Fold wttrms and Λ into trms

• reformulate the LinearMixedModel type by incorporating the wttrms and Λ members into the trms member.
• create AbstractTerm with subtypes MatrixTerm and FactorReTerm
• add some benchmarks using the BenchmarkTools package
• remove some of the instances of method definitions for functions from Base with signatures of Base classes only
  • BLAS-like in-place linear algebra with scalar multipliers are now called e.g. αβA_mul_Bc!
  • the remaining problematic methods are operations with Diagonal for which I plan to create a PR on the julia repository after consulting with Tony and Andreas

Lower Cholesky formulation

Travis failures are timeouts on julia-0.6.0-pre. Once the dust settles on the julia new release I will check for bottlenecks.

Last release before v0.8.0

Incorporate a couple of commits on the master branch prior to major changes from merging the LowerCholesky branch.

Allow 3 or more nested factors

v0.7.6

Fix correlation store in bootstrap!

Return a DataFrame from bootstrap

The bootstrap function now returns a data frame with columns corresponding to individual parameters.

Correct the calculation of the conditional std. dev. of the r.e.

Correct the calculation of conditional std dev of r.e.

* Initialize pars to optsum.initial, not optsum.final

* clean up logic in optimize for GLMM - still needs work

* Restore model at the end of the bootstrap

* Use Cholesky factor not product in `condVar`

* Need to square diagonals of Cholesky factor

Failures on v0.6.0-dev are new and likely not to be unique to this package.

Fix bug introduced in v0.7.2

In the fit! method for LinearMixedModel objects the parameters were initialized to optsum.final not optsum.initial. This is not a problem for newly created objects because final is a copy of initial, But it does cause a problem for simulations such as a parametric bootstrap.

Extend OptSummary

• extend OptSummary to include more information and to convey settings for the algorithm
• add show method for OptSummary
• clean up logic on last evaluation of objective to ensure the structure is consistent with xmin
• clean up code in fit! methods
• add Ac_mul_B methods for ScalarReTerm and VectorReTerm combinations.

Provide nAGQ=0 option for fitting GLMMs

Using

fit!(glmm(...), nAGQ = 0)

provides a faster optimization algorithm at the expense of some small loss of accuracy.

For models with a large number of fixed-effects parameters relative to the number of covariance parameters there can be considerable speedup because nAGQ=0 profiles out the fixed-effects parameters as part of the PIRLS iterations, thereby reducing the dimension of the constrained, nonlinear optimization problem passed to an NLopt optimizer.

Speed increase

Allocation profiling showed that two downdate! methods were being slowed down because the compiler had insufficient type information on ReMat types. Expanding the template parameters for these types resulted in considerable speed improvement for examples like InstEval.