h1

Applying Monte Carlo simulation to Sloan’s and Wolfendale’s use of Forbush decrease data

September 5, 2008

* Introduction
* Sloan’s and Wolfendale’s approach
* Considering standard deviation 
* Applying understanding of standard deviation to the lower cloud cover data
* Monte Carlo simulation
* Conclusions drawn from Monte Carlo Simultion

Introduction

It’s been five months since the widely remarked upon paper, Testing the proposed causal link between cosmic rays and cloud cover, by Sloan and Wolfendale, was published in the journal Environmental Research Letters.  In that short time this paper has been enshrined in the hall of fame of anthropogenic global warming dogma.  The thing that makes this paper a favorite of the climate change alarmists is its purported proof that there is no possible significant relationship between galactic cosmic ray flux, as modulated by the sun’s magnetic field and climate change.  This is important to the alarmists, because anything that might even partially exonerate CO2 must be stifled.

I spent many long hours deciphering this paper after several acquaintances alerted me of its conclusion.  I was dismayed to find out, after my careful study, that none of these acquaintances had ever actually read the paper!  I don’t know why this surprised me, they were simply victims of the global warming echo chamber.  Some of them read articles in the popular press or in web blogs that cited this paper and uncritically repeated its conclusions.

But are the conclusions that Sloan and Wolfendale based on the study of Forbush decreases (see below) correct?  Simply put, No.

The Sloan and Wolfendale paper concerns the possible link between cosmic ray flux and climate. The primary proponent of the the cosmic ray flux/climate link is Henrik Svensmark.  in his 2007 Astronomy and Geophysics paper, Cosmoclimatology, Svensmark says:

 Data on cloud cover from satellites, compared with counts of galactic cosmic rays from a ground station, suggested that an increase in cosmic rays makes the world cloudier. This empirical finding introduced a novel connection between astronomical and terrestrial events, making weather on Earth subject to the cosmic-ray accelerators of supernova remnants in the Milky Way.

This much is certain, the cosmic ray flux is modulated by the changing magnetic conditions of the sun.  Svensmark asserts that this makes the sun a very big player in changing climate conditions on the Earth.  This concept is anathema to the folks who envision gigantic climate catastrophes as a result of human activity.  Hence, the rapid dissemination of Sloan and Wolfendale’s conclusion by people who are probably very foggy on the paper’s logic.

Sloan’s and Wolfendale’s approach

A bare bones description of Sloan’s and Wolfendale’s argument is: Yes, there does appear to be some correlation between cosmic ray flux and lower cloud cover (LCC), but No, there is no causal link.  They show the correlation between cosmic ray flux (as modulated by solar magnetic activity)  and lower cloud cover with their figures 1 & 2, shown below. 

Figure 1.
123

Figure 2. Sloan and Wolfendale's original caption says "The seasonally corrected LCC amount as a function of the Climax neutron monitor count rate (both monthly averaged) during cycle 22 (1983–1996). " Red annotation was added by Moriarty and will be referred to later in the text of this post.

After presenting figures 1 and 2, where, according to the authors, “The good correlation is evident,” Sloan and Wolfendale work to show that this correlation is not due to the type of causation hypothesized by Svensmark and others.  They focused on Forbush decreases, which are short term (usually several days to two weeks) reductions (a few percent to about 15%) in cosmic ray flux.   Their reasoning is that if the correlation shown in figure 2 is actually due to causation, then there should be a measureable decrease in lower cloud cover as a consequence of the the reduced cosmic ray flux during a Forbush decrease.

So, they plotted the lower cloud cover anomaly percentage change vs. the percent change in cosmic ray flux decrease during twenty-three Forbush decreases.  Their results are reproduced below in figure 3.  They claim that if the lower cloud cover is controlled by the cosmic ray flux, then the points in figure 3 should be along the line labeled “LCC CR.”  The points don’t seem to make a good fit to the line, so they conclude that there is no causation.

 One quick point before going any further.  Forbush decreases are short duration events, from a few days to a couple of weeks.  It seems odd that for 19 out of 23 the events that Sloan and Wolfendale studied they used monthly averaged lower cloud cover data.  If there is a correlation between cosmic ray decreases during Forbush events and changes in lower cloud cover, it doesn’t seem likely that monthly averaged cloud cover data is going to reveal it.

