Mast Inference and Forecasting (mastif)

James S. Clark

2018-10-22

citation:

Clark, J.S., C. Nunes, and B. Tomasek. 2018. Pulsed-resource mast systems and the movement, demographic storage, and diet breadth of consumers, in review.

website


summary

mastif uses seed counts from seed traps to estimate seed productivity by trees and seed dispersion. Attributes of individual trees and their local environments could explain their differences in fecundity. Inference requires information on locations of trees and seed traps, and predictors (covariates and factors) that could explain source strength.

Many trees are not reproductively mature, and reproductive status is typically unknown, or partially known. For individuals of unknown maturation status, the maturation status must be estimated together with fecundity.

Predictors of fecundity include individual traits, site characteristics, climate, synchronicity with other trees, and lag effects. For studies that involve multiple plots and years, there can be effects of climate between sites and through time. There may be synchronicity between individuals both within and between plots.

Observations incorporate two types of redistribution. The first type is a redistribution from species to seed types. Not all seed types can be definitely assigned to species. The contribution of trees of each species to multiple seed types must be estimated.

The second redistribution from source trees to seed traps. Within a plot-year there is dependence between all seed traps induced by a redistribution kernel, which describes how the seeds produced by trees are dispersed in space. Because ‘seed shadows’ of trees overlap, seeds are assigned to trees only in a probabilistic sense. The capacity to obtain useful estimates depends on the number of trees, the number of seed traps, the dispersal distances of seeds, and the number of species contributing to a given seed type.

In addition to covariates and factors, the process model for fecundity can involve random tree effects, random group effects (e.g., species-location), year effects, and autoregressive lags limited only by the duration of data. Dispersal estimtes can include species and site differences (Clark et al. 2013).

mastif simulates the posterior distributions of parameters and latent variables using Gibbs sampling, with direct sampling, Hamiltonian Markov chain updates, and Metropolis updates. mastif is implemented in R and C++ with Rcpp and RcppArmadillo libraries. An algorithm summary is provided at the end of this vignette (appendix).


mastif inputs

The model requires a minimum of six inputs directly as arguments to the function mastif. There are two formulas, two character vectors, and four data.frames:

Table 2. Basic inputs

object mode components
formulaFec formula fecundity model
formulaRep formula maturation model
specNames character vector names in column species of treeData to include in model
seedNames character vector names of seed types to include in model: match column names in seedData
treeData data.frame tree-year by variables, includes columns plot, tree, year, diam, repr
seedData data.frame trap-year by seed types, includes columns plot, trap, year, one for each specNames
xytree data.frame tree by location, includes columns plot, tree, x, y
xytrap data.frame trap by location, includes columns plot, trap, x, y

I introduce additional inputs in the sections that follow, but start with this basic model.


simulated data

Data simulation is recommended as the place to start. A simulator allows me to evaluate how well parameters from be identified from my data. The inputs to the simulator can specify the following:

Table 3. Simulation inputs to mastSim

mastSim input explanation
nplot number of plots
nyr mean no. years per plot
ntree mean no. trees per plot
ntrap mean no. traps per plot
specNames tree species codes used in treeData column
seedNames seed type codes: a column name in seedData

Most of these inputs are stochasiticized to vary structure of each data set. Here are inputs to the simulator for a species I’ll call acerRubr:

seedNames  <- specNames  <- 'acerRubr'
sim <- list(nplot=5, nyr=10, ntree=30,  ntrap=40,
               specNames = specNames, seedNames = seedNames)

The list sim holds objects needed for simulation. Here is the simulation:

inputs     <- mastSim(sim)        # simulate dispersal data
seedData   <- inputs$seedData     # year, plot, trap, seed counts
treeData   <- inputs$treeData     # year, plot, tree data
xytree     <- inputs$xytree       # tree locations
xytrap     <- inputs$xytrap       # trap locations
formulaFec <- inputs$formulaFec   # fecundity model
formulaRep <- inputs$formulaRep   # maturation model
trueValues <- inputs$trueValues   # true states and parameter values

I first summarize these model inputs generated by mastSim.

mastSim

mastSim generates inputs needed for model fitting with mast. These objects are simulated data, including the four data.frames (treeData, seedData, xytree, xytrap) and formulas (formulaFec, formulaRep). Other objects are “true” parameter values and latent states in the list truevalues used to simulate the data (fec,repr,betaFec, betaRep, upar, R). I want to see if mast can recover these parameter values for the type of data I simulated.

Here is a mapping of objects created by mastSim to the model summarized above:

Table 4. Some of the objects from the list mastSim.

mastSim object variable explanation
trueValues$fec \(\psi_{ij,t}\) conditional fecundity
trueValues$repr \(\rho_{ij,t}\) true maturation status
treeData$repr \(z_{ij,t}\) observed maturation status (with NA)
trueValues$betaFec \(\beta^x\) coefficients for fecundity
trueValues$betaRep \(\beta^v\) coefficients for maturation
trueValues$R \(\mathbf{r}\) specNames to seedNames matrix, rows = \(\mathbf{r}_h\)
seedData$active in \(A_{sj,t}\) fraction of time trap is active
seedData$area in \(A_{sj,t}\) trap area

mastSim assumes that predictors of maturation and fecundity are limited to intercept and diameter. Thus, the formulaFec and formulaRep are identical. Here is fecundity:

formulaFec

Before fitting the model I say a bit more about input data.

treeData and xytree

The information on individual trees is held in the tree-years by variables data.frame treeData. Here are a few lines of treeData generated by mastSim,

head(treeData)

treeData has a row for each tree-year. Several of these columns are required:

data.frame treeData columns

treeData column explanation
plot plot name
tree identifier for each tree, unique within a plot
year observation year
species species name
diam diameter is a predictor for fecundity in \(\mathbf{X}\) and maturation in \(\mathbf{V}\)
repr immature (0), mature (1), or unknown (NA)

There are additional columns here, but these are not required for model fitting. The variable repr is reproduction status, often NA.

There is a corresponding data.frame xytree that holds tree locations.

Table 5. data.frame xytree columns

xytree column explanation
plot plot name
tree identifier for each tree, unique within a plot
x, y locations in the sample plot

xytree has fewer rows than treeData, because it is not repeated each year–it assumes that tree locations are fixed. They are cross-referenced by plot and tree:

head(xytree, 5)

seedData and xytrap

Seed counts are held in the data.frame seedData as trap-years by seed types. Here are a few lines of seedData,

head(seedData)

These columns are required.

Table 6. data.frame seedData columns

seedData column explanation
plot plot name
trap identifier for each trap, unique within a plot
year seed year
active fraction of collecting period that the trap was active
area collection area of trap (e.g., m\(^2\))
acerRubr,… seedNames for columns with seed counts

Seed trap location data are held in the data.frame xytrap.

Table 7. data.frame xytrap columns

xytrap column explanation
plot plot name
trap identifier for each trap, unique within a plot
x, y map location

There is no year in xytrap, because locations are assumed to be fixed. If a seed trap location is moved during a study, simply assign it a new trap name. Then seed counts for years when it is not present are assigned NA (missing data) will be imputed.


data summary and maps

Because they are stochastic, not all simulations from mastSim generate data with sufficient pattern to allow model fitting. I can look at the relationship between tree diameters and seeds on a map for plot mapPlot and year mapYear:

dataTab <- table(treeData$plot,treeData$year)

w <- which(dataTab > 0,arr.ind=T) # a plot-year with observations
w <- w[sample(nrow(w),1),]

mapYears <- as.numeric( colnames(dataTab)[w[2]] )
mapPlot  <- rownames(dataTab)[w[1]]
inputs$mapPlot    <- mapPlot
inputs$mapYears   <- mapYears
inputs$treeSymbol <- treeData$diam
inputs$SCALEBAR   <- T

mastMap(inputs)

Depending on the numbers of maps I adust treeScale and trapScale to see how seed accumulation compares with tree size or fecundity. In the above map, trees are shown as green circles and traps as gray squares. The size of the circle comes from the input variable treeSymbol, which is set to tree diameter. I could instead set it to the ‘true’ number of seeds produced by each tree, an output from mastSim.

inputs$treeSymbol <- trueValues$fec
inputs$treeScale  <- 2
inputs$trapScale  <- 1

mastMap(inputs)

Which are reproductively mature? Here are true values, showing just trees that are mature:

inputs$treeSymbol <- trueValues$repr
inputs$treeScale  <- .5
mastMap(inputs)

To fit the data, there must be sufficient seed traps, and sources must not be so dense such that overlapping seed shadows make the individual contributions undetectable (Appendix).

Here is the frequency distribution of seeds (observed) and of fecundities (unknown) from the simulated data:

par( mfrow=c(2,1),bty='n', mar=c(4,4,1,1) )
seedData <- inputs$seedData
seedNames <- inputs$seedNames

hist( as.matrix(seedData[,seedNames]) ,nclass=100, 
      xlab = 'seed count', ylab='per trap', main='' )
hist( trueValues$fec,nclass=100, xlab = 'seeds produced', ylab = 'per tree',
      main = '')

In data of this type, most seed counts and most tree fecundities are zero.


model fitting

Model fitting requires specification of the number of MCMC iterations ng and the burnin. Here is an analysis using the simulated inputs from mastSim with a small number of iterations:

output   <- mastif( inputs = inputs, ng = 1500, burnin = 500 )

The fitted object output contains MCMC chains, estimates, and predictions, summarized in the next section.

output summary, lists, and plots

output summary

Sample size, parameter estimates, goodness of fit are all provides as tables by summary.

summary( output )

By default, this summary is sent to the console. It can also be saved to a list, e.g., outputSummary <- summary( output ).

estimates, and predictions in output

The main objects returned by mastif include several lists summarized in Table 9. parameters are estimated as part of model fitting. predictions are generated by the fitted model, as predictive distributions. Note that the latent states \(\psi_{ij,t}\) and \(\rho_{ij,t}\) are both estimated and predicted.

