BuiltWithNOF
7. Randomisation

Randomisation is essential for any experiment as it minimises the chance of a biased result. It is the ability to randomise subjects to treatments that distinguishes experiments from surveys. Experiments are designed to detect causation (i.e. whether the treatment causes an effect), surveys can only detect association.

                                                  

About randomisation


Why randomisation is essential
Methods of randomisation
Restricted randomisation
"Improving" on randomisation

animal4
C57BL/6 mice trying to randomise (but you could do it better).
 
 
 

 
Why randomisation is essential


Randomisation is fundamental to designing good experiments as it substantially reduces the chance of bias which could arise if the animals on one treatment are in some way different from those on another treatment. It should be done formally, not semi-subjectively.

However, it is not good enough just to randomise the animals (or other experimental subjects) to the treatment groups. It must be done throughout the whole experiment. Animal cages should be housed in a random order on the shelves and all measurements should be done in random order, preferably with the investigator being blind to the treatment groups.

It is not acceptable to add another treatment group to an experiment which has already been started, as in this case randomisation to all treatment groups will be impossible.

Sometimes it is necessary to split the experiment up into smaller sub- or mini-experiments. For example, it may be impossible to make all the measurements at one time, or there may be bottlenecks such as the need to use a centrifuge with a limited number of wells. In this case a randomised block experimental design may be appropriate (see 12. Experimental Designs). Restricted randomisation is then appropriate (see below).


Methods of randomisation


Physical randomisation is easy for small experiments.  Assuming the aim is to randomise, say, 30 animals to three treatments (10 animals per treatment), then ten ones, ten twos and ten threes are written on 30 pieces of paper, these are folded and placed in a receptacle which is shaken. A paper is withdrawn, and the first animal is assigned to the indicated treatment, and so on. This can be done at a desk by recording animal number and assigned treatment on paper before going to the animal house. Some computer programs will place a column of numbers in random order, so could be used in a very similar way to physical randomisation. A third method is to use a table of random numbers which are available (with directions for use) in many statistical textbooks.


Restricted randomisation


Randomised block experimental designs are ones where the experiment has been broken down into smaller "mini-experiments" or blocks. This is done, for example, if the experiment is too large to handle conveniently, or if groups of animals differ markedly in age or weight, or there is some sort of natural structure to the experimental units, such as coming in litters.
Randomisation  is done within each block .

As an example, an experiment was to be set up to compare three treatments using a randomised block design with four blocks. Three folded bits of paper labelled "Control", "Dose 1" and "Dose 2" were picked  out of a shaken receptacle and the three animals, in order caught, were assigned accordingly for block 1. These were then replaced and picked again for block 2, etc. as shown in the table below:

Block                Treatment of each mouse  

1

Dose 1

Control

Dose 2

2

Control

Dose 1

Dose 2

3

Dose 2

Dose 1

Control

3

Dose 2

Control

Dose 1


Typically the block size will be the same as the number of treatments (as in the above case).  Similar restricted randomisation is necessary with other designs such as Latin square experiments, and crossover (within-subject) experimental designs.

"Improving" on randomisation


Some scientists have developed programs which "improve" on the randomisation by moving animals between groups until body weight is the same in each group. This is not recommended as the effect is to increase the within-group variation which will decrease the power of the experiment.

However, following randomisation to treatment groups a "statistically significant" difference in body weight etc. between the groups may be observed (although it is not usual to do a statistical analysis at this stage). In fact, by definition, this will occur in about 5% of experiments (assuming p=0.05 is the critical value). In such cases it would be sensible to re-randomise the animals.