Gives t-tests for all pairwise differences of means from a regression or generalized linear model (J.T.N.M. Thissen & P.W. Goedhart).

### Options

`PRINT` = string tokens |
What to print (`differences` , `sed` , `tvalues` , `tprobabilities` ); default `diff` , `sed` , `tval` |
---|---|

`SORT` = string token |
Whether to sort the means into ascending order (`no` , `yes` ); default `no` |

`COMBINATIONS` = string token |
Which combinations of factors in the current model to include (`full` , `present` , `estimable` ); default `esti` (similar to the `PREDICT` directive) |

`ADJUSTMENT` = string token |
Type of adjustment with linear regression models (`marginal` , `equal` ); default `marg` (similar to the `PREDICT` directive) |

`WEIGHTS` = table |
Weights classified by some or all standardizing factors; default `*` (similar to the `PREDICT` directive) |

`METHOD` = string token |
Method of forming margin (`mean` , `total` ); default mean (similar to the `PREDICT` directive) |

`ALIASING` = string token |
How to deal with aliased parameters (`fault` , `ignore` ); default `faul` (similar to the `PREDICT` directive) |

`SAVE` = identifier |
Specifies save structure of model to display; default `*` (i.e. that of the latest model fitted) |

### Parameters

`TREATFACTORS` = pointers |
Each pointer contains a list of treatment factors classifying the table of means to be compared (the right-most factor changes fastest, then the second from the right, etc.); this parameter must be set |
---|---|

`LABELS` = texts |
Structures containing strings to label rows (and columns) of the symmetric matrices of pairwise differences etc; the length of the text must equal the product of the numbers of factor levels as implied by the factor list in the `TREATFACTORS` pointer |

`NEWLABELS` = texts |
To save the row labels of the `DIFFERENCES` , `SED` , `TVALUES` and `TPROBABILITIES` matrices |

`DIFFERENCES` = symmetric matrices |
To save pairwise differences (treatment means on the diagonal) |

`SED` = symmetric matrices |
To save standard errors of the pairwise differences (missing values on the diagonal) |

`TVALUES` = symmetric matrices |
To save t-values (missing values on the diagonal) |

`TPROBABILITIES` = symmetric matrices |
To save t-probabilities (missing values on the diagonal) |

### Description

When analysing a (non-orthogonal) analysis of variance model or a generalized linear model (GLM) with the regression directives `FIT`

, `ADD`

etc., effects of factors in the model and their interactions if required, may be assessed from a suitable analysis of variance (deviance) table. With the `PREDICT`

directive tables of estimated means and their standard errors can be obtained, but not standard errors of differences of means. The `RPAIR`

procedure provides additional information on such tables by calculating t-values and corresponding two-sided t-probabilities for tests of all pairwise differences of means.

The t-statistics used are based on the residual variance (deviance) and its degrees of freedom from the current regression model. However, if the `DISPERSION`

option of the `MODEL`

directive has been set to a numerical value (as is by default the case with a GLM with binomial, poisson or multinomial distribution), the degrees of freedom are set to 10000, approximating to the normal distribution.

It is assumed that the `MODEL`

statement for the regression has defined only one response variate.

The `TREATFACTORS`

parameter must be set to a pointer containing a list of factors classifying the table of means which are to be compared.

The `PRINT`

option controls the output. By default a symmetric matrix of pairwise differences of means is printed with the means themselves down the diagonal. With a GLM these means and their pairwise differences are always calculated on the linear scale. The corresponding symmetric matrices of standard errors and of t-values are printed by default too.

The matrix rows (and columns) are ordered such that the right-most factor changes fastest, then the second from the right, etc. This default order can be changed by setting the `SORT`

option to `yes`

, in which case rows and columns of all matrices are rearranged to put the means on the diagonal of the matrix of differences into ascending order.

The `LABELS`

parameter can be used to label the rows and columns of the matrices, which are then taken in default order. When the `LABELS`

parameter has not been set and the `TREATFACTORS`

pointer contains just one factor, by default the labels or levels of the factor are used for labeling; when the pointer contains more than one factor, the default row (and column) labels are combinations of factor settings indicated by the first letter of the factor identifier followed by an ordinal level.

The `DIFFERENCES`

, `SED`

, `TVALUES`

and `TPROBABILITIES`

parameters can be used to save the output. The row labels of these matrices can be saved through the `NEWLABELS`

parameter.

The `COMBINATIONS`

, `ADJUSTMENT`

, `WEIGHTS`

, `METHOD`

, `ALIASING`

and `SAVE`

options are as in the `PREDICT`

directive.

