Discussion: View Thread

I have to advise a student on how she should analyse a dataset (already collected) for her PhD. I am not sure of the best and most straightforward to implement method and would appreciate any thoughts.

Outcome measure - photosynthetic pigments in moss

Explanatory variables - various environmental variables eg temperature, UV exposure, moisture content

Variables all continuous quantitative, some discrete. All reasonably normally distributed

3 different random samples from one site measured every few days for 22 days of measurement over a summer. Therefore the unit of measurement is day with 3 random samples within each day.

The student has performed multiple regression having considered each sample as independent, she wants to determine which variables predict the pigments.

Clearly the regression is not appropriate as this does not account for the unit of measurement, or the time effect which is likely to be  autoregressive. What sort of method would suit this data ? a mixed model.

The student needs to perform the analysis so I need the most straightforward method hopefully which can be implemented with pull down menus in JMP or SPSS, would be prepared to teach the student in SAS if this is easier. R is not an option.

Any thoughts greatly appreciated, this is not the usual sort of data I work with.
thanks
Marijka 

-------------------------------------------
Dr Marijka Batterham MMedStat PhD AStat AdvAPD
Statistical Consulting Centre
University of Wollongong, Australia
-------------------------------------------

-------------------------------------------
Raymond Hoffmann, PhD
Professor of Biostatistics
Associate Director of QHS
Medical College of Wisconsin
-------------------------------------------
You really need to describe the outcome variable. Any method greatly depends on that.  Also examine the average of the 3 outcomes over time to see whether the temporal response is linear or not. Averaging will take out the potential correlation withing a day for the first steps to identify an appropriate model.  I assume you mean an autoregressive error over time (possible, but not necessarily probable), not an autoregressive model for the main effect (which is extremely unlikely).  Again if she establishes a model, then the errors over days can be examined with an autocorrelation plot. To model the entire data set, a linear or non-linear mixed model or generalized linear mixed model, depending on the outcome.  I find it helpful to think of the with day and between days correlations separately.  One is led to a marginal or random-effects model. 
   One of the simplest introductions to this is Analysis of Longitudinal Data by Diggle, Liang and Zeger (and Hagerty in the most recent edition).
   An alternative is Applied Longitudinal Analysis by Fitmaurice, Laird and Ware (not quite a simple).  For code examples, the SAS for Mixed Models by Littell et al.

Ray

You mention several variables (e.g., pigment, temperature, moisture content).

What data is collected on each sample?  Is it data on two variables, or more than two?


-------------------------------------------
Joel Wiesen
Director
Applied Personnel Research
-------------------------------------------







Original Message:
Sent: 08-07-2013 00:55
From: Raymond Hoffmann
Subject: which method to use for student consult?



-------------------------------------------
Raymond Hoffmann, PhD
Professor of Biostatistics
Associate Director of QHS
Medical College of Wisconsin
-------------------------------------------
You really need to describe the outcome variable. Any method greatly depends on that.  Also examine the average of the 3 outcomes over time to see whether the temporal response is linear or not. Averaging will take out the potential correlation withing a day for the first steps to identify an appropriate model.  I assume you mean an autoregressive error over time (possible, but not necessarily probable), not an autoregressive model for the main effect (which is extremely unlikely).  Again if she establishes a model, then the errors over days can be examined with an autocorrelation plot. To model the entire data set, a linear or non-linear mixed model or generalized linear mixed model, depending on the outcome.  I find it helpful to think of the with day and between days correlations separately.  One is led to a marginal or random-effects model. 
   One of the simplest introductions to this is Analysis of Longitudinal Data by Diggle, Liang and Zeger (and Hagerty in the most recent edition).
   An alternative is Applied Longitudinal Analysis by Fitmaurice, Laird and Ware (not quite a simple).  For code examples, the SAS for Mixed Models by Littell et al.

Ray

There is one outcome variable photosynthetic pigment it is an absorbance reading in nm, continuous normally distributed and

max wind km/hr
cloud cover (%)
moisture content in the moss
hours of sunlight (to one decimal place)
precipitation (mm)
mean atmospheric pressure
mean humidity
max UV light

are the independent variables, she wants to know which of these are related to the pigment
regards
Marijka









Original Message:
Sent: 08-07-2013 09:18
From: Joel Wiesen
Subject: which method to use for student consult?

You mention several variables (e.g., pigment, temperature, moisture content).

What data is collected on each sample?  Is it data on two variables, or more than two?


-------------------------------------------
Joel Wiesen
Director
Applied Personnel Research
-------------------------------------------

Thanks for your response Ray
The outcome variable is quantitative and normally distributed and I would expect it to be a linear model.
We will look at the averages to start with.
I did mean the error term was autoregressive, I have the Littel et al SAS Mixed Models text so will start with that. I am used to running linear mixed models for repeated measures data on individual subjects however in this case the samples differ (as they are selected randomly within the site) each day, I am not sure how to specify this.
Thanks
Marijka





Original Message:
Sent: 08-07-2013 00:55
From: Raymond Hoffmann
Subject: which method to use for student consult?



-------------------------------------------
Raymond Hoffmann, PhD
Professor of Biostatistics
Associate Director of QHS
Medical College of Wisconsin
-------------------------------------------
You really need to describe the outcome variable. Any method greatly depends on that.  Also examine the average of the 3 outcomes over time to see whether the temporal response is linear or not. Averaging will take out the potential correlation withing a day for the first steps to identify an appropriate model.  I assume you mean an autoregressive error over time (possible, but not necessarily probable), not an autoregressive model for the main effect (which is extremely unlikely).  Again if she establishes a model, then the errors over days can be examined with an autocorrelation plot. To model the entire data set, a linear or non-linear mixed model or generalized linear mixed model, depending on the outcome.  I find it helpful to think of the with day and between days correlations separately.  One is led to a marginal or random-effects model. 
   One of the simplest introductions to this is Analysis of Longitudinal Data by Diggle, Liang and Zeger (and Hagerty in the most recent edition).
   An alternative is Applied Longitudinal Analysis by Fitmaurice, Laird and Ware (not quite a simple).  For code examples, the SAS for Mixed Models by Littell et al.

Ray