Regression Model Observation-Wise Statistics

The RELIABILITY Procedure

Regression Model Observation-Wise Statistics

For regression models that are fit using the MODEL statement, you can specify a variety of statistics to be computed for each observation in the input data set. This section describes the method of computation for each statistic. See Table 30.21 and Table 30.22 for the syntax for requesting these statistics.

Predicted Values

The linear predictor is

$\hat{\mu}_{i}={x_{i}}'{{\beta}}$

where x_i is the vector of explanatory variables for the ith observation.

Percentiles

An estimator of the p×100% percentile x_p for the ith observation for the extreme value, normal, and logistic distributions is

$\hat{x}_{i,p}={x_{i}}'\hat{{\beta}}+ z_{p}\hat{\sigma}$

where z_p=G^-1(p), G is the standardized CDF, and $\sigma$ is the distribution scale parameter.

An estimator of the p×100% percentile t_p for the ith observation for the Weibull, lognormal, and log-logistic distributions is

$\hat{t}_{i,p}=\exp[{x_{i}}'\hat{{\beta}}+ z_{p}\hat{\sigma}]$

where G is the standardized CDF of the extreme value, normal, or logistic distribution that corresponds to the logarithm of the lifetime, and $\sigma$ is the distribution scale parameter.

The percentile of the lognormal (base 10) distribution is

$\hat{t}_{i,p}=10^{[{x_{i}}'\hat{{\beta}}+ z_{p}\hat{\sigma}]}$

where G is the CDF of the standard normal distribution.

An estimator of the p×100% percentile t_p for the ith observation for the generalized gamma distribution is

$\hat{t}_{i,p} = \exp[{x_{i}}'\hat{{\beta}}+ w_{\lambda,p}\hat{\sigma}]$

where

$w_{\lambda,p} =\frac{1}{\lambda} \log(\frac{\lambda^2}2\chi^2_{(2/\lambda^2),p})$

and $\chi^2_{k,p}$ is the p×100% percentile of the chi-squared distribution with k degrees of freedom.

Standard Errors of Percentile Estimator

For the extreme value, normal, and logistic distributions, the standard error of the estimator of the p×100% percentile is computed as

$\sigma_{i,p}=\sqrt{z'{\Sigma z}}$

where

$z=[ x_{i} \ z_{p} ]$

and ${\Sigma}$ is the covariance matrix of $({\hat{\beta}}, \hat{\sigma})$ .

For the Weibull, lognormal, and log-logistic distributions, the standard error is computed as

$\sigma_{i,p}=\exp(x_{i,p})\sqrt{z'{\Sigma z}}$

where x_i,p is the percentile computed from the extreme value, normal, or logistic distribution that corresponds to the logarithm of the lifetime. The standard error for the lognormal (base 10) distribution is computed as

$\sigma_{i,p}=10^{x_{i,p}}\sqrt{z'{\Sigma z}}$

The standard error for the generalized gamma distribution percentile is computed as

$\sigma_{i,p} = \exp[{x_{i}}'\hat{{\beta}}+ w_{\lambda,p}\hat{\sigma}]\sqrt{z'{\Sigma z}}$

where

$z=[ x_{i} \ w_{\lambda,p} \ \hat{\sigma}\frac{\partial w_{\lambda,p}}{\partial \lambda} ]$

${\Sigma}$ is the covariance matrix of $({\hat{\beta}}, \hat{\sigma}, \hat{\lambda})$ , ${\beta}$ is the vector of regression parameters, $\sigma$ is the scale parameter, and $\lambda$ is the shape parameter.

Confidence Limits for Percentiles

Two-sided approximate $100\gamma\%$ confidence limits for x_i,p for the extreme value, normal, and logistic distributions are computed as

$x_{L} & = & \hat{x}_{i,p} - K_{\gamma}\sigma_{i,p} \ x_{U} & = & \hat{x}_{i,p} + K_{\gamma}\sigma_{i,p}$

where $K_{\gamma}$ represents the $100(1+\gamma)/2 x 100\%$ percentile of the standard normal distribution.

