I have reported life satisfaction as my dependent variable and many independent variables of different kinds. Of them, one is the area in which the individual lives (urban/rural) and other is the access to public provided water service. When the area variable is included in the model, the second variable is non significant. However, when it is excluded, the public service gains enough significance for a 95% level of confidence. The two variables are moderately and negatively correlated (r= -0.45).
What possible explanations do you see for this phenomenom?
Mr. Valdivieso. I think the Multicollinearity is almost definitely the problem. Try the simple VIF test. Then try to change the way you measure the variables Area & Access. Good luck.
Probably colinearity though I can't say for certain. A description of your data and the research question might be helpful. Best, D. Booth
Perhaps the more urban the place the more likelihood of access to public water(vice versa) or the more rural the less access to public water, hence higher risk on human life exposed to that factor.
By removing the area, and the second becoming significant shows the greater effect of the factor "public water" without being confounded by "area".
If there is no access the rural or urban areas, the risk will be high is a naive explanation.
Colinear, yes! Being negatively correlated implies you need to exclude area or rather use area as a blocking factor, then test for significance. Factor Analysis helps in reducing factors needed to explain the variation in a dataset.
Mr. Valdivieso. I think the Multicollinearity is almost definitely the problem. Try the simple VIF test. Then try to change the way you measure the variables Area & Access. Good luck.
Near multicollinearity is not always a near linear relationship between two variables but can also arise if there is a near linear relationship between three of more of your variables. I would recommend Belsey, Kuh and Welsch (2004), Regression Diagnostics - Identifying Influential Data and Sources of Collinearity, Wiley, for a good treatment of this subject. Most statistical/econometric software packages can now do the tests described in this book
I think there is Interaction between these tow independent variables, so that you can make another independent variable called interaction.
I agree with John that multiple variables can be involved. Also, if your sample size is small, that can impact confidence interval and 'significance.'
To consider how well you predict with or without that variable, you could plot both models on the same graphical residual analysis scatterplot for comparison. If the more complex model looks better, some kind of cross-validation would be good to see if you are overfitting.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Applied Statistics