This help page documents the statistical and algorithmic details of blocking in FrF2
Ulrike Groemping
Blocking is done with the purpose to balance the design with respect to a factor that is known or strongly suspected to have an influence but is not in itself of interest, and it is usually assumed that block factors do not interact with experimental factors. Examples are batches of material that are not large enough to accomodate the complete experiment so that e.g. half the experiment is done on the first batch and the other half on the second batch (two blocks). The block factor should be orthogonal to the experimental factors, at least to their main effects. Per default, it is also requested that the block factor is orthogonal to the 2-factor interactions. This can be changed by the user, if no such design can be found.
Blocking is currently implemented for regular fractional factorial designs only.
There are two principal ways to handle blocked designs, manual definition
(i.e. the user specifies exactly which columns are to be used for which purpose) and automatic
definition. Each situation has its specifics. These are detailed below. For users with
not so much mathematical/statistical background, it will often be best to use the automatic way,
specifying the treatment factors of interest via nfactors or factor.names
and a single number for blocks or WPs.
Users with more mathematical background may want to use the manual definitions, perhaps
in conjunction with published catalogues of good block designs, or
after inspecting possibilities with functions blockpick, blockpick.big (default before version 2 for large settings) or colpick (default since version 2 for large settings or settings with estimability requirements).
The user can start from a design with a number of factors and manually specify which factors or
interactions are to be used as block generators. If this route is chosen, blocks
can be a vector of factor names or factor letters, or of the same form as generators, except that
not only base factors but all factors can be used and single factors are permitted
(which would lead to resolution II designs if used in generators). For example,
block = Letters[c(2,4,5)]
or
block = list(2,4,5)
specify that the 2nd, 4th and 5th factor are to be used as block generators, while
block = c("Day","Shift")
indicates that the named factors “Day” and “Shift” specified in factor.names
are to be treated as blocking factors). In this case, the number of blocks is calculated,
and a new factor with the default name “Blocks” (in general the name chosen in
option block.name) is generated, which would for example contain as levels
the Day/Shift combinations. It is also possible to choose interaction effects rather than factors themselves
as block generators, e.g.
block = c("ABCD","EFGH")
or
block = list(c(1,2,3,4),c(5,6,7,8)) .
Finally, it is also possible to specify choice of blocks using a vector of Yates column numbers,
in order to be able to use catalogued blocking structures of this form, e.g. from Sitter, Chen and Feder
(1997).
The block main effects are defined by the k.block specified effect
and all interactions between them. The specified block effects are required to be independent from each other,
which implies that they generate 2^k.block blocks.
CAUTION: If the user manually generates a blocked design, it is his/her responsibility to ensure a
good choice of design (e.g. by using a catalogued design from Bisgaard 1994,
Sun, Wu and Chen 1997, Sitter, Chen and Feder (1997), or Cheng and Wu 2002).
Since version 2 of package FrF2, manual blocking is also checked for confounding
of the block factor with main effects or two-factor interactions;
this implies that some earlier code will now require the additional specification
of argument alias.block.2fis=TRUE in order to avoid errors.
If the user only specifies the number of blocks required for the experiment, function FrF2
automatically generates the blocks. For full factorial designs, function FrF2 uses function colpick
with subsequent blockgencreate, except where the Sun, Wu and Chen (1997) catalogue of blocked designs
contains suitable block generators for a design without estimability requirements
(implemented in function blockpick, which also calls colpick, if that
catalogue does not offer a solution).
Otherwise, depending on the situation,
function FrF2 uses function blockpick or function colpick
with subsequent blockgencreate;
function blockpick treats smaller problems (choose(nruns-1-nfactors,k.block) < 100000)
without estimability requirements and with force.godolphin=FALSE (the latter is per default set to TRUE
whenever alias.block.2fis=TRUE),
other problems are treated by function colpick.
Use of the earlier default function blockpick.big for large cases or the earlier behavior for full factorial designs
can be requested with the argument
block.old=TRUE; this should only be done for reproducing earlier results, as the new methodology is definitely superior.
The search for an appropriate blocked design starts with the overall best unblocked design
(in terms of aberration or MaxC2, if requested).
If this best design does not yield an adequate blocking possibility, the search continues with
the next best design and so forth (exception: with an estimability requirement, only a single design, prefiltered for the estimability requirement, is subjected to the blocking algorithm).
For the smaller problems, function blockpick
looks for k.block independent subsets among the eligible columns of the design.
(The eligible columns are all columns of the Yates matrix that are neither occupied
by treatment main effects nor by 2fis among treatments (if alias.block.2fis=FALSE,
which is the default), or all columns of the Yates matrix that are not occupied by treatment main effects
(if alias.block.2fis=TRUE). Note that no effort is made to avoid aliasing with 2-factor interactions,
if alias.block.2fis=TRUE is chosen.
For the larger problems, or blocking in combination with requiring some 2fis to be clear of aliasing,
or per default for blocking with permitting 2fis to be aliased with blocks,
function colpick creates a \(q \times n\) X matrix for creating
blocks of size \(2^q\) based on the approach described by Godolphin (2021);
function blockgencreate creates block generators from this matrix.
