I use JMP programme to run statistic on trial results, and I have been thinking if it is possible to set the ranges within a stat programme, for example because I have few reps for each traetment, I assume that I may have a wider range of results from pen to pen, and fewer replications to get an average and statistical significance- so if I have one pen that has a really high Feed conversion Ratio (FCR) for example that is not an outlier on JMP, it’s going to skew the results more and reduce my statistical significance- therefore I was wondering should I be lowering the FCR range so that this would then be considered an outlier and therefore I could exclude it from the results? I know this would then reduce my reps even further but they may be more true to the actual result? Please let me know if I can set up any ranges in JMP, and how to do it.
The question of outlier identification is at once easy, ie., using some criteria, set a numerical value(s) for low and high end if needed, and then if an observation falls outside the values, by definition the value is an outlier. The challenge is setting the values. There are various numerically based constructs which can be used. JMP has several, such as a quantile approach, and multivariate methods.
But the real question I ask is one of practicality and domain expertise...not one of statistical method. Since you are using the data to make statistical inferences, selectively NOT including certain observations should have some valid practical reason for exclusion, not just a simple, 'oh the value is beyond a range...therefore we'll delete it.' For example, 'Oh we know the measurement system failed when that specific object was measured...we have an assignable cause, so it makes sense to exclude the observation'. But if you are just deleting observations that fall beyond a certain range of values for that reason alone...this could lead to, and mind you I'm not accusing you of it, up to and including data fraud so as to acheive a perdetermined outcome.
I dont want to manipulate the data and this is not my intention, although it would be beneficial to check what would be the overall outcome with and without outlier and see does it impact on the overall conclusion.
Occaisonlly there is a big varaiability within the data and I am thinking how to deal with this "noise " .
I would like to be able to look at the results from different "persepctive" using JMP..