I've got the following equation:

T = T0 + K* [(gammma)^n]

I've used LOG to linearize this like: log(T-T0) = log(K) + n*[log(gamma)]

Is that ok to call log(T-T0) = y , log(K) = b , n*[log(gamma)] = a*x

Solve it (y = b + a*x) by means of a regression and then estimate T0 ?

Or would it be of so much error ?

ps: I know T0 might be between 1 and 2.

  • $\begingroup$ Welcome to MSE. Please use MathJax. $\endgroup$ – José Carlos Santos Oct 30 '17 at 16:24
  • $\begingroup$ As you don't know $T_0$ how do you compute the $y$'s? A low tech solution is take your best guess at $T_0$ and fit to $b$ and $a$. Use this to make a better guess at $T_0$ and then refit to $b$ and $a$ and so on until it converges to a stable value (if ever!). There are more sophisticated methods available but maybe this is good enough. $\endgroup$ – user121049 Oct 30 '17 at 17:36
  • $\begingroup$ I'm gonna learn how to use this, Sr Santos. I promisse. I'm using the eye method to estimate T0. Trust me, there's no other way available and this is the point of the entire research. Cheers ! $\endgroup$ – Gilmar Neves Dec 19 '17 at 17:58

Assuming that you have $n$ data points$(\gamma_i,T_i)$ and that you want to fit the model $$T=T_0+K \gamma^n$$, remember that what you want is to minimize $$SSQ=\sum_{i=1}^n\left(T_0+K \gamma_i^n -T_i\right)^2$$ So, there are two possible ways to do it.

First way

As you thought about it, fix $T_0$ at a given value and define $y_i=\log(T_i-T_0)$ which makes the model to be linear $$y=\log(K)+n\log(\gamma)=a+n\log(\gamma)$$ and a linear regression will give you $a$ (then $K=e^a$) and $n$. When this is done, compute the predicted values for the $T$'s and compute $SSQ$ which is a function of $T_0$. Try a few values and plot the results; when you notice a minimum of $SSQ$, you have all elements to start a nonlinear regression with the real model.

Second way

Guess a value for $n$ and define $z_i=\gamma_i^n$. This makes again a linear model $$T=T_0+K z$$ and, again, a linear regression will give you the parameters $T_0$ and $K$. Repeat the same process changing $n$ until you see a minimum of $SSQ$. At this point, you have all elements to start a nonlinear regression with the real model.

The thing which is important to remember is that what is measured is $T$ and not any of its possible transforms.

  • $\begingroup$ Thank you! That's the way I'm doing now (the first one). Since I have no T0 to begin with, I'm performing some estimatives based on an old method (looking at the graph and guessing the interception). I know it's terrible wrong. And for some data I use a previous linear regression and have the intercept value as T0. $\endgroup$ – Gilmar Neves Dec 19 '17 at 18:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.