Part III.
Mathematical Modeling of Annual Average Temperature Time
Series Data From Wichita, KS
copyright © 2007 Paolo B. DePetrillo, MD
Full
Model
T = + P1 x Cosine {(Year+ P2) *3.1415) / P3 "PDO" term
+ P4 x Cosine {(Year +P5) *3.1415) / P6 "ENSO" term
+ P7 x {Cosine {(Year +P8) *3.1415) / P9} Sunspot Cycle
* (LogSun,lag 0 years + 2*LogSun, lag 1 years + 2*LogSun,
lag 2 years +2*LogSun, lag 3 years+2*LogSun, lag 4 years +
LogSun, lag 5 years )"Sunspot number" term
+ P10 baseline temperature
value [standard error]
P1 = 0.40 [ 0.09]
P2 = -1520 [ 252 ]
P3 = 32 [ 3 ] "PDO" term cycle length
P4 = 0.16 [ 0.09 ]
P5 = -151[ 22]
P6 = 2.02 [ 0.02 ] "ENSO" term cycle length
P7 = -0.009 [ 0.002 ] * {Sunspot Number Term)
P8= 1372 [ 8 ]
P9 = 5.6 [ 0.08 ] Sunspot cycle length
P10 = 13.74 [ 0.06]
T = predicted mean annual temperature
SSE 43.02 DFE 103 MSE 0.41 RMSE 0.64
Comparing to
linear fit
Linear Fit
Model
T = P1 * Year
There is no intercept term since it did not add anything to
the model.
P1 = 0.00550 [0.00004]
SSE 86.9 DFE 126 MSE 0.69 RMSE 0.83 which is clearly better
than a "mean fit" with SSE 93.4
Compare models with the corrected Akaike's Information
Criteria
Linear Fit Full Model
Sum-of-squares 86.9 43.02
Number of data points 113 113
Number of parameters 2 13
Akaike's Information Criteria (corrected, AICc) -23.46
-76.84
Probability model is correct 0.00% 100.00%
Difference in AICc 53.38
Information ratio 391004117144.12
Full Model has a lower AICc than Linear Fit so is more
likely to be the correct model.
It is more than one million times more likely to be correct
than Linear Fit.
Compare models with F test
Model SS DF
Linear Fit (null) 86.9 111
Full Model (alternative) 43.02 100
Difference 43.88 11
Percentage Difference 102.00% 11.00%
Ratio (F) 9.27
P value <0.0001
If Linear Fit (the null hypothesis) were true, there would
be a 0.00% chance of obtaining results that fit Full Model
(the alternative hypothesis) so well.
Since the P value is less than the traditional significance
level of 5%, you can conclude that the data fit
significantly better to Full Model than to Linear Fit.
Thanks to the nice folks at GraphPad
Now, let's try the following. We are going to take the Full
Model above, and ask the question: is there a linear trend
once the cycles are taken into account. This is the key
question!
It looks about the same as the full model. However, I added
an extra parameter (the slope of the trend, if any) and the
SSE = 43.02. This is not different from the Full Model
without the linear trend and you don't have to do anything
except say that there is, in the Cyclical Full Model, NO
LINEAR TREND.
T = P1 x Cosine {(Year+ P2) *3.1415) / P3 "PDO" term
+ P4 x Cosine {(Year +P5) *3.1415) / P6 "ENSO" term
+ P7 x {Cosine {(Year +P8) *3.1415) / P9} Sunspot Cycle
* (LogSun,lag 0 years + 2*LogSun, lag 1 years + 2*LogSun,
lag 2 years +2*LogSun, lag 3 years+2*LogSun, lag 4 years +
LogSun, lag 5 years )"Sunspot number" term
+P10*Year Linear trend parameter
+ P11 baseline temperature
value [standard error]
P1 = 0.40 [ 0.09]
P2 = -1524 [ 306 ]
P3 = 32 [ 3 ] "PDO" term cycle length
P4 = -0.16 [ 0.09 ]
P5 = -151[ 28]
P6 = 2.02 [ 0.03 ] "ENSO" term cycle length
P7 = 0.009 [ 0.003 ] * {Sunspot Number Term)
P8= 1372 [ 7 ]
P9 = 5.6 [ 0.07 ] Sunspot cycle length
P10=0.0001[0.002]
<--------------------------------------The linear
trend(slope) is 0.0001 +/- 0.002 C / year. Which is of
course nonsense since the error term is an order of
magnitude higher!
P11 = 13.5 [ 4]
T = predicted mean annual temperature
SSE 43.02 DFE 102 MSE 0.42 RMSE 0.65 <-------- The error
here also increases! Suggesting that in this model, THERE
IS NO LINEAR TREND IN TEMPERATURE SINCE THE 1880'S!
But you may ask if the sunspot cycle theory is an artifact,
which is a reasonable question. So here is a nice summary
table of all the different models, in order of complexity.
1) Model Parameters T = SSE = Error
2) Mean fit P1 62.67
3) Linear only P1*Year 64.18 (OOPS!)
