Hierarchical models, including generalized linear models with mixed random and fixed effects, are increasingly popular. The rapid expansion of applications is largely due to the advancement of the Markov Chain Monte Carlo (MCMC) algorithms and related software. Data cloning is a statistical computing method introduced by Lele et al. 2007^{1} and 2010^{2}. It exploits the computational simplicity of the MCMC algorithms used in the Bayesian statistical framework, but it provides the maximum likelihood point estimates and their standard errors for complex hierarchical models. The use of the data cloning algorithm is especially valuable for complex models, where the number of unknowns increases with sample size (i.e. with latent variables), because inference and prediction procedures are often hard to implement in such situations.
Imagine a hypothetical situation where an experiment is repeated by $k$ different observers, and all $k$ experiments happen to result in exactly the same set of observations, $y^{(k)} = \left(y,y,\ldots,y\right)$. The likelihood function based on the combination of the data from these $k$ experiments is $L(\theta, y^{\left(k\right)}) = \left[L\left(\theta, y\right)\right]^k$. The location of the maximum of $L(\theta,y^{(k)})$ exactly equals the location of the maximum of the function $L\left(\theta, y\right)$, and the Fisher information matrix based on this likelihood is $k$ times the Fisher information matrix based on $L\left(\theta, y\right)$.
One can use MCMC methods to calculate the posterior distribution of the model parameters ($\theta$) conditional on the data. Under regularity conditions, if $k$ is large, the posterior distribution corresponding to $k$ clones of the observations is approximately normal with mean $\hat{\theta}$ and variance $1/k$ times the inverse of the Fisher information matrix. When $k$ is large, the mean of this posterior distribution is the maximum likelihood estimate and $k$ times the posterior variance is the corresponding asymptotic variance of the maximum likelihood estimate if the parameter space is continuous.
Data cloning is a computational algorithm to compute maximum likelihood estimates and the inverse of the Fisher information matrix, and is related to simulated annealing. By using data cloning, the statistical accuracy of the estimator remains a function of the sample size and not of the number of cloned copies. Data cloning does not improve the statistical accuracy of the estimator by artificially increasing the sample size. The data cloning procedure avoids the analytical or numerical evaluation of high dimensional integrals, numerical optimization of the likelihood function, and numerical computation of the curvature of the likelihood function.
The dclone
R package explained in Sólymos 2010^{3} provides infrastructure for data cloning.
Users who are familiar with Bayesian methodology can instantly use the package for maximum likelihood
inference and prediction. Developers of R packages can build on the low level functionality provided by
the package to implement more specific higher level estimation procedures
for users who are not familiar with Bayesian methodology.
This paper demonstrates the implementation of the data cloning algorithm, and presents a case study on
how to write high level functions for specific modeling problems.
Lele, S.R., B. Dennis and F. Lutscher, 2007. Data cloning: easy maximum likelihood estimation for complex ecological models using Bayesian Markov chain Monte Carlo methods. Ecology Letters 10, 551–563. PDF ↩
Lele, S. R., Nadeem, K., and Schmuland, B., 2010. Estimability and likelihood inference for generalized linear mixed models using data cloning. Journal of the American Statistical Association 105, 1617–1625. PDF ↩
Sólymos, P., 2010. dclone: Data Cloning in R. The R Journal 2(2), 29–37. PDF ↩