I am struggling at the moment with how to determine links between different sets of variables. I have data on land use/cover changes in a number of regions in a period of 20 years. They are all expressed in percentages of area surface changed. I have data on so called drivers of change (population change, employment levels, education structure changes, slope, altitude etc) also expressed in percentages or indices. I have a third type of variables - a questionnaire survey.

Basically what I did was multiple regression analysis on each independent variable (land cover change - forest change, grassland expansion or reduction, arable land expansion or reduction etc.) with all of the drivers of change. So I would get R for the connection between e.g. deforestation on one side and population change and altitude on the other (other drivers were eliminated through backward stepwise regression).

So after I would determine that forest cover change is affected by changes in population number or altitude, I would use data from the questionnaire survey to interpret the reasons for such connection.

Is there any better way to model these variables? I am dabbling into canonical correlation - could it be useful? Land cover variables on one side, and the drivers on the other? And could I include the data from the questionnaire survey in the canonical correlation in any way? Thank you very much for your help.


A few points:

  1. If you have the same regions for multiple years then your data are not independent, thus violating one assumption of regression. There are different ways to deal with this; the one I am most familiar with is mixed models (aka multilevel models and a variety of other terms) but you do have to deal with it. Another method is generalized estimating equations (GEE).

  2. You have multiple independent variables; so, you can include more than one in a regression. There is a whole art to model building, extensively discussed here and in books.

  3. Since your dependent variable is a percentage, you may need a nonlinear model. If all the percentages are near the middle, then you might not, but you probably should use one anyway.

So, you have a fairly complex problem.

On a much less complex level, you can (and should) make lots of graphs of the variables against each other.

  • $\begingroup$ Actually, I have data for 1991 and 2011, and every variable is just a difference between the first and the last year in the period. it's just one region. I also excluded highly correlated variables (e.g. population density and aging population) before multiple regression. I was hoping that canonical correlation could solve my problems, because it works with two sets of variables, and that's exactly what I have. $\endgroup$ – Miren Sanov Feb 15 '14 at 22:19
  • $\begingroup$ OK, canonical correlation might work for that. You might also want to look into multivariate regression (more than one DV) $\endgroup$ – Peter Flom - Reinstate Monica Feb 15 '14 at 22:30

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.