Upload
ddeepakkumar
View
138
Download
0
Embed Size (px)
DESCRIPTION
Andrew Gelman
Citation preview
Eectiveness Ubiquity Way of life
Fitting and understanding multilevel (hierarchical) modelsAndrew Gelman Department of Statistics and Department of Political Science Columbia University 8 December 2004
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Making more use of existing information
The problem: not enough data to estimate eects with condence The solution: make your studies broader and deeperBroader: extend to other countries, other years, other outcomes, . . . Deeper: inferences for individual states, demographic subgroups, components of outcomes, . . .
The solution: multilevel modelingRegression with coecients grouped into batches No such thing as too many predictors
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Making more use of existing information
The problem: not enough data to estimate eects with condence The solution: make your studies broader and deeperBroader: extend to other countries, other years, other outcomes, . . . Deeper: inferences for individual states, demographic subgroups, components of outcomes, . . .
The solution: multilevel modelingRegression with coecients grouped into batches No such thing as too many predictors
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Making more use of existing information
The problem: not enough data to estimate eects with condence The solution: make your studies broader and deeperBroader: extend to other countries, other years, other outcomes, . . . Deeper: inferences for individual states, demographic subgroups, components of outcomes, . . .
The solution: multilevel modelingRegression with coecients grouped into batches No such thing as too many predictors
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Making more use of existing information
The problem: not enough data to estimate eects with condence The solution: make your studies broader and deeperBroader: extend to other countries, other years, other outcomes, . . . Deeper: inferences for individual states, demographic subgroups, components of outcomes, . . .
The solution: multilevel modelingRegression with coecients grouped into batches No such thing as too many predictors
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Making more use of existing information
The problem: not enough data to estimate eects with condence The solution: make your studies broader and deeperBroader: extend to other countries, other years, other outcomes, . . . Deeper: inferences for individual states, demographic subgroups, components of outcomes, . . .
The solution: multilevel modelingRegression with coecients grouped into batches No such thing as too many predictors
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Making more use of existing information
The problem: not enough data to estimate eects with condence The solution: make your studies broader and deeperBroader: extend to other countries, other years, other outcomes, . . . Deeper: inferences for individual states, demographic subgroups, components of outcomes, . . .
The solution: multilevel modelingRegression with coecients grouped into batches No such thing as too many predictors
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Making more use of existing information
The problem: not enough data to estimate eects with condence The solution: make your studies broader and deeperBroader: extend to other countries, other years, other outcomes, . . . Deeper: inferences for individual states, demographic subgroups, components of outcomes, . . .
The solution: multilevel modelingRegression with coecients grouped into batches No such thing as too many predictors
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Making more use of existing information
The problem: not enough data to estimate eects with condence The solution: make your studies broader and deeperBroader: extend to other countries, other years, other outcomes, . . . Deeper: inferences for individual states, demographic subgroups, components of outcomes, . . .
The solution: multilevel modelingRegression with coecients grouped into batches No such thing as too many predictors
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Fitting and understanding multilevel modelsThe eectiveness of multilevel models Multilevel models in unexpected places Multilevel models as a way of life collaborators:Iain Pardoe, Dept of Decision Sciences, University of Oregon David Park, Dept of Political Science, Washington University Joe Bafumi, Dept of Political Science, Columbia University Boris Shor, Dept of Political Science, Columbia University Noah Kaplan, Dept of Political Science, University of Houston Shouhao Zhao, Dept of Statistics, Columbia University Zaiying Huang, Circulation, New York Times
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Fitting and understanding multilevel modelsThe eectiveness of multilevel models Multilevel models in unexpected places Multilevel models as a way of life collaborators:Iain Pardoe, Dept of Decision Sciences, University of Oregon David Park, Dept of Political Science, Washington University Joe Bafumi, Dept of Political Science, Columbia University Boris Shor, Dept of Political Science, Columbia University Noah Kaplan, Dept of Political Science, University of Houston Shouhao Zhao, Dept of Statistics, Columbia University Zaiying Huang, Circulation, New York Times
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Fitting and understanding multilevel modelsThe eectiveness of multilevel models Multilevel models in unexpected places Multilevel models as a way of life collaborators:Iain Pardoe, Dept of Decision Sciences, University of Oregon David Park, Dept of Political Science, Washington University Joe Bafumi, Dept of Political Science, Columbia University Boris Shor, Dept of Political Science, Columbia University Noah Kaplan, Dept of Political Science, University of Houston Shouhao Zhao, Dept of Statistics, Columbia University Zaiying Huang, Circulation, New York Times
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Fitting and understanding multilevel modelsThe eectiveness of multilevel models Multilevel models in unexpected places Multilevel models as a way of life collaborators:Iain Pardoe, Dept of Decision Sciences, University of Oregon David Park, Dept of Political Science, Washington University Joe Bafumi, Dept of Political Science, Columbia University Boris Shor, Dept of Political Science, Columbia University Noah Kaplan, Dept of Political Science, University of Houston Shouhao Zhao, Dept of Statistics, Columbia University Zaiying Huang, Circulation, New York Times
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Fitting and understanding multilevel modelsThe eectiveness of multilevel models Multilevel models in unexpected places Multilevel models as a way of life collaborators:Iain Pardoe, Dept of Decision Sciences, University of Oregon David Park, Dept of Political Science, Washington University Joe Bafumi, Dept of Political Science, Columbia University Boris Shor, Dept of Political Science, Columbia University Noah Kaplan, Dept of Political Science, University of Houston Shouhao Zhao, Dept of Statistics, Columbia University Zaiying Huang, Circulation, New York Times
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Outline of talk
The eectiveness of multilevel modelsState-level opinions from national polls (crossed multilevel modeling and poststratication)
Multilevel models in unexpected placesEstimating incumbency advantage and its variation Before-after studies
Multilevel models as a way of lifeBuilding and tting models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Outline of talk
The eectiveness of multilevel modelsState-level opinions from national polls (crossed multilevel modeling and poststratication)
Multilevel models in unexpected placesEstimating incumbency advantage and its variation Before-after studies
Multilevel models as a way of lifeBuilding and tting models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Outline of talk
The eectiveness of multilevel modelsState-level opinions from national polls (crossed multilevel modeling and poststratication)
Multilevel models in unexpected placesEstimating incumbency advantage and its variation Before-after studies
Multilevel models as a way of lifeBuilding and tting models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Outline of talk