Table 9. The list created by function mast.

list in output summary contents
inputs from inputs with additions includes distall (trap by tree distance)
chains MCMC chains agibbs (if random effects, the covariance matrix \(\mathbf{A}\)), bfec (\(\boldsymbol{\beta}^x\)), brep (\(\boldsymbol{\beta}^v\)), bygibbs (\(\alpha_l\) or \(\gamma_t\) if yearEffect included), rgibbs (\(\mathbf{R}\) if multiple seed types), sgibbs (\(\sigma^2\), RMSPE, deviance), ugibbs (\(u\) parameter)
data data attributes intermediate objects used in fitting, prediction
fit diagnostics DIC, RMSPE, scoreStates, predictScore
parameters posterior summaries alphaMu and alphaSe (mean and se for \(\mathbf{A}\), if included), aMu and aSe (\(\mathbf{\beta^w}_{ij}\)), betaFec (\(\boldsymbol{\beta}^x\)), betaRep (\(\boldsymbol{\beta}^v\)), betaYrMu and betaYrSe (mean and se for year or lag coefficients), upars and dpars (dispersal parameters \(u\) and \(d\)), rMu and rSe (mean and se for \(\mathbf{R}\)), sigma (\(\sigma^2\)), acfMat (group by lag empirical correlation or ACF), pacfMat (group by lag partial correlation or PACF), pacfSe (se for pacfMat), pacsMat (group by lag PACF for seed data)
prediction predictive distributions fecPred (maturation \(\rho_{ij,t}\) and fecundity \(\psi_{ij,t}\) estimates and predictions), seedPred (seed counts per trap and predictions per m\(^2\)), seedPredGrid predictions for seeds on the space-time prediction grid, treePredGrid predictions for maturation and fecundity corresponding to seedPredGrid.

Prediction scores for seed-trap observations are provided in output$fit based on the estimated fecundities \([\mathbf{y} | \phi, \rho]\) as scoreStates and on the estimated parameters \([\mathbf{y} | \phi, \rho][\phi, \rho| \boldsymbol{\beta}^x, \boldsymbol{\beta}^w, \dots]\) as predictScore. The former will be substantially higher than the latter in cases where estimates of states \(\phi, \rho\) can be found that predict seed data, but the variables in \(\mathbf{X}, \mathbf{V}\) do not predict those states well. These proper scoring rules are bases on the log likelihood for the Poisson distribution.

Predictions for a location-year prediction grid are invoked when predList is specified in the call to mastif.

plots of output

Here are plots of output, with the list plotPars passing the trueValues for these simulated data:

plotPars <- list(trueValues = trueValues)
mastPlot(output, plotPars)

Because this is a simulated data set, I pass the trueValues in the list plotPars.

By inspecting chains I decide that it is not converged and continue, with output from the previous fit being the new inputs:

output   <- mastif( inputs = output, ng = 5000, burnin = 3000 )

Other arguments passed to mastPlot in the list plotPars are given as examples below and listed at help(mastPlot).

mastPlot generates the following plots:

  • maturation: chains for maturation parameters in \(\beta^v\).

  • fecundity: chains for fecundity parameters in \(\beta^x\).

  • dispersal parameter: chains for dispersal parameter \(u\).

  • variance sigma: chains for error variance \(\sigma^2\), RMSPE, and deviance.

  • fecundity, maturation: posterior 68% (boxes) and 95% (whiskers) for \(\beta^x\) and \(\beta^v\).

  • maturation, fecundity by diameter: estimates (dots) with \(95%\) predictive means for latent states.

  • seed shadow: with 90% predictive interval

  • prediction: seed data predicted from estimates of latent maturation and fecundity (a) and from parameter estimates. If trueValues are supplied from mastSim, then panel (c) includes true versus predicted values. There is an important distinction between (a) and (b)–good predictions in (a) indicate that combinations of seed sources can be found to predict the seed data, without necessarily meaning that the process model can predict maturation and fecundity. Good predictions in (b) face the steeper challenge that the process model must predict both maturation and fecundity, which, in turn, predict seed rain.

  • parameter recovery: if trueValues are supplied from mastSim, then this plot is provided comparing true and estimated values for \(\beta^v\) and \(\beta^x\).

  • predicted fecundity, seed data: maps show predicted fecundity of trees (sizes of green circles) with seed data (gray squares). If predList is supplied to mastif, then the predicted seed surface is shown as shaded contours.

  • production by plot and year: predictions showing variation across plot area

  • partial ACF: autocorrelation function for fecundity (a) and in seed counts (b).

  • tree correlation in time: pairwise correlations between trees on the same plot.

The plots displayed by mastPlot include the MCMC chains that are not yet converged. Again, I can restart where I left off by using output as the inputs to mast. In addition, I predict the seed surface from the fitted model for a plot and year, as specified in predList.

predList <- list( mapMeters = 3, plots = mapPlot, years = mapYears ) 
output   <- mastif( inputs = output, ng = 2000, burnin = 1000, 
                    predList = predList)

To look closer at the predicted plot-year I generate a new map:

#output$treeSymbol <- output$prediction$fecPred$fecEstMu
output$mapPlot    <- mapPlot
output$mapYears   <- mapYears
output$treeScale  <- 1.5
output$trapScale  <- .8
output$PREDICT <- T
output$scaleValue <- 20
output$plotScale  <- 1
output$COLORSCALE <- T
output$LEGEND     <- T
  
mastMap(output)

The maps show seed counts (squares) and fecundity predictions (circles). For predicted plot years there is also shown a seed prediction surface. The surface is seeds per m\(^2\).

Depending on the simulation, convergence may require more iterations. The partial autocorrelation for years should be weak, because none are simulated in mastSim. However, actual data will contain autocorrelation. The fecundity-time correlations show modal values near zero, because simulated data do not include individual covariance. Positive spatial covariance is imposed by dispersal.

To send output to a single Rmarkdown file and html or pdf, I use this:

plotPars$RMD <- 'pdf'
mastPlot(output, plotPars)

This option will generate a file mastifOutput.Rmd, which can be opened in Rstudio, edited, and knitted to html format. It contains data and posterior summaries generated by summary and mastPlot. This pdf will not include stand maps. Maps will be included in the html version, obtained by setting plotPars$RMD <- 'html'.


slow convergence?

Seed shadow models confront convoluted likelihood surfaces, in the sense that we expect local optima. These surfaces are hard to traverse, because maturation status is a binary state that must be proposed and accepted together with latent fecundity. Posterior simulation can get bogged down when fecundity estimates converge for a combination of mature and immature trees that is locally but not globally optimal. Especially when there are more trees of a species than there are seed traps, we expect many iterations before the algorithm can ‘find’ the specific combination of trees that together best describe the specific combination of seed counts in many seed traps. Compounding the challenge is the fact that, because both maturation and fecundity are latent variables for each individual, the redistribution kernel must be constructed anew at each MCMC step. Yet, analysis of simulated data shows that convergence to the correct posterior distribution is common. It just may take trial and error with prior parameter values (see below) and long chains.

Examples in this vignette assign enough interations to show this progress toward convergence, sufficiently few to avoid long waiting times. To evaluate convergence, consider the plots for chains of sigma (the residual variance \(\sigma\) on log fecundity), the rspse (the seed count residual), and the deviance. Finally, the plot of seed observed vs predicted gives a sense of progress.

As demonstrated above, restart mastif with the fitted object.


multiple seed types per species

Often a seed type could have come from trees of more than one species. Seeds that are only identified to genus level include in seedNames the character string UNKN. In this simulation the three species pinuTaeda, pinuEchi, and pinuVirg contribute most seeds to the type pinuUNKN. Important: there can be only one element of seedNames containing the string UNKN.

Here are some inputs for the simulation:

specNames <- c('pinuTaeda','pinuEchi','pinuVirg')
seedNames <- c('pinuTaeda','pinuEchi','pinuVirg','pinuUNKN')
sim    <- list(nyr=4, ntree=25, nplot=10, ntrap=50, specNames = specNames,
                  seedNames = seedNames)

There is a specNames-by-seedNames matrix that is estimated as part of the posterior distribution. For this example seedNames containing the string UNKN refers to a seed type that cannot be differentiated beyond the level of the genus pinu.

Here is the simulation with output objects with 2/3 of all seeds identified only to genus level:

inputs <- mastSim(sim)        # simulate dispersal data
R      <- inputs$trueValues$R # species to seedNames probability matrix
round(R, 2)

The matrix R stacks the length-\(R\) vectors \(\mathbf{r}'_s\) discussed in the model description as a species-by-seed type matrix. There is a matrix for each plot, here stacked as a single matrix. This is the matrix of values used in simulation. mastif will estimate this matrix as part of the posterior distribution.

Here is a model fit:

output <- mastif( inputs = inputs, ng = 2000, burnin = 1000) 

The model summary now includes estimates for the unknown R as the “species to seed type matrix”:

summary( output )

Output plots will include chains for estimates in R. There will also be estimates for each species included in specNames:

plotPars <- list(trueValues = inputs$trueValues, RMD = 'pdf') 
mastPlot(output, plotPars)

Again, the .Rmd file can be knitted to a pdf file.

Here is a restart, now with plots and years specified for prediction in predList:

tab   <- with( inputs$seedData, table(plot, year) )
years <- as.numeric( colnames(tab)[tab[1,] > 0] ) # years for 1st plot
predList <- list( plots = 'p1', years = years ) 
output <- mastif( inputs = output, ng = 3000, burnin = 1500, 
                  predList = predList)
mastPlot(output)

prior parameter values

In previous examples, the prior distribution was non-informative, because I did not specify parameter values, and the default parameters make the distribution non-informative. I can specify prior values as inputs through the inputs$priorList or inputs$priorTable holding parameters named in Table 8.

Table 8. Components of inputs with default values.

object explanation
priorDist = 20 prior mean dispersal distance parameter (\(m\))
priorVDist = 10 prior variance dispersal parameter
minDist = 2 minimum mean dispersal distance (\(m\))
maxDist = 40 maximum mean dispersal distance (\(m\))
maxF = 1e+8 maximum fecundity (seeds per tree-year)
minDiam = 10 minimum diam below which a tree can mature (\(cm\))
maxDiam = 40 maximum diam above which a tree can be immature (\(cm\))
sigmaMu = 1 prior mean residual variance \(\sigma^2\)
sigmaWt = sqrt(nrow(treeData)) weight on prior mean (no. of observations)

The mean and variance of the dispersal kernel, priorDist and priorVDist, refer to the dispersal parameter \(u\), transformed to the mean distance for the dispersal kernel (\(\bar{d}\) in the Appendix). The parameters minDist and maxDist place minimum and maximum bounds on \(d\).
Fecundity \(\phi\) has a prior maximum value of maxF.

minDiam and maxDiam bound diameter ranges where a tree of unknown maturation status can be mature or not. This prior range is overridden by values in the treeData$repr column that establish an observed maturation state for a tree-year (0 - immature, 1 - mature).

In general, large-seeded species, especially those that can be dispersed by vertebrates, generate noisy seed-trap data, despite the fact that the bulk of the counts still occur close to the parent, and the mean dispersal distance is relatively low. Large-seeded species produce few fruits/seeds. Long-distance dispersal cannot be estimated from inventory plots, regardless of plot size, because the fit is dominated by locally-derived seed. Consider values for maximum fecundity as low as maxF = 10000, minimum dispersal minDist = 2, and maximum dispersal maxDist = 12. Recall that the latter values refer to the dispersal kernel parameter value, not the maximum distance a seed can travel, which is un-constrained.

