[bip] Reproducible research
James Taylor
james at jamestaylor.org
Mon Mar 9 08:47:08 PDT 2009
>> Um... no. If nobody can reproduce it, it is not science.
>
> How reproducible are astronomical observations? Are they
> science? What about those observations of rare deep-sea
> animals seen only once, which I've seen published in
> Nature?
>
> One of the theories of science is that the ability to predict,
> with falsifiability, is the key to science. Reproducibility is
> a subset of that.
Fair enough. This discussion was specifically about computational
approaches to analyzing data. For those I believe exact
reproducibility is necessary (but, to be clear, not sufficient for
interestingness).
For observational data the criteria needs to be different. This
applies to your examples (events that are so rare that even if I try I
may never seem them again) but also to experimental data where I may
be able to reproduce the steps, but interpretation is required to
decide whether I have reproduced the results.
Again, taking the astronomical observations as an example, the
observed data may be completely unique, but it also needs to undergo
processing and filtering and statistical analysis to become
understandable. The details of that process should be precisely
documented and reproducible.
You are making me question this though, why is computation / analysis
different? I experience tells us that methods and analysis details
really do matter. Tiny parameter changes can completely alter the
significance of results. Often the significance of certain assumptions
isn't even known until years later when more data is collected. When
going back and evaluating past work, I need to be 1) trust that the
observed data isn't falsified (a combination of knowing how it was
produced and inspecting it), and 2) understand exactly how the authors
derived their conclusions from that data and how they assessed the
significance of their conclusions.
More information about the biology-in-python
mailing list