The eectiveness of multilevel modelsState-level opinions from national polls (crossed multilevel modeling and poststratication)
Multilevel models in unexpected placesEstimating incumbency advantage and its variation Before-after studies
Multilevel models as a way of lifeBuilding and tting models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Outline of talk
The eectiveness of multilevel modelsState-level opinions from national polls (crossed multilevel modeling and poststratication)
Multilevel models in unexpected placesEstimating incumbency advantage and its variation Before-after studies
Multilevel models as a way of lifeBuilding and tting models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Outline of talk
The eectiveness of multilevel modelsState-level opinions from national polls (crossed multilevel modeling and poststratication)
Multilevel models in unexpected placesEstimating incumbency advantage and its variation Before-after studies
Multilevel models as a way of lifeBuilding and tting models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Outline of talk
The eectiveness of multilevel modelsState-level opinions from national polls (crossed multilevel modeling and poststratication)
Multilevel models in unexpected placesEstimating incumbency advantage and its variation Before-after studies
Multilevel models as a way of lifeBuilding and tting models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Outline of talk
The eectiveness of multilevel modelsState-level opinions from national polls (crossed multilevel modeling and poststratication)
Multilevel models in unexpected placesEstimating incumbency advantage and its variation Before-after studies
Multilevel models as a way of lifeBuilding and tting models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Outline of talk
The eectiveness of multilevel modelsState-level opinions from national polls (crossed multilevel modeling and poststratication)
Multilevel models in unexpected placesEstimating incumbency advantage and its variation Before-after studies
Multilevel models as a way of lifeBuilding and tting models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
National opinion trendsPercentage support for the death penalty 50 60 70 80 1940
1950
1960
1970 Year
1980
1990
2000
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
State-level opinion trends
Goal: estimating time series within each state One poll at a time: small-area estimation It works! Validated for pre-election polls Combining surveys: model for parallel time series Multilevel modeling + poststratication Poststratication cells: sex ethnicity age education state
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
State-level opinion trends
Goal: estimating time series within each state One poll at a time: small-area estimation It works! Validated for pre-election polls Combining surveys: model for parallel time series Multilevel modeling + poststratication Poststratication cells: sex ethnicity age education state
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
State-level opinion trends
Goal: estimating time series within each state One poll at a time: small-area estimation It works! Validated for pre-election polls Combining surveys: model for parallel time series Multilevel modeling + poststratication Poststratication cells: sex ethnicity age education state
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
State-level opinion trends
Goal: estimating time series within each state One poll at a time: small-area estimation It works! Validated for pre-election polls Combining surveys: model for parallel time series Multilevel modeling + poststratication Poststratication cells: sex ethnicity age education state
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
State-level opinion trends
Goal: estimating time series within each state One poll at a time: small-area estimation It works! Validated for pre-election polls Combining surveys: model for parallel time series Multilevel modeling + poststratication Poststratication cells: sex ethnicity age education state
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
State-level opinion trends
Goal: estimating time series within each state One poll at a time: small-area estimation It works! Validated for pre-election polls Combining surveys: model for parallel time series Multilevel modeling + poststratication Poststratication cells: sex ethnicity age education state
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Multilevel modeling of opinionsLogistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Bayesian inference, summarize by posterior simulations of : Simulation 1 75 1 ** ** . . . .. . . . . . . . 1000 ** **
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Multilevel modeling of opinionsLogistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Bayesian inference, summarize by posterior simulations of : Simulation 1 75 1 ** ** . . . .. . . . . . . . 1000 ** **
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Multilevel modeling of opinionsLogistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Bayesian inference, summarize by posterior simulations of : Simulation 1 75 1 ** ** . . . .. . . . . . . . 1000 ** **
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Multilevel modeling of opinionsLogistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Bayesian inference, summarize by posterior simulations of : Simulation 1 75 1 ** ** . . . .. . . . . . . . 1000 ** **
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Multilevel modeling of opinionsLogistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Bayesian inference, summarize by posterior simulations of : Simulation 1 75 1 ** ** . . . .. . . . . . . . 1000 ** **
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Multilevel modeling of opinionsLogistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Bayesian inference, summarize by posterior simulations of : Simulation 1 75 1 ** ** . . . .. . . . . . . . 1000 ** **
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Interlude: why multilevel = hierarchical
Logistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Crossed (nonnested) structure of age, education, state Several overlapping hierarchies
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Interlude: why multilevel = hierarchical
Logistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Crossed (nonnested) structure of age, education, state Several overlapping hierarchies
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Interlude: why multilevel = hierarchical
Logistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Crossed (nonnested) structure of age, education, state Several overlapping hierarchies
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Interlude: why multilevel = hierarchical
Logistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Crossed (nonnested) structure of age, education, state Several overlapping hierarchies
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Interlude: why multilevel = hierarchical
Logistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Crossed (nonnested) structure of age, education, state Several overlapping hierarchies
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Interlude: why multilevel = hierarchical
Logistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Crossed (nonnested) structure of age, education, state Several overlapping hierarchies
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Interlude: why multilevel = hierarchical
Logistic regression: Pr(yi = 1) = logit1 ((X )i ) X includes demographic and geographic predictors Group-level model for the 16 age education predictors Group-level model for the 50 state predictors Crossed (nonnested) structure of age, education, state Several overlapping hierarchies
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Poststratication to estimate state opinions
Implied inference for j = logit1 (X ) in each of 3264 cells j (e.g., black female, age 1829, college graduate, Massachusetts) PoststraticationWithin each state s, average over 64 cells: js Nj j js Nj Nj = population in cell j (from Census) 1000 simulation draws propagate to uncertainty for each j
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Poststratication to estimate state opinions
Implied inference for j = logit1 (X ) in each of 3264 cells j (e.g., black female, age 1829, college graduate, Massachusetts) PoststraticationWithin each state s, average over 64 cells: js Nj j js Nj Nj = population in cell j (from Census) 1000 simulation draws propagate to uncertainty for each j
