Aslak Grinsted >
Jens Morten Hansen (JMH) and co-authors have recently published a study where they use sine-regression to fit 5 oscillations plus a linear trend to a 160-year sea level record from waters near Denmark (a stack of local GIA corrected tide gauge records). They then observe that the sines-plus-trend model correlates highly with the record it was fitted to, after applying a 19-year moving average. This should be unsurprising as the procedure guarantees high correlation, regardless of input data (see figure). This result is clearly not significant in any meaningful sense.
After having hand tuned their model to fit the data, they proceed to post-rationalize the fitted model frequencies in terms of the nearest half-multiples of the lunar nodal cycle (18.6 years). Yet again, and equally unsurprising, a high correlation is observed (after 19-year averaging). Again, this key result is clearly not significant.
It is therefore not surprising that it has been hard to get this study published. After multiple rejections, it was eventually published in Journal of Coastal Research. They then began to publicize it in primarily danish media. Here are some of the headlines it generated:
This persistent push for publicity have prompted us to voice our critical concerns regarding this study. This is not a case of scientific censorship. This study has fundamental flaws.
Our criticism - over fitting
We wrote a critical comment highlighting the flaws in their analysis. It can be found here (in danish - TODO add link when it goes online.). We demonstrate by example the the non-significance by showing how you get high correlations regardless of data set, and periods used (see figure). We could have calculated a p-value using a monte carlo significance test against auto-correlated noise.
The plot on this page shows how we can fit the population in Denmark, and the frequency of the words "Frankenstein", and "hamlet" in books using the exact same lunar frequencies they used in their study. We can similarly predict sea level using the orbital periods of exoplanets. High-correlations are guaranteed even when there clearly is no causal link. I.e. the procedure cannot be used to support any hypothesis.
The basic issue is over-fitting: Their statistical model has 17 degrees of freedom (5 periods, 5 phases, 5 amplitudes, 1 trend, 1 constant). They test against 19-year moving averages of a 160-year series. There is at most 8-9 independent points in the smoothed series (and that is being generous as the time-series is auto-correlated). Ofcourse it will fit.
A set of frequencies drop out of the iterative hand tuning of the model parameters. They need an explanation for these frequencies in order to have a tidy story. But there is no need for explanations for this correlation as high correlations are guaranteed regardless of what data is used. The journal article has a long weaving argument for the lunar nodal resonance hypothesis. I see this as a post-rationalization of a (very) insignificant correlation. I am not convinced.
In a radio show from the danish national broadcasting company (DR), JMH lets us know that he also tried to push the story to a large Swedish newspaper. It was rejected with a note that the newspaper never accepted letters opening with references to the Emperors new clothes. I find that reference ironic.
The data in the figure comes from these sources: Sea level digitized from JMHs paper. Word frequencies from google ngrams. DK population from statistikbanken.dk. exoplanet periods from openexoplanetcatalogue.com Tamino
Here's Tamino on another case of sea level cycle fitting.
Bamber and Aspinall 2013 published a formal expert elicitation on the future of our two ice sheets, and found a large spread between individual expert answers. It has been argued by them, and others that there are outlier experts in the data, and that they drive the long-tails of the pooled uncertainty distribution. I wonder if that is really true or whether it could be a visual artifact from drawing samples from a long tailed distribution. When you plot samples from a long tailed distribution then of course the tail values will appear to stand out, but that does not necessarily make them outliers. The question is really whether there are two groups of experts optimists vs. pessimists or whether they all have been sampled from a continuum of experts.
I focus on the expert assessment of the rate of future ice loss from the West Antarctic Ice Sheet (WAIS), as this is a key uncertainty. In particular i use the 2012 estimates shown in red in their figure 1 reproduced on this page. I have digitized this figure to obtain the low and high values of each expert range listed in the bottom of this page.
When I rank these values from low to high and plot them then I see that there is no evidence for a discontinuity in the neither experts upper, nor lower estimates. There is no split between ice sheet optimists and ice sheet pessimists. For comparison I also plot some surrogate data sampled from a log-normal distribution, so that you can see an example of how it looks when you draw a small sample from a continuous long tailed distribution.
See also the final remarks at realclimate in relation to the Horton et al. paper: "There is no split into two groups that could be termed “alarmists” and “skeptics” – this idea can thus be regarded as empirically falsified"
Data plotted in graph:
s=[3.0 0.5;5.0 25.0;1.5 0.5;2.0 0.2;-0.5 2.0;4.0 30.0;-0.5 10.0;0.0 2.0;2.0 8.0;0.5 2.5;1.0 20.0;0.0 4.0;0.3 2.0];
Technical detail: if you draw 13 samples from the surrogate distribution, then there is a 51% chance that the highest value will be greater than 5. The maximum value in the surrogate plot is thus very close to the median value for this distribution. It is not an outlier.
Precision Experiment Data Records (PEDR) records. PEDRs are generated from the raw altimetry profile data with precision orbit corrections applied. They are time-ordered binary tables with attached PDS labels. You can download the MOLA PEDR data here.
Perhaps somebody will find this useful. The function is attached below.
