Determining and excluding the outliers of a dataset I am trying to model biological processes using Ordinary Differential Equations. I have a (pretty large) model that I am trying to parameterize using software (Copasi's implementation of the Genetic Algorithm if your interested). I am running the estimation multiple times such that I have several vectors containing estimates for the models parameters that are ordered by the lowest (therefore best) sum of squares value. As part of my analysis of this data I want figure out which of these parameters are correlating with other parameters and this should give an indication as to which parameters are not identifiable. The problem is that some of the data has such a large spread that the resultant scatter plots show nothing. Therefore, my question is: how would it be best to determine the outliers of my data so that I could exclude (say) 20% of my data points and replot? 
The first two columns of my data (truncated):
(v1).Kcat   (v1).km
1.16567 0.0141031
0.0351003   153.682
0.920923    84.1231
0.403507    0.76484
0.0118404   0.38389
0.01    0.01
0.219686    0.169275
314.67  0.80948

Thanks
 A: Some methods of outlier detection.
A formal answer to your question is that you could get rid
of 5% of 'outliers' from each variable, and still have at least 90%
of your bivariate observations left.
There are many methods of identifying outliers in an 'automatic' or formal
way. From the fragment of data you give, it is not possible to
recommend a specific method. 
(a) One method would be to sort each variable
from smallest to largest and omit a certain percentage from each 'tail': maybe
the top 2.5% and the bottom 2.5%. Because you give only eight
values for each variable, it is not possible to illustrate this exact
method with the data you give (2.5% of 8 is a small fraction
of an observation, so nothing gets removed). 
(b) Another method is to use
boxplots, which have a criterion for identifying outliers that
depends (roughly speaking) on how far lower and higher observations
are from the middle half of the data. There is no way to predict
in advance how many low or high values will be indicated (if any).
In the data fragment you give, boxplots indicate only the
largest observation in each variable as an outlier.
(c) You say you are making scatterplots of the data. That opens
up the possibility that you have regression analysis in mind.
In that case, you can use standard methods of regression diagnostics
to identify points that fall relatively far from a regression line,
or that have an 'undue influence' on how the line is calculated.
Using your data fragment, two points were called out as
exceptional (0.0351003, 153.682) and (314.67,  0.80948). These
happen to contain the values identified by univariate boxplots 
as outliers.
However, I always feel uncomfortable using such formal approaches
alone. In your case, I would recommend you think what might be
causing the 'outliers'. Do they represent errors or unimportant
exceptions? If so, it may be OK to eliminate them. 
In methods (b) and (c), you may need a couple of iterations.
Once the most extreme outliers are taken out, the reduced dataset
may show additional outliers because the data-derived criteria have now shifted. I suspect the few observations you showed may not be
typical, but chosen to illustrate the presence of outliers. 
But I have to say that several iterations of outlier removals by
each method left nothing worth plotting in your data.
(d) Another common method of 'taming' data with a 'large spread'
is to take logarithms of the data. Applied to your bit of data,
this method did decrease spread nicely, but did not lead to a
scatterplot with any clear trend.
Cautionalry comments.
By contrast,
in some situations, the outliers contain important information,
and you should not use some formal method to get rid of them, but
find some method to take their 'message' into account. For
example, in earthquake data it is only the outliers that matter
to the general public; seismic events below about 3 on the Richter
scale may hardly be noticed even by people very near the epicenters.
In some emerging biotech and computer technologies,  outliers in data
may signal profitable new methods.
Trying to do statistical consulting without getting to know
processes and measurement method well is about as irresponsible
as 'doctors' doing diagnoses before prescribing various popular prescription
drugs on the basis of an online multiple choice survey. I am
only trying to give an overview of methods that might be used,
but not making any specific recommendations.