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Poststratication to estimate state opinions
Implied inference for j = logit1 (X ) in each of 3264 cells j (e.g., black female, age 1829, college graduate, Massachusetts) PoststraticationWithin each state s, average over 64 cells: js Nj j js Nj Nj = population in cell j (from Census) 1000 simulation draws propagate to uncertainty for each j
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Poststratication to estimate state opinions
Implied inference for j = logit1 (X ) in each of 3264 cells j (e.g., black female, age 1829, college graduate, Massachusetts) PoststraticationWithin each state s, average over 64 cells: js Nj j js Nj Nj = population in cell j (from Census) 1000 simulation draws propagate to uncertainty for each j
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Poststratication to estimate state opinions
Implied inference for j = logit1 (X ) in each of 3264 cells j (e.g., black female, age 1829, college graduate, Massachusetts) PoststraticationWithin each state s, average over 64 cells: js Nj j js Nj Nj = population in cell j (from Census) 1000 simulation draws propagate to uncertainty for each j
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Poststratication to estimate state opinions
Implied inference for j = logit1 (X ) in each of 3264 cells j (e.g., black female, age 1829, college graduate, Massachusetts) PoststraticationWithin each state s, average over 64 cells: js Nj j js Nj Nj = population in cell j (from Census) 1000 simulation draws propagate to uncertainty for each j
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
CBS/New York Times pre-election polls from 1988
Validation study: t model on poll data and compare to election results Competing estimates:No pooling: separate estimate within each state Complete pooling: no state predictors Hierarchical model and poststratify
Mean absolute state errors:No pooling: 10.4% Complete pooling: 5.4% Hierarchical model with poststratication: 4.5%
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
CBS/New York Times pre-election polls from 1988
Validation study: t model on poll data and compare to election results Competing estimates:No pooling: separate estimate within each state Complete pooling: no state predictors Hierarchical model and poststratify
Mean absolute state errors:No pooling: 10.4% Complete pooling: 5.4% Hierarchical model with poststratication: 4.5%
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
CBS/New York Times pre-election polls from 1988
Validation study: t model on poll data and compare to election results Competing estimates:No pooling: separate estimate within each state Complete pooling: no state predictors Hierarchical model and poststratify
Mean absolute state errors:No pooling: 10.4% Complete pooling: 5.4% Hierarchical model with poststratication: 4.5%
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
CBS/New York Times pre-election polls from 1988
Validation study: t model on poll data and compare to election results Competing estimates:No pooling: separate estimate within each state Complete pooling: no state predictors Hierarchical model and poststratify
Mean absolute state errors:No pooling: 10.4% Complete pooling: 5.4% Hierarchical model with poststratication: 4.5%
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
CBS/New York Times pre-election polls from 1988
Validation study: t model on poll data and compare to election results Competing estimates:No pooling: separate estimate within each state Complete pooling: no state predictors Hierarchical model and poststratify
Mean absolute state errors:No pooling: 10.4% Complete pooling: 5.4% Hierarchical model with poststratication: 4.5%
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
CBS/New York Times pre-election polls from 1988
Validation study: t model on poll data and compare to election results Competing estimates:No pooling: separate estimate within each state Complete pooling: no state predictors Hierarchical model and poststratify
Mean absolute state errors:No pooling: 10.4% Complete pooling: 5.4% Hierarchical model with poststratication: 4.5%
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
CBS/New York Times pre-election polls from 1988
Validation study: t model on poll data and compare to election results Competing estimates:No pooling: separate estimate within each state Complete pooling: no state predictors Hierarchical model and poststratify
Mean absolute state errors:No pooling: 10.4% Complete pooling: 5.4% Hierarchical model with poststratication: 4.5%
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
CBS/New York Times pre-election polls from 1988
Validation study: t model on poll data and compare to election results Competing estimates:No pooling: separate estimate within each state Complete pooling: no state predictors Hierarchical model and poststratify
Mean absolute state errors:No pooling: 10.4% Complete pooling: 5.4% Hierarchical model with poststratication: 4.5%
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
CBS/New York Times pre-election polls from 1988
Validation study: t model on poll data and compare to election results Competing estimates:No pooling: separate estimate within each state Complete pooling: no state predictors Hierarchical model and poststratify
Mean absolute state errors:No pooling: 10.4% Complete pooling: 5.4% Hierarchical model with poststratication: 4.5%
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
CBS/New York Times pre-election polls from 1988
Validation study: t model on poll data and compare to election results Competing estimates:No pooling: separate estimate within each state Complete pooling: no state predictors Hierarchical model and poststratify
Mean absolute state errors:No pooling: 10.4% Complete pooling: 5.4% Hierarchical model with poststratication: 4.5%
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
State-level opinions from national polls Poststratication Validation
Validation study: comparison of state errors1988 election outcome vs. poll estimateno pooling of state effects Actual election outcome 0.2 0.4 0.6 0.8 1.0 complete pooling (no state effects) Actual election outcome 0.2 0.4 0.6 0.8 1.0 Actual election outcome 0.2 0.4 0.6 0.8 1.0 multilevel model
0.0
0.0
0.0
0.2 0.4 0.6 0.8 1.0 Estimated Bush support
0.0
0.2 0.4 0.6 0.8 1.0 Estimated Bush support
0.0 0.0
0.2 0.4 0.6 0.8 1.0 Estimated Bush support
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel models always
Anything worth doing is worth doing repeatedly A method is any procedure applied more than once City planningOutward expansion: tting a model to other countries, other years, other outcomes, . . . Inlling: inferences for individual states, demographic subgroups, components of data, . . .
Frequentist statistical theory of repeated inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel models always
Anything worth doing is worth doing repeatedly A method is any procedure applied more than once City planningOutward expansion: tting a model to other countries, other years, other outcomes, . . . Inlling: inferences for individual states, demographic subgroups, components of data, . . .
Frequentist statistical theory of repeated inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel models always
Anything worth doing is worth doing repeatedly A method is any procedure applied more than once City planningOutward expansion: tting a model to other countries, other years, other outcomes, . . . Inlling: inferences for individual states, demographic subgroups, components of data, . . .
Frequentist statistical theory of repeated inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel models always
Anything worth doing is worth doing repeatedly A method is any procedure applied more than once City planningOutward expansion: tting a model to other countries, other years, other outcomes, . . . Inlling: inferences for individual states, demographic subgroups, components of data, . . .
Frequentist statistical theory of repeated inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel models always
Anything worth doing is worth doing repeatedly A method is any procedure applied more than once City planningOutward expansion: tting a model to other countries, other years, other outcomes, . . . Inlling: inferences for individual states, demographic subgroups, components of data, . . .