Figure 3. Sloan and Wolfendales original caption says "The measured change in the LCC plotted against the change in the Oulu neutron monitor count rate during the measurement time of 1 month for the D2 data (solid circles) and 1 week for the D1 data (open squares). The solid line shows the values expected from the smooth curve shown in figure 2. The Oulu count was observed to change by 17% due to the solar modulation during solar cycle 22." The red annotation was added by Moriarty and will be referred to later in the text of this post.

 

Considering Standard Deviation

To evaluate Sloan and Wolfendale’s conclusions, it it necessary to have an understanding of standard deviations.  If you are mathematically savvy, I don’t want to talk down to you.  So if you are clear on standard deviation, just skip this section.

The red annotation in figures 2 and 3 highlight the standard deviation of the lower cloud cover anomaly data.  Standard deviation is a statistic that tells how tightly measurements of the same thing are grouped together.  Suppose, for example, you measured the voltage between two points 50 times and the average was 3 volts, and about 2/3 of the measurements were between 2.9 and 3.1 volts, and about 95% or the measurements were between 2.8 volts and 3.2 Volts.  then the standard deviation of the measurements was about 0.1 Volts.  In general, 68% of measurements are within one standard deviation of the average and 95% of the measurements are within two standard deviations.  The voltage example is illustrated in figure 4, below.

Figure 4.

Figure 4. Example of voltage measurements. The average is 3 V and the standard deviation is 0.1 V.

Applying understanding of standard deviation to the lower cloud cover data.

The line labeled “LCC CR” in figure 3 represents, in simple terms, the line upon which the lower cloud cover anomaly change would reside for a given cosmic ray count rate decrease if the correlation were due entirely to causation and the standard deviation for the lower cloud cover measurements were zero.  It has a slope of -0.051.  So for a 10% decrease in cosmic ray flux we would expect a -0.51% change in lower cloud cover.  This idea is shown in figure 5, below.   The x value (cosmic ray count rate decrease) for each point is the same as the corresponding point in figure 3.  The y value (Lower Cloud Cover Anomaly Change) is adjusted to reside on the line.

Figure 5.

Figure 5. Same as figure 3, with all the same cosmic ray count rate decreases, but the lower cloud cover changes moved to the line that would be expected if changes in the lower cloud cover were 100% controlled by the cosmic ray flux and the standard deviation of the lower cloud cover measurement were zero.

If we retain the assumption that that the lower cloud cover changes are governed by changes in the cosmic ray flux, but we increase the standard deviation of the measurements of the lower cloud cover, then we can model the effect by having the points in figure 5 “wander” away form the line in the y direction.  The amount that each point is allowed to wander will be determined by the choice of standard deviation.  For example, if we choose a standard deviation of 0.1% for the measurement of the lower cloud cover change, then we will allow the points to move in such a way that 68% of them end up less than 0.1% (one standard deviation) away form the line, and 95% of the points end up less than 0.2% (two standard deviations) away from the line.  There are an infinite number of ways to do this – four of them are shown in figure 6.  In each of the four cases the newly scattered data is fit to a line which is very similar to the “LCC CR” line shown in figure 3 and figure 5.

figure 6

figure 6. Each one of these four examples show data scattered away from the prefect correlation line with a standard deviation of 0.1% in the measurement of the lower cloud cover anomaly change.

For the record, the data in figure 6  and the following figures were generated using a “Gaussian white noise generator.” A Gaussian distribution is the name for the distribution shown in figure 4, above.  You may have heard of it referred to as the “normal’ distribution. 

Now we can incrementally increase the standard deviation of the lower cloud cover data.  Figure 7 shows four examples as the standard deviation goes through 0.2%, 0.4%, 0.6%, and 0.8%.  As the standard deviation increases the best line fit for each set of scattered data tends to look less and less like the lines shown in figures 3 and 5.

figure 7

figure 7. Four examples of data scattered away from the prefect correlation line with standard deviations of 0.2%, 0.4%, 0.6%. and 0.8% in the measurement of the lower cloud cover anomaly change.

