Discussion Topic |
|
This thread has been locked |
k-man
Gym climber
SCruz
|
|
|
Topic Author's Reply - Sep 5, 2014 - 09:33am PT
|
Actually Sketch, I'm just playing with you.
We know that your graph is completely irrelevant. Even as a layperson, I can understand enough of what Ed wrote to see that your graph is meaningless when it comes to applying it to climate change.
I was just poking you to see if you had the character to admit you were mistaken when you posted it, that you actually had no clue as to what it was you were posting.
Now you can go back to your name calling. It's a predictable behavior of yours when it's been shown that you don't know what the hell you're talking about.
|
|
raymond phule
climber
|
|
I think Ed's confidence intervals are overblown from applying the standard deviation of the data points from the trend line to each data point, rather than figuring out a variance/standard deviation measure for the actual measurements for the data points (which are themselves averages of many sensors with who knows what justification for them accurately representing global temperature changes, so certainly there is some variance/standard deviation which could be used there from existing data), which would logically be smaller than the variance/standard deviation Ed used.
This really doesn't make much sense. I believe that Ed either made a mistake or used a method I didn't understand.
It is possibly to use that data and calculate confidence interval for the trend without doing what you propose (and that wouldn't make any sense either)
In either case though the broader point of that particular data not justifying either a warming or cooling trend in particular is likely valid however, though that just means we can't say either warming nor cooling definitively happened over that time period.
Yes, that is true.
The trouble in not using a linear trend is it then becomes hard to justify if warming or cooling happened, so it would not be very useful for an argument such as this. With a linear trend line, the sign on the slope tells all.
The thing is that I believe that it often exist better methods to analyze variable data than linear trends. Large variability and short term intervals are always problematic though.
The sign of the slope neither tells all or much at all if the confidence includes zero with a great margin.
|
|
k-man
Gym climber
SCruz
|
|
|
Topic Author's Reply - Sep 5, 2014 - 11:38am PT
|
... being a dick ...
Name calling, just as predicted...
BTW Sketch, I didn't just show up. I started the thread. You are the one who showed up and started to throw around passive-aggressive attacks (or have you forgotten that too?).
This latest episode of yours started when you posted a graph that's obvious BS, and when I (and others) called you on it, your true colors bled through.
I asked you a couple of times to tell us why your graph has any relevance to our discussion. Instead of trying to explain why it has merit, you start slinging crap.
It's such a common sight from you.
|
|
raymond phule
climber
|
|
I really doubt that you have seen that movie.
Can you give an example of a nightmare scenario in the immediate future from that movie?
|
|
Ed Hartouni
Trad climber
Livermore, CA
|
|
the temperature anomaly uncertainties due to the measurements are likely to be smaller, but then you'd get a disagreement with the hypothesis that the straight line was a good model to the fit.
What I did was to use the data variance from the straight line and then make a model of the variance, that it was normal distributed...
when I do that, it scales the covariance matrix from the least squares fit and pops out the uncertainty of both the intercept and the slope...
the way to think about it is that you could do the observations over again, but rearranging the data randomly according to the gaussian... then you'd get different slopes, the range of those slopes you'd expect to get is given by those dashed lines.
If you don't think I did it correctly you could re-do the analysis and find my error... that's what it's all about, after all, and why I explained it in detail. You could reproduce the calculations independently, no need to have an opinion. You could have a supported criticism.
I'll check it again... it was late last night after a beer....
I found an error and I'll track it down...
|
|
raymond phule
climber
|
|
Sorry, but your post seems like gibberish to me.
I am pretty sure that Ed didn't do like you propose. The variability in the data is not due to measurement errors in the sensors but actually mostly differences in the climate signal (weather). I am not talking about fitting a squiggly line but using more advanced method like filtering or arma (arima?) modeling to actual modeling the data instead of just using a linear least square trend to the data.
One thing that seem to be often overseen (among skeptics atleast) in is that a trend analysis or other methods used to analyse the temperature data is to try to decompose the signal into the underlying signal and the variability/noise.
So T=T_0+V, were T is the temperature, T_0 the "real" temperature (the signal that we want to determine) and V is natural variability or noise.
One important thing is that T_0 should be continuous. That is often not the case with simple trend analysis if you look at different time intervals.
Monolith has many times posted graphs that show what I mean. It really doesn't make sense to talk about a pause because a trend is zero when the trend up to the start of the "pause" has a much lower value.
The underlying signal cant be for example
 in 1980
.2 in 1990
.3 in the end of 1994
.6 in the beginning of 1995
.6 in 1995-2010
|
|
Ed Hartouni
Trad climber
Livermore, CA
|
|
the trend calculator does a running average to filter the data, which I didn't do, and neither does the trend calculator for WoodForTrees.com (unless you ask for that)...
when I correct my error (using 1/sigma rather than sigma on the covariance matrix) I get an uncertainty of ±0.002251șC/yr which is the "1-sigma" uncertainty...
