where , y is t for the extreme value, normal, and logistic distributions or for the Weibull, lognormal or loglogistic distributions, and are location and scale parameter estimates for mode i, and and are the standard () survival function and density function for the specified distribution.
Twosided approximate pointwise confidence intervals are computed as in Meeker and Escobar (1998, section 3.6) as

where

where and is the pth quantile of the standard normal distribution.