Prior parameter values can be passed as a list, e.g.,

inputs$priorList <- list(minDiam = 25, maxDiam = 60)

Alternatively, prior parameter values can be passed as a table, with parameters for column names for specNames for rownames. Here is a file holding parameter values and the function mastPriors to obtain a table for several species in the genus \(Pinus\),

d <- "https://github.com/jimclarkatduke/mast/blob/master/priorParameters.txt?raw=True"
download.file(d, destfile="priorParameters.txt")

specNames <- c("pinuEchi","pinuRigi","pinuStro","pinuTaed","pinuVirg")
priorTable <- mastPriors("priorParameters.txt", specNames, 
                         code = 'code4', genus = 'pinus')

The argument code = 'code4' specifies the column in priorParameters.txt holding the codes corresponding to specNames. The genus is provided to allow for genus-level parameters in cases where priors for individual species have not been specified in priorParameters.txt.

Note that betaPrior specifies predictor effects by sign. The use of prior distributions that are flat, but truncated, is two-fold (Clark et al. 2013). First, where prior information exists, it often is limited to a range of values. For regressions coefficients (e.g., \(\boldsymbol{\beta}\)) we typically have a prior belief about the sign of the effect (positive or negative), but not its magnitude or appropriate weight. Second, within bounds placed by the prior, the posterior has the shape of the liklelihood, unaffected by a prior weight. Prior weight is hard to specify sensibly for this hierarchical, non-linear model. An example using betaPrior is provided below.


specNames and seedNames

Based on applications to data sets from a large number of labs, mastif adopts conventions for misidentification errors applied to seeds. Some data sets include different groups of species in specNames and seedNames that could be mistaken for one another. In these cases, estimates of the matrix R reflect potential misidentifications.

In many data sets, the seeds of multiple species are not differentiated. To estimate true seed identity (and ID error), the undifferentiated seed type occupies a column in seedData having a column name that includes the string UNKN. For a concrete example, if specNames on a plot include trees in treeData$species with the specNames pinuTaed and pinuEchi, then many seeds might be indentified as seedNames pinuUNKN. In these cases, estimates of the matrix R reflect the contributions of each species to the UNKN type. This could be specified in several ways:

specNames <- c('pinuTaed', 'pinuEchi')

#seeds never differentiated:
seedNames <- c('pinuUNKN') 

#one species sometimes differentiated:
seedNames <- c('pinuTaed', 'pinuUNKN')    

#both species sometimes differentiated:
seedNames <- c('pinuTaed', 'pinuEchi', 'pinuUNKN')

mastPlot written to files

The function mastPlot generates summary plots of output. The user has access to all objects used to generate these plots, as discussed in the previous section. By default plots in mastPlot can be rendered to the console. If plotPars$SAVEPLOTS = T is included in the list plotPars passed to mastPlot, then each plot is saved to a .pdf file. Plots can be consolidated into a single .Rmd file, which can be knitted to .html or .pdf with plotPars$RMD = html or plotPars$RMD = pdf.


my data

For illustration I use sample data analyzed by Clark et al. (2013), with data collection that has continued through 2017. It consists of multiple years and sites.

Here I load data for species with a single recognized seed type, Liriodendron tulipifera. Here is a map, with seed traps scaled by seed counts, and trees scaled by diameter,

library(repmis)
d <- "https://github.com/jimclarkatduke/mast/blob/master/liriodendronExample.rData?raw=True"
source_data(d)
mapList <- list( treeData = treeData, seedData = seedData, 
                 specNames = specNames, seedNames = seedNames, 
                 xytree = xytree, xytrap = xytrap, mapPlot = 'DUKE_EW', 
                 mapYears = 2011:2014, treeSymbol = treeData$diam, 
                 treeScale = 1.2, trapScale = 1.5, plotScale = 2, 
                 SCALEBAR=T, scaleValue=50)
mastMap(mapList)

Here are a few lines of treeData, which has been trimmed down to this single species,

head(treeData, 3)

Here are a few lines of seedData,

head(seedData, 3)

diameter effect

Here I fit the model using log(diam); the diam column will be fitted as log diameter. In addition, I specify prior parameter values directly in the inputs list.

formulaFec <- as.formula( ~ log(diam))    # fecundity model
formulaRep <- as.formula( ~ log(diam))    # maturation model

inputs   <- list(specNames = specNames, seedNames = seedNames, 
                 treeData = treeData, seedData = seedData, xytree = xytree, 
                 xytrap = xytrap, priorTable = priorTable)
output <- mastif( formulaFec, formulaRep, inputs = inputs,  ng = 1000, 
                  burnin = 500 )

Following this short sequence, I fit a longer one and predict one of the plots:

predList <- list( mapMeters = 10, plots = 'DUKE_EW', years = 2010:2015 ) 
output <- mastif( inputs = output, ng = 3000, burnin = 1000, 
                  predList = predList )

Here are some plots followed by comments on display panels. Notice that when it progresses to the maps for plot DUKE_HW, they will include the predicted seed shadows:

mastPlot(output)

From fecundity chains, still more iterations are needed for convergence.

Both canopy and log(diam) contribute to fecundity. log(diam) contributes to maturation.

From seed shadow, seed rain beneath a 20-cm diameter tree averages near 0.2 seeds per m\(^2\).

Although a combination of maturation/fecundity can be found to predict the seed data in prediction (part a), the process model does not well predict the maturation/fecundity combination (part b). In other words, the maturation/fecundity/dispersal aspect of the model is effective, whereas the design does not yet include variables that help explain maturation/fecundity.

From the many maps in predicted fecundity, seed data, note that the DUKE_EW site includes prediction surfaces, as specfied in predList.

Here is the summary:

summary( output )

At this point in the the Gibbs sampler, the DIC and root mean square prediction error are:

output$fit

As mentioned above, prediction scores for seed-trap observations are based on the estimated fecundities \([\mathbf{y} | \phi, \rho]\) as scoreStates and on the estimated parameters \([\mathbf{y} | \phi, \rho][\phi, \rho| \boldsymbol{\beta}^x, \boldsymbol{\beta}^w, \dots]\) as predictScore.

Here is an example using betaPrior to specify a quadratic diameter effect. First, here is the priorTable:

d <- "https://github.com/jimclarkatduke/mast/blob/master/pinusExample.rData?raw=True"
repmis::source_data(d)

priorTable <- mastPriors("priorParameters.txt", specNames, 
                         code = 'code4', genus = 'pinus')

Here I specify formulas, prior bounds for diam (quadratic), and fit the model:

formulaRep <- as.formula( ~ diam )
formulaFec <- as.formula( ~ diam + I(diam^2) )   

betaPrior <- list(pos = 'diam', neg = 'I(diam^2)' )
                  
inputs <- list( treeData = treeData, seedData = seedData, xytree = xytree,
                xytrap = xytrap, specNames = specNames, seedNames = seedNames,
                betaPrior = betaPrior, priorTable = priorTable)

output <- mastif(inputs = inputs, formulaFec, formulaRep, 
                 ng = 1000, burnin = 500)

# restart
output <- mastif(output, formulaFec, formulaRep, 
                 ng = 3000, burnin = 1000)
mastPlot(output)

Many of the seed ID maxtrix elements are well identified (trace plots labeled “M matrix plot”, others are uncertain due to small numbers of the different species that could contribute to a seed type. These are presented as bar and whisker plots for elements of matrix \(\mathbf{M}\) (fraction of seed from a species that is counted as each seed type) in the plot labeled species -> seed type. Bar plots show the fraction of unknown seeds that derive from each species, labeled species -> seed type. The distribution of diameters in the data will limit the ability to estimate its effect on maturation and fecundity.

Note that the estimates for fecundity coefficients, \(\boldsymbol{\beta}^x\), are positive for diam and negative for diam^2. If there is no evidence in the data for a decrease \(\frac{\partial{\psi}}{\partial{diam}}\), then the quadratic term is near zero. In this case, the predicted fecundity in the plot labelled maturation, fecundity by diameter will increase exponentially.

maturation/fecundity data, if available

Most individuals produce no seed, due to resource limitation, especially light, in crowded stands. As trees increase in size and resource access they mature, and become capable of seed production. Maturation is hard to observe, so it must be modelled. In our data sets, maturation status is observed, but most values are NA; it is only assigned if certain. A value of 1 means that reproduction is observed. A value of 0 means that the entire crown is visible during the fruiting season, and it is clear that no reproduction is present. Needless to say, most observations from beneath closed canopies are NA. The observations are entered in the column repr in the data.frame treeData. Above I showed an example where minDiam is set as a prior that trees of smaller diameters are immature and maxDiam is the prior that trees above this diameter are mature.

mastif further admits estimates of cone production. The column treeData$coneCount refers to the cones that will open and contribute to trap counts in the year treeData$year. There is another column treeData$coneFraction, which refers to the fraction of the tree canopy that is represented by the cone count. When cone counts are used, a specNames \(\times 1\) matrix must be supplied with seeds per cone. The rownames of this matrix are specNames.

count count example needed?

when trees are sampled less frequently than seeds

Seed studies often include seed collections in years when trees are not censused. For example, tree data might be collected every 2 to 5 years, whereas seed data are available as annual counts. In this case, there there are years in seedData$year that are missing from treeData$year. mastif needs a tree year for each trap year. If tree data are missing from some trap years, I need to constuct a complete treeData data.frame that includes these missing years. This amended version of treeData must be accessible to the user, to allow for the addition of covariates such as weather variables, soil type, and so forth.

The function mastFillCensus allows the user to access the filled-in version of treeData that will be fitted by mastif. mastFillCensus accepts the same list of inputs that is passed to the function mastif. The missing years are inserted for each tree with interpolated diameters. The list inputs is returned with objects updated to include the missing census years and modified slightly for analysis by mastif. inputs$treeData can now be annotated with the covariates that will be included in the model. Here is the example from the help file. First I read a file that has complete years, so I randomly remove years:

d <- "https://github.com/jimclarkatduke/mast/blob/master/pinusExample.rData?raw=True"
repmis::source_data(d)

# randomly remove years for this example:
years <- sort(unique(treeData$year))
sy    <- sample(years,5)
treeData <- treeData[treeData$year %in% sy,]
treeData[1:10,]

Note the missing years, as is typical of mast data sets. Here is the file after filling missing years:

inputs   <- list( specNames = specNames, seedNames = seedNames, 
                  treeData = treeData, seedData = seedData, 
                  xytree = xytree, xytrap = xytrap, priorTable = priorTable)
inputs <- mastFillCensus(inputs, beforeFirst=10, afterLast=10)
inputs$treeData[1:10,]

The missing plotYr combinations in treeData have been filled to match thos in seedData. Now columns can be added to inputs$treeData, as needed in formulaFec or formulaRep.

