Sharada Ramadass
2017-May-14 02:33 UTC
[R] variable scale and transform confusion with glmm
Hello, I am a complete newbie to GLMM and R. I do understand some bit of statistics though I am in no-way a core statistician. So, here are my doubts and I would really appreciate if someone can provide some inputs. I have looked up for prior responses on various lists and could not come up with satisfactory results that clear my confusion. 1. My problem is an ecological problem and I am trying to model growth rate in trees as a response to various predictors (fixed and random). So far, so good. 2. Literature tells me that people use RGR (relative growth rate) to look at growth to account for girth size classes. 3. My AGR or RGR are very small values (mathemetically in terms of numbers) since my timeline for the data is very short. That is my limitation. 4. Some predictors have large values (orders of magnitude, mathematically) while some other others have smaller values. 5. So I have very small values for my growth rate, very large values for some predictors and all the other predictors are in a similar range of values, mathematically. Here are my questions: 1. Does using AGR (absolute growth rate) introduce any bias or inflation in the model if we use AGR instead of RGR? One paper (stoll 1990) did mention the use of AGR over RGR to avoid skewness. 2. I get 'large variance' errors when running lmer on the model with the raw data (both response and predictors). Is that a problem? 3.If I had to transform the data, should I transform it for all predictors and response (independent of which ones are extreme in their values in orders of magnitude)? 4. If I did apply some kind of transformation, how do you interpret the parameter estimates? Do you need to undo the transformation to get correct values? Some posts seem to indicate you need to un-transform the results. 5. For transformation/scaling, I am confused as to what should be done. Some posts suggested simply scaling the variables up/down my multiplicative factors. Again should this be done for all predictors? If done for only select few, do we need to interpret their parameter estimates differently? 6. The scale function in R has also been suggested as a way to do the scaling. This seems to center the mean and not necessarily have just a multiplicative effect? Is this the function to use for transform? Again, only for some variables or for all? 7. Can the response alone be transformed (log or scale) and results interpreted as-is? 8. Is there a certain log transform only that should be applied (to which base)? Again, some posts indicate you can transform to base 10 or natural log while others indicate log transform is natural log only. Thanks and Regards, Sharada
This list is about R programming not statistics, so your post is OT. Try stats.stackexchange.com instead. However, given your admitted statistical ignorance, I think you need a local consultant to lead you through the statistical wilderness, not a remote internet list. Note that, e.g. "which base" to use for logs,is always irrelevant (other than as a matter of convention, possibly). Cheers Bert Bert Gunter "The trouble with having an open mind is that people keep coming along and sticking things into it." -- Opus (aka Berkeley Breathed in his "Bloom County" comic strip ) On Sat, May 13, 2017 at 7:33 PM, Sharada Ramadass <sharada.ramadass at gmail.com> wrote:> Hello, > I am a complete newbie to GLMM and R. I do understand some bit of > statistics though I am in no-way a core statistician. So, here are my > doubts and I would really appreciate if someone can provide some > inputs. > I have looked up for prior responses on various lists and could not > come up with satisfactory results that clear my confusion. > 1. My problem is an ecological problem and I am trying to model growth > rate in trees as a response to various predictors (fixed and random). > So far, so good. > 2. Literature tells me that people use RGR (relative growth rate) to > look at growth to account for girth size classes. > 3. My AGR or RGR are very small values (mathemetically in terms of > numbers) since my timeline for the data is very short. That is my > limitation. > 4. Some predictors have large values (orders of magnitude, > mathematically) while some other others have smaller values. > 5. So I have very small values for my growth rate, very large values > for some predictors and all the other predictors are in a similar > range of values, mathematically. > > Here are my questions: > 1. Does using AGR (absolute growth rate) introduce any bias or > inflation in the model if we use AGR instead of RGR? One paper (stoll > 1990) did mention the use of AGR over RGR to avoid skewness. > 2. I get 'large variance' errors when running lmer on the model with > the raw data (both response and predictors). Is that a problem? > 3.If I had to transform the data, should I transform it for all > predictors and response (independent of which ones are extreme in > their values in orders of magnitude)? > 4. If I did apply some kind of transformation, how do you interpret > the parameter estimates? Do you need to undo the transformation to get > correct values? Some posts seem to indicate you need to un-transform > the results. > 5. For transformation/scaling, I am confused as to what should be > done. Some posts suggested simply scaling the variables up/down my > multiplicative factors. Again should this be done for all predictors? > If done for only select few, do we need to interpret their parameter > estimates differently? > 6. The scale function in R has also been suggested as a way to do the > scaling. This seems to center the mean and not necessarily have just a > multiplicative effect? Is this the function to use for transform? > Again, only for some variables or for all? > 7. Can the response alone be transformed (log or scale) and results > interpreted as-is? > 8. Is there a certain log transform only that should be applied (to > which base)? Again, some posts indicate you can transform to base 10 > or natural log while others indicate log transform is natural log > only. > > Thanks and Regards, > Sharada > > ______________________________________________ > R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide http://www.R-project.org/posting-guide.html > and provide commented, minimal, self-contained, reproducible code.