After re-reading all three reports, and obtaining additional information from Murphy (2004), I began to refine my model inputs and realized it was no longer appropriate to refer to my models as reflecting the state of knowledge of others, as I was begining to form my own knowledge base using information from the various reports and other publications. I am currently rereading Meldrum (2006), and reading, for the first time, Glickman (1998). It therefore seems inappropriate to put anyone's name to my model except my own. It can thus be known as the Gc Model, but it changes fairly frequently, so the day's version of the model should be specified. The lens focal length for a particular model should also be specified.
I post here yesterday's version of my model, Gc Model (07-20-09). I assume the lens has a 15-mm focal length, but my Excel worksheet is configured such that I can easily substitute a different focal length. If I want to assess the effects of model uncertainty, I can define the focal length to not only be uncertain with regard to manufacturing tolerances, but also with regard to what lens was used. The model estimates Patty's walking height in frame 352 of the Patterson-Gimlin film. The model is easily extended to provide an estimate of standing height, but, given the uncertainties in walking height alone, I want to limit discussion to estimating walking height before moving on. For the same reason I want to limit discussion to frame 352 at this time.
The model presentation is organized as follows:
- - The equation, presented algebraically with all input and output variables defined qualitatively;
- The input variables, defined quantitatively and graphically, including assumptions the quantitative definitions are based on;
- The output variable, including a graphical representation along with statistical description; and,
- The sensitivity analysis showing a graphical representation of which input variables are driving the uncertainty in the output variable.
During the course of elicting inputs for model refinement, it is easy to separate the people who are problem solvers from those who are problems. Problem solvers follow up any criticism of a set of assumptions or rationale for them with an alternative set of assumptions and rationale. I am aware that there are other approaches that can be used to estimate Patty's walking height. I am not interested in hearing about other approaches in this thread.
The software I'm using to do the modeling is Oracle's Crystal Ball (www.oracle.com/crystalball), which is an add-on to Microsoft's Excel (www.microsoft.com/office/2007-rlt/en-US/Excel).
In developing probability distributions to represent uncertain input variables, I use unbiased estimates of the population standard deviation (en.wikipedia.org/wiki/Unbiased_estimation_of_standard_deviation) to prevent under-representing uncertainty.
I've tried to keep this post brief, but I've obviously failed. However, I could easily drone on about the methodology. So, if there are any questions on my methodology that aren't addressed herein, I'll be happy to answer them.
Equation
Bill's optics equation is simply:
O ft = (D ft)(A mm)/(F mm),
where O is the subject's height (in this case, Patty's walking height in PGF frame 352), D is the distance from the camera lens to the subject, A is the height of the image of the subject on the film, and F is the camera lens focal length. This equation strictly applies only when the lens is focused at infinity, but because we are considering distances on the order of 100 ft, any correction factor for Patty being less than an infinite distance from the camera lens is infinitessimally small and is thus ignored here.
Each of the input variables is represented as an uncertain value described by a probability distribution, and these distributions are presented in the next section. However, Crystal Ball does not allow me to define F directly as an uncertain variable using Bill's bench test results for a 25-mm lens (this matter is discussed below). So, I need to multiply F by an external uncertainty factor, uF. Because the uncertainty factor is bi-directional (positive half the time, negative the other half), I need to multiply the external uncertainty factor by a direction factor, d. As a result, my working version of the above equation becomes:
O ft = (D ft)(A mm)/[(F mm)(1-uFd)].
These two additional variables are unitless.
Input Variables
D, Distance from Patterson's Camera Lens to Patty in Frame 352.
Murphy (2004) reports that Grover Krantz estimated D to be 102 ft. He also reports that he found an estimate by Rene Dahinden, after Rene's death, of 101 ft. Dahinden's estimate is based on a stick he obtained from the site that Patty stepped on or near in the PGF. Thus, he had an exact measurement of the length of the stick. Exactly how these two estimates were derived, I don't know. But the mean and standard deviation of these two estimates can be calculated: 101.5 ft, and 0.8660 ft, respectively. When one's knowledge constraints consist of a mean and standard deviation (in an arithmetic framework), the maximum entropy solution (i.e., the most uncertain probability distribution possible) is a normal distribution, which I denote here as N(101.5, 0.8660) ft, and is depicted as:
.
