X-Message-Number: 26802 References: <> From: David Stodolsky <> Subject: Re: Predicted Cryonics Institute patients to 2044 Date: Fri, 12 Aug 2005 22:20:49 +0200 On 10 Aug 2005, at 14:15, Brent Fox wrote: > Re: Predicted Cryonics Institute patients to 2044 > > In CryoNet Message #26782 David Stodolsky wrote: > > >> The earlier years were eliminated, >> because the slope of the line on a log plot was clearly different >> from that starting in 1990. However, it doesn't appear that including >> that data would make much difference (note that the Predicted for >> earlier years is reasonable). >> > > The more historical data that is incorporated would provide a more > accurate > forecast. Depends upon whether the underlying parameters are stationary. Log plot suggests not: <http://cryin.secureid.org/stories/storyReader$52> Most of this earlier data would be excluded from an exponential fit by the model building program, in any case. Given that the model uses only 2 parameters and we have 15 data points remaining, the power of the test is more than adequate - most models we are looking at are highly significant. > > >> The patients received from ACS are excluded, since we don't know >> their year of suspension and the growth curve for ACS may have been >> different, in any case. >> > > The inclusion of ACS patients should be no different than a last > minute > signup/suspension. ACS is a separate organization and the > suspension dates > would be irrelevant in regards as to when CI accepted the patients. The question is whether we want to model the data or the underlying process. If we want to model the growth of people in suspension or if we want to model CI growth, then the ACS patients should not be included (until we get date data from ACS). However, it doesn't seem like inclusion would make much difference. If we include ACS patients as a group, the observed actually is closer to the current predicted. The only situation in which it would make sense to include the ACS input is if we assume that there will be similar transfers in the future. It seems unlikely that CI will get groups of patients on a regular basis as other organizations terminate their storage activity. > > >> I looked at few other models. The cubic >> function appeared to be better than a quadratic. However, the >> straight line on the log plot indicated an exponential relationship >> is most likely. >> > > In running just the limited historical data from 1990 (being fair > to David, > excluding the ACS patients acquired in 2004), and projecting to > 2050 yields > a more realistic view, and supports a Quadratic Trend growth model. > The > Exponential Growth Model promoted fails accuracy measures (MAPE, > MAD, MSD) > compared with the Quadratic Trend model. MSD is a measure of error of the prediction to the individual data points. I used Root Mean Square Error, which is a measure of the error in the predicted trend. This is more suitable for modeling the process, as opposed to the data. See: Evaluating Goodness-of-Fit in Comparison of Models to Data: <http://66.102.9.104/search?q=cache:fO9bGmDRgL8J:www.lrdc.pitt.edu/ schunn/gof/GOF.doc+%22Mean+Squared+Deviation%22+%22Root+Mean+Square +Error%22&hl=en&client=safari> dss David Stodolsky Skype: davidstodolsky Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=26802