Data frame with continuous time-block variable in first column, ordered by dates in second column, followed by columns of taxa abundance time-series
variables
A vector as long as the number of columns in data indicating how each taxon column should be assigned
for the analysis (first two values for the time-block and date columns should be 0):
0 : not included
1 : included as a variate
2 : included as a covariate
Alternatively, an object of class MAR resulting from a previous call to
run.mar from which to extract variable assignments
restrictions
A matrix with n variate rows and n variate + n covariate columns of values indicating
the potential of a direct effect of each column variable on each row variable:
0.5 : possible (may be included in model)
0 : unlikely/implausible (won't be included in model)
1 : probable (will be included in model)
Alternatively, on object of class MAR resulting from a previous call to
run.mar from which to extract interaction restrictions
search
A character string indicating the type of search that should be used to find the best-fit model; either "random" (default), "exhaustive", "fwdstep", or "exhaustive.true"; see "Details" section below for descriptions of search types
boot
Either an integer indicating the number of iterations that should be performed in the model bootstrap
or FALSE to skip bootstrapping
ntop
If search="random", "exhaustive", or "exhaustive.true", the number of top best-fit models from the random search to be returned for potential comparison to the selected best-fit model
export
If set to TRUE, a call to export.MAR is executed at the end of the analysis. export.MAR creates a new directory and saves all components of the MAR model object in that directory as csv files
Details
Variables and Constraints:
If the variables or restrictions arguments are not provided, the function creates
windows that allow the user to pick which column variables in data should be included in the MAR model as variates or covariates and to set restrictions on potential interactions between model variables. Unless the variables argument is provided in the function call, the restrictions argument must be NULL.
Search Types:
If search="random", which is the default, a random search is performed to find the best-fit model (as determined by AIC) for the included variate time-series. For each variate, 100 random models are constructed according to the restrictions that were set, and the model with the lowest AIC of these models is retained. This process is repeated 100 times, resulting in 100 "best-of-100" models. If any variable occurs in less than 15 of the 100 "best-of-100" models, that variable is discarded (i.e., the probability of that variable occurring in the random search is set to 0) and the search is repeated until the number of variables in successive searches remains constant (resulting in at least 2 search iterations per variate unless all variables are retained in the first iteration). The model with the lowest AIC of the final 100 "best-of-100" models is retained.
If search="exhaustive", a search through possible models for each variate with respect to restrictions is performed using a leap and bound algorithm (Furnival and Wilson,1974) to find the "best-fit" (lowest AIC) model of all potential variable combinations without explicitly examining all possible subsets.
If search="fwdstep", the best-fit model for each variate is built up from the NULL intercept model by sequentially adding whichever variable most improves the model AIC from the pool of potential variables. The model from the series with the lowest AIC is retained.
If search="exhaustive.true", a true exhaustive search through all potential variable combinations with respect to restrictions is performed.
Statistics:
The coefficients of the final "best-fit" MAR model for all variates are attained using least-squares estimation. The coefficients of the B- and C-matrices represent interaction strengths of the column variables on the row variables. If the data were z-scored prior to analysis (see prepare.data), the A intercept values will not be significantly different from 0. Estimates of the stationary distribution mean and covariance for each variate, of the process errors, and of community stability (resilience and reactivity) are calculated following Ives (2003). If bootstrap is not set to FALSE, these statistics are also calculated for the bootstrapped model.
Value
Returns a list of class MAR containing:
variables.selected
corresponds to variables argument
restrictions.set
corresponds to restrictions argument
search.type
corresponds to search argument
search.time.s
time (in seconds) the best-fit model search took
And for each of $bestfit and $bootstrap:
A
each row is the a-value for the variate
B
B-matrix interaction coefficients of columns on rows
C
C-matrix interaction coefficients of columns on rows
log.likelihood
log.likelihood value for model
AIC
AIC value for model
BIC
BIC value for model
R2.values
R^2 and conditional R^2 values for each variate
stationary.distribution
mean means of variates' stationary distributions covariance covariance matrix of stationary distribution
process.errors
residuals E covariance sigma corrmatrix correlation matrix
stability
resilience
eigB eigenvalues of the B matrix detB determinant of the B matrix maxeigB max eigenvalue of B matrix maxeigkrB max eigenvalue of B matrix kronecker products
reactivity
sigma.over.Vinf -tr(sigma)/tr(Vinf) maxeigBxB max eigenvalue of B'B matrix ("worst-case" reactivity)
If bootstrapping is not performed, $bootstrap will be NULL. Otherwise, in addition to the statistics above, $bootstrap will also contain a $limits list with the upper and lower 95% confidence limits of the best-fit model elements.
If search="random", "exhaustive", or "exhaustive.true", the result will also contain $top.bestfit, an array of the top best-fit models tested during the model search (the first of which is the best-fit model that was selected). The number of top models returned may be less than the value set for the ntop argument if ntop exceeds the potential number of model configurations that can be tested for the selected variables and search method. The dimension of the array in which each top best-fit model is stored is named by its respective AIC value.
Warning
The "random" model search may select different best-fit models when run multiple times on the same data, particularly for searches including a large number of variables.
The "exhaustive.true" search can become very time-consuming for models with more than 12 variables.
Author(s)
LP Scheef
References
Furnival GM, Wilson Jr RW (1974) Regressions by leaps and bounds. Technometrics 16:499-511
Ives AR, Dennis B, Cottingham KL, and Carpenter SR (2003) Estimating community stability and ecological interactions from time-series data. Ecological Monographs 73:301-330
See Also
prepare.data, plot.MAR, export.MAR
Packages used for exhaustive search methods: leaps, bestglm
Examples
## Not run:
## These examples take 1-2 minutes to run
## construct a MAR model using 'run.mar' arguments to set variables and restrictions ##
data(L4.mar)
myvar<-c(0,0,0,1,1,0,0,0,1,1,1,1,0,0,1,1,0,0,2,2,2) # 8 variates, 3 covariates
myres<-matrix(0.5,nrow=length(which(myvar==1)),
ncol=length(which(myvar!=0))) # no restrictions (all 0.5)
run1<-run.mar(L4.mar, variables=myvar, restrictions=myres, search="exhaustive")
run1 # only some elements of the object are printed
str(run1) # to see all elements
summary(run1) # some summary statistics for the model
plot(run1)
# set a few restrictions on taxa interactions
myres[1,c(1,6,9)]<-c(1,0,0) # included, not included, not included
# re-run the analysis with same variates as 'run1' and new restrictions
run1b<-run.mar(L4.mar,run1,myres,"exhaustive")
plot(run1,run1b)
# 'run1' variables and restrictions with a different search method
run1c<-run.mar(L4.mar,run1,run1,"fwdstep")
plot(run1,run1c,legend=TRUE) # plot with legend
## construct a MAR model using windows to select variables and restrictions ##
run2<-run.mar(L4.mar,search="exhaustive")
run2
summary(run2)
plot(run2)
## End(Not run)