spatFilt.bms {spatBMS}R Documentation

Bayesian Model Sampling and Averaging with uncertain Spatial Effects

Description

Given data and prior information, this function samples all possible model combinations via MC3 and returns aggregate results.

Usage

spatFilt.bms(X.data, WList,burn = 1000, iter = NA, nmodel = 100, mcmc = "bd",
  g = "UIP", mprior = "random", mprior.size = NA, user.int = TRUE, 
    start.value = NA, logfile = FALSE, logstep = 10000)

Arguments

X.data a data frame or a matrix, with the dependent variable in the first column, followed by the covariates (alternatively, X.data can also be provided as a formula). Note that bms automatically estimates a constant, therefore including constant terms is not necessary.
WList a list object with length corresponding to the number of weight matrices you use (length(WList) must be >1)). Each element of the list contains a matrix with the extracted eigenvectors, where the matrices do not have to have the same column dimension.
burn The (positive integer) number of burn-in draws for the MC3 sampler, defaults to 1000. (Not taken into account if mcmc="enumerate")
iter If mcmc is set to an MC3 sampler, then this is the number of iteration draws to be sampled (ex burn-ins), default 3000 draws.
nmodel the number of best models for which information is stored (default 500). Best models are used for convergence analysis between likelihoods and MCMC frequencies, as well as likelihood-based inference.
Note that a very high value for nmodel slows down the sampler significantly. Set nmodel=0 to speed up sampling (if best model information is not needed).
mcmc a character denoting the model sampler to be used.
The MC3 sampler mcmc="bd" corresponds to a birth/death MCMC algogrithm. mcmc="rev.jump" enacts a reversible jump algorithm adding a "swap" step to the birth / death steps from "bd".
Alternatively, the entire model space may be fully enumerated by setting mcmc="enumerate" which will iterate all possible regressor combinations (Note: consider that this means 2^K iterations, where K is the number of covariates.)
Cf. section 'Details' for more options.
g the hyperparameter on Zellner's g-prior for the regression coefficients.
g="UIP" corresponds to g=N, the number of observations (default);
g="BRIC" corresponds to the benchmark prior suggested by Fernandez, Ley and Steel (2001), i.e g=max(N, K^2), where K is the total number of covariates;
g="RIC" sets g=K^2 and conforms to the risk inflation criterion by George and Foster (1994)
g="HQ" sets g=log(N)^3 and asymptotically mimics the Hannan-Quinn criterion with C_{HQ}=3 (cf. Fernandez, Ley and Steel, 2001, p.395)
mprior a character denoting the model prior choice, defaulting to "random":
mprior="fixed" denotes fixed common prior inclusion probabilities for each regressor as e.g. in Sala-i-Martin, Doppelhofer, and Miller(2004) - for their fine-tuning, cf. mprior.size. Preferable to mcmc="random" if strong prior information on model size exists;
mprior="random" (default) triggers the 'random theta' prior by Ley and Steel (2008), who suggest a binomial-beta hyperprior on the a priori inclusion probability;
mprior="uniform" employs the uniform model prior;
mprior="customk" allows for custom model size priors (cf. mprior.size);
mprior="pip" allows for custom prior inclusion probabilities (cf. mprior.size);
Note that the prior on models with more than N-3 regressors is automatically zero: these models will not be sampled.
mprior.size if mprior is "fixed" or "random", mprior.size is a scalar that denotes the prior expected value of the model size prior (default K/2).
If mprior="customk" then a custom model size prior can be provided as a K+1 vector detailing the priors from model size 0 to K (e.g. rep(1,K+1) for the uniform model prior);
if mprior="pip", then custom prior inclusion probabilities can be provided as a vector of size K, with elements in the interval (0,1)
user.int 'interactive mode': print out results to console after ending the routine and plots a chart (default TRUE).
start.value specifies the starting model of the iteration chain. For instance a specific model by the corresponding column indices (e.g. starting.model=numeric(K) starts from the null model including solely a constant term) or start.value=c(3,6) for a starting model only including covariates 3 and 6.
If start.model is set to an integer (e.g. start.model=15) then that number of covariates (here: 15 covariates) is randomly chosen and the starting model is identified by those regressors with an OLS t-statistic>0.2.
The default value start.value=NA corresponds to start.value=min(ncol(X.data),nrow(X.data)-3). Note that start.value=0 or start.value=NULL starts from the null model.
If mcmc="enumerate" then start.value is the index to start the iteration (default: 0, the null model) . Any number between 0 and K^2-1 is admissible.
logfile setting logfile=TRUE produces a logfile named "test.log" in your current working directory, in order to keep track of the sampling procedure. logfile equal to some filepath (like logfile="subfolder/log.txt") puts the logfile into that specified position. (default: logfile=FALSE). Note that logfile="" implies log printouts on the console.
logstep specifies at which number of posterior draws information is written to the log file; default: 10 000 iterations

Details

Ad mcmc:
Interaction sampler: adding an ".int" to an MC3 sampler (e.g. "mcmc="bd.int") provides for special treatment of interaction terms. Interaction terms will only be sampled along with their component variables: In the colnumn names of X.data, interaction terms need to be denominated by names consisting of the base terms separated by # (e.g. an interaction term of base variables "A", "B" and "C" needs column name "A#B#C"). Then variable "A#B#C" will only be included in a model if all of the component variables ("A", "B", and "C") are included.

The MC3 samplers "bd", "rev.jump", "bd.int" and "rev.jump.int", iterate away from a starting model by adding, droping or swapping (only in the case of rev.jump) covariates.

In an MCMC fashion, they thus randomly draw a candidate model and then move to it in case its marginal likelihood (marg.lik.) is superior to the marg.lik. of the current model.

