Purpose
Relationships between the health of redcedar and urban heat were
explored with generalized linear mixed effect models.
Models
- List of Models
- All
Areas Analyses
- Model Selection - Temperature Time Series Comparison
- binomial.daily <- glmmTMB(binary.tree.canopy.symptoms ~
dist.from.mean.daily + (1|Area),family=binomial,data=data)
- binomial.am <- glmmTMB(binary.tree.canopy.symptoms ~
dist.from.mean.am + (1|Area),data=data,family=binomial)
- binomial.af <- glmmTMB(binary.tree.canopy.symptoms ~
dist.from.mean.af + (1|Area),data=data,family=binomial)
- binomial.pm <- glmmTMB(binary.tree.canopy.symptoms ~
dist.from.mean.pm + (1|Area),data=data,family=binomial)
- Core Response Variables
- Response 1.0 - Binary Health Response (healthy,
unhealthy)
- binomial.af <- glmmTMB(binary.tree.canopy.symptoms ~
dist.from.mean.af + (1|Area),data=data,family=binomial)
- Response 2.0 - Binary Top-Dieback (healthy, dead
top)
- top.dieback.binomial.af <- glmmTMB(top.dieback ~
dist.from.mean.af + (1|Area),data=data,family=binomial)
- Response 3.0 - Binary Thinning Response (healthy,
thinning)
- thinning.binomial.af <- glmmTMB(thinning ~ dist.from.mean.af +
(1|Area),data=data,family=binomial)
- Co-Factors
- Co-Factor Model 1 - including addition of location
- Response 1.1 - Binary Health Response (healthy,
unhealthy)
- binomial.af.size <- glmmTMB(binary.tree.canopy.symptoms ~
dist.from.mean.af + tree.size.simplified
+(1|Area),data=data.size.site.filter,family=binomial)
- Response 2.1 - Binary Top-Dieback (healthy, dead
top)
- Response 3.1 - Binary Thinning Response (healthy,
thinning)
- Co-Factor Model 2 - including addition of tree size
- Response 1.2 - Binary Health Response (healthy,
unhealthy)
- binomial.af.site <- glmmTMB(binary.tree.canopy.symptoms ~
dist.from.mean.af + tree.size.simplified
+(1|Area),data=data.size.site.filter,family=binomial)
- Response 2.2 - Binary Top-Dieback (healthy, dead
top)
- Response 3.2 - Binary Thinning Response (healthy,
thinning)
- Portland
Analyses
- Model Selection - Temperature Time Series Comparison
- binomial.daily <- glmmTMB(binary.tree.canopy.symptoms ~
mean.temp.daily ,family=binomial,data=data)
- binomial.am <- glmmTMB(binary.tree.canopy.symptoms ~ DN_AM1
,data=data,family=binomial)
- binomial.af <- glmmTMB(binary.tree.canopy.symptoms ~ DN_AF1
,data=data,family=binomial)
- binomial.pm <- glmmTMB(binary.tree.canopy.symptoms ~
DN_PM1,data=data,family=binomial)
- Response 1.3 - Binary Health Response (healthy,
unhealthy)
- binomial.af <- glmmTMB(binary.tree.canopy.symptoms ~ DN_AF1
,data=data,family=binomial)
- Response 2.3 - Binary Top Dieback (healthy, dead
top)
- top.dieback.binomial.af <- glmmTMB(top.dieback ~
DN_AF1,data=data,family=binomial)
- Response 3.3 - Binary Thinning Response (healthy,
thinning)
- thinning.binomial.af <- glmmTMB(thinning ~
DN_AF1,data=data,family=binomial)
- Tacoma
Analyses
- Model Selection - Temperature Time Series Comparison
- binomial.daily <- glmmTMB(binary.tree.canopy.symptoms ~
mean.temp.daily ,family=binomial,data=data)
- binomial.am <- glmmTMB(binary.tree.canopy.symptoms ~ DN_AM1
,data=data,family=binomial)
- binomial.af <- glmmTMB(binary.tree.canopy.symptoms ~ DN_AF1
,data=data,family=binomial)
- binomial.pm <- glmmTMB(binary.tree.canopy.symptoms ~
DN_PM1,data=data,family=binomial)
- Response 1.4 - Binary Health Response (healthy,
unhealthy)
- binomial.af <- glmmTMB(binary.tree.canopy.symptoms ~ DN_AF1
,data=data,family=binomial)
- Response 2.4 - Binary Top Dieback (healthy, dead
top)
- top.dieback.binomial.af <- glmmTMB(top.dieback ~
DN_AF1,data=data,family=binomial)
- Response 3.4 - Binary Thinning Response (healthy,
thinning)
- thinning.binomial.af <- glmmTMB(thinning ~
DN_AF1,data=data,family=binomial)
- King
County Analyses
- Model Selection - Temperature Time Series Comparison
- binomial.daily <- glmmTMB(binary.tree.canopy.symptoms ~
mean.temp.daily ,family=binomial,data=data)
- binomial.am <- glmmTMB(binary.tree.canopy.symptoms ~ DN_AM1
,data=data,family=binomial)
- binomial.af <- glmmTMB(binary.tree.canopy.symptoms ~ DN_AF1
,data=data,family=binomial)
- binomial.pm <- glmmTMB(binary.tree.canopy.symptoms ~
DN_PM1,data=data,family=binomial)