Frequentist statistical theory of repeated inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel models always
Anything worth doing is worth doing repeatedly A method is any procedure applied more than once City planningOutward expansion: tting a model to other countries, other years, other outcomes, . . . Inlling: inferences for individual states, demographic subgroups, components of data, . . .
Frequentist statistical theory of repeated inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Incumbency advantage in U.S. House elections
Regression approach (Gelman and King, 1990):For any year, compare districts with and without incs running Control for vote in previous election Control for incumbent party vit = 0 + 1 vi,t1 + 2 Pit + Iit + it
Other estimates (sophomore surge, etc.) have selection bias
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Incumbency advantage in U.S. House elections
Regression approach (Gelman and King, 1990):For any year, compare districts with and without incs running Control for vote in previous election Control for incumbent party vit = 0 + 1 vi,t1 + 2 Pit + Iit + it
Other estimates (sophomore surge, etc.) have selection bias
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Incumbency advantage in U.S. House elections
Regression approach (Gelman and King, 1990):For any year, compare districts with and without incs running Control for vote in previous election Control for incumbent party vit = 0 + 1 vi,t1 + 2 Pit + Iit + it
Other estimates (sophomore surge, etc.) have selection bias
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Incumbency advantage in U.S. House elections
Regression approach (Gelman and King, 1990):For any year, compare districts with and without incs running Control for vote in previous election Control for incumbent party vit = 0 + 1 vi,t1 + 2 Pit + Iit + it
Other estimates (sophomore surge, etc.) have selection bias
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Incumbency advantage in U.S. House elections
Regression approach (Gelman and King, 1990):For any year, compare districts with and without incs running Control for vote in previous election Control for incumbent party vit = 0 + 1 vi,t1 + 2 Pit + Iit + it
Other estimates (sophomore surge, etc.) have selection bias
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Incumbency advantage in U.S. House elections
Regression approach (Gelman and King, 1990):For any year, compare districts with and without incs running Control for vote in previous election Control for incumbent party vit = 0 + 1 vi,t1 + 2 Pit + Iit + it
Other estimates (sophomore surge, etc.) have selection bias
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Estimated incumbency advantage from lagged regressionsEst inc adv from lagged regression 0.0 0.05 0.10 0.15 1900
1920
1940 1960 Year
1980
2000
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Can we do better?
Regression estimate: vit = 0 + 1 vi,t1 + 2 Pit + Iit +
it
Political science problem: is assumed to be same in all districts Statistics problem: the model doesnt t the data Well show pictures of the model not tting Well set up a model allowing inc advantage to vary
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Can we do better?
Regression estimate: vit = 0 + 1 vi,t1 + 2 Pit + Iit +
it
Political science problem: is assumed to be same in all districts Statistics problem: the model doesnt t the data Well show pictures of the model not tting Well set up a model allowing inc advantage to vary
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Can we do better?
Regression estimate: vit = 0 + 1 vi,t1 + 2 Pit + Iit +
it
Political science problem: is assumed to be same in all districts Statistics problem: the model doesnt t the data Well show pictures of the model not tting Well set up a model allowing inc advantage to vary
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Can we do better?
Regression estimate: vit = 0 + 1 vi,t1 + 2 Pit + Iit +
it
Political science problem: is assumed to be same in all districts Statistics problem: the model doesnt t the data Well show pictures of the model not tting Well set up a model allowing inc advantage to vary
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Can we do better?
Regression estimate: vit = 0 + 1 vi,t1 + 2 Pit + Iit +
it
Political science problem: is assumed to be same in all districts Statistics problem: the model doesnt t the data Well show pictures of the model not tting Well set up a model allowing inc advantage to vary
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Model mistUnder the model, parallel lines are tted to the circles (open seats) and dots (incs running for reelection)Democratic vote in 1988 0.2 0.4 0.6 0.8 1.0
o o o o oo oo o o o o oo o o
o oo o o
o
Coefficients for lagged vote 0.2 0.4 0.6 0.8 1.0 1.2
0.0
incumbents running open seats
0.0
0.2 0.4 0.6 0.8 1.0 Democratic vote in 1986
1900
1920
1940 1960 Year
1980
2000
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel model
for t = 1, 2: vit = 0.5 + t + i + it Iit +
it
2 1 is the national vote swing i is the normal vote for district i: mean 0, sd . it is the inc advantage in district i at time t: mean , sd it s are independent errors: mean 0 and sd .
Candidate-level incumbency eects:If the same incumbent is running in years 1 and 2, then i2 i1 Otherwise, i1 and i2 are independent
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel model
for t = 1, 2: vit = 0.5 + t + i + it Iit +
it
2 1 is the national vote swing i is the normal vote for district i: mean 0, sd . it is the inc advantage in district i at time t: mean , sd it s are independent errors: mean 0 and sd .
Candidate-level incumbency eects:If the same incumbent is running in years 1 and 2, then i2 i1 Otherwise, i1 and i2 are independent
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel model
for t = 1, 2: vit = 0.5 + t + i + it Iit +
it
2 1 is the national vote swing i is the normal vote for district i: mean 0, sd . it is the inc advantage in district i at time t: mean , sd it s are independent errors: mean 0 and sd .
Candidate-level incumbency eects:If the same incumbent is running in years 1 and 2, then i2 i1 Otherwise, i1 and i2 are independent
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel model
for t = 1, 2: vit = 0.5 + t + i + it Iit +
it
2 1 is the national vote swing i is the normal vote for district i: mean 0, sd . it is the inc advantage in district i at time t: mean , sd it s are independent errors: mean 0 and sd .
Candidate-level incumbency eects:If the same incumbent is running in years 1 and 2, then i2 i1 Otherwise, i1 and i2 are independent
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel model
for t = 1, 2: vit = 0.5 + t + i + it Iit +
it
2 1 is the national vote swing i is the normal vote for district i: mean 0, sd . it is the inc advantage in district i at time t: mean , sd it s are independent errors: mean 0 and sd .
Candidate-level incumbency eects:If the same incumbent is running in years 1 and 2, then i2 i1 Otherwise, i1 and i2 are independent
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel model
for t = 1, 2: vit = 0.5 + t + i + it Iit +
it
2 1 is the national vote swing i is the normal vote for district i: mean 0, sd . it is the inc advantage in district i at time t: mean , sd it s are independent errors: mean 0 and sd .
