Beyond GLM: The potential for a generic likelihood toolbox Peter Dalgaard Department of Biostatistics University of Copenhagen Royal Statistical Society, Nottingham, September 2008 1 / 14
Introduction ◮ Developments in statistics are driven by its tools ◮ Tools that we have available are much stronger than what we currently use them for ◮ It is feasible to use features of the R language to work with models at a level that transcends the current practice ◮ (However, this is all still at a very preliminary stage) ◮ Overview ◮ Pre-history/motivation ◮ Current capabilities in R ◮ Ideas for extensions 2 / 14
Introduction ◮ Developments in statistics are driven by its tools ◮ Tools that we have available are much stronger than what we currently use them for ◮ It is feasible to use features of the R language to work with models at a level that transcends the current practice ◮ (However, this is all still at a very preliminary stage) ◮ Overview ◮ Pre-history/motivation ◮ Current capabilities in R ◮ Ideas for extensions 2 / 14
Introduction ◮ Developments in statistics are driven by its tools ◮ Tools that we have available are much stronger than what we currently use them for ◮ It is feasible to use features of the R language to work with models at a level that transcends the current practice ◮ (However, this is all still at a very preliminary stage) ◮ Overview ◮ Pre-history/motivation ◮ Current capabilities in R ◮ Ideas for extensions 2 / 14
Introduction ◮ Developments in statistics are driven by its tools ◮ Tools that we have available are much stronger than what we currently use them for ◮ It is feasible to use features of the R language to work with models at a level that transcends the current practice ◮ (However, this is all still at a very preliminary stage) ◮ Overview ◮ Pre-history/motivation ◮ Current capabilities in R ◮ Ideas for extensions 2 / 14
Introduction ◮ Developments in statistics are driven by its tools ◮ Tools that we have available are much stronger than what we currently use them for ◮ It is feasible to use features of the R language to work with models at a level that transcends the current practice ◮ (However, this is all still at a very preliminary stage) ◮ Overview ◮ Pre-history/motivation ◮ Current capabilities in R ◮ Ideas for extensions 2 / 14
Introduction ◮ Developments in statistics are driven by its tools ◮ Tools that we have available are much stronger than what we currently use them for ◮ It is feasible to use features of the R language to work with models at a level that transcends the current practice ◮ (However, this is all still at a very preliminary stage) ◮ Overview ◮ Pre-history/motivation ◮ Current capabilities in R ◮ Ideas for extensions 2 / 14
Introduction ◮ Developments in statistics are driven by its tools ◮ Tools that we have available are much stronger than what we currently use them for ◮ It is feasible to use features of the R language to work with models at a level that transcends the current practice ◮ (However, this is all still at a very preliminary stage) ◮ Overview ◮ Pre-history/motivation ◮ Current capabilities in R ◮ Ideas for extensions 2 / 14
Introduction ◮ Developments in statistics are driven by its tools ◮ Tools that we have available are much stronger than what we currently use them for ◮ It is feasible to use features of the R language to work with models at a level that transcends the current practice ◮ (However, this is all still at a very preliminary stage) ◮ Overview ◮ Pre-history/motivation ◮ Current capabilities in R ◮ Ideas for extensions 2 / 14
Generalized Linear Models ◮ 3 dozen years ago generalized linear models were born ◮ This extended the capabilities of regression analysis and brought major modes of analysis into the same framework ◮ Multiple regression ◮ Logistic regression ◮ Poisson (rate) regression ◮ (Cox regression) ◮ Encapsulated in GLIM and Genstat software (c.1974) 3 / 14
Generalized Linear Models ◮ 3 dozen years ago generalized linear models were born ◮ This extended the capabilities of regression analysis and brought major modes of analysis into the same framework ◮ Multiple regression ◮ Logistic regression ◮ Poisson (rate) regression ◮ (Cox regression) ◮ Encapsulated in GLIM and Genstat software (c.1974) 3 / 14
