# GLMs must be 'linear in the parameters'

I am experiencing some cognitive dissonance about what 'linear in the parameters' means. For example, here and here.

For example, my understanding is $y_i = \beta_0 + \beta_1\beta_2x_1 + \exp(\beta_3)(x_2)^2 + \epsilon$ is not linear in the parameters, because it has two parameter variables multiplied together (namely ${\beta_1, \beta_2}$).

If $\beta_1$ (say) was replaced with $\gamma_1$, a constant, it would be.

Appreciate if someone could clarify this point.

-
While you Are correct the function is not linear in the parameters, it can be made so with a log transformation – Repmat Mar 18 at 8:42
@Repmat How so? How will log transformation help here? – Nick Cox Mar 18 at 13:28
I don't see that anything, linear or nonlinear, will make $\beta_1, \beta_2$ separately estimable here. More positively, watch that GLM in different contexts means general linear models and generalized linear models, which overlap but are by no means identical classes. – Nick Cox Mar 18 at 13:29
Removing the product and the exp – Repmat Mar 18 at 13:31
OK, so you are reparameterising. That's not transformation (of variables), which was what I was inferring. – Nick Cox Mar 18 at 13:33

Your example model can be reëxpressed to be linear in the parameters $\alpha=\beta_1\beta_2$ & $\zeta=\exp\beta_3$:

$$g(\operatorname{E} Y) = \beta_0 + \alpha x_1 + \zeta x_2^2$$

(Clearly $\beta_1$ & $\beta_2$ aren't separately estimable; a non-linear model wouldn't help there. And note that $\hat\zeta$ must be constrained to be positive.) Some models can't be so reëxpressed:

$$g(\operatorname{E} Y) = \beta_0 + \beta_1 x_1 + x_2^{\beta_2}$$

Some can be though it's not obvious at first: http://stats.stackexchange.com/a/60504/17230.

There's a very thorough discussion of different meanings of "linear" at How to tell the difference between linear and non-linear regression models?.

-
+1 for the diæresis. – amoeba Mar 18 at 12:52
+1 for the ligature. (In fact I use second-hand tremas, which you can pick up cheap now.) – Scortchi Mar 18 at 12:53
The reëxpression loses the information that $\zeta$ is constrained to be positive. – Juho Kokkala Mar 18 at 15:14
@JuhoKokkala: Good point - I'll note that. – Scortchi Mar 18 at 15:16

Linear in the parameters means that you can write your prediction as

$$\beta_0+\sum_{j=1}^px_{ij}\beta_j$$

For some definition of $x_{ij}$. But these x's need not be linear functions of your data. For example, ploynomial fitting of a time series has $x_{ij}=t_i^j$ where $t_i$ is the time associated with data point $i$. The prediction is a non linear function of time, but it is linear in the betas.

UPDATE

In response to the comment, the answer is "sort of". If $\beta_2$ was constant, then the predictor is linear in $\beta_0,\beta_1,\exp (\beta_3)$. It is not linear in $\beta_3$, but a transformation of $\beta_3$. In terms of least squares estimates it doesn't make much difference here.

-
Thanks so much for replying. Perhaps my question isn't clarified by including a transformation of the x's. I am asking about the beta's (parameters), not transforms of x's. Perhaps if you could comment on my specific example above. – Ben S Mar 18 at 6:35

I think it's better for you to understand the three components of the GLM. Esp, you need understand how link function is defined.

You can refer to the page 7 in the slides below. 'linear in the parameters' is true after being transformed by the link function.