PROC PROBIT does not use any observations having missing values for any of the independent variables, the response variables, or the weight variable. If only the response variables are missing, statistics requested in the OUTPUT statement are computed.
For binary response data, PROC PROBIT fits the following model by default,
where p is the probability of the response level identified as the first level in the Weighted Frequency Counts for the Ordered Response Categories table in the output and is the normal cumulative distribution function. By default, the covariate vector x contains an intercept term . This is sometimes called Abbot s formula.
Because of the symmetry of the normal (and logistic) distribution, the effect of reversing the order of the two response values is to change the signs of ² in the preceding equation.
By default, response levels appear in ascending , sorted order (that is, the lowest level appears first and then the next lowest, and so on). There are a number of ways that you can control the sort order of the response categories and, therefore, which level is assigned the first ordered level. One of the most common sets of response levels is {0,1}, with 1 representing the event with the probability that is to be modeled .
Consider the example where Y takes the values 1 and 0 for event and nonevent, respectively, and EXPOSURE is the explanatory variable. By default, PROC PROBIT assigns the first ordered level to response level 0, causing the probability of the nonevent to be modeled. There are several ways to change this.
Besides recoding the variable Y, you can
assign a format to Y such that the first formatted value (when the formatted values are put in sorted order) corresponds to the event. For this example, Y=0 could be assigned formatted value ˜nonevent and Y=1 could be assigned formatted value ˜event. Since ORDER=FORMATTED by default, Y=1 becomes the first ordered level. See Example 60.3 for an illustration of this method.
proc format; value disease 1=event 0=nonevent; run; proc probit; model y=exposure; format y disease.; run;
arrange the input data set so that Y=1 appears first and use the ORDER=DATA option in the PROC PROBIT statement. Since ORDER=DATA sorts levels in order of their appearance in the data set, Y=1 becomes the first ordered level. Note that this option causes class variables to be sorted by their order of appearance in the data set, also.
The log- likelihood function is maximized by means of a ridge-stabilized Newton-Raphson algorithm. Initial regression parameter estimates are set to zero. The INITIAL= and INTERCEPT= options in the MODEL statement can be used to give nonzero initial estimates.
The log-likelihood function, L , is computed as
where the sum is over the observations in the data set, w i is the weight for the i th observation, and p i is the modeled probability of the observed response. In the case of the events/trials syntax in the MODEL statement, each observation contributes two terms corresponding to the probability of the event and the probability of its complement:
where r i is the number of events and n i is the number of trials for observation i . This log-likelihood function differs from the log-likelihood function for a binomial or multinomial distribution by additive terms consisting of the log of binomial or multinomial coefficients. These terms are parameter-independent and do not affect the model estimation or the standard errors and tests.
The estimated covariance matrix, V , of the parameter estimates is computed as the negative inverse of the information matrix of second derivatives of L with respect to the parameters evaluated at the final parameter estimates. Thus, the estimated covariance matrix is derived from the observed information matrix rather than the expected information matrix (these are generally not the same). The standard error estimates for the parameter estimates are taken as the square roots of the corresponding diagonal elements of V .
If convergence of the maximum likelihood estimates is attained, a Type III chi-square test statistic is computed for each effect, testing whether there is any contribution from any of the levels of the effect. This statistic is computed as a quadratic form in the appropriate parameter estimates using the corresponding submatrix of the asymptotic covariance matrix estimate. Refer to Chapter 32, The GLM Procedure, and Chapter 11, The Four Types of Estimable Functions, for more information about Type III estimable functions.
The asymptotic covariance matrix is computed as the inverse of the observed information matrix. Note that if the NOINT option is specified and class variables are used, the first class variable contains a contribution from an intercept term. The results are displayed in an ODS table named Type3Analysis .
Chi-square tests for individual parameters are Wald tests based on the observed information matrix and the parameter estimates. If an effect has a single degree of freedom in the parameter estimates table, the chi-square test for this parameter is equivalent to the Type III test for this effect.
