Hi:
I have a cross-sectional dataset where the observations are for regions nested in countries (There are 16 countries which have a mean of 9.2 regions). I´m fitting a multilevel model where regions are level 1 and countries are level 2.
The data includes a number of predictor variables, some of which may be expected to be very highly correlated.
I have a question about assessing the correlations between the predictor variables…
When I assess correlations across the entire dataset (i.e. across all regions), the between-predictor correlations tend to be reasonably low. For instance, for variables A and B the correlation coefficient is -0.32.
But, when I assess the correlations within each country (i.e. across regions within each country), there are instances where the coefficients are quite high. For instance, for variables A and B, the largest (as absolute values) three are: 0.98, 0.94, 0.81; and, the smallest are: 0.63, 0.32, 0.30.
In terms of potential consequences of collinearity, when fitting the model there were no evident large changes in the parameters (in terms of magnitude, sign or standard error).
My question is: when assessing the between-predictor correlations in order to consider their implications for predictor coefficients in the fitted model, is it sufficient to consider the correlations across the dataset as a whole, or, do I also need to consider the within-country correlations, too?
Any advice would be greatly appreciated.
Cheers
I have a cross-sectional dataset where the observations are for regions nested in countries (There are 16 countries which have a mean of 9.2 regions). I´m fitting a multilevel model where regions are level 1 and countries are level 2.
The data includes a number of predictor variables, some of which may be expected to be very highly correlated.
I have a question about assessing the correlations between the predictor variables…
When I assess correlations across the entire dataset (i.e. across all regions), the between-predictor correlations tend to be reasonably low. For instance, for variables A and B the correlation coefficient is -0.32.
But, when I assess the correlations within each country (i.e. across regions within each country), there are instances where the coefficients are quite high. For instance, for variables A and B, the largest (as absolute values) three are: 0.98, 0.94, 0.81; and, the smallest are: 0.63, 0.32, 0.30.
In terms of potential consequences of collinearity, when fitting the model there were no evident large changes in the parameters (in terms of magnitude, sign or standard error).
My question is: when assessing the between-predictor correlations in order to consider their implications for predictor coefficients in the fitted model, is it sufficient to consider the correlations across the dataset as a whole, or, do I also need to consider the within-country correlations, too?
Any advice would be greatly appreciated.
Cheers
