Slightly off the topic, but still potentially relevant in terms of realistic 
experimental error: when dealing with the small volumes typically used in 
crystallization (say 1 uL + 1 uL drops), and using a 10 uL pipette, the errors 
are fairly high (more like 30% than 5-10%), leading to a lot of 
non-reproducibility in the experiment- even when setting up the same exact 
solution many times.  Going to robotics helps with the reproducibility in 
liquid transfer, but doesn't necessarily help with the reproducibility of 
crystallization (an example of this can be found in: 
http://journals.iucr.org/d/issues/2007/07/00/bw5202/ ).

Cheers,  tom

From: CCP4 bulletin board [mailto:[email protected]] On Behalf Of 
[email protected]
Sent: Thursday, 14 March 2013 12:07 PM
To: [email protected]
Subject: Re: [ccp4bb] statistical or systematic? bias or noise?

I think that in statistics you can build a model that describes (and predicts) 
the uncertainty. So if you have done similar (!) replicate experiments, from 
which you can build the model, you can apply it to a single observation and 
provide a reasonably good guess for the value that you were measuring and its 
variance. Of course that guess would not be as good as the average value and 
variance from true replicates.

With protein crystals (or solutions for that matter), the sample is often too 
precious to redo the experiment and it is worth thinking about doing replicate 
experiments with a "cheap one", build the model, and then apply it to single 
"expensive" observations. That would be statistically justified (provided that 
the model is valid for all sets of experiments). I have not built such models, 
but we know that pipetting isn't really as good as we believe. If you randomly 
dial to a particular value on your pipetteman (say 5 uL), you will get a 
certain pattern of "errors" (which is really not a good word for it), while if 
you consistently dial either from a low (1uL) or a high (10uL) value towards 
the value you want, you will get another pattern. Those two patterns are not 
representative of each other, I don't think, and you would need to understand 
how to do experiments consistently to stay within your error-model (bad word).

Among many other things, statisticians try to come up with models that explain 
the uncertainty so that you know what to think, even if your set of observation 
is too small to say for sure, with n=1 being the ultimate too small. (Maybe not 
ultimate, n=0 is really too small.)

Mark



-----Original Message-----
From: Alexander Aleshin 
<[email protected]<mailto:[email protected]>>
To: CCP4BB <[email protected]<mailto:[email protected]>>
Sent: Wed, Mar 13, 2013 3:05 pm
Subject: Re: [ccp4bb] statistical or systematic? bias or noise?

On Mar 13, 2013, at 1:36 PM, Ed Pozharski wrote:


But what if I only have one measurement worth of sample?

Is it proper to use statistical analysis for a single measurement? I thought 
statistics, by definition, means multiple measurements.

Alex

Reply via email to