[R] Simulation - Natrual Selection

[Ben Ward]

On 05/01/2011 17:40, Bert Gunter wrote:
>> My hypothesis was specified before I did my experiment. Whilst far from
>> perfect, I've tried to do the best I can to assess rise in resistance,
>> without going into genetics as it's not possible. (Although may be at the
>> next institution I've applied for MSc).
>>
>> With my hypothesis (I mentioned it below), I was of the frame of mind that a
>> nonsignificant p-value on the cleaner variable (for now - experiment is far
>> from over), indicated a lack of evidence for rejecting the null. And so at
>> the minute, it looks like the type of cleaner makes no difference.
> I have no fundamental objection, but be careful. I would simply
> qualify your last sentence by saying that it means that the
> experimental noise is to great to precisely determine the size of the
> cleaner effect. Scientific reality tells us that it is never exactly
> 0; what your results show is that your uncertainty about the value of
> the difference encompasses both positive and negative values. This
> does NOT mean that the difference might not be scientifically large
> enough to be of interest -- a confidence interval for the difference
> (MUCH better than a P value) would help you determine that. If the
> interval is narrow enough that the difference, positive or negative,
> is too small to be of scientific interest, then you're done. If the
> linterval is large, then it tells you that you need more data, a
> better experiment (less noisy) etc.
>
> -- Bert
>
At the moment I wouldn't call the confidence interval small, it's 
definately wide, and at the minute the confidence interval covers zero. 
My R-squared at the minite is also 0.5, this is mostly due to the few 
extreme cases of adaptation as I mentioned before, but I'm hesitant to 
remove it as papers in my literature study which also evolve bacteria 
show that there is often (sometimes wide) variation in the paths 
populations take. So whilst mathematically a bit undesirable, and makes 
me and the model uncertain, it does fall into place with what is known, 
or has been previously shown of the reality of selection. Again if I 
include the data from the bacteria dropped from the study, all that 
"improves", and uncertainty is reduced.

It may also be worth me mentioning, I am also taking a more traditional 
approach (by that I mean a more "Statistics 101" approach, indeed that 
is all the stats tuition covered in my course as a taught element), 
incase what I've described above did not work or was not ideal, because 
we (me and my supervisor) did forsee a model report may contain a lot of 
uncertainty. Indeed we did expect some populations to adapt and some to 
not etc. So I've also been collecting data on the width of the zones of 
inhibition shown by putting disks of cleaner on plates of growth, and 
measuring the dead zone that results. I can get lots of data from this 
with only a few plates, and doing this at the start of the study, a few 
times in the middle, and at the end. Will allow me to do more 
traditional analysis, for example t.test on the dead zone widths at the 
end of the study, between cleaner a and b.  Or a non parametric 
equivalent, maybe even a permutation test. The modelling stuff is 
already beyond what my supervisor expects of me, but I felt it would add 
value and a lot more insight to the study, allowing more variables to be 
accounted for, than a more short-sighted traditional "test".