2

A model $Y=(\beta_0+\beta_1x)^{-1}+\epsilon$, where $\epsilon \sim N(0,\sigma^2)$ is to be fitted to the data $(x_1,Y_1), (x_2,Y_2), \dots (x_n,Y_n)$.

Is this a linear or non linear model, and why?

Show how it can be made into a GLM and state the link function?

Any help would be great!

Luvpsi
  • 81
  • 5
  • It is linear in the parameters, So it is a linear model. However it is undefined at x=0. – Michael R. Chernick Aug 07 '17 at 00:57
  • Please add the actual question to the body of post. – Firebug Aug 07 '17 at 11:45
  • 5
    @Michael Please explain what meaning you ascribe to "linear." As far as I can tell, this model is not linear in *any* of the parameters or its variable $x$ in the standard mathematical or statistical meaning of the term. And why, exactly, would it be undefined at $x=0$? Except when $\beta_0$ also is zero, one can readily make sense of and compute $(\beta_0+\beta_1(0))^{-1}$ Luvspi, have you seen https://stats.stackexchange.com/questions/148638? It explains the meaning of "linear" regression model. – whuber Aug 07 '17 at 13:47
  • 1
    I thought the parentheses were missing in the original post. So only x was raised to the -1 power. – Michael R. Chernick Aug 07 '17 at 16:44
  • @whuber - My apologies. – Digio Aug 07 '17 at 22:21
  • 1
    @Digio No need for that: whatever you might think of the outcome, your analysis was interesting. – whuber Aug 07 '17 at 22:41
  • Updated the question. The entire content of the parenthesis is raised to -1 power (inverse?). – Luvpsi Aug 07 '17 at 22:42
  • @whuber I had a look at the post above that explains the difference and it is making some sense now. How would I go about getting this model into GLM? – Luvpsi Aug 08 '17 at 22:30

1 Answers1

7

This is just a somewhat unusual generalised linear model with a Gaussian response $Y\sim N(\mu,\sigma^2)$ where the mean of the response $\mu$ is linked to the linear predictor $\eta$ via $$ g(\mu)=\underbrace{\beta_0 + \beta_1 x}_\eta. $$ where $g$ is the inverse link function $g(\mu)=1/\mu$,

The linear predictor $\eta$ is clearly linear in $x$ and, as implied by its name, in the regression coefficients $\beta_0$ and $\beta_1$. However, because of the non-linear link function, $\mu=EY$ is non-linear in the model parameters and in $x$.

To fit this in R do glm(y ~ x, gaussian(link = "inverse")).

Jarle Tufto
  • 7,989
  • 1
  • 20
  • 36