From: Barbo on
I have a survival dataset (see below). I want to fit to Weibull function
S(t)=EXP(-alpha*Time^gamma). If I use Excel Solver, alpha=0.00367 and
gamma=2.32. If I use a method tranforming probability to Ln(Ln(1/S(t)) which
will be a linear function of Ln (t). But I got a different alpha and gamma.
When I compared two alphas adn gammas, the Solver results had a better
goodness of fit. What is the problem? THanks

Month Probability of survival
0 1
1 1
2 0.92
3 0.90
4 0.88
5 0.83
6 0.76
7 0.74
8 0.69
9 0.57
10 0.48
11 0.39
12 0.27
13 0.27
14 0.13

From: Mike Middleton on
Barbo -

> What is the problem? <

There is no problem. You should expect a difference.

Your Solver method (which I prefer) minimizes sum of squared deviations
between actual S and fitted S.

The other method uses transformed values, so it does not yield a better fit.

(Excel's trendline features use transformations to fit the logarithmic,
power, and exponential functions. The approach appears to be a computational
convenience. Better fits are obtained using Solver.)

- Mike
http://www.MikeMiddleton.com
Mike(a)DecisionToolworks.com


"Barbo" <Barbo(a)discussions.microsoft.com> wrote in message
news:CFCE454F-D077-4526-BFE5-66902FFE0A12(a)microsoft.com...
> I have a survival dataset (see below). I want to fit to Weibull function
> S(t)=EXP(-alpha*Time^gamma). If I use Excel Solver, alpha=0.00367 and
> gamma=2.32. If I use a method tranforming probability to Ln(Ln(1/S(t))
> which
> will be a linear function of Ln (t). But I got a different alpha and
> gamma.
> When I compared two alphas adn gammas, the Solver results had a better
> goodness of fit. What is the problem? THanks
>
> Month Probability of survival
> 0 1
> 1 1
> 2 0.92
> 3 0.90
> 4 0.88
> 5 0.83
> 6 0.76
> 7 0.74
> 8 0.69
> 9 0.57
> 10 0.48
> 11 0.39
> 12 0.27
> 13 0.27
> 14 0.13
>
From: Barbo on
Hi Mike, thanks for the reply. So if I want to find a function with best fit,
adding a trendline is not an proper method.

I got the weibull function estimates from someone else and then try to
replicate the results in Excel. When I used Solver, most of the time I can
get the parameter estimates with the best fit. I was told there is no
feasible solution. How to properly use Solver? Thanks

Barbo


"Mike Middleton" wrote:

> Barbo -
>
> > What is the problem? <
>
> There is no problem. You should expect a difference.
>
> Your Solver method (which I prefer) minimizes sum of squared deviations
> between actual S and fitted S.
>
> The other method uses transformed values, so it does not yield a better fit.
>
> (Excel's trendline features use transformations to fit the logarithmic,
> power, and exponential functions. The approach appears to be a computational
> convenience. Better fits are obtained using Solver.)
>
> - Mike
> http://www.MikeMiddleton.com
> Mike(a)DecisionToolworks.com
>
>
> "Barbo" <Barbo(a)discussions.microsoft.com> wrote in message
> news:CFCE454F-D077-4526-BFE5-66902FFE0A12(a)microsoft.com...
> > I have a survival dataset (see below). I want to fit to Weibull function
> > S(t)=EXP(-alpha*Time^gamma). If I use Excel Solver, alpha=0.00367 and
> > gamma=2.32. If I use a method tranforming probability to Ln(Ln(1/S(t))
> > which
> > will be a linear function of Ln (t). But I got a different alpha and
> > gamma.
> > When I compared two alphas adn gammas, the Solver results had a better
> > goodness of fit. What is the problem? THanks
> >
> > Month Probability of survival
> > 0 1
> > 1 1
> > 2 0.92
> > 3 0.90
> > 4 0.88
> > 5 0.83
> > 6 0.76
> > 7 0.74
> > 8 0.69
> > 9 0.57
> > 10 0.48
> > 11 0.39
> > 12 0.27
> > 13 0.27
> > 14 0.13
> >
> .
>
From: Mike Middleton on
Barbo -

> How to properly use Solver? <

Solver's success with nonlinear functions may depend on the initial values
for the changing cells.

For simple functions, like the Weibull, you may have some idea of reasonable
initial values.

- Mike
http://www.MikeMiddleton.com
Mike(a)DecisionToolworks.com



"Barbo" <Barbo(a)discussions.microsoft.com> wrote in message
news:BD2F239D-B24A-49BD-9274-B09576D1F741(a)microsoft.com...
> Hi Mike, thanks for the reply. So if I want to find a function with best
> fit,
> adding a trendline is not an proper method.
>
> I got the weibull function estimates from someone else and then try to
> replicate the results in Excel. When I used Solver, most of the time I can
> get the parameter estimates with the best fit. I was told there is no
> feasible solution. How to properly use Solver? Thanks
>
> Barbo
>
>
> "Mike Middleton" wrote:
>
>> Barbo -
>>
>> > What is the problem? <
>>
>> There is no problem. You should expect a difference.
>>
>> Your Solver method (which I prefer) minimizes sum of squared deviations
>> between actual S and fitted S.
>>
>> The other method uses transformed values, so it does not yield a better
>> fit.
>>
>> (Excel's trendline features use transformations to fit the logarithmic,
>> power, and exponential functions. The approach appears to be a
>> computational
>> convenience. Better fits are obtained using Solver.)
>>
>> - Mike
>> http://www.MikeMiddleton.com
>> Mike(a)DecisionToolworks.com
>>
>>
>> "Barbo" <Barbo(a)discussions.microsoft.com> wrote in message
>> news:CFCE454F-D077-4526-BFE5-66902FFE0A12(a)microsoft.com...
>> > I have a survival dataset (see below). I want to fit to Weibull
>> > function
>> > S(t)=EXP(-alpha*Time^gamma). If I use Excel Solver, alpha=0.00367 and
>> > gamma=2.32. If I use a method tranforming probability to Ln(Ln(1/S(t))
>> > which
>> > will be a linear function of Ln (t). But I got a different alpha and
>> > gamma.
>> > When I compared two alphas adn gammas, the Solver results had a better
>> > goodness of fit. What is the problem? THanks
>> >
>> > Month Probability of survival
>> > 0 1
>> > 1 1
>> > 2 0.92
>> > 3 0.90
>> > 4 0.88
>> > 5 0.83
>> > 6 0.76
>> > 7 0.74
>> > 8 0.69
>> > 9 0.57
>> > 10 0.48
>> > 11 0.39
>> > 12 0.27
>> > 13 0.27
>> > 14 0.13
>> >
>> .
>>