In cases where tree censuses start after seed trapping begins or tree censuses end before seed trapping ends, it may be reasonable to assume that the same trees are producing seed in those years pre-trapping or post-trapping years. In these cases, mastFillCensus can accommodate these early and/or late seed trap data by extrapolating treeData beforeFirst years before seed trapping begins or afterLast years after seed tapping ends.

adding covariates

Continuing with the \(Pinus\) example, here I want to add covariates for missing years as needed for an AR(\(p\)) model, in this example \(p\) = 4. First, consider the tree-years included in this sample, before and after in-filling:

# original data
table(treeData$year)

# filled census
table(inputs$treeData$year)
table(seedData$year)

Note that the infilled version of tree data in the list inputs has the missing years, corresponding to those included in seedData. Here I set up the year effects model and infill to allow for p = 3,

p <- 3
inputs   <- mastFillCensus(inputs, p = p)
treeData <- inputs$treeData

Here are regions for random effects in the AR(3) model:

region = list(sApps = c("CWT_118", "CWT_218"),
              piedmont = c("DUKE_BW", "DUKE_EW", "MARSF", "MARSP"))
treeData$region <- 'sApps'
for(j in 1:length(region)){
  wj <- which( as.character(treeData$plot) %in% region[[j]])
  treeData$region[wj] <- names(region)[j]
}
inputs$treeData <- treeData
yearEffect <- list(groups = c('species','region'), p = p)

I add the climatic deficit (monthly precipitation minus PET) as a covariant. First, here is a data file,

d <- "https://github.com/jimclarkatduke/mast/blob/master/def.csv?raw=True"
download.file(d, destfile="def.csv")

The format for the file dev.csv is plot by year_month. I have saved it in my local directory. The function mastClimate returns a list holding three vectors, each having length equal to nrow(treeData). Here is an example for the cumulative moisture deficit for the previous summer. I provide the file name, the vector of plot names, the vector of previous years, and the months of the year. To get the cumulative deficit, I use FUN = 'sum':

treeData <- inputs$treeData
deficit  <- mastClimate( file = 'def.csv', plots = treeData$plot, 
                         years = treeData$year - 1, months = 6:8, 
                         FUN = 'sum', vname='def')
treeData <- cbind(treeData, deficit)
summary(deficit)

The first column in deficit is the variable itself for each tree-year. The second column holds the site mean value for the variable. The third column is the difference between the first two columns All three can be useful covariates, each capturing different effects (Clark et al. 2014). I could append any or all of them as columns to treeData.

Here is an example using minimum temperature of the preceeding winter. I obtain the minimum with two calls to mastClimate, first for Dec of the previous year (years = treeData$year - 1, months = 12), then from Jan, Feb, Mar for the current year (years = treeData$year, months = 1:3). I then take the minimum of the two values:

# include min winter temperature
d <- "https://github.com/jimclarkatduke/mast/blob/master/tmin.csv?raw=True"
download.file(d, destfile="tmin.csv")

# minimum winter temperature December through March of previous winter

t1 <- mastClimate( file = 'tmin.csv', plots = treeData$plot, 
                   years = treeData$year - 1, months = 12, FUN = 'min',
                   vname = 'tmin')
t2 <- mastClimate( file = 'tmin.csv', plots = treeData$plot, 
                   years = treeData$year, months = 1:3, FUN = 'min',
                   vname = 'tmin')
tmin <- apply( cbind(t1[,1], t2[,1]), 1, min)
treeData$tminDecJanFebMar <- tmin
inputs$treeData <- treeData

Here is a model using several of these variables:

formulaRep <- as.formula( ~ diam )
formulaFec <- as.formula( ~ diam + defJunJulAugAnom + tminDecJanFebMar )   
output <- mastif(inputs = inputs, formulaFec, formulaRep, 
                 yearEffect = yearEffect,
                 ng = 2500, burnin = 1000)

Data for future years can be a scenario, based assumptions of status quo in the mean and variance, an assumed rate of climate change, and so on.


flexibility

Seed production is volatile, with order of magnitude variation from year-to-year. There is synchronicity among individuals of the same species, the “masting” phenomenon. There are large difference between individuals, that are not explained by environmental variables. In this section I discuss extensions to random effects, year effects, lag effects, and fitting multiple species having seed types that are not always identifiable to species.

random individual effects

Random individual effects can include random intercepts and slopes for the fecundities of each individual tree that is imputed to be in the mature state for at least 3 years. [There are no random effects on maturation, because they would be hard to identify from seed trap data for this binary response.]

I include in the inputs list the list randomEffect, which includes the column name for the random group. Typically this would be a unique identifier for a tree within a plot, e.g., randomEffect$randGroups = 'tree':

randomEffect <- list(randGroups = 'tree', 
                     formulaRan = as.formula( ~ diam ) )

However, randGroups could be the plot name. This column is interpreted as a factor, each level being a group. Random effects will not be fitted on individuals that are in the mature state less than 3 years.

The formulaRan is the random effects model. It includes a subset of variables specified in fecundity model formulaFec. [Because random effects are centered on zero, terms in formulaRan must already be included in the fecundity design formulaFec (see appendix).] Here is a brief description of what’s happening:

There is a design vector \(\mathbf{w}_{ij,t}\) that can include all or some of the columns in \(\mathbf{x}_{ij,t}\). There is an individual-effects coefficient matrix \(\boldsymbol{\beta}^w_{ij}\), which enters the mean structure this way:

\[\log \psi_{ij,t} \sim N \left( \mathbf{x}'_{ij,t} \boldsymbol{\beta}^x + \mathbf{w}'_{ij,t} \boldsymbol{\beta}^w_{ij}, \sigma^2 \right)\]

Here is a fit with random effects.

formulaFec <- as.formula( ~ diam)    # fecundity model
formulaRep <- as.formula( ~ diam)    # maturation model 
randomEffect <- list(randGroups = 'tree', 
                     formulaRan = as.formula( ~ diam) )
output <- mastif( formulaFec, formulaRep, inputs = inputs, 
                  ng = 2000, burnin = 1000,
                  randomEffect = randomEffect )

Here is a restart:

output <- mastif( inputs = output, ng = 2000, burnin = 1000)
mastPlot(output)

There are several things to note:

output$fit

year and lag effects

The model admits year effects and lag effects, the latter as an AR(\(p\)) model. Year effects assign a coefficient to each year \(t = [1, \dots, T_j]\). Lag effects assign a coefficient to each of \(j \in \{1, \dots, p\}\) plags, where the maximum lag \(p\) should be substantially less than the number of years in the study. Year effects can be organized in random groups. Specification of random groups is done in the same way for year effects and for lag effects.

random groups for years and lags

I define random groups for year and lag effects by species, by plots, or both. When there are multiple species that contribute to the modeled seed types, I expect the year effects to depend on which species is actually producing the seed. When there are multiple plots sufficiently distant from one another, I might allow for the fact that year effects or lag effects differ by group; yearEffect$groups allows that they need not mast in the same years. In the example below, I’ll use the term region for plots in the same plotGroup.

Here is a breakdown for this data set by region:

with(treeData, colSums( table(plot, region)) )

Here are year effects structured by random groups of plots, given by the column region in treeData:

yearEffect <- list(groups = 'region')

This option will fit a year effect for both provinces and years having sufficient individuals estimated to be in the mature state. Here is the model with random year effects,

\[\log \psi_{ij,t} \sim N \left( \mathbf{x}'_{ij,t} \mathbf{\beta}^x + \gamma_t + \gamma_{g[i],t}, \sigma^2 \right)\] where the year effect \(\gamma_{g[i],t}\) is shared by trees in all plots defined by yearEffect$province, \(ij \in g\). Year effects are sampled directly from conditional posteriors (see Appendix).

inputs$treeData  <- treeData 
output <- mastif(formulaFec, formulaRep, inputs = inputs, 
                 ng = 1500, burnin = 500, 
                 randomEffect = randomEffect, yearEffect = yearEffect)

Here is a restart, with predictions for one of the plots:

predList <- list( mapMeters = 10, plots = 'DUKE_BW', years = 1998:2014 ) 
output <- mastif(inputs = output, predList = predList, ng = 3000, burnin = 1000)
mastPlot(output)

Note that still more iterations are needed for convergence. Here are some comments on mastPlot:


random groups and individuals when there are multiple species

Most data sets have multiple seed types that complicate estimation of mast production by each species. This example considers Pinus spp, seeds of which cannot be confidently assigned to species. Here I load the data and generate a sample of maps from several years, including all species and seed types.

d <- "https://github.com/jimclarkatduke/mast/blob/master/pinusExample.rData?raw=True"
repmis::source_data(d)

mapList <- list( treeData = treeData, seedData = seedData, 
                 specNames = specNames, seedNames = seedNames, 
                 xytree = xytree, xytrap = xytrap, mapPlot = 'DUKE_EW',
                 mapYears = c(2007:2010), treeSymbol = treeData$diam,
                 treeScale = .6, trapScale=1.4,
                 plotScale = 1.2, LEGEND=T)
mastMap(mapList)

Note the tendency for high seed accumulation (large green squares) near dense, large trees (large brown circles).

Here is an AR(p) model by species with individual tree random effects.


fitting AR(p) and random effects with multiple species

In this example, I again model seed production as a function of log diameter, diam, and tree growth rate, canopy, now for multiple species and seed types. This is an AR(p) model, because I include the number of lag terms yearEffect$p = 5,

\[\log \psi_{ij,t} \sim N \left( \mathbf{x}'_{ij,t} \mathbf{\beta}^x + \sum^p_{l=1} (\alpha_l + \alpha_{g[i],l}) \psi_{ij,t-l}, \sigma^2 \right)\] Only years \(p < t \le T_i\) are used for fitting. Samples are drawn directly from the conditional posterior distribution.

In the table printed at the outset are trees by plot and year, i.e., the groups assigned in inputs$yearEffect. The zeros indicate either absence of trees or that no plots were sampled in those years. (These are not the same thing, mastif knows the difference).

