[Vision2020] GISS Director Gavin Schmidt 5-12-18 "Transparency in climate science"

Ted Moffett starbliss at gmail.com
Fri May 18 18:57:56 PDT 2018

A long comment section discussion follows this article at website below.
Vision2020 Post: Ted Moffett
in climate science
Filed under:

   - Climate modelling
   - Climate Science
   - Instrumental Record
   - Paleoclimate

— gavin @ 12 May 2018

Good thing? Of course.*

I was invited to give a short presentation to a committee at the National
Academies last week on issues of reproducibility and replicability in
climate science for a report they have been asked to prepare by Congress
slides <http://www.realclimate.org/images//NRC_RR_Schmidt.pdf> give a brief
overview of the points I made, but basically the issue is *not* that there
isn’t enough data being made available, but rather there is too much!

A small selection of climate data sources is given on our (cleverly
named) “Data
Sources <http://www.realclimate.org/index.php/data-sources/>” page and
these and others are enormously rich repositories of useful stuff that
climate scientists and the interested public have been diving into for
years. Claims that have persisted for decades that “data” aren’t available
are mostly bogus (to save the commenters the trouble of angrily demanding
it, here is a link for data from the original hockey stick paper
You’re welcome!).

The issues worth talking about are however a little more subtle. First off,
what definitions are being used here. This committee has decided that

   - *Reproducibility* is the ability to test a result using independent
   methods and alternate choices in data processing. This is akin to a
   different laboratory testing an experimental result or a different climate
   model showing the same phenomena etc.
   - *Replicability* is the ability to check and rerun the analysis and get
   the same answer.

[Note that these definitions are sometimes swapped in other discussions.]
The two ideas are probably best described as checking the robustness of a
result, or rerunning the analysis. Both are useful in different ways.
Robustness is key if you want to make a case that any particular result is
relevant to the real world (though that is necessary, not sufficient) and
if a result is robust, there’s not much to be gained from rerunning the
specifics of one person’s/one group’s analysis. For sure, rerunning the
analysis is useful for checking the conclusions stemmed from the raw data,
and is a great platform for subsequently testing its robustness (by making
different choices for input data, analysis methods, etc.) as efficiently as

So what issues are worth talking about? First, the big success in climate
science with respect to robustness/reproducibility is the Coupled Model
Intercomparison Project <https://www.wcrp-climate.org/wgcm-cmip> – all of
the climate models from labs across the world running the same basic
experiments with an open data platform that makes it easy to compare and
contrast many aspects of the simulations. However, this data set is growing
very quickly and the tools to analyse it have not scaled as well. So, while
everything is testable in theory, bandwidth and computational restrictions
make it difficult to do so in practice. This could be improved with
appropriate server-side analytics (which are promised this time around) and
the organized archiving of intermediate and derived data. Analysis code
sharing in a more organized way would also be useful.

One minor issue is that while climate models are bit-reproducible at the
local scale (something essential for testing and debugging), the
environments for which that is true are fragile. Compilers, libraries, and
operating systems change over time and preclude taking a code from say 2000
and the input files and getting exactly the same results (bit-for-bit) with
simulations that are sensitive to initial conditions (like climate models).
The emergent properties should be robust, and that is worth testing. There
are ways to archive the run environment in digital ‘containers’, so this
isn’t necessarily always going to be a problem, but this has not yet become
standard practice. Most GCM codes are freely available (for instance, GISS
ModelE <https://www.giss.nasa.gov/tools/modelE/>, and the officially open
source DOE E3SM <https://github.com/E3SM-Project>).

There is more to climate science than GCMs of course. There are operational
products (like GISTEMP <http://data.giss.nasa.gov/gistemp> – which is both
replicable and reproducible), and paleo-climate records (such as are put
together in projects like PAGES2K
<http://www.pages-igbp.org/ini/wg/2k-network/intro>). Discussions on what
the right standards are for those projects are being actively discussed
(see this string of comments
<https://www.clim-past.net/14/593/2018/cp-14-593-2018-discussion.html> or
the LiPD project <https://lipd.net/> for instance).

In all of the real discussions, the issue is not *whether* to strive for
R&R, but *how* to do it efficiently, usably, and without unfairly burdening
data producers. The costs (if any) of making an analysis replicable are
borne by the original scientists, while the benefits are shared across the
community. Conversely, the costs of reproducing research is borne by the
community, while benefits accrue to the original authors (if the research
is robust) or to the community (if it isn’t).

One aspect that is perhaps under-appreciated is that if research is done
knowing from the start that there will be a code and data archive, it is
much easier to build that into your workflow. Creating usable archives as
an after thought is much harder. This lesson is one that is also true for
specific communities – if we build an expectation for organized community
archives and repositories it’s much easier for everyone to do the right

[*Update:* My fault I expect, but for folks not completely familiar with
the history here, this is an old discussion – for instance, “On Replication
from 2009, a suggestion for a online replication journal
last year, multiple posts focused on replicating previously published work (

* For the record, this does not imply support for the new EPA proposed rule
on ‘transparency’**. This is an appallingly crafted ‘solution’ in search of
a problem, promoted by people who really think that that the science of air
pollution impacts on health can be disappeared by adding arbitrary hoops
for researchers to jump through. They are wrong

** Obviously this is my personal opinion, not an official statement.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.fsr.com/pipermail/vision2020/attachments/20180518/1b64e919/attachment.html>

More information about the Vision2020 mailing list