Candidate-level incumbency eects:If the same incumbent is running in years 1 and 2, then i2 i1 Otherwise, i1 and i2 are independent
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel model
for t = 1, 2: vit = 0.5 + t + i + it Iit +
it
2 1 is the national vote swing i is the normal vote for district i: mean 0, sd . it is the inc advantage in district i at time t: mean , sd it s are independent errors: mean 0 and sd .
Candidate-level incumbency eects:If the same incumbent is running in years 1 and 2, then i2 i1 Otherwise, i1 and i2 are independent
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Multilevel model
for t = 1, 2: vit = 0.5 + t + i + it Iit +
it
2 1 is the national vote swing i is the normal vote for district i: mean 0, sd . it is the inc advantage in district i at time t: mean , sd it s are independent errors: mean 0 and sd .
Candidate-level incumbency eects:If the same incumbent is running in years 1 and 2, then i2 i1 Otherwise, i1 and i2 are independent
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Fitting the multilevel modelBayesian inference Linear parameters: national vote swings, district eects, incumbency eects 3 variance parameters: district eects, incumbency eects, residual errors Need to model a selection eect: information provided by the incumbent party at time 1 Solve analytically for Pr(inclusion), include factor in the likelihood Gibbs-Metropolis sampling, program in Splus
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Fitting the multilevel modelBayesian inference Linear parameters: national vote swings, district eects, incumbency eects 3 variance parameters: district eects, incumbency eects, residual errors Need to model a selection eect: information provided by the incumbent party at time 1 Solve analytically for Pr(inclusion), include factor in the likelihood Gibbs-Metropolis sampling, program in Splus
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Fitting the multilevel modelBayesian inference Linear parameters: national vote swings, district eects, incumbency eects 3 variance parameters: district eects, incumbency eects, residual errors Need to model a selection eect: information provided by the incumbent party at time 1 Solve analytically for Pr(inclusion), include factor in the likelihood Gibbs-Metropolis sampling, program in Splus
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Fitting the multilevel modelBayesian inference Linear parameters: national vote swings, district eects, incumbency eects 3 variance parameters: district eects, incumbency eects, residual errors Need to model a selection eect: information provided by the incumbent party at time 1 Solve analytically for Pr(inclusion), include factor in the likelihood Gibbs-Metropolis sampling, program in Splus
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Fitting the multilevel modelBayesian inference Linear parameters: national vote swings, district eects, incumbency eects 3 variance parameters: district eects, incumbency eects, residual errors Need to model a selection eect: information provided by the incumbent party at time 1 Solve analytically for Pr(inclusion), include factor in the likelihood Gibbs-Metropolis sampling, program in Splus
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Fitting the multilevel modelBayesian inference Linear parameters: national vote swings, district eects, incumbency eects 3 variance parameters: district eects, incumbency eects, residual errors Need to model a selection eect: information provided by the incumbent party at time 1 Solve analytically for Pr(inclusion), include factor in the likelihood Gibbs-Metropolis sampling, program in Splus
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Estimated incumbency advantage and its variationAverage Incumbency Advantage 0.12 SD of District Effects 1900 1920 1940 1960 1980 2000 0.08 0.15 0.0 1900 0.05 0.10
0.0
0.04
1920
1940
1960
1980
2000
Year
Year
Residual SD of Election Results 1900 1920 1940 1960 1980 2000
SD of Incumbency Advantage
0.06
0.04
0.02
0.0
0.0 1900
0.02
0.04
0.06
1920
1940
1960
1980
2000
Year
Year
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Est inc adv from lagged regression 0.0 0.05 0.10 0.15 1900
Compare old and new estimates
1920
1940 1960 Year
1980
2000
Average Incumbency Advantage
0.0 1900
0.04
0.08
0.12
1920
1940
1960
1980
2000
Year
Andrew Gelman
Fitting and understanding multilevel models
ntage
0.06
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
No-interaction modelBefore-after data with treatment and control groups Default model: constant treatment eectsFishers classical null hyp: eect is zero for all cases Regression model: yi = Ti + Xi + i"after" measurement, ytreatment
control
"before" measurement, x
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
No-interaction modelBefore-after data with treatment and control groups Default model: constant treatment eectsFishers classical null hyp: eect is zero for all cases Regression model: yi = Ti + Xi + i"after" measurement, ytreatment
control
"before" measurement, x
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
No-interaction modelBefore-after data with treatment and control groups Default model: constant treatment eectsFishers classical null hyp: eect is zero for all cases Regression model: yi = Ti + Xi + i"after" measurement, ytreatment
control
"before" measurement, x
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
No-interaction modelBefore-after data with treatment and control groups Default model: constant treatment eectsFishers classical null hyp: eect is zero for all cases Regression model: yi = Ti + Xi + i"after" measurement, ytreatment
control
"before" measurement, x
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Actual data show interactions
Treatment interacts with before measurement Before-after correlation is higher for controls than for treated units ExamplesAn observational study of legislative redistricting An experiment with pre-test, post-test data Congressional elections with incumbents and open seats
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Actual data show interactions
Treatment interacts with before measurement Before-after correlation is higher for controls than for treated units ExamplesAn observational study of legislative redistricting An experiment with pre-test, post-test data Congressional elections with incumbents and open seats
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Actual data show interactions
Treatment interacts with before measurement Before-after correlation is higher for controls than for treated units ExamplesAn observational study of legislative redistricting An experiment with pre-test, post-test data Congressional elections with incumbents and open seats
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Actual data show interactions
Treatment interacts with before measurement Before-after correlation is higher for controls than for treated units ExamplesAn observational study of legislative redistricting An experiment with pre-test, post-test data Congressional elections with incumbents and open seats
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Actual data show interactions
Treatment interacts with before measurement Before-after correlation is higher for controls than for treated units ExamplesAn observational study of legislative redistricting An experiment with pre-test, post-test data Congressional elections with incumbents and open seats
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Actual data show interactions
Treatment interacts with before measurement Before-after correlation is higher for controls than for treated units ExamplesAn observational study of legislative redistricting An experiment with pre-test, post-test data Congressional elections with incumbents and open seats
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Observational study of legislative redistricting before-after data(favors Democrats)
.. . . .. . ... . . . .. . . .. . .. . . .. .. .. .. . .. . ... . o .. . o . . .. . . . .x .o... .. .. .o . . o . . .x. . o x ... x . .... . . . . . . . .o . x . x . x . . . x .. . ... x . . . . . . . . . .x . .