Pachauri has been accused of sexual harassment. You can read about it here:
Pachauri's initial defense was to claim that he didn't do it, but that his phone was hacked. That is not entirely implausible considering the dirty lengths that climate obstructionists would go to (e.g. the CRU hack). Hence I initially withheld any commentary. The climate community I follow on twitter has also mostly been quiet on this.
Since then more news has surfaced. We have seen the content of some of the messages in the case, and more women have spoken out against Pachauri:
I don't know if he is guilty or not, but these additional reports push me towards the womens account.
I don't usually comment on gender issues, but I think it is important to publicly make it clear that sexual harassment is not OK regardless of who you are, or how much bad publicity it might give for climate science.
Chocolatey is a package manager sort of like apt-get for windows, or a command-line version of ninite. You can do stuff like:
Usually you have to run choco as an administrator. That means starting the command prompt as an admin. That can get annoying pretty fast, so I recommend installing the sudo package as the next step.
choco install sudo
After this you can install a bunch of other stuff using "sudo choco install xxxx" install, without starting an administrative shell.
A better terminal, and putty
I installed putty manually instead of using chocolatey because that makes it easier to get conemu to cooperate with putty. see also comments here
For scientific python you might want to try Anaconda instead of installing python yourself. That is not in the chocolatey.
Note this python section is from memory and may not be optimal.
If you already have python2, then you might want to remove it first using the control panel.
If PIP is not found, then it is most likely a problem with your path.
Below is a pdf with some exercises for that i use in my course on inverse problems.
Problem 0: Goal understand why it is attractive to use loglikelihoods over likelihoods.
Problem 1: An exact model with a single model parameter. This makes easy to plot.
Problem 2: A variant of problem 1 with an inexact model.
Problem 3: Goal: Show that it is easier to calculate derived quantities such as "expected value" of the posterior when you use mcmc.
Problem 4: A problem where it is easy to understand why is the likelihood called a likelihood.
Problem 5: A problem where you have to grasp many concepts.
Exercises 4 & 5 are probably the most fun.
I have been attempting to track ice motion in landsat 8 files. I have only been able to find the L1T files. These are terrain corrected files. Unfortunately for my study region (the Renland ice cap, East Greenland) the DEM used in this processing is of too poor quality and which introduces large geodetic errors. DEM errors can be seen as residual perspective effects [Click image on the right to expand to an animated GIF showing flipping between two images].
I would therefore really like to have the L1G product and account for the terrain correction my self (using gimpdem).
Another advantage of L1G over L1T is that there is less resampling which degrades feature tracking performance.
(This work is partly motivated by testing a new open source feature tracking & georeferencing toolbox for matlab we have been working on. I am looking forwards to putting it on github.)
West Antarctic Ice Sheet collapse has been discussed over time. It is not an exhaustive list.
See also these posts:
IPCC uncertainty guideline note (see table 1). However, I just realized that this range was actually calculated as the 5-95% range from CMIP5. From table 1 I would have called that the very likely range. Can anybody explain to me the motivation for calling it the likely range?
[EDIT2: I asked Jonathan Gregory, and he has been very helpful in explaining the motivation to me. He says: "[Sect 126.96.36.199 explains that] the 5-95% range of CMIP5 models coincides with the assessed likely range of TCR. That's the main reason for proceeding this way. It implies, as you say, that the CMIP5 models do not cover the entire range which is considered "very likely" but we do not have sufficient confidence to quantify a "very likely" range for projections."]
This also affects other parts of the sea level chapter. For example they say that only marine instabilities could cause a rise greater than the likely range. But does that mean that there is a 33% or 10% or chance of that? I have no idea... (or indeed 5% as they appear to use 5-95% for the likely range in this chapter).
The supplementary material to the sea level chapter says that they follow section 188.8.131.52 when they use 5-95% as the likely range.
[EDIT: I just found an explanation in an SPM footnote: "Calculated from projections as 5−95% model ranges. These ranges are then assessed to be likely ranges after accounting for additional uncertainties or different levels of confidence in models. For projections of global mean sea level rise confidence is medium for both time horizons."]
If you have any comments then please tweet me at @agrinsted or mail me. I'd greatly appreciate it.
A recent observational study found an observed sustained increase in ice discharge from the Amundsen Sea Embayment, West Antarctica, from 1973 to 2013 (Mouginot et al. 2014). This prompted this reaction from Eric Steig:
Models from Mengel & Levermann (2014) also show that "We have probably overestimated the stability of East Antarctica so far". This is early work and inconclusive, but it highlights a need for further study. [Remark: At EGU2014 Vermeersen and Pollard gave presentations showing WAIS collapse in a few centuries, and a large EAIS response on longer time scales. I assume that this work will be out really soon, so i will update this post when I know more. ]
Crumbling ice figure courtesy Jan Åström, CSC, Finland
For adaptation planning we really need the full unconditional uncertainty distribution. The above studies shows that the AR5 conditionality on no marine instability may be excluding a important part of the pdf. It appears to me to be much more than just a very remote possibility. This is also reflected in a recent expert elicitation from Bamber & Aspinall (2013). I question how useful the AR5 projections can be when used as-is for local adaptation planning.
1-10 of 70