We can repeat the process illustrated in figures 6 and 7, but now with a standard deviation equal to the actual standard deviation of the lower cloud cover data, about 1%, as shown in the red annotation of figures 2 and 3.  Then we can compare the results to the actual measured data in figure 3.  In figure 8, below, four cases where a standard deviation of about 1% has been applied to the lower cloud cover data with the same Gaussian white noise generator.

Figure 8.

Figure 8. Each one of these four examples shows data scattered away from the prefect correlation line with a standard deviation of 1.0% in the measurement of the lower cloud cover anomaly change.

Notice that when the standard deviation is about 1%, the best fit lines can look much different than the original line (see figures 3 and 5) from which they were scattered.  Just compare the lines in figure 8 with the lines in figures 3 and 5.  Every one of the cases in figure 8 are completely consistent with a the lower cloud cover being entirely controlled by the cosmic ray flux, but with a 1% standard deviation in the measurement of the cloud cover.  What can we say about the comparison of these four cases in figure 8 and the actual data shown in figure 3?

Monte Carlo Simulation

The four modeled plots in figure 8 are quite different from each other and different from the actual measured data in figure 3.  Yet we know the four plots in figure 8 are consistent with the lower cloud cover being controlled by the cosmic ray flux but having a standard deviation of 1%, because we created them to be that way.  The important question here is: Is the measured data, shown in figure 3, also consistent with a lower cloud cover actually being controlled by the cosmic ray flux, but having a standard deviation of 1 percent?

A Monte Carlo simulation is a perfect tool for answering this question.  A Monte Carlo simulation is the repeated calculation of some model or algorithm with random inputs from the domain of all possible inputs.  Then conclusions are drawn from the accumulated results of the repeated calculation.  A classic example is to use the Monte Carlo method to calculate Pi

I have used the Monte Carlo method by repeating the calculations to generate the type of data shown in figure 8.  I didn’t repeat the calculation just 4 times – I repeated it 10,000 times.  Of course, it is totally impractical and not very enlightening to show all 10,000 plots.  Instead, for each of the 10,000 sets of data I have calculated the slope and the intercept of the best fit lines.  Then these 10,000 slopes and intercepts may be analyzed and compared to the slope and intercept of the actual measured data shown in figure 3.  The best fit line for the point shown in figure 3 has a slope of 0.001 and and intercept of 0.11.

Figure 9 shows a histogram of the slopes of the best fit lines for each of the 10,000 generated sets of data.  The average slope of all of these lines is -0.051, exactly the same as the slope of the lines in figures 3 and 5, as would be expected.  The spread of the slopes yields a standard deviation of 0.055. 

Similarly, figure 10 shows a histogram of the intercepts of the best fit lines for each of the 10,000 generated sets of data.  The average intercept was, of course, 0.0, as expected.  The standard deviation of the intercepts was 0.36.

Figure 9.

Figure 9. Histogram of the slopes of the best fit lines for each of the 10,000 cases generated with the Monte Carlo simulation. The mean slope is -0.051, the same as Sloan and Wolfendale's "LCC CR" line shown if figure 3. 68% of all cases have slopes within 0.055 of the mean (between -0.106 and +0.004). So the standard deviation of the slopes is 0.055

Figure 10.

Figure 10. Histogram of the intercepts of the best fit lines for each of the 10,000 cases generated with the Monte Carlo simulation. The mean intercept is 0.0, the same as Sloan and Wolfendale's "LCC CR" line shown if figure 3. 68% of all cases have intercepts within 0.36 of the mean. So the standard deviation of the intercepts is 0.055

 

Conclusions drawn form Monte Carlo simulation

The pertinent question is this: Is the measured data, shown in figure 3, consistent with a lower cloud cover being controlled by the cosmic ray flux, but having a standard deviation of 1 percent?

Figure 11 shows the intercept vs. the slope for all 10,000 iterations of the Monte Carlo simulation.  It shows a wide range of possible outcomes: some with large slopes and some with small slopes, some with large intercepts and some with small intercepts.  They are arrayed about the mean slope and intercept of -0.051 and 0.0 respectively, the values that represent the “perfect fit” line labeled “LCC CR” in figures 3 and 5.  The red “X” marks the location of the slope and intercept of the 23 data points used by Sloan and Wolfendale in figure 3.  The points within the yellow region represent Monte Carlo cases where the combination (root sum square) of the slope and intercept make them closer to the “perfect fit” than the data used by Sloan and Wolfendale. Only 26% of all cases fall within the yellow region.