The 2.5th percentile of this distribution is 99.80 ft; the 97.5th percentile is 103.2 ft.
A decent, relatively simple introduction to maximum entropy is provided in Harr (1997). The principle was first discovered by Shannon (1948), and later expounded upon by Jaynes (1957, and 2003).
Many of the inputs are reported to only two significant digits (D here is reported to three significant digits); thus, the output must be limited to two significant digits. To avoid significant rounding error in the calculation, however, I will retain four significant digits for each of the inputs.
A, Height of Film Image of Patty in Frame 352.
Munns (2009) reports the height of a full frame of 16 mm film to be 0.292 in. He further demonstrates that Patty's image is 15% of a full frame. Thus:
A = (0.15)(0.292 in)(25.4 mm/in) = 1.113 mm.
Gf (2009) provides three estimates of A. In one he notes the height of the film frame to be 7.6 mm, and Patty's image (not in frame 352, but in a relatively nearby frame) to be 16%. Thus:
A = (0.16)(7.6 mm) = 1.216 mm.
In the same report he later states the film frame height (looks like frame 352 to me, but he says it's frame 350) is 1,786 pixels, while Patty's walking height is 268 pixels. Assuming the frame height is 7.6 mm:
A = (268/1,768)(7.6 mm) = 1.152 mm.
Finally, in http://www.bigfootforums.com/index.php?sho...6877&st=363, post #371, Gf shows a copy of frame 352 in which the film height is 961 pixels and Patty's image is 156 pixels. With his previous estimate of the film frame height being 7.6 mm:
A = (156/961)(7.6 mm) = 1.234 mm.
The above four estimates of A -- 1.113, 1.152, 1.216, and 1.234 mm -- have a mean of 1.179 mm and a standard deviation of 0.06099 mm. The maximum entropy solution is N(1.179, 0.06099) mm:
.
The 2.5th percentile of this distribution is 1.059 mm; the 97.5th percentile is 1.299 mm.
F, Focal Length of Patterson's Camera Lens.
As my efforts presented herein started as an attempt to quantify the uncertainty in Bill Munns's estimate of Patty's walking height, I am using a focal length of 15.00 mm. The model can easily be run with a 25-mm lens assumed, and I will post results for that after getting this version posted.
The uncertainty factor for F I derived from Bill Munn's bench test of a 25-mm lens. The maximum entropy solution is a double exponential distribution, which is not an option in Crystal Ball. Therefore, I have to create the uncertainty factor distribution by multiplying two other distributions together, then multiplying the result (actually, 1 less the result) by F. The two other distributions are defined below.
uF, Uncertainty Factor for F.
Bill Munns, in post #217 at www.bigfootforums.com/index.php?showtopic=26877&st=198, bench tested the focal length of a 25-mm lens and arrived at an effective focal length of 25.4 mm. Therefore, the error relative to the 25-mm specification is:
(25.4-25)/25 = 0.01600.
Because negative errors are possible, we take the absolute value to fit a distributioni. If this value represents our best estimate of the uncertainty, and if we further assume that the lower bound of possible errors is 0 (a physical constraint), the maximum entropy solution is an exponential distribution, denoted e(0.01600):
.
To accomodate negative errors, the distribution should be negative half the time. This is accounted for in the next variable.
The 2.5th percentile of this distribution is 0.0003945; the 97.5th percentile is 0.05855.
d, Direction Factor for uF.
To make uF negative half the time and postive the other half, it needs to be multiplied by -1 50% of the time and by 1 the other 50%. This is accomplished with by defining a custom distribution in Crystal Ball and I denote that distribution as CD(-1, 0.500; 1, 0.500):
.
The 2.5th and 97.5th percentiles are -1.000 and 1.000, respectively.
When F is multiplied by (1-uFd), the resulting uncertain representation of F for a 15.00-mm lens is:
.
The 2.5th percentile of F is 14.28 mm; the 97.5th percentile is 15.72 mm.
Output Variable
The model was evaluated by means of a 1,000,000-trial Latin hypercube simulation. The results for O are:
.