In releases previous to Version 8.2, a multiple degree of freedom statistic was computed for each effect to test for contribution from any level of the effect. In general, the Type III test statistic in a main effect only model (no interaction terms) will be equal to the previously computed effect statistic, unless there are collinearities among the effects. If there are collinearities, the Type III statistic will adjust for them, and the value of the Type III statistic and the number of degrees of freedom might not be equal to those of the previous effect statistic.
The theory behind these tests assumes large samples. If the samples are not large, it may be better to base the tests on log-likelihood ratios. These changes in log likelihood can be obtained by fitting the model twice, once with all the parameters of interest and once leaving out the parameters to be tested . Refer to Cox and Oakes (1984) for a discussion of the merits of some possible test methods .
If some of the independent variables are perfectly correlated with the response pattern, then the theoretical parameter estimates may be infinite. Although fitted probabilities of 0 and 1 are not especially pathological, infinite parameter estimates are required to yield these probabilities. Due to the finite precision of computer arithmetic, the actual parameter estimates are not infinite. Indeed, since the tails of the distributions allowed in the PROBIT procedure become small rapidly , an argument to the cumulative distribution function of around 20 becomes effectively infinite. In the case of such parameter estimates, the standard error estimates and the corresponding chi-square tests are not trustworthy.
The distributions, F ( x ), allowed in the PROBIT procedure are specified with the DISTRIBUTION= option in the model statement. The cumulative distribution functions for the available distributions are
The variances of these three distributions are not all equal to 1, and their means are not all equal to zero. Their means and variances are shown in the following table, where ³ is the Euler constant.
Distribution | Mean | Variance |
---|---|---|
Normal |
| 1 |
Logistic |
| 2 /3 |
extreme value or Gompertz | ˆ’ ³ | 2 /6 |
When comparing parameter estimates using different distributions, you need to take into account the different scalings and, for the extreme value (or Gompertz) distribution, a possible shift in location. For example, if the fitted probabilities are in the neighborhood of 0.1 to 0.9, then the parameter estimates from the logistic model should be about / larger than the estimates from the probit model.
The INEST= data set names a SAS data set that specifies initial estimates for all the parameters in the model.
The INEST= data set must contain the intercept variables (named Intercept for binary response model and Intercept , Intercept2 , Intercept3 , and so forth, for multinomial response models) and all independent variables in the MODEL statement.
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. If there is more than one observation in a BY group, the first one read is used for that BY group .
If the INEST= data set also contains the _ TYPE_ variable, only observations with _ TYPE_ value PARMS are used as starting values. Combining the INEST= data set and the option MAXIT= in the MODEL statement, partial scoring can be done, such as predicting on a validation data set by using the model built from a training data set.
You can specify starting values for the iterative algorithm in the INEST= data set. This data set overwrites the INITIAL= option in the MODEL statement, which is a little difficult to use for models with multilevel interaction effects. The INEST= data set has the same structure as the OUTEST= SAS-data-set on page 3762, but is not required to have all the variables or observations that appear in the OUTEST= data set. One simple use of the INEST= option is passing the previous OUTEST= data set directly to the next model as an INEST= data set, assuming that the two models have the same parameterization.
For a two-level response, the probability that the lesser response occurs is modeled by the probit equation as
The probability of the other (complementary) event is 1 ˆ’ p .
For a multilevel response with outcomes labeled l i for i =1 , 2, , k , the probability, p j , of observing level l j is as follows .
Thus, for a k -level response, there are k ˆ’ 2 additional parameters, a 2 , a 3 , , a k “ 1 , estimated. These parameters are denoted by Intercept j , j = 2, 3, , k ˆ’ 1 in the output.
An intercept parameter is always added to the set of independent variables as the first term in the model unless the NOINT option is specified in the MODEL statement. If a classification variable taking on k levels is used as one of the independent variables, asetof k indicator variables is generated to model the effect of this variable. Because of the presence of the intercept term, there are at most k ˆ’ 1 degrees of freedom for this effect in the model.
Two goodness-of-fit tests can be requested from the PROBIT procedure: a Pearson chi-square test and a log-likelihood ratio chi-square test.