 So, when we start with the assumption lower cloud cover exactly follows the cosmic ray count, as indicated by the lines in figures 3 and 5, but add a standard deviaton of about 1% to the lower cloud data, then 74% of the time the rusult will be worse than the real data used by Sloan and Wolfendale.   This means, quite simply, that the measured data, shown in figure 3, is consistent with a lower cloud cover being controlled by the cosmic ray flux, but having a standard deviation of 1%. 

figure 11.

figure 11. Figure 11. Intercept vs. slope for each of the 10,000 interations of the Monte Carlo simulation. The vertical and horizontal lines mark the means of the slope and intercept, which are the same as the slope and intercept of Sloan and Wolfendale's "LCC CR" line in figure 3. The yellow region, which contains 26% of all iterations, highlights the cases where the root sum square of the deviations from the slope and intercept means are less than the deviations for the 23 Forbush decreases shown in figure 3 (shown by the red "X."

If fact, the badly scattered data is in figure 3 is still a better fit to the “LCC CR” line than would be expected, given the standard deviation of the lower cloud cover measurements. 

Because the standard deviation of the lower cloud cover data is so high (nearly 1%), Sloan’s and Wolfendale’s attempt to to use Forbush decreases to verify or reject the possibility that cosmic ray flux changes induce lower cloud cover changes was poorly conceived.  The data is far to noisy to tell one way or the other.  Claims in the press and the blogosphere that Sloan and Wolfendale have used Forbush decreases to show there is “no link between cosmic rays and global warming” are just plain wrong.

17 comments

  1. So in the end, it was yet another study where they claim conclusions that the accuracy of their data cannot support. Why am I not suprised.


  2. Actually, you had me with Sloan and Wolfendale’s use of monthly averages of lower cloud cover data for 19 out of 23 Forbush decreases they studied.

    Usually, when one discovers that the data is so severely inadequate, the response is to find another thesis topic. Students under time pressure can be forgiven for sticking to such a “doomed” project, especially if it’s a master’s thesis.

    However, that explanation doesn’t apply here. They pressed on, knowing that they can’t get the data they need. Why waste everyone’s time — and kill trees to print noise?

    The answer is left as a homework assignment to the reader.


  3. If you read Svensmark he covers the use of Forbush decreases in attempts to falsify his theory. The problem is that the Forbush decreases do not block the higher energy GCR that are apparently key in cloud formation, just as fluctuation in earth’s magnetic field do not meaningfully affect the GCR that are key in cloud formation.

    Interesting to see the results of the CLOUD project at CERN which is to kick off this year and is expected to verify or falsify the cloud formation theories of Svensmark. If CLOUD validates Svensmark’s theory of cloud formation then we can start looking for correlations.


  4. Bullocks. The data do not show any (even remotely) relation between CR and clouds during the Forbush events. It is as simple as this.
    You could save all your intellectual energy (with Monte Carlo computations for simple linear regression. Why not just using the linear equations of the regression to make the probability statements) by just pointing to the noisy cloud data and that they even for decades will not completely rule out a CR connection. But if you really want to promote this idea you have to realize that it is the duty of people like Svensmark to bring positive proofs for their theory and not show that the idea cannot completely be ruled out from a statistical point of view.


  5. Georg Hoffman,

    Thank you for your comment.

    Do you have the impression that I have claimed that the data shows a relationship between clouds and Forbush events? If you do, then you should re-read the post.

    In particular, the last paragraph states…

    “Because the standard deviation of the lower cloud cover data is so high (nearly 1%), Sloan’s and Wolfendale’s attempt to use Forbush decreases to VERIFY OR REJECT the possibility that cosmic ray flux changes induce lower cloud cover changes was poorly conceived. The data is far too noisy to tell ONE WAY OR THE OTHER. Claims in the press and the blogosphere that Sloan and Wolfendale have used Forbush decreases to show there is “no link between cosmic rays and global warming” are just plain wrong.”