4) Linear + intercept P1*Year + P2 61.83
5) Sunspot Only P1 x {Cosine {(Year+ P2) *3.1415) / P3}xST
59.57
6)"PDO" Cycle Only P1 x Cosine {(Year+ P2) *3.1415) / P3
54.65 (Good start!)
7) "PDO" + "ENSO" P1 x Cosine {(Year+ P2) *3.1415) / P3
+ P4 x Cosine {(Year +P5) *3.1415) / P6 50.09
8) Full Model P1 x Cosine {(Year+ P2) *3.1415) / P3
+ P4 x Cosine {(Year +P5) *3.1415) / P6
+ P7 x {Cosine {(Year +P8) *3.1415) / P9}x ST
P10 - baseline 43.02
Models 5 - 8 all have the P10 baseline term. Nothing works
without it!
ST = (LogSun,lag 0 years + 2*LogSun, lag 1 years +
2*LogSun, lag 2 years +2*LogSun, lag 3 years+2*LogSun, lag
4 years + LogSun, lag 5 years )
or the "Area under the curve" of the sunspot vs time graph
done by trapezoidal method.
Addendum
After staring at the Full model graph, I
noticed that there was a fairly strong cycle with a period
of 24 years (12.06 x 2). I do not know if this is an
artifact, or a climactic cycle, but for those who are
interested, I have repeated some of the analysis.
To keep it interesting I have also outlined the longest
cycle, which may be evidence of the Pacific Decadal
Oscillation (PDO) using the blue lines. Depending on where
you start measuring trends, you can get quite different
answers for the slope!
T = + P1 x Cosine {(Year+ P2) *3.1415) / P3 "PDO" term
+ P4 x Cosine {(Year +P5) *3.1415) / P6 "24-year cycle?"
+P5 x Cosine {(Year +P6) *3.1415) / P7 "ENSO" term
+ P8 x {Cosine {(Year +P9) *3.1415) / P10} Sunspot Cycle
* (LogSun,lag 0 years + 2*LogSun, lag 1 years + 2*LogSun,
lag 2 years +2*LogSun, lag 3 years+2*LogSun, lag 4 years +
LogSun, lag 5 years )"Sunspot number" term
+ P13 baseline temperature
value [standard error]
P1 = 0.41 [ 0.09] "PDO" term
P2 = -1419 [ 261 ]
P3 = 31 [ 2 ] "PDO" term cycle period years 31 x 2
P4 = 0.23 [ 0.09 ]
P5 = 1283[ 31]
P6 = 12.1 [ 0.6 ]
P7 = -0.16[ 0.08 ])
P8= -153 [ 27 ]
P9 =2.02 [ 0.03 ] "ENSO" term cycle period years 2.02 x 2
P10=0.009[0.002] * {Sunspot Number Term
P11=1374[8]
P12=5.59[0.07] Sunspot cycle period years 5.59 x 2
P10 = 13.74 [ 0.06]
T = predicted mean annual temperature
SSE 40.22 DFE 100 MSE 0.40 RMSE 0.63
I added three extra parameters, but got a better fit - I
will not present the math again!
Conclusions
In Wichita, KS, since 1889, these may have influenced the
mean annual temperature: A cycle of about 64 years that
looks like the PDO, a cycle of about 4 years that looks
like ENSO. A cycle of about 11 years (sunspot cycle) and
the effect of average sunspot number lagged from 2 to 4
years, and possibly a cycle of 24 years which is not named
at this time. I'll leave it to the experts to figure out if
it is an artifact or a real cycle.
Sunspot cycles probably influencer temperature variations
and should probably be tested in GCM models. OK, this is a
stretch, but if other locations show the same pattern, it
strengthens the argument.
According to a linear fit model, we can conclude that since
1889, in Wichita, KS there may be a warming trend of
0.00702 +/- 0.00004 +/- SE Celsius degrees/ Year. However,
the caveat is that this when we remove cyclical influences
with the Full Model, this trend is no longer seen.
This raises the question if the linear trend is an artifact
of attempting a linear fit on a cyclical time series.
The Full Model predicts a gradual decrease in average
annual temperature for the next 20-25 years in Wichita, KS,
based on entering the "cold" half of the PDO cycle as well
as the cold part of the sunspot cycle. I have simply
continued the past 11 years of sunspots,so prediction of
solar activity might help make this a better model. I have
not included the "24 year" cycle in the predictive model
since I don't know of any natural climactic temperature
cycle, like PDO or ENSO, that fits the periodicity of 24
years. If you know of one, please let me know!
Thanks to the folks at the station, there is no missing
data. Even through the Great Depression.
.
Limitations
Data is from 1889 to present. This limits the conclusions
to this time period, namely the magnitude of many of the
parameters.
I have not ruled out an effect of carbon dioxide on
temperature. In stat speak, I do not have the power to rule
it out at some reasonable confidence level. And I don't
reach a reasonable level of significance to rule it in.
Hence, the question is not answered. Like everything else,
it would have been nice to have a lot more data.
There are obviously other factors influencing the
temperature, and of course it is possible that carbon
dioxide may itself influence the length of some of the
cycles. That is an empiric question which needs expert
study.
Next page