Hi,
I am using the nlm() function to fit the following exponential function to
some data by minimizing squared differences between predicted and observed
values:
csexponential<- function(x, t1, ti, p){
ti + abs(t1 - ti)*(exp(-(p*(x-1))))
}
As background, the data is performance measured across time. As you might
imagine, we get rapid improvement across the first couple of time points and
then the improvement becomes more gradual. In psychology this is known as
the power law of practice.
For some cases the learning is so rapid that the function appears to have a
notch (imagine an "L" shaped function). In these cases the parameter
estimate for the power, p, is large (typically p>15).
I have repeated the fitting procedure on the same set of data and
"appear"
to have found that with the same starting values and arguments to nlm I get
somewhat different values for p in those cases of extremely rapid learning
described above. For example, one time p=21 and another p=23. The relative
change is not huge, but I would like the parameter estimates to be stable
across replications with the same data/settings.
It is certainly possible that I changed something in the code inadvertantly
and that is why I am observing these discrepancies. However, it is also
possible that there may be some random decision making within nlm. So that
if nlm finds a space where the fits are equally good, it may return slightly
different values each time it is run because it lands in a different
location.
I suspect my later interpretation is false because the estimated values do
not change after each and every replication of the analysis. But I thought I
might ask.
Thanks for any insight you can provide,
Steve Lacey
[[alternative HTML version deleted]]
Some things to try are:
1. the nls function
2. replacing p with 1/p
Steven Lacey <slacey <at> umich.edu> writes:
:
: Hi,
:
: I am using the nlm() function to fit the following exponential function to
: some data by minimizing squared differences between predicted and observed
: values:
:
: csexponential<- function(x, t1, ti, p){
: ti + abs(t1 - ti)*(exp(-(p*(x-1))))
: }
:
: As background, the data is performance measured across time. As you might
: imagine, we get rapid improvement across the first couple of time points and
: then the improvement becomes more gradual. In psychology this is known as
: the power law of practice.
:
: For some cases the learning is so rapid that the function appears to have a
: notch (imagine an "L" shaped function). In these cases the parameter
: estimate for the power, p, is large (typically p>15).
:
: I have repeated the fitting procedure on the same set of data and
"appear"
: to have found that with the same starting values and arguments to nlm I get
: somewhat different values for p in those cases of extremely rapid learning
: described above. For example, one time p=21 and another p=23. The relative
: change is not huge, but I would like the parameter estimates to be stable
: across replications with the same data/settings.
:
: It is certainly possible that I changed something in the code inadvertantly
: and that is why I am observing these discrepancies. However, it is also
: possible that there may be some random decision making within nlm. So that
: if nlm finds a space where the fits are equally good, it may return slightly
: different values each time it is run because it lands in a different
: location.
:
: I suspect my later interpretation is false because the estimated values do
: not change after each and every replication of the analysis. But I thought I
: might ask.
:
: Thanks for any insight you can provide,
: Steve Lacey