I'll check the effect of the 12 month running average...
thanks for jammer, raymond phule and monolith for catching and checking my calculation.
As you can see from the corrected analysis, the trend line could be positive of negative given the data...
the results for a running average are not that different...
(I wonder how "running average" is defined in the analysis above).
|
|
Chiloe
Trad climber
Lee, NH
|
|
The hourglass-shaped confidence bands that Monolith shows above look like standard linear-regression confidence intervals. These come in two flavors of vanilla:
(1) Confidence interval for the conditional mean of y given x.
(2) Confidence interval for individual predicted values of y given x (which will be wider than (1)).
These answer two different questions. Both are given in any text on regression. As Raymond notes there also are less vanilla approaches including time series regression which we aren't using here, but might be needed in real research. Also as noted by both Ed and Raymond there are many, almost infinitely many, alternatives to straight-line models. Temperature does not change in a straight line so we know at the start a linear model is at best a crude approximation -- and often, it's visibly wrong. In a lot of my examples I use something called lowess regression instead, which is very nonlinear, but that's another story (and it's not physically based either).
Just to complicate things further, all of these "trend lines" we're seeing here (I think) use a method called ordinary least squares (OLS). There are many other ways to fit a straight line model too, and OLS is not always the best.
|
|
Ed Hartouni
Trad climber
Livermore, CA
|
|
I'm using least squares to obtain my fits...
and from that obtaining the "covariance matrix" which I scale by the deduced variance of the data to obtain an estimate on the uncertainties of the fit parameters...
I don't know how monolith's confidence bands are being calculated... though I suspect I could construct a pseudo-chi^2 map and find the 95% CL contours (2 sigma)
|
|
Ed Hartouni
Trad climber
Livermore, CA
|
|
so jammer look up the paper(s) that regularizes the temperature measurements use to construct the "global mean" and report back what the stated uncertainties are...
|
|
Chiloe
Trad climber
Lee, NH
|
|
I can post the formulas for regression confidence bands later today if someone else hasn't (I've got a conference call coming in right now). Basically the standard errors for either (1) or (2) in my note above use the standard deviation of the residuals, se, times an expression that becomes larger with squared distance from the mean of x. That's why they are narrowest in the middle (precisely, where x = xmean) and widen out toward the extremes of the x distribution. Put another way, we have most confidence in our predictions near the center or densest part of the distribution, and less confidence toward the tails.
Incidentally they widen out lots when you go beyond the range of observed x values. So even if the model is true (which with temperature time series it obviously is not!) out-of-sample predictions become highly uncertain.
|
|
raymond phule
climber
|
|
The squiggly line is the best most scientifically agreed upon "model" of the differences in temperature from year to year (to argue it is a direct measurement of temperature, raymond, is a HUGE stretch since that would assume we have and have had a MASSIVE number of sensors all over.
Then I guess that it is good that I don't think that.
It is an average of agreed upon indicators, and that is a "model" since statistical arguments are used to justify certain readings as more reliable and/or to infer missing data), and to interpret what the model is saying about the temperature difference overall we must use a linear model. Otherwise we would just look at that squiggly line and the answer would be clear...
Why do we need to use a linear model?
|
|
AndyMan
Sport climber
CA
|
|
Global warming scammers will keep raking in the cash from gullible Supertopoers, despite:
Arctic sea ice is now back to normal
Antarctic sea ice is now at record HIGHs
No warming for the past 2 decades
The Medieval, Roman Warm Periods and the Holocene were all much warmer than now.
Cyclones are at a 30 year LOW
Rate of sea level rise is DECREASING.
The rate of warming between 1910 and 1940 was greater than any subsequent period.
The rate of warming at the start of man's first civilisation has been 16 times as great as the past 100 years.
|
|
k-man
Gym climber
SCruz
|
|
|
Topic Author's Reply - Sep 5, 2014 - 03:57pm PT
|
Regurgitate and Digest.
|
|
monolith
climber
SF bay area
|
|
Why did the 'pause' increase the long term warming rate?
|
|
TGT
Social climber
So Cal
|
|
|
|
rick sumner
Trad climber
reno, nevada/ wasilla alaska
|
|
Bad, bad, bad The Chief. You're posting unauthorized graphics again. Don't you know that the cyclic nature of warming and cooling apparent in your graph has been smoothed out of official existence by the Mann and cohorts. And, sin of sins, it indicates much of the warming coming out of the LIA occurred prior to CO2's influence.
|
|
rick sumner
Trad climber
reno, nevada/ wasilla alaska
|
|
Yes Chief. It's well known their have been decadal scale temp spikes upwards or downwards as large as 10-15 c going into and out of the Bolling Alerod and Younger Dryas. The magnitude and rapidity of these events make the natural warming coming out of the LIA to the turn of the millenium look absolutely enemic. This measly .8c modern rise is the molehill the charlatans have made a mountain of . As old Ed has said, " the anthropogenic signal is feeble compared to the range of natural variation.
|
|
|
SuperTopo on the Web
|
Let us know!