    The fact is that this data has circulated around the blogosphere as PROOF that Svensmark and others are wrong.

    As for wasting intellectual energy – It seems that it was Sloan and Wolfendale who take who take up journal space with an ill conceived “test” involving Forbush decreases. Their test sheds no light because the uncertainty in the data insures there is can be no pass/fail criteria.

    I struggled to find a way to make this point clear to the majority of people surfing the blogosphere who are not necessarily up to speed when it comes to dealing with the data. These people become “victims” of other people’s interpretation of that data. This is unfortunate.

    Best regards,
    Tom


  6. Tom
    I dont knwo how you could think I havent read carefully your post. Figure 3 shows the best known relation between CR and clouds during about 20 Forbush event and the relation is simply not existing. It is up to the people propagating the CR hypothesis to demonstrate convincingly such a strong control of CRs on our climate. The paper of Sloan and Wolfendale (the latter by the way one of the godfathers of Solar Physics and Astronomer Royal in the 90s) is just another nail in the cuffin.
    The entire idea of “proof” is already a phylosophical misconception of yours. Anyhow to compute the error margins of a linear regression with given sigma in x and y is just one formula in an introduction into statistics textbook and I have still no idea why you need a Monte Carlo simulation for that.
    Cheers
    Georg


  7. By the by
    I know at least one other group (Joerg Beer, Zuerich) trying the same thing, I think with better cloud data. They’ve found the same thing, no relation between CR and clouds whatsoever.
    I will ask them if they published their results.
    Georg


  8. Georg Hoffmann,

    OK. Your answer, above, says

    “Figure 3 shows the best known relation between CR and clouds during about 20 Forbush event and the relation is simply not existing.”

    This statement makes it clear that you do not understand. You are claiming that figure 3 shows that there IS NOT a relationship between cosmic ray flux and lower cloud cover.

    The entire point of the post is that the data in figure three is consistent with cosmic ray flux completely controlling lower cloud cover, but with a 1% standard deviation in the lower cloud cover data. So are all the very different artificially generated cases in figure 8. This does not say that the cosmic ray flux DOES control the lower cloud cover, only that the data is CONSISTENT with the cosmic ray flux controlling the lower cloud cover.

    Because the standard deviation of the lower cloud cover is so large, no conclusion about the existence or non-existence of a relationship between cosmic ray flux and lower cloud cover can be drawn from this data. Your statement that I quoted shows that you have, in fact, drawn such a conclusion (“the relation is simply not existing”). You cannot draw that conclusion from this data. If you do draw such a conclusion, then you are in error. Period.

    As for Wolfendale being “one of the godfathers of Solar Physics and the Astronomer Royal of the 1990s,” that is an “appeal to authority” argument. Sometimes such an argument can carry weight, but it is considered weak when opposed to concrete evidence, such as I have provided. Your use of this argument is especially egregious because you are using the authority to bolster his own position.

    I hope you are correct about the other group having “better cloud data.” I will keep my eyes open for that. However, it seems that this is a rather futile approach if the ISCCP D2 data is used instead of the D1 data. A paper by Su Yeon Oh, et. al., “Globally nonsimultaneous Forbush decrease events and their implications,” Journal of Geophysical Research, Vol. 113, 2008, lists the dates for 49 Forbush events all with decreases greater than 3.5% Although Oh’s paper has nothing to do with cloud cover, it lists 5 times as many Forbush decreases above 3.5% than do Sloan and Wolfendale. So it seems there is more usefull CR data available.

    best regards,
    Tom


  9. Tom
    ok. I think your repeated statement makes it perfectly clear that you are simply a) not understanding the data you are showing and b) you are unfortunately not able to understand how scientific reasoning works and c) you obviously didnt know how to compute errors bars in linear regressions. That is quite a lot given the rather bold statements you are making.
    Error bars exist everywhere in science. If only statements would be considered which are in a agreement with 99% hypothesis testing there wouldn’t be much science left. I dont know who gave you the idea that actually Sloan&Wolfendale are in the position to make a proof. There tested a straightforward consequence of the CR hypothesis and they couldnt find any relationship between CR and cloud cover. Rather sad for the most important climate mechanism ever since the age of the dinosaurs to 20th centure climate warming dont you think? It is to people like Svensmark to demonstrate something. For the larger CR decreases during the Forbush events the cloud cover shows even the opposite sign of what Svensmarks hypothesis claims. It is now the nth refutation of different aspects of the CR theory. Take it at least as given that there are actually very very few people working in cloud physics considering the entire CR hypoythesis as something remotely interesting. Actually though I know quite a lot of them there is none.
    I mentioned the functions and reputation of Arnold Wolfendale not as an argument for whatsoever but only for your education.

