2
$\begingroup$

I have a data set, total_data and I applied a model to it. For instance, the model has one parameter $\beta$, and I calculated the log-likelihood of the fitted model (using maximum likelihood method). In the mean time, I have a stratification variable that could divide total_data into two subsets: high_risk_data and low_risk_data. Now, I could apply the same model to each subset, obtaining $\beta_1$ and $\beta_2$. I could sum up the log-likelihood from each fitting and obtain some kind of overall log-likelihood for the total_data under model 2.

Model 1: parameter $\beta$ fitted on total_data;
Model 2: parameters $\beta_1$ and $\beta_2$ fitted on each subset of data.

Can I consider them as nested models, since I have one fewer parameter in model 1? Can I apply the likelihood ratio test to select a better model?

I know a more standard way to obtain something similar to model2: introduce a second stratification variable-group and apply the model ($\beta$,$\beta_{group}$) to the total_data. The likelihood ratio test would be to test against H0:$\beta_{group}=0$. Is this the same as the method I described above?

$\endgroup$
5
  • $\begingroup$ What is 'maxlog-likelihood'? Is it the maximum of the log-likelihood, or the maximizer of it? Whichever it is, why can you sum them? Do you mean instead that the log-likelihoods for the subsets can be added to obtain the overall log-likelihood, or that the maximum value of the overall log-likelihood must always be the sum of the two maximum values obtained on the subsets? $\endgroup$
    – Glen_b
    Jun 20 '13 at 9:40
  • $\begingroup$ Yes, this is what I suppose in this method: the log-likelihoods for the subsets can be added to obtain the overall log-likelihood. I don't know whether this is valid and I need some expert to point out what's the problem with this. $\endgroup$ Jun 20 '13 at 10:17
  • $\begingroup$ You still need to resolve the question of what you mean by "maxlog-likelihood" earlier; I suggest making an edit to clarify both parts of your question. $\endgroup$
    – Glen_b
    Jun 20 '13 at 10:19
  • $\begingroup$ I know a more standard way to obtain something similar to model2: introduce a second stratification variable-group and apply the model ($\beta$,$\beta_{group}$) to the total_data. The likelihood ratio test would be to test against H0:$\beta_{group}=0$. Is this the same as the method I described above? $\endgroup$ Jun 20 '13 at 10:20
  • $\begingroup$ That's a very good question to ask. You should edit to add that to your question. $\endgroup$
    – Glen_b
    Jun 20 '13 at 10:27
1
$\begingroup$

To answer my questions. Yes, they are nested models. Here is an example of linear model. Fitting two separate models to each subgroup:

model 1: $Y=\alpha'+\beta'X_{1}+\epsilon$ (when $X_{2}=group1$)

model 2: $Y=\alpha''+\beta''X_{1}+\epsilon$ (when $X_{2}=group2$)

equals

model3: $Y=\alpha+\beta_{1}X_{1}+\beta_{2}X_{2}+\beta_{3}X_{1}X_{2}+\epsilon$

It turns out that:

$\alpha'=\alpha$

$\beta'=\beta_{1}$

$\alpha''=\alpha+\beta_{2}$

$\beta''=\beta_{1}+\beta_{3}$

and $RSS_{model3}=RSS_{model1}+RSS_{model2}$

Therefore, $Y=\alpha+\beta X_{1}+\epsilon$ is nested in the fitting by using model 1 and 2.

$\endgroup$

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.