Re: [Phys-L] statistics, or not

[John Denker <jsd@av8n.com>]

On 07/06/2017 03:57 AM, David Bowman wrote:

> This discussion has caused me to be reminded of a quote attributed to Ernst Rutherford:
>
> "If your experiment needs statistics, you ought to have done a better experiment."

Let's talk about that.

I recommend a balanced approach: 
   A reasonable amount of data-taking, plus
   a reasonable amount of data-analysis.

Data-taking can be very expensive.  You have a duty to the folks who
funded you to do a decent job with the analysis.

To say the same thing the other way, if you screw up the experiment,
you can't fix it with statistics ... and conversely, if you screw up
the analysis, you can't fix it by taking more data.

--------

We might also talk about the distinction between "statistics" and
"fancy statistics".  In my book, fitting a straight line to a bunch
of points counts as statistics.  I don't think anybody would think
of that as being so overly sophisticated as to reflect poorly on
one's data-taking skills.

Another issue is that fancy statistics books tend to be so full of
jargon as to be almost impenetrable.

Another issue is that the academic statistics community has often
been far too set in its ways, to such a degree that important
advances often come from outside the community, e.g.
 -- a brewery operator (Gosset)
 -- an astronomer turned farmer (Fisher)
 -- an electrical engineer (Shannon)
 -- various cryptographers
 -- various pattern recognition / machine learning guys
 ++ et cetera.

It's strange to watch Big Name statisticians say something is
impossible, only to have somebody inform them it's already being
done on an industrial scale.

Bottom line:  I can understand why some people don't like statistics.
However, you have to do a certain amount of it, like it or not.