In the code below I specify formulas, AR(\(p\)), and random effects, and some prior values. Due to the large number of trees, convergence is slow. Because I do not assume that trees of different species necessarily mast in the same years, I allow them to differ through random groups on the AR(\(p\)) terms.

formulaFec <- as.formula( ~ diam )   # fecundity model
formulaRep <- as.formula( ~ diam )            # maturation model

yearEffect   <- list(groups = 'species', p = 4)   # AR(4)
randomEffect <- list(randGroups = 'tree', 
                     formulaRan = as.formula( ~ diam ) )

seedMass <- matrix( c(0.0170,0.0270,0.0167,0.0070,0.0080), ncol=1)
rownames(seedMass) <- c('pinuEchi','pinuRigi','pinuStro','pinuTaed','pinuVirg')
colnames(seedMass) <- 'gmPerSeed'

inputs   <- list( specNames = specNames, seedNames = seedNames, 
                  treeData = treeData, seedData = seedData, 
                  xytree = xytree, xytrap = xytrap, priorDist = 20,
                  priorVDist = 5, minDist = 15, maxDist = 30, 
                  minDiam = 12, maxDiam = 40,
                  maxF = 1e+6, seedMass = seedMass)
output <- mastif(formulaFec, formulaRep, inputs = inputs, ng = 300, 
               burnin = 100, yearEffect = yearEffect, 
               randomEffect = randomEffect)

Here is a restart:

 output <- mastif(inputs = output, ng = 2000, 
               burnin = 1000)
plotPars <- list(MAPS = F)
mastPlot(output, plotPars = plotPars)

Again, convergence will require more iterations. The AR(p) coefficients in the lag effect group panel shows the coefficients by random group. They are also shown in a separate panel, with each group plotted separately. In the ACF eigenvalues panel are shown the eigenvalues for AR lag coefficients on the real (horizontal) and imaginary (vertical) scales with the unit circle, within which oscillations are damped. The imaginary axis describes oscillations.

Here’s a restart with predictions:

plots <- c('DUKE_EW','CWT_118')
years <- 1980:2025
predList <- list( mapMeters = 10, plots = plots, years = years ) 
output <- mastif(inputs = output, predList = predList, ng = 3000, 
                 burnin = 1000)

and updated plots:

mastPlot( output, plotPars = list(MAPS=F) )

Note that convergence requires additional iterations (larger ng). The predictions of seed production will progressively improve with convergence.

mapList <- output
mapList$mapPlot <- 'DUKE_EW'
mapList$mapYears <- c(2011:2012)
mapList$PREDICT <- T
mapList$treeScale <- 1.2
mapList$trapScale <- .8
mapList$LEGEND <- T
mapList$scaleValue <- 50
mapList$plotScale <- 2
mapList$COLORSCALE <- T
mapList$mfrow <- c(2,1)

mastMap( mapList )

Or a larger view of a single map:

mapList$mapPlot <- 'CWT_118'
mapList$mapYears <- 2015
mapList$PREDICT <- T
mapList$treeScale <- 1.5
mapList$trapScale <- .8
mapList$LEGEND <- T
mapList$scaleValue <- 50
mapList$plotScale <- 2
mapList$COLORSCALE <- T
mapList$mfrow <- c(1,1)
mastMap( mapList )

Here is a summary of parameter estimates:

summary( output )

The names of species and seed types can be modified by mast to remove characters that can be confused or are reserved for specific purposes. mast also corrects for some inconsistencies, usually with a warning. For example, see the section on sample years.


fitting year effects

To substitute year effects for AR(\(p\)) effects, simply remove the argument specification of p in the yearEffect list:

yearEffect   <- list(groups = c('species','region'))   # year effects


no predictors

In cases where there are no predictors that explain variation among individuals the formula can be limited to an intercept, using ~ 1. In this case, it can be valuable to estimate fecundity even when there are no good predictors. For this intercept-only model, random effects and/or year effects can provide variation:

d <- "https://github.com/jimclarkatduke/mast/blob/master/liriodendronExample.rData?raw=True"
repmis::source_data(d)

#group plots in regions for year effects
region <- rep('sApps',nrow(treeData))
region[ as.character(treeData$plot) == 'DUKE_EW' ] <- 'piedmont'

treeData$region <- region

formulaFec   <- as.formula(~ diam)
formulaRep   <- as.formula( ~ diam ) 
yearEffect   <- list(groups = 'region')
randomEffect <- list(randGroups = 'treeID', 
                     formulaRan = as.formula( ~ 1 ) )
inputs <- list(specNames = specNames, seedNames = seedNames, 
               treeData = treeData, seedData = seedData, 
               xytree = xytree, xytrap = xytrap,
               priorDist = 10, priorVDist = 5, maxDist = 50, minDist = 5,
               minDiam = 25, maxF = 1e+6)
output <- mastif(inputs, formulaFec, formulaRep, ng = 2000, burnin = 1000,
               randomEffect = randomEffect, yearEffect = yearEffect )
mastPlot(output)

Without predictors, the fitted variation is coming from random effects and year effects.


convergence

R code is highly vectorized. Unavoidable loops are written in C++ and exploit the C++ library Armadillo, available through RcppArmadillo. Alternating with Metropolis are Hamiltonian MC steps to encourage large movements.

Despite extensive vectorization and C++ for cases where loops are unavoidable, convergence can be slow. A collection of plots inventoried over dozens of years can generate in excess of \(10^6\) tree-year observations and \(10^4\) trap year observations. For reasons discussed in the appendix, there is no escaping the requirement of large numbers of indirectly sampled latent variables. If random effects are included, there are (obviously) as many random groups as there are trees. All tree fecundities must be imputed.


trouble shooting

Because seed-trap studies involve multiple data sets (seed traps, trees, covariates) that are collected over a number of years and multiple sites, combining them can expose inconsistencies that are not immediately evident. Of course, a proper analysis depends on alignment of trees, seed traps, and covariates with unique tree names (treeData$tree) and trap names (seedData$trap) in each plot (treeData$plot, seedData$plot).

Notes are displayed by mastif at execution summarizing aspects of the data that might trigger warnings. All of these issues have arisen in data sets I have encountered from colleagues:

Alignment of data frames. The unique trees in each plot supplied in treeData must also appear with x and y in xytree. The unique traps in each plot supplied in seedData must also appear with x and y in xytrap. Problems generate a note:

Note: treeData includes trees not present in xytree

Spatial coordinates. Because tree censuses and seed traps are often done at different times, by different people, the grids often disagree. Spatial range tables are displayed for (x, y) coordinates in xytree and xytrap.

Unidentified seeds. Seeds that cannot be identified to species contain the character string UNKN. If there are species in specNames that do not appear in seedNames, then the UNKN seed type must be included in seedNames and in columns of seedData. For example, if caryGlab, caryTome appear in specNames, and caryGlab, caryUNKN appear in seedNames (and as columns in seedData), then caryUNKN will be the imputed fate for all seeds emanating from caryOvat and some seeds from caryGlab.

This note will be displayed:

Note: unknown seed type is caryUNKN

If there are seedNames that do not appear in specNames, this note is given:

Note: seedNames not in specNames and not "UNKN": caryCord

Moved caryCord to "UNKN" class

Design issues. The design can seem confusing, because there are multiple species on multiple plots in multiple years. There is a design matrix that can be found here for fecundity:

output$inputs$setupData$xfec

and here for maturation:

output$inputs$setupData$xrep

(Variables are standardized, because fitting is done that way, but coefficients are reported on their original scales. Unstandardized versions of design matrices are xfecU and xrepU.)

There should not be missing values in the columns of treeData that will be used as predictors (covariates or factors). If there are missing values, a note will be generated:

Fix missing values in these variables: [1] "yearlyPETO.tmin1, yearlyPETO.tmin2, flowering.covs.pr.data, flowering.covs.tmin.data, s.PETO"

There can be missing seed counts in seedData–missing values will be imputed.

A table containing the Variance Inflation Factor (VIF), range of each variable, and correlation matrix will be generated at execution. VIF values > 10 and high correlations between covariates are taken as evidence of redundancy. A table will be generated for each species separately. However, in xfec and xrep they are treated as a single matrix.

Year effects by random group require replication within groups. Here is a note for the AR model showing sizes of groups defined by species and region (NE, piedmont, sApps):

no. trees with > plag years, by group:

  `caryGlab-NE caryGlab-piedmont    caryGlab-sApps `
  `          1               497               361 `
  
  `caryOvat-piedmont caryTome-piedmont    caryTome-sApps `
  `        122               569                77 `

small group: caryGlab-NE

There is only one tree in the caryGlab-NE group, suggesting insufficient replication and a different aggregation scheme.

For the AR(\(p\)) model, values are imputed for \(p\) years before and after a tree is observed, and only trees observed for > \(p\) years will contribute to parameter estimates. If the study lasts 3 years, then the model should not specify yearEffect$p = 5. A note will be generated to inform on the number of observations included in parameter estimates:

Number of full observations with AR model is: [1] 21235

Prediction. If predList is supplied, then fecundity and seed density will be predicted for specified plot-years. The size of the prediction grid is displayed as a table of prediction nodes by plot and year. Large prediction grids slow execution. To reduce the size of the grid, increase the inputs$predList$mapMeters (the default is 5 m by 5 m).

When covariates are added as columns to treeData, they must align with treeData$plot, treeData$year, and, if they are tree-level covariates, with treeData$tree. The required column treeData$diam is an example of the latter.

appendix

model overview

Consider an inventory plot where trees produce seeds, depending on size, resource availability, competition, climate, and so on. The capacity to produce seed can increase with tree size. Individuals that have developed this capacity have reached maturation. The amount of seed produced in a year by a mature individual is termed fecundity. Seed traps accumulate seeds, depending on the combined fecundities of nearby trees, their dispersal capacities, and their distances from traps. Typically, data collection will involve multiple plots and multiple years. In order to estimate fecundity, I need to know the tree and seed trap locations, and I need to estimate dispersal together with seed production of each tree. For trees of unknown maturation status, that too must be estimated. Predictors that help to explain maturation and fecundity will improve estimates.

observed quantities

Let \(h = 1, \dots, H\) designate species, and \(r = 1, \dots, R\) designate the seed types that could have been produced by any of the \(H\) species. Again, this distinction is needed, because not all seeds collected in traps can be identified to species. For example, if the plot includes trees of Carya glabra and C. tomentosa, then seed types might include C. glabra, C. tomentosa, and Carya spp, the latter including all seeds that could not be identified to species.

Here are the main elements of the model:

dynamic process

The process model is a state-space model for log fecundity and maturation, with joint distribution \([\psi_{ij,t}, \rho_{ij,t}]\). Log fecundity is continuous, \(\psi_{ij,t} \in (-\infty, \infty)\). True maturation status is the indicator \(\rho_{ij,t} \in \{0, 1\}\), subject to the constraint that maturation is a one-way process, \([\rho_{ij,t+1} = 1|\rho_{ij,t} = 1] = 1\), and \([\rho_{ij,t} = 1|\rho_{ij,t+1} = 0] = 0\). Fecundity is

\[f_{ij,t} = \left \{ \begin{matrix} \ \psi_{ij,t} & \rho_{ij,t} = 1\\ \ 0, & \rho_{ij,t} = 0 \end{matrix} \right.\] The latent states \(\psi\) and \(\rho\) are imputed and linked to observations with error.