Estimated partisan bias (adjusted for state) -0.05 0.0 0.05
no redistricting
Dem. redistrict bipartisan redistrict Rep. redistrict
(favors Republicans)
-0.05
0.0
0.05
Estimated partisan bias in previous election
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Experiment: correlation between pre-test and post-test data for controls and for treated units1.0 controls
0.8
correlation 0.9
treated 1 2 grade 3 4
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Correlation between two successive Congressional elections for incumbents running (controls) and open seats (treated)0.8 incumbents
correlation 0.4 0.6
0.0
0.2
open seats
1900
1920
1940 year
1960
1980
2000
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Interactions as variance components
Unit-level error term i For control units, i persists from time 1 to time 2 For treatment units, i changes:Subtractive treatment error (i only at time 1) Additive treatment error (i only at time 2) Replacement treatment error
Under all these models, the before-after correlation is higher for controls than treated units
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Interactions as variance components
Unit-level error term i For control units, i persists from time 1 to time 2 For treatment units, i changes:Subtractive treatment error (i only at time 1) Additive treatment error (i only at time 2) Replacement treatment error
Under all these models, the before-after correlation is higher for controls than treated units
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Interactions as variance components
Unit-level error term i For control units, i persists from time 1 to time 2 For treatment units, i changes:Subtractive treatment error (i only at time 1) Additive treatment error (i only at time 2) Replacement treatment error
Under all these models, the before-after correlation is higher for controls than treated units
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Interactions as variance components
Unit-level error term i For control units, i persists from time 1 to time 2 For treatment units, i changes:Subtractive treatment error (i only at time 1) Additive treatment error (i only at time 2) Replacement treatment error
Under all these models, the before-after correlation is higher for controls than treated units
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Interactions as variance components
Unit-level error term i For control units, i persists from time 1 to time 2 For treatment units, i changes:Subtractive treatment error (i only at time 1) Additive treatment error (i only at time 2) Replacement treatment error
Under all these models, the before-after correlation is higher for controls than treated units
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Interactions as variance components
Unit-level error term i For control units, i persists from time 1 to time 2 For treatment units, i changes:Subtractive treatment error (i only at time 1) Additive treatment error (i only at time 2) Replacement treatment error
Under all these models, the before-after correlation is higher for controls than treated units
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
General framework Estimating incumbency advantage and its variation Interactions in before-after studies
Interactions as variance components
Unit-level error term i For control units, i persists from time 1 to time 2 For treatment units, i changes:Subtractive treatment error (i only at time 1) Additive treatment error (i only at time 2) Replacement treatment error
Under all these models, the before-after correlation is higher for controls than treated units
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Some new tools
Building and tting multilevel models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Some new tools
Building and tting multilevel models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Some new tools
Building and tting multilevel models Displaying and summarizing inferences
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Building and tting multilevel models
A reparameterization can change a model (even if it leaves the likelihood unchanged) Redundant additive parameterization Redundant multiplicative parameterization Weakly-informative prior distribution for group-level variance parameters
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Building and tting multilevel models
A reparameterization can change a model (even if it leaves the likelihood unchanged) Redundant additive parameterization Redundant multiplicative parameterization Weakly-informative prior distribution for group-level variance parameters
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Building and tting multilevel models
A reparameterization can change a model (even if it leaves the likelihood unchanged) Redundant additive parameterization Redundant multiplicative parameterization Weakly-informative prior distribution for group-level variance parameters
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Building and tting multilevel models
A reparameterization can change a model (even if it leaves the likelihood unchanged) Redundant additive parameterization Redundant multiplicative parameterization Weakly-informative prior distribution for group-level variance parameters
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Building and tting multilevel models
A reparameterization can change a model (even if it leaves the likelihood unchanged) Redundant additive parameterization Redundant multiplicative parameterization Weakly-informative prior distribution for group-level variance parameters
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant parameterizationage state Data model: Pr(yi = 1) = logit1 0 + age(i) + state(i)
Usual model for the coecients:2 jage N(0, age ), for j = 1, . . . , 4 2 jstate N(0, state ), for j = 1, . . . , 50
Additively redundant model:2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Why add the redundant age , state ?Iterative algorithm moves more smoothlyAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant parameterizationage state Data model: Pr(yi = 1) = logit1 0 + age(i) + state(i)
Usual model for the coecients:2 jage N(0, age ), for j = 1, . . . , 4 2 jstate N(0, state ), for j = 1, . . . , 50
Additively redundant model:2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Why add the redundant age , state ?Iterative algorithm moves more smoothlyAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant parameterizationage state Data model: Pr(yi = 1) = logit1 0 + age(i) + state(i)
Usual model for the coecients:2 jage N(0, age ), for j = 1, . . . , 4 2 jstate N(0, state ), for j = 1, . . . , 50
Additively redundant model:2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Why add the redundant age , state ?Iterative algorithm moves more smoothlyAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant parameterizationage state Data model: Pr(yi = 1) = logit1 0 + age(i) + state(i)
