Can i use OLS regression?

[SeanSweeney]

Hello all,

I have field data collected from 60 forest plots.  Within these plots we measured the proportion of species colonized my two  groups of mycorrhizae fungi that we call AM and ECM.  So for each plot I have the proportion of species colonized by AM (percent AM).  This is the criterion variable.  
I have satellite data collected from 4 dates at irregular intervals, early spring, summer, early Fall, and late Fall.  I'm only using the reflectance from one of several bands, band 4 which measures the level of photosynthetic activity.  So, my variables are as follows:

DV:  Percent AM
IV(1):  B4 reflectance measured in T1
IV(2):  B4 reflectance measured in T2
IV(3):  B4 reflectance measured in T3 and
IV(4):  B4 reflectance measured in T4

The zero-order correlation between my DV and two IVs are low and not significant (-.175 and -.0112) however the partials suggest that they act as controls for the remaining two.  The model falls apart with their absence.  
I have high correlation between some of my IVs.  There are collinearity problems, VIFs range from 2.1 to 8.1 for the four IVs.  

The R2 from the model is 0.75.  An examination of residuals versus predicted indicate no systematic deviation from the reference line that would indicate non-linearity in one or more variables.  The normal prob plot of residuals reveals no serious departure from normality.  I want to use this model to generate a predictive surface of percent AM, to map the mycorrhizae gradient.  

I'm not super concerned with collinearity since this is a predictive model but I'm worried that this is not the appropriate model given that I'm using multiple measures of the same IV.

Any help or suggestions is much appreciated.

Cheers,
Sean

[Bruce Weaver]

Hi Sean.  I have a couple questions.

1. Is percent AM measured just once (presumably at the end), or is it measured on each of the 4 occasions where B4 reflectance is measured?

2. What value are you typically seeing for percent AM?  (My concern is that if they are too close to 0% or 100%, a regression model that treats % AM as continous will not work very well.)

Another option that occurs to me (assuming % AM is measured just once) would be to use GENLIN to run a binomial logistic regression--i.e., a model that uses "Events of Trials" as the outcome rather than a single binary outcome variable.  Events would be a variable recording the number of species (i.e., a count) colonized by AM, and the Trials would be another variable giving the total number of species in that forest plot.  Exp(B) from this model would give you an odds ratio.  

HTH.

p.s. - At the time I replied, your post was visible in Nabble, but had not yet been forwarded to the mailing list.  I wonder if that suggests a problem with your subscription to the mailing list?

SeanSweeney wrote

Hello all,

I have field data collected from 60 forest plots.  Within these plots we measured the proportion of species colonized my two  groups of mycorrhizae fungi that we call AM and ECM.  So for each plot I have the proportion of species colonized by AM (percent AM).  This is the criterion variable.  
I have satellite data collected from 4 dates at irregular intervals, early spring, summer, early Fall, and late Fall.  I'm only using the reflectance from one of several bands, band 4 which measures the level of photosynthetic activity.  So, my variables are as follows:

DV:  Percent AM
IV(1):  B4 reflectance measured in T1
IV(2):  B4 reflectance measured in T2
IV(3):  B4 reflectance measured in T3 and
IV(4):  B4 reflectance measured in T4

The zero-order correlation between my DV and two IVs are low and not significant (-.175 and -.0112) however the partials suggest that they act as controls for the remaining two.  The model falls apart with their absence.  
I have high correlation between some of my IVs.  There are collinearity problems, VIFs range from 2.1 to 8.1 for the four IVs.  

The R2 from the model is 0.75.  An examination of residuals versus predicted indicate no systematic deviation from the reference line that would indicate non-linearity in one or more variables.  The normal prob plot of residuals reveals no serious departure from normality.  I want to use this model to generate a predictive surface of percent AM, to map the mycorrhizae gradient.  

I'm not super concerned with collinearity since this is a predictive model but I'm worried that this is not the appropriate model given that I'm using multiple measures of the same IV.

Any help or suggestions is much appreciated.

Cheers,
Sean