There are predictors in \(\mathbf{v}_{ij,t}\) that can account for maturation, including resource access and crowding. Predictors in \(\mathbf{x}_{ij,t}\) explain diffences in seed production. Generatively, the model can start with a maturation status, modeled as a probit. Let \(\rho_{ij,t} = 1\) be the event that individual \(ij\) is mature at year \(t\), with probability,

\[[\rho_{ij,t} = 1] = \Phi \left( \mathbf{v}'_{ij,t} \mathbf{\beta}^v \right) \] where \(\Phi()\) is the standard normal distribution function. Mature individuals produce seeds at log fecundity

\[\log \psi_{ij,t} \sim N \left( \mathbf{x}'_{ij,t} \mathbf{\beta}^x, \sigma^2 \right)\]

The fecundity model is, in fact, flexible (Table 1).

Table 1. Process models for log fecundity.

terms form
fixed effects \(\mathbf{x}'_{ij,t} \mathbf{\beta}^x\)
fixed year by random group \(\gamma_{t} + \gamma_{g,t}\)
AR(\(p\)) fixed lag by random group \(\sum^p_{l=1} (\alpha_l + \alpha_{g,l}) \log \psi_{gij,t-l}\)
random individual effect \(\mathbf{w}'_{ij,t} \mathbf{\beta}^w_{ij}\)

The first line in Table 1 holds fixed effects, which apply to all models. At minimum, the fixed effect is an intercept, but there will typically be covariates and/or factors.

Year effects \(\gamma_t\) assign a coefficient to each year. Year effects are not recommended if there are year variables in the design matrix; year effects will compete with year variables in \(\mathbf{x}_{ij,t}\). When multiple random groups are specified, there is a random year effect, \(\gamma_{g,t}\) for group \(g\).

AR(\(p\)) is the autoregressive model with \(l = 1, \dots, p\) lag terms, with coefficients \(\alpha_l\). If there are multiple groups, then these are random by group (subscript \(g\)). Groups can be defined by plots within a region and/or species. Year effects cannot be used with AR effects.

Random individual effects \(\beta_{ij}^w\) can be used in combination with year effects or lag (AR) effects. Random individual require that individuals are observed repeatedly. They are a good idea when there are more than four years of data. They may not be a good idea if there are a number of individual variables in the design matrix (e.g., tree size, tree canopy area).

All models include process error, with variance \(\sigma^2\).

observation model

The observation model includes i) the uncertain maturation status of trees, ii) minimum and maximum visual seed counts, fecundity estimates on trees, iii) the uncertain assignment of seeds to species, and iv) Poisson sampling of seeds that have been redistributed to seed traps by dispersal.

maturation status observations

The maturation observation model recognizes uncertainty in the assignment of maturation (fruits are often unobservable in crowded canopies) and the fact that trees are not observed in many years. Let \(z_{ij,t}\) be the observed status, which can be mature (fruits observed, \(z_{ij,t} = 1\)), uncertain (fruits not observed, canopy obscure), and immature (\(z_{ij,t} = 0\)). \(t_{ij,l}\) is the last year in which individual \(ij\) was observed to be immature. \(t_{ij,m}\) is the first year \(ij\) was observed in the mature state. Status is known to be mature any time after the tree is first observed to be mature and to be immature any time before the last time it is observed to have been immature. Between these times, the status is unknown and modeled with a probit:

\[ \begin{matrix} \ z_{ij,t_m} = 0 \rightarrow & \rho_{ij,t} = 0, \forall t \leq t_m\\ \ z_{ij,t_l} = 1 \rightarrow & \rho_{ij,t} = 1, \forall t \geq t_l\\ \ t_l < t < t_m \rightarrow & [\rho_{ij,t} = 1] =\Phi \left( \mathbf{v}'_{ijt} \mathbf{\beta}^v \right) \end{matrix} \] where \(t_l\) is an observation year earlier than \(t\), and \(t_m\) is an observation year after year \(t\)

fecundity estimates

If there are visual estimates of seeds on a tree, these bound estimates of fecundity. For example, if cones are counted on trees, these counts can be multiplied by the expected range of seeds per cone to bound fecundity estimates.

seed counts and dispersal

Many of the seeds counted in traps could have been produced by more than one species. Uncertainty in species assigment for seeds must be estimated in the form of a length-\(R\) composition vector \(\mathbf{r}_h\), where \(R\) is the number of seed types to which species \(h\) contributes seed. There is one vector for each species. Element \(r\) is the fraction of seeds produced by trees of species \(h\) that are expected as seed type \(r\). Elements of \(\mathbf{r}_h\) sum to 1. The vector corresponding to individual \(i\) is designated with the notation \(\mathbf{r}_{i[h]}\). The expected fecundity is

\[\mathbf{f}_{ij,t} = \mathbf{r}_{h[i]}\rho_{ij,t} \psi_{ij,t} \] where \(\mathbf{f}_{ij,t}\) is the length-\(R\) vector of seed production for each seed type \(r\). Thus, the composition vector \(\mathbf{r}_{h[i]}\) redistributes the seed produced by individual \(ij,t\) among \(R\) seed types.

Seed dispersal induces dependence in seed trap counts by a \(S_{j,t} \times n_{j,t}\) redistribution kernel \(\mathbf{S}_{s,t}\). Expected seeds per m\(^2\) at seed traps \(s = 1, \dots, S_{j,t}\) is

\[\mathbf{\Lambda}_{j,t} = \mathbf{S}_{j,t} \mathbf{F}_{j,t}\] \(\mathbf{\Lambda}_{j,t}\) is the \(S_{j,t} \times R\) matrix of expected seed counts, and \(\mathbf{F}_{j,t}\) is the \(n_{j,t} \times R\) matrix of seed production from all trees on the plot.

The data model has likelihood

\[y_{rsj,t} \sim Poi(A_{sj,t} \lambda_{rsj,t})\]

where \(y_{rsj,t}\) is the seed count for type \(r\) in trap \(s\) on plot \(j\) in year \(t\), \(A_{sj,t}\) is the area (e.g., meters) of the seed trap times the fraction of collecting time that trap \(sj,t\) was active–seed traps are subject to damage.

The redistribution kernel \(\mathbf{S}_{j,t}\) has elements

\[\mathbf{S}_{j,t[s,i]} = \frac{u_{h[i]}}{\pi \left(u_{h[i]} + d^2_{s,i} \right)^2}\] for distance \(d_{s,i}\) and fitted dispersal parameter \(u_h\) species (or random group) \(h\) corresponding to tree \(i\). This is a two-dimensional Student’s \(t\) distribution (Clark et al. 1999) and a special case of the Matern distribution, often used in spatial analysis models (e.g., Banerjee et al. CC). The mean dispersal distance is

\[\bar{d}_h = \frac{\pi \sqrt{u_h}}{2}\] (Clark et al. 1999).

why is fecundity a latent variable?

A latent-variable treatment of fecundity is needed to allow for the way in which fecundity is related to seed counts, through a dispersal kernel. The expected seed intensity is the product of dispersal and fecundity,

\[\mathbf{\Lambda} = \mathbf{S} \mathbf{F}\]

Clearly, I can write out a solution for the \(n \times R\) matrix \(\mathbf{F}\). However, there is a unique solution only if there are more seed traps in the \(S \times R\) matrix \(\mathbf{\Lambda}\) than there are sources in \(\mathbf{F}\), i.e., \(S > n\). A bigger issue is the fact that this solution would undoubtedly include negative source values. Prior knowledge tells me that this is impossible. I thus require a model for \(\mathbf{F}\), which could be as simple as a prior distribution, which might impose nothing more than non-negative values. Or it can bring in predictors, in which case \(\mathbf{F}\) could be replaced by a function of predictors. This is the approach taken by many seed-shadow models. Conditional fecundity \(\psi\) and maturation status \(\rho\) were introduced as latent variables hierarchically at least as early as Clark et al. (2004) to allow for knowledge of suitable range and maturation schedules and the large heterogeneity among trees. A hierarchical specification can be quite general, while still providing stability over suitable parameter ranges.

where is the information?

Uncertainty in maturation and fecundity estimates depends on both the seed data (through the dispersal kernel) and the fecundity model. Trees that are distant from all seed traps are uninfluenced by seed data and derive all information from the maturation/fecundity model, relying on individual-scale covariates, such as tree diameter. Model fitting thus ultimately depends on the trees are are sufficiently close to affect seed counts. These are the trees that contribute to the coefficients in the process model, which, in turn, control fecundity estimates for distant trees. For this reason, uncertainty in maturation/fecundity estimates tends to be highest for trees having no nearby seed traps (Clark et al. 2004, 2010).

algorithm notes

initialization

Model fitting begins with an Expectation-Maximumization (EM) step to initialize fecundities and the dispersal parameter \(u\), ignoring the fecundity model. The idea here is to find suitable initial states that predict the seed data, without regard to the regression model. This EM step focusses on trees that are close to seed traps, because these are the ones that contribute most to the data. Once this initialization of nearby trees is completed, remaining trees are initialized by predicting from the initial fecundity model fitted to these nearby trees.

From extensive simulation experiments it’s clear that convergence can occur from a range of initial states, but proper initialization accelerates it (Clark et al. 2004). Again, trees distant from seed traps are essentially ‘predicted’ from the fecundity model, rather than contributing much to the estimates themselves. The algorithm can be stabilized with prior distributions that bound fecundity and dispersal estimates to reasonable, albeit flexible, ranges (see below). The caveat here is that nothing useful will come from data sets where all plots are limited to few, distant trees. Conversely, crowded monocultures cannot inform about dispersal distance or individual fecundities. They can inform about averages.

In these notes I summarize algorithms used for posterior simulation and prediction.

latent states

mastif is a state space model, with observed seeds \(y_{sj,t}\) and maturation status \(z_{ij,t}\). Observed states are linked by data models to latent fecundity \(\psi_{ij,t}\) and true maturation status \(\rho_{ij,t}\). To enhance mixing, the latent states are sampled with a combination of alternating methods. The first method samples maturation and fecundity jointly as a Metropolis random walk, with details depending on the process model. The second method conditions on maturation state and samples from the Hamiltonian.

joint fecundity and maturation updates

To impute states, fecundity and maturation status are proposed jointly as

\[[\psi^*_{ij,t}, \rho^*_{ij,t}] = [\psi^*_{ij,t} | \rho^*_{ij,t}][\rho^*_{ij,t}]\] Maturation year is a random walk, centered on the currently imputed maturation year and subject to constraints imposed by observed or currently imputed states (mature individuals cannot become immature). Possible maturation years range from the last year in which an individual was observed in the immature state to the first year in which that individual was observed in the mature state. If there are no maturation observations, then all information on maturation status comes through the seed data. Maturation is proposed and accepted jointly with fecundity for all trees on a plot year. This blocking is necessitated by the fact that the likelihood for each trap-year conditionally depends on all tree-years for that plot (Clark et al. 2004).