    All the best
    Georg


  10. Not very surprisingly S&W spotted the four strongest Forbush events as well. That is what they wrote:
    “However,
    the four points above a counting rate change of 9% have a mean LCC anomaly change of
    0.68±0.45%. This is compatible with the dashed line showing no correlation between the LCC
    and CR rate changes but it is 2.8 standard deviations above the value of −0.6% expected had
    there been a correlation similar to that seen in figure 2.”
    Concluding
    “In conclusion, it is statistically improbable that the Forbush decreases are compatible with
    the hypothesis of a correlation between LCC and ionization as proposed in [1,2]. Hence Forbush
    decreases do not provide evidence which can be used to corroborate such a hypothesis.”
    which reminds everyone again (as I said) who has to bring evidence for the CR hypothesis.

    Georg


  11. Georg Huffmann,

    Thank you for your additional comments and your persistence.

    You stated “I dont know who gave you the idea that actually Sloan&Wolfendale are in the position to make a proof.”

    I am not sure what gives you the idea that I expect any kind of “proof” from Sloan and Wolfendale.

    However, the title of their paper is “Testing the proposed causal link…” Therefore, while I don’t expect a “proof,” I do expect a test. A good test has some criteria for distinguishing between passing and failure. Sloan’s and Wolfendale’s Forbush decrease test does not provide such a criteria. Given the domain of Forbush decrease data that they used and the standard deviation of the cloud cover data, their results are consistent with lower cloud cover being modulated by cosmic ray flux. But it does not provide evidence one way or the other for such a link. It is simply not a good test.

    Next, you point out “the four points above a counting rate change of 9%…” If you are agreeing that the other 20 points below 9% are not actually relevant, then we have truly made some progress.

    By my count there are at least 15 Forbush events with cosmic ray decreases over 9% during the period considered. Why do only four of them show up in this set of data. Why isn’t the corresponding cloud cover data from the ISCCP D1 (3 hour) data used in all cases, instead of the D2 (monthly averaged) data? Doesn’t that just make sense when looking for changes over time periods that are less than a month? I am not implying any nefarious motive, just simple sloppiness.

    Finally, for now: Another interesting and relevant point. The ISCCP says the following about the uncertainties for their cloud cover data:

    “For cloud cover, determined for regions about 280 km in size, the estimated random error is 10% for instantaneous values and  3% for monthly mean values. Random errors are larger in the polar regions, at least 20-30% for instantaneous values.”

    These random errors are considerably larger than the error I used in my analysis. Increasing the standard deviation in my analysis just makes my point all the more valid.

    I look forward to you further comments.

    Best regards
    Tom


  12. An interesting and related point concerning the recent magnetic activity of the sun


  13. Ok just to finish that.
    A Test ist not a statistical test. It can be but it must not. S&W actually delivered both.
    I got the diploma work done at the University of Zuerich. They use daily data and do a lot of statistical tests. It confirms what I said before. No significance and no relationship.
    The paper will come out soon.
    Georg


  14. Got the link to here from a discussion over at http://wattsupwiththat.com/2008/10/04/speckwatch/#comment-45771 . I look frward to reading your next.


  15. How come that a study of this low quality passes peer review, whilst much better one (e.g. from R Spencer) is withheld? Do magazines apply censorship these days?


  16. Do magazines apply censorship these days?

    Yes, and so do “peer review” journals closely tied to the AGW propaganda industry.


  17. […] Lubos Motls inlägg om hans senaste studie. (Relevant kritik av Sloan är utöver den Motl nämner en av Tom Moriaty.) Denna effekt av solen är naturligtvis inte det enda som styr klimatet. T ex styrs de senaste […]



Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: