Combining parametric, semi-parametric, and non-parametric survival\n models with stacked survival models

For estimating conditional survival functions, non-parametric estimators can\nbe preferred to parametric and semi-parametric estimators due to relaxed\nassumptions that enable robust estimation. Yet, even when misspecified,\nparametric and semi-parametric estimators can possess better operating\ncharacteristics in small sample sizes due to smaller variance than\nnon-parametric estimators. Fundamentally, this is a bias-variance tradeoff\nsituation in that the sample size is not large enough to take advantage of the\nlow bias of non-parametric estimation. Stacked survival models estimate an\noptimally weighted combination of models that can span parametric,\nsemi-parametric, and non-parametric models by minimizing prediction error. An\nextensive simulation study demonstrates that stacked survival models\nconsistently perform well across a wide range of scenarios by adaptively\nbalancing the strengths and weaknesses of individual candidate survival models.\nIn addition, stacked survival models perform as good as, or better than, the\nmodel selected through cross-validation. Lastly, stacked survival models are\napplied to a well-known German breast cancer study.\n