Limits for the Weibull, lognormal, and log-logistic percentiles are computed as

$t_{L} & = & \exp(x_{L}) \ t_{U} & = & \exp(x_{U})$

where x_L and x_U are computed from the corresponding distributions for the logarithms of the lifetimes. For the lognormal (base 10) distribution,

$t_{L} & = & 10^{x_{L}} \ t_{U} & = & 10^{x_{U}}$

Limits for the generalized gamma distribution percentiles are computed as

$t_{L} = \exp[{x_{i}}'\hat{{\beta}}+ w_{\lambda,p}\hat{\sigma} - K_{\gamma}\sqrt{... ...x_{i}}'\hat{{\beta}}+ w_{\lambda,p}\hat{\sigma} + K_{\gamma}\sqrt{z'{\Sigma z}}]$

Reliability Function

For the extreme value, normal, and logistic distributions, an estimate of the reliability function evaluated at the response y_i is computed as

$R(y_{i}) = 1 - G(\frac{y_{i}-{x_{i}}'\hat{{\beta}}}{\hat{\sigma}})$

where G(x) is the standardized CDF of the distribution from Table 30.47.

Estimates of the reliability function evaluated at the response t_i for the Weibull, lognormal, log-logistic, and generalized gamma distributions are computed as

$R(t_{i}) = 1 - G(\frac{\log(t_{i})-{x_{i}}'\hat{{\beta}}}{\hat{\sigma}})$

where G(x) is the standardized CDF of the corresponding extreme value, normal, logistic, or generalized log-gamma distributions.

Residuals

The RELIABILITY procedure computes several different kinds of residuals. In the following equations, y_i represents the ith response value if the extreme value, normal, or logistic distributions are specified. If t_i is the ith response and if the Weibull, lognormal, log-logistic, or generalized gamma distributions are specified, then y_i represents the logarithm of the response y_i = log(t_i). If the lognormal (base 10) distribution is specified, then y_i = log₁₀(t_i).

Raw Residuals

The raw residual is computed as

$r_{Ri}=y_{i}-{x_{i}}'\hat{{\beta}}$

Standardized Residuals

The standardized residual is computed as

$r_{Si}=\frac{y_{i}-{x_{i}}'\hat{{\beta}}}{\hat{\sigma}}$

Adjusted Residuals

If an observation is right censored, then the standardized residual for that observation is also right censored. Adjusted residuals adjust censored standardized residuals upward by adding a percentile of the residual lifetime distribution, given that the standardized residual exceeds the censoring value. The default percentile is the median (50th percentile), but you can, optionally, specify a $\gammax 100\%$ percentile using the RESIDALPHA= $\gamma$ option in MODEL statement. The $\gamma x 100$ percentile residual life is computed as in Joe and Proschan (1984). The adjusted residual is computed as

$r_{Ai}= \{ G^{-1}[1-(1-\gamma)S(u_{i})] & { for right-censored observations } \ u_{i} & { for uncensored observations } .$

where G is the standard CDF,

S(u)=1-G(u)

is the reliability function, and

$u_{i} = \frac{y_{i}-{x_{i}}'\hat{{\beta}}}{\hat{\sigma}}$

If the generalized gamma distribution is specified, the standardized CDF and reliability functions include the estimated shape parameter $\hat{\lambda}$ .

Modified Cox-Snell Residuals

Let

$\delta_{i}= \{ 1 & { for uncensored observations } \ 0 & { for right-censored observations } .$

The Cox-Snell residual is defined as

r_Ci = -log(R(y_i))

where

$R(y)=1-G(\frac{y-{x_{i}}'\hat{{\beta}}}{\hat{\sigma}})$

is the reliability function. The modified Cox-Snell residual is computed as in Collett (1994, p.152):

$r_{Ci}^{'} = r_{Ci} + (1-\delta_{i})\alpha$

where $\alpha$ is an adjustment factor. If the fitted model is correct, the Cox-Snell residual has approximately a standard exponential distribution for uncensored observations. If an observation is censored, the residual evaluated at the censoring time is not as large as the residual evaluated at the (unknown) failure time. The adjustment factor $\alpha$ adjusts the censored residuals upward to account for the censoring. The default is $\alpha=0.693$ , the median of the standard exponential distribution. You can, optionally, specify any adjustment factor by using the MODEL statement option RESIDADJ= $\alpha$ . Another commonly used value is the mean of the standard exponential distribution, $\alpha=1$ .

Deviance Residuals

Deviance residuals are a zero-mean, symmetrized version of modified Cox-Snell residuals. Deviance residuals are computed as in Collett (1994, p.153):

$r_{Di}={\rm sgn}(\delta_{i}-r_{Ci}) \{-2[\delta_{i}-r_{Ci}+\delta_{i}\log(r_{Ci})]\}^{1/2}$

where

${sgn}(u) = \{ -1 & { if } u \lt 0 \ 1 & { if } u \ge 0 .$

Chapter Contents
Previous
Next
Top