Given proposed \(\rho^*_{ij,t}\), fecundity \(\psi^*_{ij,t}|\rho^*_{ij,t}\) is proposed from a normal distribution, with censoring imposed by the proposed \(\rho^*_{ij,t}\),

\[\psi^*_{ij,t} | \rho^*_{ij,t} \sim N \left( \psi_{ij,t}, s_{ij,t} \right) I(r_1 < \psi^*_{ij,t} \leq r_2)\] where lower and upper bounds are \((r_1, r_2) = (0, \infty)\) for \(\rho^*_{ijt} = 1\) and \((-\infty, 0]\) for \(\rho^*_{ijt} = 0\).

Again, proposal acceptance is done in a plot-year block, summarized this way:

\[[\boldsymbol{\psi}_{j,t}, \boldsymbol{\rho_{j,t}} | \mathbf{y}_{j,t}, \mathbf{z}_{j,t}] \propto P_1 \times P_2 \times P_3\] where

\[\begin{align} P_1 &= \prod_{s=1}^{S_j} \prod_{r=1}^R Poi \left(y_{rkj,t} | A_{sj} \lambda_{rsj,t}(\boldsymbol{\psi}_{j,t}, \boldsymbol{\rho}_{j,t}, \mathbf{r})\right) \\ P_2 &= \prod_{i=1}^{n_i} [z_{ij,t} | \rho_{ij,t}] \\ P_3 &= \prod_{i=1}^{n_i} [\psi_{ij,t}| \rho_{ij,t}] \end{align}\]

(Conditioning on parameters is omitted to improve clarity.)

Sampling is done in either of two ways, depending on the process model, both beginning with proposed \(\{\rho^*_{ij,t} | \rho_{j,t-1}, \rho_{ij,t+1} \}_{i=1}^{n_j}\) for all trees in plot-year \((j, t)\) from a random walk. Then:

Method 1:

  1. propose all \(\psi^*_{ij,t} | \rho^*_{ij,t}, \psi_{ij,t}\) for the plot-year from a normal distribution centered on the currently imputed \(\psi_{ij,t}\) with an adaptive variance parameter.
  2. accept/reject \(\boldsymbol{\rho}^*_{j,t}, \boldsymbol{\psi}^*_{j,t}\) as a block with probability \(P_1 P_2 P_3\).

Method 2:

  1. propose all \(\psi^*_{ij,t} | \rho^*_{ij,t}\) for the plot-year from the conditional normal distribution obtained from \(P_3\). Details are limited to most complex case of the AR(p) model (see below).
  2. accept/reject \(\boldsymbol{\rho}^*_{j,t}, \boldsymbol{\psi}^*_{j,t}\) as a block with probability \(P_1 \times P_2\).

Hamiltonian updates

Hamiltonian updates cannot be used with discrete maturation status. However, by conditioning on maturation state, mixing of fecundity is accelerated with Hamiltonian updates for currently imputed mature individuals.

Each observation is a length-\(R\) vector \(\mathbf{y}_{sj,t}\) with Poisson intensity \(A_{sj} \lambda_{rsj,t} = A_{sj} \sum_{i=1}^{n_j,t} \mathbf{S}_{[sjt,i]} e^{\psi_{ij,t}} \mathbf{r}_{i[h],r}\), where \(\mathbf{r}_{i[h],r}\) is the row vector \(\mathbf{r}\) corresponding to the species of individual \(ij,t\) and the column for seed type \(r\). The \(S_{j,t} \times n_{j,t}\) kernel matrix \(\mathbf{S}_{j,t}\) has elements \(\mathbf{S}_{[sjt,i]}\) for row \(s\) and column \(i\). The Hamiltonian can be written as

\[ H(\psi_{ij,t}, p) = B(\psi_{ij,t}) + C(p) \]

where \(\psi_{ij,t} = \{\psi_{1,j,t}, \dots, \psi_{n_j,j,t} \}\) is log fecundity for an individual on plot-year \((j,t)\), and \(C(p) = \sum_{i=1}^{n_j} \frac{p_i^2}{2m_i}\) is the kinetic energy, taken as a quadratic function of momentum variables \(m_i\), which are tuned to optimize performance (Neal 2011). The first term incorporates the conditional distribution,

\[ \begin{align} B(\psi_{ij,t}) &= -\log \left(\pi(\psi_{ij,t} | \mathbf{y}_{j,t}, \mu_{ij,t},\sigma^2) \right) \\ & \propto \sum_{r,s} \left( - y_{rsj,t} \log \lambda_{rsj,t} + A_{sj} \lambda_{rsj,t} \right) + \frac{1}{2\sigma^2} (\psi_{ij,t} - \mu_{ij,t})^2 \end{align} \] The gradient is used to direct proposals efficiently:

\[\frac{\partial B}{\partial \psi_{ij,t}} = e^{\psi_{ij,t}}\sum_{r} \mathbf{r}_{i[h],r} \sum_s \mathbf{S}_{[sjt,i]} \left(- \frac{y_{rsj,t}}{\lambda_{rsj,t}} + A_{sj} \right) + \frac{1}{\sigma^2} \left( \psi_{ij,t} - \mu_{ij,t} \right)\]

Hamiltonian updates are individually slow, but affect larger steps then a Metrolis random walk, especially with large data sets. The two methods are mixed stochastically in the Gibbs sampler.

coefficients

Direct sampling of coefficients in \(\boldsymbol{\beta}^x\) and \(\boldsymbol{\beta}^v\) is available from Gaussian conditional posterior distributions. Gaussian prior distributions are non-informative. For \(\boldsymbol{\beta}^x\) conditional distributions marginalize random effects (see below). The variance \(\sigma^2\) has an inverse gamma prior – and is sampled directly from the conjugate inverse gamma posterior.

random individual effects

A few notes here on the random effects algorithm. Let \(\mathbf{w}_{ij,t}\) be a design vector holding all or some of the columns in \(\mathbf{x}_{ij,t}\). There is an individual-effects coefficient matrix \(\boldsymbol{\beta}^w_{ij}\),

\[\psi_{ij,t} \sim N \left( \mathbf{x}'_{ij,t} \boldsymbol{\beta}^x + \mathbf{w}'_{ij,t} \boldsymbol{\beta}^w_{ij}, \sigma^2 \right)\] The prior distribution includes

\[ \begin{align} \boldsymbol{\beta}^w_{ij}|\mathbf{B}_{w} &\sim MVN(\mathbf{0},\mathbf{B}_{w}) \\ \mathbf{B}_{w} &\sim IW \left( \tilde{\mathbf{B}}, df \right) \end{align} \]

where \(df = Q^w + 2\), \(Q^w\) is the number of columns in \(\mathbf{w}_{ij,t}\), and \(\tilde{\mathbf{A}} = \mathbf{I}_r\) is a prior diagonal matrix. The conditional posterior matrix is

\[\boldsymbol{\beta}^w_{ij}|\boldsymbol{\beta}^x, \mathbf{B}_w \sim MVN(\mathbf{V}_{ij}\mathbf{v}_{ij},\mathbf{V}_{ij})\] where

\[\begin{align} \mathbf{v}_{ij} &= \frac{1}{\sigma^2} \sum_{t \in \{t_i\}} \mathbf{w}_{ijt} (\psi_{ijt} - \mathbf{x}'_{ijt} \boldsymbol{\beta}^x ) \\ \mathbf{V}_{ij} &= \frac{1}{\sigma^2} \sum_{t \in \{t_i\}} \mathbf{w}_{ijt} \mathbf{w}'_{ijt} + \mathbf{B}_w^{-1} \end{align}\]

The summations are taken over all observation years for an individual \(i\), the set \(\{ t_i \}\) for which the individual is mature. Here is the conditional for the covariance,

\[\mathbf{B}_w|\{ \boldsymbol{\beta}^w_{ij}\} \sim IW \left( \sum \boldsymbol{\beta}^w_{ij} \boldsymbol{\beta}^{w'}_{ij} + df \times \tilde{\mathbf{B}}, \sum_j n_j + df \right)\]

random groups in the year and AR(\(p\)) models

The year and AR(\(p\)) models allow for group random effects on year and lag coefficients, respectively–if groups are defined by the user, they will be treated as random. This is done, because year and lag terms across groups are highly unbalanced. Plot-species groups can hold different numbers of plots and trees in different years. Plots can be established at different times, have different plot areas, and support very different communities of species. For a given species, abundance across plots may range from zero to high. Within plots, numbers of mature individuals vary across years with recruitment, maturation, and death. Within posterior simulation, their imputed maturation statuses change by tree and year. The sizes of design matrices are thus dynamic.

Given this imbalance, treating groups as random provides the advantage that no arbitrary rules are needed to catch computation errors that would result from plot-years that are at some iterations imputed to have mature trees and other iterations not.

AR(\(p\)) model

The AR(\(p\)) model allows for the dependence of the current states of \(\psi_t\) on \(p\) previous states. The process is homogeneous in time, because the lag coefficients \(\alpha_p\) are constant. I start with a few words on structure.

imputed past, predicted future

AR models handle the early years in different ways. There is no AR(\(p\)) estimate for years \(t \in \{1, \dots, p \}\). One of the more common ways to deal with these years is to simply condition on them. This seems like a big price to pay. Because mastif is a state-space model, and I am imputing fecundity and maturation anyway, it makes sense to imput fecundity/maturation for years \(t-p, \dots, t-1\).

So while I am imputing the past, it makes sense to predict the future. Conditionally, fecundity in the final year \(T_i\) depends on the future, up to year \(T_i + p\). To accommodate past and future, mastif imputes backward \(p\) years from the first observation and predicts forward \(p\) years beyond the last observed year.

A consistent treatment would appear to demand that AR effects be restricted to individuals that have been mature for the past \(t - p\) years. Note that this would not be a concern if maturation state was known. I adopt this rule, so lag effect estimates are not biased downward by the inclusion of trees that might have been immature and, thus, having no effect.

Although fecundity is imputed for all years, including before observations began, sampling of coefficients for fixed effects and lag effects is restricted to years in which trees were observed and mature.

AR(p) model structure

To avoid further notation, the description that follows should be understood to apply only to tree-years for which the mature state extends back to \(t - p\) years. Also to simplify notation I initially omit the subscript \(j\). Note that multiple plots \(j\) might fall within a group \(g\).