Statistics: Forecast values
Trials 1,000,000
Mean 8.0
Median 8.0
Standard Deviation 0.46
Variance 0.21
Skewness 0.074
Kurtosis 3.1
Coeff. of Variability 0.057
Minimum 5.8
Maximum 11
Range Width 4.9
Mean Std. Error 0.00046
Percentiles: Forecast values
2.5% 7.1
50% 8.0
97.5% 8.9.
With 95% of the values spanning a 1.8-ft range, this cannot be regarded as a precise estimate. There is no easy way to validate the results, but they can be compared to some historical values. One must remember that the
Murphy (2004) reports that Patterson's initial impression of Patty's height was that she was 6 ft or greater; Gimlin's is reported as 6.5 to 7 ft. Murphy doesn't specify whether these are walking height or standing height estimates, but because Patty was walking for most of the time she was observed, and because Patterson's horse fell when Patty was first encountered, at which time she stood and looked at her intruders for a while before walking away, I'm assuming that these initial impressions are based on her walking height. Only 1.4% of the values of O presented above fall into the range of 6 to 7 ft.
Murphy also reports six independent historical height estimates (in inches):
- 72 (Krantz),
75.5 (Leclerc),
77 (Grieve),
78 (Bayanov & Bourtsev),
80 (Green), and
87.5 (Glickman).
These six values fit a lognormal distribution with an arithmetic mean of 6.532 ft and an arithmetic standard deviation of 0.4500 ft, denoted LN(6.532, 0.4500) ft. The 2.5th percentile of this distribution is 5.7 ft; the 97.5th percentile is 7.5 ft. The estimate of O above differs from this distribution by an average of 1.4 ft. A difference of 0 lies at the 1.3rd percentile of the distribution of the difference, indicating that O is significantly higher than the distribution derived from the historical values.
Sensitivity Analysis
The sensitivity analysis shows which input variables the model output is most sensitive to. The highly sensitive inputs are the best candidates for refinement, as any reduction of uncertainty in them will result in a significant reduction of uncertainty in the output. The measure of sensitivity is rank correlation between a given input and the output. By squaring the rank correlation coefficient, the contribution to uncertainty is obtained. These contributions are additive, so the contribution to uncertainty in O due to F can be obtained by summing the contributions from uF and d.
The sensitivity analysis is presented below as a chart:
.
The model is most sensitive, by far to A. Because this is the input variable that I would think should be the easiest to measure, I am hoping that careful measurements can be made to refine A substantially. But, regardless of the degree of refinement, O is clearly sensitive to small changes in A.
The focal length, F is the input the model is next most sensitive to. Gf and MANGLER have both stated that manufacturing tolerance for lenses is typically 10%, but no documentation has yet been offered. If their position is true, the uncertainty in F will be will become greater than it is now.
Distance from lens to Patty is currently the most certain variable, but I expect that reviewers may well find me guilty of being overconfident about how D is currently defined. If so, refinement may increase the uncertainty in D and, thus, in O.
Literature Cited
Gigantofootecus, 2009, Review of "The Bill Munns Report": a PG Film Analysis, www.readclip.com/crypto/review.htm.
Glickman, J., 1998, Toward a Resolution of the Bigfoot Phenomenon, North American Science Institute.
Harr, M.E., 1987, Reliability-Based Design in Civil Engineering, McGraw-Hill.
Jaynes, E.T., 1957, Information Theory and Statistical Mechanics, Physical Review 106:620-630.
Jaynes, E.T., 2003, Probability Theory: The Logic of Science, Cambridge University Press.
MANGLER, 2009, Personal Perspective of the PGF, notthemunnsreport.com.
Meldrum, J., 2006, Sasquatch: Legend Meets Science, Tom Doherty Associates.
Munns, B., 2009, The Munns Report, www.themunnsreport.com.
Murphy, C.L., 2004, Meet the Sasquatch, Hancock House Publishers.
Shannon, C.E., 1948, A Mathematical Theory of Communication, Bell System Technical Journal 27: 379-423 and 623-656; cm.bell-labs.com/cm/ms/what/shannonday/shannon1948.pdf.