To compute the test statistics, you can use the AGGREGATE or AGGREGATE= option grouping the observations into subpopulations. If neither AGGREGATE nor AGGREGATE= is specified, PROC PROBIT assumes that each observation is from a separate subpopulation and computes the goodness-of-fit test statistics only for the events/trials syntax.
If the Pearson goodness-of-fit chi-square test is requested and the p -value for the test is too small, variances and covariances are adjusted by a heterogeneity factor (the goodness-of-fit chi-square divided by its degrees of freedom) and a critical value from the t distribution is used to compute the fiducial limits. The Pearson chi-square test statistic is computed as
where the sum on i is over grouping, the sum on j is over levels of response, the r ij is the frequency of response level j for the i th grouping, n i is the total frequency for the i th grouping, and ij is the fitted probability for the j th level at the i th grouping.
The likelihood ratio chi-square test statistic is computed as
This quantity is sometimes called the deviance . If the modeled probabilities fit the data, these statistics should be approximately distributed as chi-square with degrees of freedom equal to ( k ˆ’ 1) — m ˆ’ q , where k is the number of levels of the multinomial or binomial response, m is the number of sets of independent variable values (covariate patterns), and q is the number of parameters fit in the model.
In order for the Pearson statistic and the deviance to be distributed as chi-square, there must be sufficient replication within the groupings. When this is not true, the data are sparse, and the p -values for these statistics are not valid and should be ignored. Similarly, these statistics, divided by their degrees of freedom, cannot serve as indicators of overdispersion. A large difference between the Pearson statistic and the deviance provides some evidence that the data are too sparse to use either statistic.
One way of correcting overdispersion is to multiply the covariance matrix by a dispersion parameter. You can supply the value of the dispersion parameter directly, or you can estimate the dispersion parameter based on either the Pearson chi-square statistic or the deviance for the fitted model.
The Pearson chi-square statistic and the deviance
are defined in the section Lack of Fit Tests on page 3759. If the SCALE= option is specified in the MODEL statement, the dispersion parameter is estimated by
In order for the Pearson statistic and the deviance to be distributed as chi-square, theremustbesufficient replication within the subpopulations. When this is not true, the data are sparse, and the p -values for these statistics are not valid and should be ignored. Similarly, these statistics, divided by their degrees of freedom, cannot serve as indicators of overdispersion. A large difference between the Pearson statistic and the deviance provides some evidence that the data are too sparse to use either statistic.
You can use the AGGREGATE (or AGGREGATE=) option to define the subpopulation profiles. If you do not specify this option, each observation is regarded as coming from a separate subpopulation. For events/trials syntax, each observation represents n Bernoulli trials, where n is the value of the trials variable; for single-trial syntax, each observation represents a single trial. Without the AGGREGATE (or AGGREGATE=) option, the Pearson chi-square statistic and the deviance are calculated only for events/trials syntax.
Note that the parameter estimates are not changed by this method. However, their standard errors are adjusted for overdispersion, affecting their significance tests.
For a single independent variable, such as a dosage level, the models for the probabilities can be justified on the basis of a population with mean µ and scale parameter ƒ of tolerances for the subjects. Then, given a dose x, the probability, P , of observing a response in a particular subject is the probability that the subject s tolerance is less than the dose or
Thus, in this case, the intercept parameter, b , and the regression parameter, b 1 , are related to µ and ƒ by
Note: The parameter ƒ is not equal to the standard deviation of the population of tolerances for the logistic and extreme value distributions.
In bioassay problems, estimates of the values of the independent variables that yield a desired response are often needed. For instance, the value yielding a 50% response rate (called the ED50 or LD50) is often used. The INVERSECL option requests that confidence limits be computed for the value of the independent variable that yields a specified response. These limits are computed only for the first continuous variable effect in the model. The other variables are set either at their mean values if they are continuous or at the reference (last) level if they are discrete variables. For a discussion of inverse confidence limits, refer to Hubert, Bohidar, and Peace (1988).