Conditionally, the model for an individual \(i\) in group \(g \in \{1, \dots, G \}\) could be written as

\[\psi_{ig,t} | \mu_{ig,t}, \boldsymbol{\alpha}, \boldsymbol{\alpha}_g, \boldsymbol{\tilde{\psi}}_{ig,t} \sim N \left( m_{ig,t}, \sigma^2 \right)\]

where

\[ \begin{align} m_{ig,t} &= \mu_{ig,t} + \sum^p_{l=1} (\alpha_l + \alpha_{gl}) \psi_{ig,t-l} \\ &= \mu_{ig,t} + (\boldsymbol{\alpha} + \boldsymbol{\alpha}_{g})' \boldsymbol{\tilde{\psi}}_{ig,t} \end{align} \]

\(\mu_{ig,t} = \mathbf{x}'_{ig,t} \boldsymbol{\beta}^x\) is the fixed effect, \(\boldsymbol{\tilde{\psi}}_{ig,t} = (\psi_{ig,t-1}, \dots, \psi_{ig,t-p})'\) is the vector of lagged fecundities for \((ig,t)\), \(\boldsymbol{\alpha} = (\alpha_1, \dots, \alpha_p)'\) is the vector of fixed effects for lag \(l = 1, \dots, p\), and \(\boldsymbol{\alpha}_g = (\alpha_{g1}, \dots, \alpha_{gp})\) is the random effect for group \(g\) with prior distribution

\[\begin{align} \boldsymbol{\alpha}_{g} &\sim MVN(\mathbf{0}, \mathbf{A}_{(l)}) \\ \mathbf{A}_{(l)} &\sim IW \left( \tilde{\mathbf{A}}_{(l)}, df \right) \end{align}\]

To facilitate sampling, the fecundity values are organized into a vector \(\boldsymbol{\psi} = \{\psi_{ig,t} | i = 1, \dots, n, g = 1, \dots, G, t = 1, \dots, T_i \}\) and a corresponding matrix of \(p\) lag terms. For example, a vector with these subscripts

\[\boldsymbol{\psi} = \left( \psi_{i,g,t}, \psi_{i,g,t+1}, \dots, \psi_{i,g,T_i}, \psi_{i+1,g,t}, \dots, \right)\] has the lag matrix with matching rows and \(p\) columns,

\[ \boldsymbol{\tilde{\Psi}} = \pmatrix{ \psi_{i,g,t-1} & \dots & \psi_{i,g,t-p} \\ \psi_{i,g,t} & \dots & \psi_{i,g,t+1-p} \\ \vdots & \vdots & \vdots \\ \psi_{i,g,T_i-1} & \dots & \psi_{i,g,T_i-p} \\ \psi_{i+1,g,t} & \dots & \psi_{i+1,g,t+1-p} \\ \vdots & \vdots & \vdots }. \]

Construct the matrix:

\[ \mathbf{G} = \pmatrix{ \psi_1 & \psi_2 & \dots & \psi_{p-1} & \psi_p \\ 1 & 0 &\dots & 0 & 0 \\ 0 & 1 &\dots & 0 & 0 \\ \vdots & \ddots & \ddots & \ddots & \vdots \\ 0 & 0 & \dots & 0 & 0 \\ 0 & 0 & \dots & 1 & 0 }. \] (West and Harrison 1994). A stationary AR(\(p\)) process has all eigenvalues of less than unit modulus (whether real or complex). A quasi-periodic process has complex eigenvalues.

sample fixed effects

To sample fixed effects, I move a few terms to the left,

\[\mathbf{m} = \boldsymbol{\tilde{\Psi}} \boldsymbol{\alpha}\] where \(\mathbf{m}\) has elements \(\psi_{ig,t} - \mu_{ig,t} - \boldsymbol{\alpha}_{g[i]}' \boldsymbol{\tilde{\psi}}_{ig,t}\), and \(\boldsymbol{\alpha}_{g[i]}\) indicates the vector of lags for the group to which individual \(i\) belongs. The conditional posterior matrix for fixed effects is

\[\begin{align} \boldsymbol{\alpha}|\{\boldsymbol{\alpha}_g \} &\sim MVN(\mathbf{V}\mathbf{v},\mathbf{V}) \\ \mathbf{v} &= \sigma^{-2}\boldsymbol{\tilde{\Psi}}'\mathbf{m} \\ \mathbf{V}^{-1} &=\sigma^{-2}\boldsymbol{\tilde{\Psi}}'\boldsymbol{\tilde{\Psi}} + \mathbf{A}^{-1} \end{align}\]

random group effects

For random effects, make a slight change in \(\mathbf{m}\),

\[\begin{align} \boldsymbol{\alpha}_g &\sim MVN(\mathbf{V}_g\mathbf{v}_g,\mathbf{V}_g) \\ \mathbf{v}_g &= \frac{1}{\sigma^2} \sum_{i,t} \boldsymbol{\tilde{\psi}}'_{ig,t} m_{ig,t} \\ \mathbf{V}_g &= \frac{1}{\sigma^2} \sum_{i,t} \boldsymbol{\tilde{\psi}}_{ig,t} \boldsymbol{\tilde{\psi}}'_{ig,t} + \mathbf{A}_{(l)}^{-1} \end{align} \]

where \(m_{ig,t} = \psi_{ig,t} - \mu_{ig,t} - \boldsymbol{\alpha}' \boldsymbol{\tilde{\psi}}_{ig,t}\). The summations are taken over all observation years in which individual \(i\) has been in the mature state for the previous \(p\) years, for all individuals in group \(g\). Here is the conditional for the covariance,

\[\mathbf{A}_{(l)}|\{ \boldsymbol{\alpha}_g \} \sim IW \left( \sum_{g=1}^G \boldsymbol{\alpha}_g \boldsymbol{\alpha}'_g + df \times \tilde{\mathbf{A}}_{(l)}, G + df \right)\]

latent states

Latent states in the AR(p) model are sampled by proposing from the conditional posterior for the fecundity/maturation submodel \(\psi_t, z_t | \psi_{\{-t\}}, z_{t-1},z_{t+1}\) and accepting from the likelihood for seed data (Method 2). To reduce clutter, I now omit subscripts \(ijg\). If there are random groups in the model, then everything I say below is handled at the group level, with lag coefficient \(\alpha_l\) being replaced with \(\alpha_l + \alpha_{gl}\) for group \(g\).

The AR(p) model can be written as

\[\psi_t \sim N \left( m_t, \sigma^2 \right)\]

where

\[m_t = \mu_t + \sum^p_{l=1} \alpha_l \psi_{t-l}\]

The exponent of the conditional distribution \(\psi_t | \psi_{\{-t\}}\) can be factored this way:

\[\frac{1}{\sigma^2} \left[ \left( \psi_t - m_t \right)^2 + \sum_{k=1}^p \left( n_{t,k} - \alpha_k \psi_t \right)^2 \right]\]

where

\[n_{t,k} = \psi_{t+k} - \mu_{t+k} - \sum_{l=1}^p \alpha_l \psi_{t+k-l}I(l\neq k)\] All of this goo just isolates the terms in \(\psi_t\).

To sample latent states, I propose from

\[\psi_t \sim N \left( V v_t, V \right)\]

where

\[\begin{align} v_t &= \frac{1}{\sigma^2} \left( m_t + \sum_{k=1}^p n_{t,k} \alpha_k \right) \\ V^{-1} &= \frac{1}{\sigma^2} \left( 1 + \sum_{k=1}^p \alpha_k^2 \right) \end{align} \]

Proposals are accepted as a block for each plot-year in the data set, based on the likelihood for seed data (see Method 2).

year effects

For a single group, years effects are fixed, drawn from the conditional

\[ \begin{align} \gamma_t &\sim N(V_t v_t, V_t) \\ V_{t}^{-1} &= \frac{n_t}{\sigma^2} + 1/\tau^2 \\ v_{t} &= \frac{1}{\sigma^2} \sum_{i}(\psi_{i,t} - \mu_{i,t}) \end{align} \]

With multiple groups, there are random year effects across groups:

\[\begin{align} \gamma_{g,t} &\sim N(V_t v_t, V_t) \\ V_{g,t}^{-1} &= \frac{n_{g,t}}{\sigma^2} + 1/\tau^2_t \\ v_{g,t} &= \frac{1}{\sigma^2} \sum_{i \in g}(\psi_{i,t} - \mu_{i,t} - \gamma_t) \end{align} \]

Years have a sum-to-zero constraint imposed in Gibbs sampling on the year effects. The intercept for a given year is the overall intercept plus the year effect for that year.

The variance for random effects:

\[ \tau^2_t \sim IG \left( 2 + \frac{n_t}{2}, 1 + \frac{1}{2} \sum\gamma_{g,t}^2 \right) \] where \(n_t\) is the number of groups available in year \(t\), i.e., those having mature individuals in that year.

other parameters

The error variance \(\sigma^2\) is sampled from the conditional inverse gamma posterior distribution.

If there are no random groups, the dispersal parameter \(u\) is sampled with Metropolis, with an adaptive proposal variance and truncated normal prior distribution,

\[ [u] \propto L \times N(u | u_0, U_0) I(u > 0) \] where the likelihood \(L\) is \(\prod_{r,s,j,t} Poi \left(y_{rsj,t} | A_{sj} \lambda_{rsj,t}(u,\boldsymbol{\psi}_{j,t}, \boldsymbol{\rho}_{j,t}, \mathbf{r})\right)\), \(u_0\) and \(U_0\) are the prior mean and variance dispersal parameters.

If there are random groups, then an additional stage for the global mean. The previous distribution applies to \(u_g\) for group \(g\),

\[ [u_g] \propto L \times N(u_g | u, U) \] The \(u_g\) are proposed and accepted as a block.

The global mean and variance have conditional distributions:

\[ \begin{align} u | u_1, \dots, u_G, u_0, U,U_0 &\sim N \left(Vv,V \right) \\ V^{-1} &= \frac{G}{U} + \frac{1}{U_0} \\ v &= \frac{1}{U} \sum_g u_g + \frac{ u_0 }{U_0} \\ U |u_1, \dots, u_G, u &\sim IG \left(2 + \frac{G}{2}, 1 + \frac{1}{2} \sum_g ( u_g - u )^2 \right) \end{align} \]

references

Clark, JS, DM Bell, M Kwit, A Powell, and K Zhu. 2013. Dynamic inverse prediction and sensitivity analysis with high-dimensional responses: application to climate-change vulnerability of biodiversity. Journal of Biological, Environmental, and Agricultural Statistics 18, 376-404.

Clark, JS, C Nunes, and B Tomasek. 2018. Pulsed-resource mast systems and the movement, demographic storage and diet breadth of consumers, in review.

Clark, JS, S LaDeau, and I Ibanez. 2004. Fecundity of trees and the colonization-competition hypothesis, Ecological Monographs, 74, 415-442.

Clark, JS, M Silman, R Kern, E Macklin, and J Hille Ris Lambers. 1999. Seed dispersal near and far: generalized patterns across temperate and tropical forests. Ecology 80, 1475-1494.

Neal, R.M. 2011. MCMC using Hamiltonian dynamics. In: Handbook of Markov Chain Monte Carlo, edited by S. Brooks, A. Gelman, G. Jones, and X.-L. Meng. Chapman & Hall / CRC Press.