Usual model for the coecients:2 jage N(0, age ), for j = 1, . . . , 4 2 jstate N(0, state ), for j = 1, . . . , 50
Additively redundant model:2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Why add the redundant age , state ?Iterative algorithm moves more smoothlyAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant parameterizationage state Data model: Pr(yi = 1) = logit1 0 + age(i) + state(i)
Usual model for the coecients:2 jage N(0, age ), for j = 1, . . . , 4 2 jstate N(0, state ), for j = 1, . . . , 50
Additively redundant model:2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Why add the redundant age , state ?Iterative algorithm moves more smoothlyAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant parameterizationage state Data model: Pr(yi = 1) = logit1 0 + age(i) + state(i)
Usual model for the coecients:2 jage N(0, age ), for j = 1, . . . , 4 2 jstate N(0, state ), for j = 1, . . . , 50
Additively redundant model:2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Why add the redundant age , state ?Iterative algorithm moves more smoothlyAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant additive parameterizationModelage state Pr(yi = 1) = logit1 0 + age(i) + state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Identify using centered parameters: jage = jage age , for j = 1, . . . , 4 jstate = jstate state , for j = 1, . . . , 50 Redene the constant term: 0 = 0 + age + age
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant additive parameterizationModelage state Pr(yi = 1) = logit1 0 + age(i) + state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Identify using centered parameters: jage = jage age , for j = 1, . . . , 4 jstate = jstate state , for j = 1, . . . , 50 Redene the constant term: 0 = 0 + age + age
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant additive parameterizationModelage state Pr(yi = 1) = logit1 0 + age(i) + state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Identify using centered parameters: jage = jage age , for j = 1, . . . , 4 jstate = jstate state , for j = 1, . . . , 50 Redene the constant term: 0 = 0 + age + age
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant multiplicative parameterizationNew modelage state Pr(yi = 1) = logit1 0 + age age(i) + state state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Identify using centered and scaled parameters: jage = age (jage age ), for j = 1, . . . , 4 jstate = state jstate state , for j = 1, . . . , 50 Faster convergence More general model, connections to factor analysisAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant multiplicative parameterizationNew modelage state Pr(yi = 1) = logit1 0 + age age(i) + state state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Identify using centered and scaled parameters: jage = age (jage age ), for j = 1, . . . , 4 jstate = state jstate state , for j = 1, . . . , 50 Faster convergence More general model, connections to factor analysisAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant multiplicative parameterizationNew modelage state Pr(yi = 1) = logit1 0 + age age(i) + state state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Identify using centered and scaled parameters: jage = age (jage age ), for j = 1, . . . , 4 jstate = state jstate state , for j = 1, . . . , 50 Faster convergence More general model, connections to factor analysisAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Redundant multiplicative parameterizationNew modelage state Pr(yi = 1) = logit1 0 + age age(i) + state state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Identify using centered and scaled parameters: jage = age (jage age ), for j = 1, . . . , 4 jstate = state jstate state , for j = 1, . . . , 50 Faster convergence More general model, connections to factor analysisAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Weakly informative prior distribution for the multilevel variance parameterRedundant multiplicative parameterization:age state Pr(yi = 1) = logit1 0 + age age(i) + state state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Separate prior distributions on the and parameters:Normal on Inverse-gamma on 2
Generalizes and xes problems with the standard choices of prior distributionsAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Weakly informative prior distribution for the multilevel variance parameterRedundant multiplicative parameterization:age state Pr(yi = 1) = logit1 0 + age age(i) + state state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Separate prior distributions on the and parameters:Normal on Inverse-gamma on 2
Generalizes and xes problems with the standard choices of prior distributionsAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Weakly informative prior distribution for the multilevel variance parameterRedundant multiplicative parameterization:age state Pr(yi = 1) = logit1 0 + age age(i) + state state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Separate prior distributions on the and parameters:Normal on Inverse-gamma on 2
Generalizes and xes problems with the standard choices of prior distributionsAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Weakly informative prior distribution for the multilevel variance parameterRedundant multiplicative parameterization:age state Pr(yi = 1) = logit1 0 + age age(i) + state state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Separate prior distributions on the and parameters:Normal on Inverse-gamma on 2
Generalizes and xes problems with the standard choices of prior distributionsAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Weakly informative prior distribution for the multilevel variance parameterRedundant multiplicative parameterization:age state Pr(yi = 1) = logit1 0 + age age(i) + state state(i) 2 jage N(age , age ), for j = 1, . . . , 4 2 jstate N(state , state ), for j = 1, . . . , 50
Separate prior distributions on the and parameters:Normal on Inverse-gamma on 2
Generalizes and xes problems with the standard choices of prior distributionsAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Displaying and summarizing inferences
Displaying parameters in groups rather than as a long list Average predictive eects R 2 and partial pooling factors Analysis of variance
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Displaying and summarizing inferences
Displaying parameters in groups rather than as a long list Average predictive eects R 2 and partial pooling factors Analysis of variance
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Displaying and summarizing inferences
Displaying parameters in groups rather than as a long list Average predictive eects R 2 and partial pooling factors Analysis of variance
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Displaying and summarizing inferences
Displaying parameters in groups rather than as a long list Average predictive eects R 2 and partial pooling factors Analysis of variance