For the PROBIT procedure, the response variable is a probability. An estimate of the first continuous variable value needed to achieve a response of p is given by
where F is the cumulative distribution function used to model the probability, x * is the vector of independent variables excluding the first one, which can be specified by the XDATA= option described in the section XDATA= SAS-data-set on page 3763, b * is the vector of parameter estimates excluding the first one, and b 1 is the estimated regression coefficient for the independent variable of interest. Note that, for both binary and ordinal models, the INVERSECL option provides estimates of the value of x 1 yielding Pr(first response level)= p , for various values of p .
This estimator is given as a ratio of random variables, for example, r = a/b . Confidence limits for this ratio can be computed using Fieller s theorem. A brief description of this theorem follows. Refer to Finney (1971) for a more complete description of Fieller s theorem.
If the random variables a and b are thought to be distributed as jointly normal, then for any fixed value r the following probability statement holds if z is an ± / 2 quantile from the standard normal distribution and V is the variance-covariance matrix of a and b .
Usually the inequality can be solved for r to yield a confidence interval. The PROBIT procedure uses a value of 1.96 for z , corresponding to an ± value of 0.05, unless the goodness-of-fit p -value is less than the specified value of the HPROB= option. When this happens, the covariance matrix is scaled by the heterogeneity factor, and a t distribution quantile is used for z .
It is possible for the roots of the equation for r to be imaginary or for the confidence interval to be all points outside of an interval. In these cases, the limits are set to missing by the PROBIT procedure.
Although the normal and logistic distribution give comparable fitted values of p if the empirically observed proportions are not too extreme, they can give appreciably different values when extrapolated into the tails. Correspondingly, the estimates of the confidence limits and dose values can be different for the two distributions even when they agree quite well in the body of the data. Extrapolation outside of the range of the actual data are often sensitive to model assumptions, and caution is advised if extrapolation is necessary.
The OUTEST= data set contains parameter estimates and the log likelihood for the model. You can specify a label in the MODEL statement to distinguish between the estimates for different modeling using the PROBIT procedure. If you specify the COVOUT option, the OUTEST= data set also contains the estimated covariance matrix of the parameter estimates.
The OUTEST= data set contains each variable used as a dependent or independent variable in any MODEL statement. One observation consists of parameter values for the model with the dependent variable having the value ˆ’ 1. If you specify the COVOUT option, there are additional observations containing the rows of the estimated covariance matrix. For these observations, the dependent variable contains the parameter estimate for the corresponding row variable. The following variables are also added to the data set:
_MODEL_ | a character variable containing the label of the MODEL statement, if present, or blank otherwise |
_NAME_ | a character variable containing the name of the dependent variable for the parameter estimates observations or the name of the row for the covariance matrix estimates |
_TYPE_ | a character variable containing the type of the observation, either PARMS for parameter estimates or COV for covariance estimates |
_DIST_ | a character variable containing the name of the distribution modeled |
_LNLIKE_ | a numeric variable containing the last computed value of the log likelihood |
_C_ | a numeric variable containing the estimated threshold parameter |
INTERCEPT | a numeric variable containing the intercept parameter estimates and covariances |
Any BY variables specified are also added to the OUTEST= data set.
The XDATA= data set is used for specifiying values for the effects in the MODEL statement when predicted values and/or fiducial limits for a single continuous variable (dose variable) are required. It is also used for plots specified by the CDFPLOT, IPPPLOT, LPREDPLOT, and PREDPPLOT statement.
The XDATA= data names a SAS data set that contains user input values for all the independent variables in the MODEL statement and the variables in the CLASS statement. The XDATA= data set has the same structure as the DATA= data set but is not required to have all the variables or observations that appear in the DATA= data set.
The XDATA= data set must contain all the independent variables in the MODEL statement and variables in the CLASS statement. Even though variables in the CLASS statement may not be used in the MODEL statement, valid values are required for these variables in the XDATA= data set. Missing values are not allowed. For independent variables in the MODEL statement, although the dose variable s value is not used in the computing of predicted values and/or fiducial limits for the dose variable, missing values are not allowed in the XDATA= data set for any of the independent variables. Missing values are allowed for the dependent variables and other variables if they are included in the XDATA= data set and not listed in the CLASS statement.
