2
$\begingroup$

I have a dependent variable, which is a multinomial distribution resulting from positive and negative correlation coefficients. This is a study from the field of Dendrochronology.

Multinomial distribution

Those correlation coefficients represent the relationships between tree-growths (i.e. standardized tree-ring widths) and sums of precipitation. For example: There is tree-ring width chronology (TRWi from site 1) and there are summer precipitation in mm (for site 1). An example table is below.

enter image description here

Calculated correlation coefficient is between TRWi and Precipitation variable. Longitude, latitude and altitude are constants for this specific site. Now, there are approximately 1000 different sites and in the first step, correlation coefficients are calculated for each site. At some sites, trees respond positively to higher precipitation sums, while for others this relationship is negative. After 1000 calculations, regression table looks like one below.

enter image description here

Now, it is time for step 2 and I would like to model this variable as a function of longitudes, latitudes and altitudes.

What model would you suggest?

$\endgroup$
  • $\begingroup$ Please expand the question to include more detail of where these correlation coefficients came from. $\endgroup$ – Robert Long Apr 24 at 7:24
  • 2
    $\begingroup$ I added additional information. I hope it is more clear now. $\endgroup$ – JerryTheForester Apr 24 at 7:47
  • 1
    $\begingroup$ Is there a reason why you don't model the ring widths as a response variable, with precipitation, longitude, latitude and altitude as explanatory variables ? $\endgroup$ – Robert Long Apr 24 at 8:03
  • $\begingroup$ This is actually very interesting suggestion. If you look more closely, how each correlation coefficient is calculated, you will note that latitudes, longitudes and altitudes are constant values in this equation (see Expanded question). $\endgroup$ – JerryTheForester Apr 24 at 8:25
  • $\begingroup$ I am not sure what you mean by "constant values". Can you post a link to the actual data ? What is your research question ? $\endgroup$ – Robert Long Apr 24 at 8:37
2
$\begingroup$

Well if your research question really is about the correlation coefficients I suppose you could run a linear regression using the correlation coefficients as predicted variable and altitude, longitude and lattitude as explanatory variables.

However, if your main interest lies in how the relation between tree ring width and precipation changes with location and height I think you are better off using a linear model with interaction effects. You are then effectively estimating the equation:

$Width=\beta_0+\beta_1Precipation+\beta_2altitude+\beta_3latitude+\beta_4longitude+\beta_5Precipation*altitude+\beta_6Precipation*latitude+\beta_7Precipation*longitude$

If the relation between amount of rainfall and growth changes with say altitude that would be picked up by $\beta_5$.

$\endgroup$
  • 1
    $\begingroup$ As far as I know bimodality is not an issue for linear regression per se. You want your residuals to be (roughly) normally distributed, not the original variables. In fact if you use linear regression to compare two means (an unusual scenario, but it is possible) you'd expect your outcome variable to be bimodal $\endgroup$ – Maarten Punt Apr 24 at 14:16
  • 1
    $\begingroup$ Yes, on reflection I agree! (+1) $\endgroup$ – Robert Long Apr 24 at 14:33

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.