Generalized Linear Models ◮ 3 dozen years ago generalized linear models were born ◮ This extended the capabilities of regression analysis and brought major modes of analysis into the same framework ◮ Multiple regression ◮ Logistic regression ◮ Poisson (rate) regression ◮ (Cox regression) ◮ Encapsulated in GLIM and Genstat software (c.1974) 3 / 14
Generalized Linear Models ◮ 3 dozen years ago generalized linear models were born ◮ This extended the capabilities of regression analysis and brought major modes of analysis into the same framework ◮ Multiple regression ◮ Logistic regression ◮ Poisson (rate) regression ◮ (Cox regression) ◮ Encapsulated in GLIM and Genstat software (c.1974) 3 / 14
Generalized Linear Models ◮ 3 dozen years ago generalized linear models were born ◮ This extended the capabilities of regression analysis and brought major modes of analysis into the same framework ◮ Multiple regression ◮ Logistic regression ◮ Poisson (rate) regression ◮ (Cox regression) ◮ Encapsulated in GLIM and Genstat software (c.1974) 3 / 14
Generalized Linear Models ◮ 3 dozen years ago generalized linear models were born ◮ This extended the capabilities of regression analysis and brought major modes of analysis into the same framework ◮ Multiple regression ◮ Logistic regression ◮ Poisson (rate) regression ◮ (Cox regression) ◮ Encapsulated in GLIM and Genstat software (c.1974) 3 / 14
Generalized Linear Models ◮ 3 dozen years ago generalized linear models were born ◮ This extended the capabilities of regression analysis and brought major modes of analysis into the same framework ◮ Multiple regression ◮ Logistic regression ◮ Poisson (rate) regression ◮ (Cox regression) ◮ Encapsulated in GLIM and Genstat software (c.1974) 3 / 14
Some observations about GLMs ◮ Reuse of prior technology/ideas ◮ Weighted least squares ◮ Analysis of variance/deviance ◮ Wilkinson-Rogers formulas ◮ Innovation ◮ Use of W-R formulas by software (Genstats "TREAT" directive) ◮ Focusing ◮ Esp. on “What can we do with linear predictors” ◮ Delimiting ◮ Tendency to “forget” other options ◮ Partially false distinctions between models that “can be fitted with standard software” and “difficult models” 4 / 14
Some observations about GLMs ◮ Reuse of prior technology/ideas ◮ Weighted least squares ◮ Analysis of variance/deviance ◮ Wilkinson-Rogers formulas ◮ Innovation ◮ Use of W-R formulas by software (Genstats "TREAT" directive) ◮ Focusing ◮ Esp. on “What can we do with linear predictors” ◮ Delimiting ◮ Tendency to “forget” other options ◮ Partially false distinctions between models that “can be fitted with standard software” and “difficult models” 4 / 14
Some observations about GLMs ◮ Reuse of prior technology/ideas ◮ Weighted least squares ◮ Analysis of variance/deviance ◮ Wilkinson-Rogers formulas ◮ Innovation ◮ Use of W-R formulas by software (Genstats "TREAT" directive) ◮ Focusing ◮ Esp. on “What can we do with linear predictors” ◮ Delimiting ◮ Tendency to “forget” other options ◮ Partially false distinctions between models that “can be fitted with standard software” and “difficult models” 4 / 14
Some observations about GLMs ◮ Reuse of prior technology/ideas ◮ Weighted least squares ◮ Analysis of variance/deviance ◮ Wilkinson-Rogers formulas ◮ Innovation ◮ Use of W-R formulas by software (Genstats "TREAT" directive) ◮ Focusing ◮ Esp. on “What can we do with linear predictors” ◮ Delimiting ◮ Tendency to “forget” other options ◮ Partially false distinctions between models that “can be fitted with standard software” and “difficult models” 4 / 14
Some observations about GLMs ◮ Reuse of prior technology/ideas ◮ Weighted least squares ◮ Analysis of variance/deviance ◮ Wilkinson-Rogers formulas ◮ Innovation ◮ Use of W-R formulas by software (Genstats "TREAT" directive) ◮ Focusing ◮ Esp. on “What can we do with linear predictors” ◮ Delimiting ◮ Tendency to “forget” other options ◮ Partially false distinctions between models that “can be fitted with standard software” and “difficult models” 4 / 14
Some observations about GLMs ◮ Reuse of prior technology/ideas ◮ Weighted least squares ◮ Analysis of variance/deviance ◮ Wilkinson-Rogers formulas ◮ Innovation ◮ Use of W-R formulas by software (Genstats "TREAT" directive) ◮ Focusing ◮ Esp. on “What can we do with linear predictors” ◮ Delimiting ◮ Tendency to “forget” other options ◮ Partially false distinctions between models that “can be fitted with standard software” and “difficult models” 4 / 14
Recommend
More recommend
