Between sleep deprivation and it having been years since my last stats
class, I cannot wrap my mind around what should be an easy problem of
curve fitting.  In school I did have econometrics, but now I'm not
quite figuring out how to set this up. (Yes, if you don't use what you
learn you can forget it all. sigh.)

Can the following problem be done as a weighted least squares?  And if
not, in general can I do what I need to do in Excel (with stats
add-ins)? Most stand-alone stats programs are expensive (although some
have free trials).

I've got data on the adoption rate of technologies by year and
country. I want to estimate the adoption curve for certain countries,
and then forecast out just a few years (again, by country). The
literature on technology diffusion and the raw data itself suggest
using a logistic curve with a ceiling:

Penetration at time t  Pt = L / (1 + a exp(-bt))
L is the ceiling, a is the timing variable, b is the
slope.

For each country I have the penetration rate (per 100 or 1000 people)
of the technology for each year since the technology was introduced.
The probability of success here is this rate which is binary at the
individual level-- does a person have a computer (or internet access
or a cellular phone) or not?

I have the population growth rate, and for now will assume that the
max penetration (ceiling) is a fixed % of total population. As this is
more of an exploratory / qualitative look at the data I'm not using
any other explanatory variables, just time. In many of the countries
the technology is mature-- the growth rate is past its the inflection
point-- but none have reached their ceilings.

So essentially this is a logit with pre grouped data, or is it? 

My Gujarati textbook has an example of estimating a and b for a logit
with grouped data using WLS. For each Xi, one obtains the logit
LN(Pi/(1-Pi) = A + B Xi + Ui.  One then multiplies both sides by the
square root of the weight:
Wi = Ni(Pi*)(1-Pi*)
where Ni =  samples at X=i, Pi* = relative frequency at X=i 
to account for heteroscedasticity.  I can understand how this works
when you are grouping individual data. By with my data there isn't a
Ni per se: this is population data already. Can I still uses WLS- what
would be the weight? (As a side question, can't the inflection point
in the change of the growth rate be used to estimate total time to
full penetration, given the assumptions of logistic curves?)

And if not WLS, how should I think about approaching this data? This
sort of curve fitting + slight forecasting should be easy, if I could
just see over the mental block I've got about it. Thank you.

Catharine
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to