If BY processing is used, the XDATA= data set should also include the BY variables, and there must be at least one valid observation for each BY group. If there is more than one valid observation in one BY group, the last one read is used for that BY group.
If there is no XDATA= data set in the PROC PROBIT statement, by default, the PROBIT procedure will use overall mean for effects containing continuous variable (or variables) and the highest level of a single classification variable as reference level. The rules are summarized as follows:
If the effect contains a continuous variable (or variables), the overall mean of this effect is used.
If the effect is a single classification variable, the highest level of the variable is used.
If you request the iteration history (ITPRINT), PROC PROBIT displays
the current value of the log likelihood
the ridging parameter for the modified Newton-Raphson optimization process
the current estimate of the parameters
the current estimate of the parameter C for a natural (threshold) model
the values of the gradient and the Hessian on the last iteration
If you include CLASS variables, PROC PROBIT displays
the numbers of levels for each CLASS variable
the (ordered) values of the levels
the number of observations used
After the model is fit, PROC PROBIT displays
the name of the input data set
the name of the dependent variables
the number of observations used
the number of events and the number of trials
the final value of the log-likelihood function
the parameter estimates
the standard error estimates of the parameter estimates
approximate chi-square test statistics for the test
If you specify the COVB or CORRB options, PROC PROBIT displays
the estimated covariance matrix for the parameter estimates
the estimated correlation matrix for the parameter estimates
If you specify the LACKFIT option, PROC PROBIT displays
a count of the number of levels of the response and the number of distinct sets of independent variables
a goodness-of-fit test based on the Pearson chi-square
a goodness-of-fit test based on the likelihood-ratio chi-square
If you specify only one independent variable, the normal distribution is used to model the probabilities, and the response is binary, PROC PROBIT displays
the mean MU of the stimulus tolerance
the scale parameter SIGMA of the stimulus tolerance
the covariance matrix for MU, SIGMA, and the natural response parameter C
If you specify the INVERSECL options, PROC PROBIT also displays
the estimated dose along with the 95% fiducial limits for probability levels 0.01 to 0.10, 0.15 to 0.85 by 0.05, and 0.90 to 0.99
PROC PROBIT assigns a name to each table it creates. You can use these names to reference the table when using the Output Delivery System (ODS) to select tables and create output data sets. These names are listed in the following table. For more information on ODS, see Chapter 14, Using the Output Delivery System.
ODS Table Name | Description | Statement | Option |
---|---|---|---|
ClassLevels | Class variable levels | CLASS | default |
ConvergenceStatus | Convergence status | MODEL | default |
CorrB | Parameter estimate correlation matrix | MODEL | CORRB |
CovB | Parameter estimate covariance matrix | MODEL | COVB |
CovTolerance | Covariance matrix for location and scale | MODEL | default [ *] |
GoodnessOfFit | Goodness of fit tests | MODEL | LACKFIT |
IterHistory | Iteration history | MODEL | ITPRINT |
LagrangeStatistics | Lagrange statistics | MODEL | NOINT |
LastGrad | Last evaluation of the gradient | MODEL | ITPRINT |
LastHess | Last evaluation of the Hessian | MODEL | ITPRINT |
LogProbitAnalysis | Probit analysis for log dose | MODEL | INVERSECL |
ModelInfo | Model information | MODEL | default |
MuSigma | Location and scale | MODEL | default [ *] |
NObs | Observations Summary | PROC | default |
ParameterEstimates | Parameter estimates | MODEL | default |
ParmInfo | Parameter indices | MODEL | default |
ProbitAnalysis | Probit analysis for linear dose | MODEL | INVERSECL |
ResponseLevels | Response-covariate profile | MODEL | LACKFIT |
ResponseProfiles | Counts for ordinal data | MODEL | default |
Type3Analysis | Type 3 tests | MODEL | default [ *] |
[ *] Depends on data. |