I am investigating factors that may agricultural household's decision on whether or not will they participate in land circulation. One of the existed literature argues that individuals at the village level may be similar from one to the other hence they are going to use a random intercept model:

Here is the model specification:

$$logit(\pi_{ij})=log(\frac{\pi_{ij}}{1-\pi_{ij}})=\beta_{0j}+\beta_{1j}X_{1ij}+\beta_{2j}X_{2ij}+...+\beta_{hij}X_{hij} +\epsilon_{ij}$$ $$\beta_{0j}=\gamma_{00}+\gamma_{01}G_{1j}+\gamma_{02}G_{2j}+...+\gamma_{0k}G_{kj}+\mu_{0j}$$ $$\beta_{1j}=\gamma_{10}+\gamma_{11}G_{1j}+\gamma_{12}G_{2j}+...+\gamma_{1k}G_{kj}+\mu_{1j}$$ $$\beta_{hj}=\gamma_{h0}+\gamma_{h1}G_{1j}+\gamma_{h2}G_{2j}+...+\gamma_{hk}G_{kj}+\mu_{hj}$$

$\pi_{ij}$ represents the probability of household i in village j decides to participate in the land circulation program, h represents the number of the independent individual level variable while k represents the number of independent variable at village level.

Why can't I simply introduce village dummies and villege*individual interactive dummies and run the logit regression to get the results? What are the differences between two models?

  • $\begingroup$ @MichaelM Do you mean using dummy variables would be a better option? $\endgroup$
    – JoZ
    Nov 10, 2017 at 19:50
  • 2
    $\begingroup$ What do you mean with the village * individual interaction? Do you have more observations per individual? Otherwise, wouldn't this introduce more variables than you have observations (which would be an undefined model)? $\endgroup$
    – Gijs
    Nov 10, 2017 at 19:56
  • $\begingroup$ @MichaelM sorry I misunderstood ... $\endgroup$
    – JoZ
    Nov 10, 2017 at 20:51

1 Answer 1


Random intercept models (and multi-level models in general) allow us to relax the assumption of independent errors. They do this by having two sorts of errors (labeled $\epsilon$ and $\mu$ in your question).

  • $\begingroup$ Adding a lot of dummies would also solve this issue. $\endgroup$
    – Michael M
    Nov 10, 2017 at 18:27
  • 4
    $\begingroup$ No, it doesn't. There is still only one error term, regardless of how many dummies, and the model still assumes independent errors. $\endgroup$
    – Peter Flom
    Nov 10, 2017 at 18:42
  • $\begingroup$ Does that mean by relaxing the assumption of independent errors we can still have unbiased estimator? @MichaelM $\endgroup$
    – JoZ
    Nov 10, 2017 at 18:46
  • $\begingroup$ @PeterFlom: I will add my point as a comment to the OP. No need to speak of error terms in GLM settings anyway ;-). $\endgroup$
    – Michael M
    Nov 10, 2017 at 18:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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