2
$\begingroup$

I would like to be assisted on how to go about choosing the treatments to work with in a fractional factorial the following scenario. I have 3 factors ie 3 feeding intervals, 3 larval densities and 3 total feeds. This makes it a 3 factorial experiment.

How can I go about getting a fractional factorial design. The 27 possible observations are not possible considering that I would like to replicate twice, taking it to 54. I wanted help on how to use fractional factorial method to be able to select at least 16.

$\endgroup$
  • $\begingroup$ This is very short. Can you please give some more details and background, what is the goal of the experiment, ... $\endgroup$ – kjetil b halvorsen Nov 20 '17 at 10:48
  • 1
    $\begingroup$ the experiment aims at getting the best feeding time, density and total feed to give to my larvae for best production. The values on table are; (1) feeding intervals - 1, 2, 3; 3 larval desities - 1.0 kg of larvae/unit area, 1.5 and 2, (c)total feed to give- 90kg, 150kg and 200kg. $\endgroup$ – Davis Nov 20 '17 at 10:54
  • $\begingroup$ Can you please add additional information as an edit to the post? A full factorial has then 3*3*3=27 observations, is that impossible to run? How many runs can you afford? $\endgroup$ – kjetil b halvorsen Nov 20 '17 at 11:02
  • $\begingroup$ The 27 possible observations are not possible considering that I would like to replicate twice, taking it to 54. I wanted help on how to use fractional factorial method to be able to select atleast 16. $\endgroup$ – Davis Nov 20 '17 at 11:20
  • $\begingroup$ I dont have time for a longer answer now: fractioning with 3-level factors is not as straightforward as with 2-level factors. I would go for D-optimal design. In R use package AlgDesign, or have a look at cran.r-project.org/web/views/ExperimentalDesign.html $\endgroup$ – kjetil b halvorsen Nov 20 '17 at 11:26
1
$\begingroup$

You haven't really given enough information, but I will try to give some ideas. One method is to first construct the complete design with $n=3^3$, and then use some algorithm for optimal experimental design, say D-optimality, there are R implementations in the packages AlgDesign and OptimalDesign, on CRAN, see https://CRAN.R-project.org/view=ExperimentalDesign. But here I will exemplify another approach (maybe they can be combined), construction fractional factorial designs directly with algebraic methods , using the CRAN package planor (*). See https://r2012-bordeaux.sciencesconf.org/file/14501 for more details and examples!

library(planor)
ex1key  <-  planor.designkey(factors=c("block", LETTERS[1:3]),
                             nlevels = c(3, rep(3, 3)),
                             block = ~ block,  
                             model= ~ block+ (A+B+C)^2,
                             estimate= ~ A + B + C ,
                             nunits=3*3^2, max.sol=1)
          Preliminary step 1 : processing the model specifications
Preliminary step 2 : performing prime decompositions on the factors
*** Main step for prime p = 3 : key-matrix search
  => search for columns 2 to 4 
      first visit to column 2
      first visit to column 3
    ---    col. 3 ( j = 2) 22 selected candidates
      first visit to column 4
    ---    col. 4 ( j = 3) 18 selected candidates
The search is closed: max.sol = 1 solution(s) found 

summary(ex1key)

********** Prime  3  design **********

--- Solution  1  for prime  3  ---

TREATMENT EFFECTS CONFOUNDED WITH THE MEAN
nil

BLOCK-and-TREATMENT EFFECTS CONFOUNDED WITH THE MEAN
1 = block^2  A^2  B

WEIGHT PROFILES
Treatment effects confounded with the mean: none 
Treatment effects confounded with block effects: 2^1 
Treatment pseudo-effects confounded with the mean: none
Treatment pseudo-effects confounded with block effects: 2^1 

ex1Des
An object of class "planordesign"
Slot "design":
   block A B C
1      1 1 1 1
2      1 1 1 2
3      1 1 1 3
4      1 2 2 1
5      1 2 2 2
6      1 2 2 3
7      1 3 3 1
8      1 3 3 2
9      1 3 3 3
10     2 1 2 1
11     2 1 2 2
12     2 1 2 3
13     2 2 3 1
14     2 2 3 2
15     2 2 3 3
16     2 3 1 1
17     2 3 1 2
18     2 3 1 3
19     3 1 3 1
20     3 1 3 2
21     3 1 3 3
22     3 2 1 1
23     3 2 1 2
24     3 2 1 3
25     3 3 2 1
26     3 3 2 2
27     3 3 2 3

Slot "factors":
An object of class "designfactors"
Slot "fact.info":
      nlev block ordered model basic dummy
block    3  TRUE   FALSE  TRUE FALSE FALSE
A        3 FALSE   FALSE  TRUE FALSE FALSE
B        3 FALSE   FALSE  TRUE FALSE FALSE
C        3 FALSE   FALSE  TRUE FALSE FALSE

Slot "pseudo.info":
      parent nlev block ordered model basic dummy
block      1    3  TRUE   FALSE  TRUE FALSE FALSE
A          2    3 FALSE   FALSE  TRUE FALSE FALSE
B          3    3 FALSE   FALSE  TRUE FALSE FALSE
C          4    3 FALSE   FALSE  TRUE FALSE FALSE

Slot "levels":
$block
[1] 1 2 3

$A
[1] 1 2 3

$B
[1] 1 2 3

$C
[1] 1 2 3



Slot "model":
[[1]]
[[1]]$Model
~block + (A + B + C)^2

[[1]]$Estimate
~A + B + C



Slot "designkey":
[[1]]
An object of class keymatrix

********** Prime  3  design **********

    block A B C
*U*     1 0 1 0
*U*     0 1 1 0
*U*     0 0 0 1



Slot "nunits":
[1] 27

Slot "recursive":
[1] FALSE

(*) That package seems to be still under development and somewhat unfinished, and so is not the easiest package to use. Do not misunderstand, the package is a very ambitious project, and bound to be very useful. But at this moment it is not very userfriendly!

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.