- Response 1.5 - Binary Health Response (healthy,
unhealthy)
- binomial.af <- glmmTMB(binary.tree.canopy.symptoms ~ DN_AF1
,data=data,family=binomial)
- Response 2.5 - Binary Top Dieback (healthy, dead
top)
- top.dieback.binomial.af <- glmmTMB(top.dieback ~
DN_AF1,data=data,family=binomial)
- Response 3.5 - Binary Thinning Response (healthy,
thinning)
- thinning.binomial.af <- glmmTMB(thinning ~
DN_AF1,data=data,family=binomial)
Model Approaches Considered
- Approaches considered
- ordinal vs ordinal - chi squared
- e.g. health category vs holc grade
- categorized as 0 or 1 - binomial categorical
- e.g. probability of healthy vs unhealthy - does that increase with
afternoon temperature
- categorized as healthy, unhealthy, dead - ordinal distribution
- Increases in percent (proportion) canopy dieback - beta distribution
- likley need to account for zero inflation
- e.g. does increases in urban heat increase likelyhood of having
dieback (zi model) AND does increases in urban heat increase precent
dieback proportion for trees with some dieback (conditional model).
Binomial Distribution
- Response variables:
- healthy, unhealthy
- healthy, dead top
- healthy, thinning
Excluding dead trees might make sense biologically, because we’re not
sure what factor actually killed the tree. Also the dead trees may have
been dead for different lengths of time. Some may have been affected by
heat, where others had not. There is good biological reasoning to
exclude the dead trees.
Ordinal Distribution
Ordinal
Distribution Analysis
Response variable: category: healthy, unhealthy, dead
If we want to include three categories in our response variable
(healthy, unhealthy, dead), we may want to consider an ordinal
distribution. An ordinal distribution would be most appropriate for
these categories compared to a multinomial distribution because the
categories have a natural order.
Alternate response variable
Response variable: category: 0% Dieback, 1-29% Dieback, 30-59%
dieback, 60-99% dieback, dead
Beta Distribution
Beta
Distribution Analysis
Response variable: percent canopy dieback
Because the response variable is percent (we can consider it as
proportion too) we can do regression using a beta distribution
(comparing shapes) rather than a linear regression. We also need to use
an distribution zero or one inflated rate.
Note, beta distribution analyses only included observations where
community scientists estimated the percent canopy dieback (which was
optional).
We considered converting the categorical variable for percent canopy
affected (e.g. 1-30% is affected, 31-60% is affected, etc), but
determined it was too awkward and innacurate.
Zero inflated beta regression
- We referenced the following documents for guidance:
- Websites
- Blog Posts
- Vignettes
- Github/Stack Exchange Issue Posts
- Youtube Videos
- Background Papers
Interpreting zi Model Outputs
The conditional output is indicating that for all observations with
greater than 0 dieback proportions, distance from mean am influences the
proprtion of dieback. The zi model tells us which predictor increases
the propability of non-zero. > interesting that the distance from
mean am has a negative estimate though - hard to know if it is
observations higher than the mean or lower than the mean (ie as temp
increases, distance from mean decreases (but higher or lower than
mean))
Including a predictor in the zi= bit of the model would test whether
heat influences whether there is dieback at all.
Possible understanding from: https://stats.stackexchange.com/questions/466047/interpreting-output-for-glmmtmb-for-zero-inflated-count-data
References
- Brooks, M. E., Kristensen, K., van Benthem, K. J., Magnusson, A.,
Berg, C. W., Nielsen, A., Skaug, H. J., Machler, M., & Bolker, B. M.
(2017). glmmTMB balances speed and flexibility among packages for
Zero-inflated Generalized Linear Mixed Modeling. The R Journal, 9(2),
378-400. https://doi.org/10.32614/RJ-2017-066
- Answers reference ‘pages 382-383 explain all components of the model
summary’
- Voelkel, J., and Shandas, V. (2017). Towards Systematic Prediction
of Urban Heat Islands: Grounding Measurements, Assessing Modeling
Techniques. Climate 5, 41. doi: 10.3390/cli5020041.