Searches for an efficient partially-replicated design (R.W. Payne).
|Controls printed output (
||Levels of the treatment factor; if unset, takes the levels declared for the factor specified by the
||Number of times each treatment occurs in the design|
||Number of blocks|
||Saves the efficiency factors of the treatment term within blocks|
||Specifies the number of random starting configurations to take in the search for the best design; default 5|
||Number of designs to try from each starting configuration; default 20|
||Seed for the random numbers used to randomize the design; default 0|
||Seed for the random numbers used to select the random starting configurations; default 0|
||Whether to put the design factors into a spreadsheet (
Partially replicated designs can be used when a design is too small to provide more than one unit for every treatment. So some treatments occur on only one unit in the design, others occur on two units, and there may be others (usually control treatments) that occur on several. The designs can thus provide an effective way of screening large number of treatments.
NREPEATS option specifies the number of repeats (i.e. replicates) of each treatment in the design, in a variate with the same number of values as the number of treatments. Also, the sum of its values defines the number of units in the design.
LEVELS option can be used to define the treatment levels, as in the
FACTOR directive, and the
TREATMENTS option can save a factor containing the generated values.
LEVELS can be omitted if the
TREATMENTS factor has already been declared with the right levels.
NBLOCKS option can be used to specify the number of blocks in the design, and the
BLOCKS option can save a factor containing the generated values.
NBLOCKS can be omitted if the
BLOCKS factor has already been declared with the right number of levels. Note that, if the number of units in the design is not an exact multiple of the number blocks, some blocks will contain one fewer unit than others.
UNITS option can supply a factor to save the values generated for the unit-within-block factor (which identifies the units within each block).
SEED option allows you to set the seed to be used to randomize the design. The default setting of zero continues the sequence of random numbers from those used to select the random starting configurations. The
TRYSEED options control the way in which
AFPREP searches for the best design, as described in the Method Section.
Printed output is controlled by the
||to print the design,|
||to print the harmonic mean and the range of values of the treatment efficiency factors,|
||to print the factor values, and|
||to provide monitoring information during the search.|
During monitoring the current best design is marked by an asterisk (
You can set option
SPREADSHEET=design to put the design factors into a Genstat spreadsheet.
AFPREP uses the
AEFFICIENCY procedure to calculate the within-block efficiency factors of the treatments. If there are fewer than 201 treatments, all their efficiency factors are calculated. This is not feasible, however, when there are more treatment. So the unreplicated treatments are then ignored.
AFPREP chooses the best candidate design by firstly taking the design with the largest minimum efficiency factor (i.e. it tries to avoid having a low efficiency for any treatment contrast). Then, if there are several designs with the same minimum efficiency factor, it takes the design with the largest harmonic mean efficiency factor; this aims to minimize the (ordinary, arithmetic) average standard error of difference between pairs of treatments. The efficiency factors of the best design can be saved using the
The candidate designs are generated in a way that avoids any treatment occurring more times than are necessary in the same block. This is done by generating block and unit values with the block factor as the fastest moving factor (i.e. block values that are repeated sequences of 1, 2…), and generating the treatment factor with all the repeated levels together. So the designs that
AFPREP considers will differ according to the way in which the different levels are ordered within the treatment factor. It can run through these orderings systematically but, unless there are very few treatment levels, there will be too many orderings to examine them all. So it pursues a mixed strategy, running through a systematic sequence of orderings from several random starting arrangements. The
NSTARTS option specifies the number of random starts to make, and the
NTRIES option specifies the number of designs to examine in each sequence. The seed for the random numbers used to select the random starts is specified by the
TRYSEED option. The default value of zero continues the existing sequence of random numbers if any have been used already in this run of Genstat. Otherwise, it initializes the seed automatically.
Commands for: Design of experiments.
CAPTION 'AFPREP example',\ !t('Design for 129 treatments: 96 with a single replicate,',\ '28 with 2 replicates, 3 controls with 6 replicates,',\ 'and 2 controls with 5 replicates.'); STYLE=meta,plain VARIATE [VALUES=96(1),28(2),3(6),2(5)] nreps AFPREP [PRINT=design,monitor,efficiency; LEVELS=129;\ NREPEATS=nreps; NBLOCKS=6; TREATMENTS=Treat; BLOCKS=Block;\ SEED=104623; TRYSEED=2389634]