This approach can be used in combination with argument estimable,
as long as clear=TRUE. The implementation of this approach is described in Groemping (2021).
The argument force.godolphin of function FrF2 can enforce the Godolphin approach instead
of the default approach for small blocked designs without alias.block.2fis=TRUE,
and the Godolphin approach can be switched off for alias.block.2fis=TRUE applications by explicitly requesting
force.godolphin=FALSE. Note that the Godolphin approach solely focuses on clear 2fis
of the blocked design and does not attempt to avoid confounding of the block factor with non-clear 2fis; it may
thus confound 2fis with the block factor even if this were avoidable, maintaining the same number of clear 2fis.
For the larger problems, in versions before 2.0, which can be activated in current versions with block.old=TRUE,
function blockpick.big permutes the k~base factors of candidate designs with nfactors + k.block factors
in search of a design the first k.block~factors of which can be used for block construction. Any
specification of design (via options design or generators) is ignored. Note that function
blockpick.big is not guaranteed to find an existing blocked design.
Sun, Wu and Chen (1997) provided a catalogue of blocked designs
with a few quality criteria, and they stated that there is no single best design, but that the choice
depends on the situation. FrF2 always comes up with one specific solution design.
Comparisons to the catalogued designs in Sun, Wu and Chen (1997) have shown that
the designs found in FrF2 are often but not always isomorphic to the catalogued ones.
Differences do occur, especially if the base designs are resolution III, or if blockpick.big
has to be used. Expert users who want to be certain to use a “best” blocked design should manually
implement a specific catalogued design or inspect several solutions from functions blockpick or colpick (or, if desparate, blockpick.big).
Please contact me with any suggestions for improvements.
Bisgaard, S. (1994a). Blocking generators for small \(2^{k-p}\) designs. J. Quality Technology 26, 288-294.
Chen, J., Sun, D.X. and Wu, C.F.J. (1993) A catalogue of 2-level and 3-level orthogonal arrays. International Statistical Review 61, 131-145.
Cheng, C.-S. and Tsai, P.-W. (2009). Optimal two-level regular fractional factorial block and split-plot designs. Biometrika 96, 83-93.
Cheng, S.W. and Wu, C.F.J. (2002). Choice of optimal blocking schemes in 2-level and 3-level designs. Technometrics 44, 269-277.
Godolphin, J. (2021). Construction of Blocked Factorial Designs to Estimate Main Effects and Selected Two-Factor Interactions. J. Royal Statistical Society B 83, 5-29. tools:::Rd_expr_doi("10.1111/rssb.12397").
Groemping, U. (2019). An algorithm for blocking regular fractional factorial 2-level designs with clear two-factor interactions. Reports in Mathematics, Physics and Chemistry, Report 3/2019, Department II, Beuth University of Applied Sciences Berlin.
Sitter, R.R., Chen, J. and Feder, M. (1997). Fractional Resolution and Minimum Aberration in Blocked 2n-k Designs. Technometrics 39, 382--390.
Sun, D.X., Wu, C.F.J. and Chen, Y.Y. (1997). Optimal blocking schemes for \(2^p\) and \(2^{n-p}\) designs. Technometrics 39, 298-307.
See Also FrF2 for regular fractional factorials,
catlg for the Chen, Sun, Wu catalogue of designs
and some accessor functions,
and splitplot for the statistical aspects of split-plot designs.
########## automatic blocked designs ###################
## from a full factorial ##
FrF2(8,3,blocks=2)
## with replication
run.order(FrF2(8,3,blocks=2,wbreps=2))
run.order(FrF2(8,3,blocks=2,wbreps=2,repeat.only=TRUE))
run.order(FrF2(8,3,blocks=2,bbreps=2))
run.order(FrF2(8,3,blocks=2,bbreps=2,wbreps=2))
## automatic blocked design with fractions
FrF2(16,7,blocks=4,alias.block.2fis=TRUE)
## isomorphic non-catalogued design as basis
FrF2(16,gen=c(7,11,14),blocks=4,alias.block.2fis=TRUE)
## FrF2 uses blockpick.big and ignores the generator
FrF2(64,gen=c(7,11,14),blocks=16,alias.block.2fis=TRUE)
########## manual blocked design ####################
### example that shows why order of blocks is not randomized
### can of course be randomized by user, if appropriate
FrF2(32,9,blocks=c("Day","Shift"),alias.block.2fis=TRUE,
factor.names=list(Day=c("Wednesday","Thursday"), Shift=c("Morning","Afternoon"),
F1="",F2="",F3="",F4="",F5="",F6="",F7=""), default.levels=c("current","new"))
########## blocked design with estimable 2fis ####################
### all interactions of last two factors to be estimable clearly
### in 64 run design with blocks of size 4
### not possible with catalogue entry 9-3.1
FrF2(design="9-3.2", blocks=16, alias.block.2fis=TRUE,
factor.names = list(C1="",C2="",C3="",C4="",C5="",C6="",C7="",
N1=c("low","high"),N2=c("low","high")),
default.levels = c("current","new"),
estimable=compromise(9, 8:9)$requirement)
Run the code above in your browser using DataLab