Options: `PRINT`

, `SORT`

, `COMBINATIONS`

, `ADJUSTMENT`

, `WEIGHTS`

, `METHOD`

, `ALIASING`

, `SAVE`

.

Parameters: `TREATFACTORS`

, `LABELS`

, `NEWLABELS`

, `DIFFERENCES`

, `SED`

, `TVALUES`

, `TPROBABILITIES`

.

### Method

The procedure uses the `PREDICT`

directive to save a table of predictions and corresponding variance-covariance matrix. With a GLM the setting of option `BACKTRANSFORM`

of the `PREDICT`

directive is always `none`

.

### Action with `RESTRICT`

Any restrictions applied to vectors used in the regression apply also to the results from `RPAIR`

.

### See also

Procedures: `ALLDIFFERENCES`

, `AMCOMPARISON`

, `AUMCOMPARISON`

, `PAIRTEST`

, `PPAIR`

.

Commands for: Regression analysis.

### Example

CAPTION 'RPAIR example',\ !t('1) Data from Snedecor, G.W. & Cochran, W.G.',\ '(1976). Statistical Methods (6th edition).',\ 'Iowa State University Press. Ames. page 480.'); STYLE=meta,plain FACTOR [LEVELS=5; VALUES=18(1),45(2),6(3),9(4),15(5)] Breed FACTOR [LABELS=!T(Male,Female); VALUES=12(1),6(2),30(1),15(2),\ 4(1),2(2),6(1),3(2),10(1),5(2)] Sex VARIATE [NVALUES=93] %Dressng READ %Dressng 13.3 12.6 11.5 15.4 12.7 15.7 13.2 15.0 14.3 16.5 15.0 13.7 18.2 11.3 14.2 15.9 12.9 15.1 10.9 3.3 10.5 11.6 15.4 14.4 11.6 14.4 7.5 10.8 10.5 14.5 10.9 13.0 15.9 12.8 14.0 11.1 12.1 14.7 12.7 13.1 10.4 11.9 10.7 14.4 11.3 13.0 12.7 12.6 14.3 15.3 11.8 11.0 10.9 10.5 12.9 12.5 13.0 7.6 12.9 12.4 12.8 10.9 13.9 13.6 13.1 4.1 10.8 12.9 14.4 11.6 13.2 12.6 15.2 14.7 12.4 13.8 14.4 4.9 10.3 10.3 10.1 6.9 13.2 11.0 12.2 13.3 12.9 9.9 12.8 8.4 10.6 13.9 10.0 : MODEL %Dressng FIT Breed * Sex RPAIR [PRINT= d, s, tv, tp] !P( Breed), !P( Sex, Breed) RPAIR [PRINT= d] !P( Breed), !P( Sex, Breed);\ LABELS= !T( Breed1, Breed2, Breed3, Breed4, Breed5),\ !T( MalBr1, MalBr2, MalBr3, MalBr4, MalBr5,\ FemBr1, FemBr2, FemBr3, FemBr4, FemBr5) " The next analysis of variance has been added for comparison." TREATMENTS Breed * Sex ANOVA [FPROBABILITY= yes] %Dressng CAPTION !t('2) Data from Snedecor & Cochran (1980), Statistical Methods',\ '(7th edition), page 289; also used by M.S. Ridout',\ 'in Genstat Newsletter 20, page 22.') FACTOR [NVALUES= 25; LEVELS= 5; VALUES= 5(1...5)] Rows FACTOR [LEVELS= 5; VALUES= (1...5)5] Cols FACTOR [LABELS= !T( K, M, N, O, P)] Treatmnt VARIATE [NVALUES= 25] Wireworm READ Treatmnt, Wireworm; FREPRESENTATION= labels P 3 O 2 N 5 K 1 M 4 M 6 K 0 O 6 N 4 P 4 O 4 M 9 K 1 P 6 N 5 N 17 P 8 M 8 O 9 K 0 K 4 N 4 P 2 M 4 O 8 : MODEL [DISTRIBUTION= poisson] Wireworm FIT Rows + Cols + Treatmnt RPAIR [PRINT= d, s, tv, tp] !P(Treatmnt) " normal probabilities " MODEL [DISTRIBUTION= poisson; DISPERSION= *] Wireworm FIT Rows + Cols + Treatmnt RPAIR [PRINT= d, s, tv, tp] !P(Treatmnt) " t probabilities " RPAIR [SORT= yes; PRINT= *] !P(Treatmnt); NEWLABELS= Labels;\ DIFF= Diff; SED= Sed; TVALUES= Tval; TPROB= Pval PRINT [RLPRINT=*] Labels & [SERIAL= yes; RLPRINT=labels] Diff, Sed, Tval, Pval