In case the candidate's marg.lik is inferior, it is randomly accepted or rejected according to a probability formed by the ratio of candidate marg.lik over currrent marg.lik. Over time, the sampler should thus converge to a sensible distribution. For aggregate results based on these MC3 frequencies, the first few iterations are typically disregarded (the 'burn-ins').

Value

A list of classes bma and spatFilt, that may be displayed using e.g. summary.bma or coef.bma. The list contains the following elements:

info a list of aggregate statistics: iter is the number of iterations, burn the number of burn-ins.
The following have to be divided by cumsumweights to get posterior expected values: inccount are the posterior inclusion probabilities, b1mo and b2mo the first and second moment of coefficients, add.otherstats other statistics of interest (typically the moments of the shrinkage factor), msize is the post. expected model size, k.vec the posterior model size distribution, pos.sign the unconditional post. probability of positive coefficents, corr.pmp is the correlation between the best models' MCMC frequencies and their marg. likelihoods.
timed is the time that was needed for MCMC sampling, cons is the posterior expected value of the constant. K and N are the maximum number of covariates and the sample size, respectively.
arguments a list of the evaluated function arguments provided to bms (see above)
topmod a 'topmod' object containing the best drawn models. see topmod for more details
start.pos the positions of the starting model. If bmao is a'bma' object this corresponds to covariates bmao$reg.names[bmao$start.pos]. If bmao is a chain that resulted from several starting models (cf. c.bma, then start.pos is a list detailing all of them.
gprior.info a list detailing information on the g-prior: gtype corresponds to argument g above, is.constant is FALSE if gtype is either "hyper" or "EBL", return.g.stats corresponds to argument g.stats above, shrinkage.moments contains the first and second moments of the shrinkage factor (only if return.g.stats==TRUE), g details the fixed g (if is.constant==TRUE), hyper.parameter corresponds to the hyper-g parameter a as in Liang et al. (2008)
mprior.info a list detailing information on the model prior: origargs lists the original arguments to mprior and mprior.size above; pmp(...) is a function to calculate the prior model probability for a specific model; mp.mode corresponds to argument mprior above; mp.msize denotes the prior mode size; mp.Kdist is a (K+1) vector with the prior model size distribution from 0 to K
X.data data.frame or matrix: corresponds to argument X.data above, possibly cleaned for NAs
reg.names character vector: the covariate names to be used for X.data
spatFilt.call the original call to the spatFilt.bms function
wTopModels a binary matrix containing the best nmodel models along an index (W-Index) which W-matrix has been used.
WList corresponds to argument WList.
Wcount a vector containing the frequency with which the different weight matrices have been visited by the MCMC chain. Burnins are not counted.

Note

There are several ways to speed-up sampling: nmodel=10 saves only the ten best models, at most a marginal improvement. nmodels=0 does not save the best (500) models, however then posterior convergence and likelihood-based inference are not possible. g.stats=FALSE saves some time by not retaining the shrinkage factors for the MC3 chain (and the best models).

Author(s)

Martin Feldkircher

References

Crespo Cuaresma, Jesus and M. Feldkircher (2010): Spatial Filtering, Model Uncertainty and the Speed of Income Convergence in Europe. Working Papers 160, Oesterreichische Nationalbank.

Feldkircher, M. and S. Zeugner (2009): Benchmark Priors Revisited: On Adaptive Shrinkage and the Supermodel Effect in Bayesian Model Averaging, IMF Working Paper 09/202.

Fernandez, C. E. Ley and M. Steel (2001): Benchmark priors for Bayesian model averaging. Journal of Econometrics 100(2), 381–427

Ley, E. and M. Steel (2008): On the Effect of Prior Assumptions in Bayesian Model Averaging with Applications to Growth Regressions. working paper

Tiefelsdorf, M. and D. Griffith (2007): Semiparametric filtering of spatial autocorrelation: the eigenvector approach. Environment and Planning A, 37, 1193-1221.

See Also

coef.bma, plotModelsize and density.bma for some operations on the resulting 'bma' object, c.bma for integrating separate MC3 chains and splitting of sampling over several runs.

Check http://bms.zeugner.eu for additional help.

Examples

  library(BMS)
  data(dataBoston); data(WL.boston)
  #estimating a standard MC3 chain 
  bma1=spatFilt.bms(X.data=dataBoston,WList=WL.boston,burn=1e05,iter=1e05,
              nmodel=100,mcmc="bd",g="bric",mprior="random",mprior.size=(ncol(dataBoston)-1)/2)
              
  coef(bma1,exact=TRUE, std.coefs=TRUE) #standard coefficients based on exact likelihoods of the 100 best models
  
  #suppressing user-interactive output and  using a customized starting value (the null model)
  bma2 = spatFilt.bms(dataBoston,WList=WL.boston,burn=1000, iter=2000, nmodel=10, 
                      start.value=rep(0,ncol(dataBoston)),user.int=FALSE,)
  coef(bma2)
  
  #MC3 chain with the bric prior, saving only the 20 best models, and an alternative sampling procedure; putting a log entry to console every 1000th step
  bma3 = spatFilt.bms(dataBoston,burn=1000,WList=WL.boston, iter=5000, nmodel=20, g="bric", mcmc="rev.jump",logfile="",logstep=1000)
  image(bma3,cex=0.6) #showing the coefficient signs of the 20 best models and setting the font size of the labels in the plot to cex=0.6
  
  #using an interaction sampler for two interaction terms
  bma4 = spatFilt.bms(X.data=dataBoston,WList=WL.boston,burn=1000,iter=9000,start.value=0,mcmc="bd.int") 
  
  density(bma4,reg="NOX") # plot posterior density for covariate "NOX" with a grid
  grid()
[Package spatBMS version 0.0 Index]