Random Effects Log Gaussian Cox Process Model for Neuroimaging Coordinate-Based Meta Analysis
This repository has code for the following paper:
Pantelis Samartsidis, Claudia R. Eickhoff, Simon B. Eickhoff Tor D. Wager, Lisa Feldman Barrett, Shir Atzil, Timothy D. Johnson, Thomas E. Nichols (2018). Bayesian log-Gaussian Cox process regression: applications to meta-analysis of neuroimaging working memory studies. Journal of the Royal Statistical Society. Series C (Applied Statistics), in press.
To reproduce the results in our paper:
make
./lgcp 7000 15000 5 50 40 15 > lgcp.log 2>&1 &
You will need a computer with GPU and have to install CUDA.
You may need to issue a disown
command.
At present please consult the usage:
Usage: lgcp Burnin Iters Adjust AdjustWin Thin Save [GPU]
Burnin -- The burn-in period of the HMC
Iters -- The total number of iterations AFTER burn-in
Adjust -- How often to adjust the stepsize
AdjustWindow
-- Chain window when adjusting the stepsize
Thin -- How often to save the running sum of the GPs
Save -- How often to save snapshots of the GPs
GPU -- GPU device number (defaults to 0)
The following files are expected in the ./inputs directory:
setup.txt: Contains following values, one per line:
* Total number of elements in the initial grid. The program
will figure out how many there are in the extended grid
* Total number of points (foci)
* Total number of point patterns (contrasts/studies)
* Total number of covariates
* Total number of spatially varying covariates
* Total number of HMC leapfrog steps
* Seed
* HMC mass parameters (4 values), if one wants to see between-type
comparisons
seed.dat: 3 long integers
rho.txt: Correlation decay parameters, one for each spatially varying
covariate
sigma.txt: Marginal standard deviations, one for each spatially varying
covariate
beta.txt: Overall mean parameter, one for each covariate
gamma.txt: Standard normal variates, 144*192*144=3981312 for each spatially
varying covariate. If missing, random numbers are generated.