PDA

View Full Version : Minimizing group differences at baseline in controlled trials



whopkins59
03-11-2010, 06:43 PM
I have devised a spreadsheet to achieve the goal in the subject line of this
message, and I am now writing a paper about it for publication in
Sportscience. The draft title and abstract of the paper appear below. I'd
be grateful for offers from folks with an interest in this sort of thing to
try out the spreadsheet and review the article. Please contact me and I will
email you the spreadsheet in the first instance. Whether the article is
worth writing depends on whether anyone knows of an already published
practical method. There are published methods for minimization, but as far
as I can see they are all aimed at allocation of each subject to a treatment
group as soon as the subject is recruited. My method works in the far more
common situation in our disciplines, where the subjects have all been
recruited before allocation to treatment groups.

Will
Will G Hopkins, PhD FACSM
Contact info: http://sportsci.org/will
Sportscience: http://sportsci.org
Statistics: http://newstats.org
Be creative: break rules.

A Spreadsheet to Reduce Differences Between Group Means of Subject
Characteristics in a Controlled Trial

When subject characteristics may substantially modify the effect of a
treatment, subjects in a controlled trial should be allocated to groups in a
manner that minimizes differences between mean values of the
characteristics. Random allocation produces trivial differences between the
means only with >100 subjects per group. The spreadsheet accompanying this
article provides an alternative to randomization for use with smaller sample
sizes when the characteristics of all subjects are known before the
controlled trial begins. The spreadsheet allocates subjects to 2-5 groups,
giving most importance to minimizing differences between the means of one
characteristic, such as the baseline values of the dependent variable. Up to
five other characteristics are given equal secondary importance. In
simulations with 10 subjects per group, random allocation results typically
in small-moderate differences between group means for each characteristic,
whereas the spreadsheet produces trivial differences for the most important
characteristic and differences for the other characteristics that reach the
threshold for small only when there are >4 characteristics. With >12
subjects per group all differences in means between means produced by the
spreadsheet are typically trivial. Use of this spreadsheet for controlled
trials with small sample sizes will improve the trustworthiness of treatment
effects.