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Displaying and summarizing inferences
Displaying parameters in groups rather than as a long list Average predictive eects R 2 and partial pooling factors Analysis of variance
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Raw display of inferenceB.0 b.female b.black b.female.black B.age[1] B.age[2] B.age[3] B.age[4] B.edu[1] B.edu[2] B.edu[3] B.edu[4] B.age.edu[1,1] B.age.edu[1,2] B.age.edu[1,3] B.age.edu[1,4] B.age.edu[2,1] B.age.edu[2,2] B.age.edu[2,3] B.age.edu[2,4] B.age.edu[3,1] B.age.edu[3,2] B.age.edu[3,3] B.age.edu[3,4] B.age.edu[4,1] B.age.edu[4,2] B.age.edu[4,3] B.age.edu[4,4] B.state[1] B.state[2] mean 0.402 -0.094 -1.701 -0.143 0.084 -0.072 0.044 -0.057 -0.148 -0.022 0.148 0.023 -0.044 0.059 0.049 0.001 0.066 -0.081 -0.004 -0.042 0.034 0.007 0.033 -0.009 -0.141 -0.014 0.046 0.042 0.201 0.466 sd 2.5% 25% 0.147 0.044 0.326 0.102 -0.283 -0.162 0.305 -2.323 -1.910 0.393 -0.834 -0.383 0.088 -0.053 0.012 0.087 -0.260 -0.121 0.077 -0.105 -0.007 0.096 -0.265 -0.115 0.131 -0.436 -0.241 0.082 -0.182 -0.069 0.112 -0.032 0.065 0.090 -0.170 -0.030 0.133 -0.363 -0.104 0.123 -0.153 -0.011 0.124 -0.146 -0.023 0.116 -0.237 -0.061 0.152 -0.208 -0.008 0.127 -0.407 -0.137 0.102 -0.226 -0.048 0.108 -0.282 -0.100 0.135 -0.215 -0.030 0.102 -0.213 -0.039 0.130 -0.215 -0.029 0.109 -0.236 -0.064 0.190 -0.672 -0.224 0.119 -0.280 -0.059 0.132 -0.192 -0.024 0.142 -0.193 -0.022 0.211 -0.131 0.047 0.252 Andrew Gelman 0.008 0.310 50% 75% 97.5% Rhat n.eff 0.413 0.499 0.652 1.024 110 -0.095 -0.034 0.107 1.001 1000 -1.691 -1.486 -1.152 1.014 500 -0.155 0.104 0.620 1.007 1000 0.075 0.140 0.277 1.062 45 -0.054 -0.006 0.052 1.017 190 0.038 0.095 0.203 1.029 130 -0.052 0.001 0.133 1.076 32 -0.137 -0.044 0.053 1.074 31 -0.021 0.025 0.152 1.028 160 0.142 0.228 0.370 1.049 45 0.015 0.074 0.224 1.061 37 -0.019 0.025 0.170 1.018 1000 0.032 0.118 0.353 1.016 580 0.022 0.104 0.349 1.015 280 0.000 0.052 0.280 1.010 1000 0.032 0.124 0.449 1.022 140 -0.055 0.001 0.094 1.057 120 0.000 0.041 0.215 1.008 940 -0.026 0.014 0.157 1.017 170 0.009 0.091 0.361 1.021 230 0.003 0.052 0.220 1.019 610 0.009 0.076 0.410 1.080 61 -0.005 0.043 0.214 1.024 150 -0.086 -0.003 0.100 1.036 270 -0.008 0.033 0.239 1.017 240 0.019 0.108 0.332 1.010 210 0.016 0.095 0.377 1.015 160 0.172 0.326 0.646 1.003 960 Fitting and 0.603 understanding multilevel models 0.440 1.047 1.001 1000
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Raw graphical displayBugs model at "C:/books/multilevel/election88/model4.bug", 3 chains, each with 2001 iterations 80% interval for each chain 4 B.0 b.female b.black b.female.black B.age[1] [2] [3] [4] B.edu[1] [2] [3] [4] * B.age.edu[1,1] [1,2] [1,3] [1,4] [2,1] [2,2] [2,3] [2,4] [3,1] [3,2] * B.state[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] B.region[1] [2] [3] [4] [5] 2 0 2 1q
Rhat 1.5 2+ B.00.6 0.4 0.2 0 0.1 0 0.1 0.2 0.3 1 1.5 2 2.5 0.5 1
medians and 80% intervalsq q
q
q
b.female
q q
q
q q q q
b.black
q q
q q q q
0 b.female.black0.5 0.4 0.2 0 0.2 0.4 0.5 0 0.5 0.5 0 0.5 1 2 0 2 4 1 0.5 0 0.5 1 0.6 0.4 0.2 0 1 0.5 0
q
B.age
q q
q q q q q
q q q q q q q q q q
1 2 3 4q q q
B.edu
q q q
1 2 3 4q q q q q q q q q q q q q q q q
B.age.edu
3 1 2 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q
* B.state
q q q q q q q q q q
1 2 3 4 5 6 7 8 910 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40
B.region
q q
q
q q q
1 2 3 4 5
Sigma.age
q q q
q q q q
Sigma.edu
q q q
Andrew Gelmanq
0.3 Sigma.age.edu0.2 0.1 Fitting and 0
understanding multilevel models
q q q
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Better graphical display 1: demographics2.5female black female x black 1829 3044 4564 65+ no h.s. high school some college college grad 1829 x no h.s. 1829 x high school 1829 x some college 1829 x college grad 3044 x no h.s. 3044 x high school 3044 x some college 3044 x college grad 4564 x no h.s. 4564 x high school 4564 x some college 4564 x college grad 65+ x no h.s. 65+ x high school 65+ x some college 65+ x college grad
2
1.5
1
0.5
0
0.5
1
2.5
2
1.5
1
0.5
0
0.5
1
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Better graphical display 2: within statesAlaska Pr (support Bush) 0.0 0.4 0.8 Pr (support Bush) 0.0 0.4 0.8 Arizona Pr (support Bush) 0.0 0.4 0.8 Arkansas Pr (support Bush) 0.0 0.4 0.8 California
2.0 1.0 0.0 linear predictor Colorado
2.0 1.0 0.0 linear predictor Connecticut
2.0 1.0 0.0 linear predictor Delaware
2.0 1.0 0.0 linear predictor District of Columbia
Pr (support Bush) 0.0 0.4 0.8
Pr (support Bush) 0.0 0.4 0.8
Pr (support Bush) 0.0 0.4 0.8
2.0 1.0 0.0 linear predictor
2.0 1.0 0.0 linear predictor
2.0 1.0 0.0 linear predictor
Pr (support Bush) 0.0 0.4 0.8
2.0 1.0 0.0 linear predictor
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Better graphical display 3: between statesNortheast0.5 0.5 regression intercept regression intercept
Midwest0.5 regression intercept
Southregression intercept TN AL VA MS NC SC OK LA GA KY TXFL AR 0.5
West
UT AZ WY NV AK ID WA CO HI CA MT NM OR 0.0 0.5 0.5
0.0
0.0
NE
0.5
0.5
0.5
0.6
0.7
0.5
0.6
0.7
0.5 0.5
NJ CT VT WV PAME DE RI MA NY MD
NH
MO MI WI IAIL MN
0.0
OH
KS IN ND SD
0.6
0.7
0.6
0.7
R vote in prev elections
R vote in prev elections
R vote in prev elections
R vote in prev elections
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Average predictive eectsWhat is E (y | x1 = high) E(y | x1 = low), with all other xs held constant? y
u
vAndrew Gelman Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Average predictive eects
What is E (y | x1 = high) E(y | x1 = low), with all other xs held constant? In general, dierence can depend on x Average over distribution of x in the dataYou cant just use a central value of x
Compute APE for each input variable x Multilevel factors are categorical input variables
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Average predictive eects
What is E (y | x1 = high) E(y | x1 = low), with all other xs held constant? In general, dierence can depend on x Average over distribution of x in the dataYou cant just use a central value of x
Compute APE for each input variable x Multilevel factors are categorical input variables
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Average predictive eects
What is E (y | x1 = high) E(y | x1 = low), with all other xs held constant? In general, dierence can depend on x Average over distribution of x in the dataYou cant just use a central value of x
Compute APE for each input variable x Multilevel factors are categorical input variables
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Average predictive eects
What is E (y | x1 = high) E(y | x1 = low), with all other xs held constant? In general, dierence can depend on x Average over distribution of x in the dataYou cant just use a central value of x
Compute APE for each input variable x Multilevel factors are categorical input variables
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Average predictive eects
What is E (y | x1 = high) E(y | x1 = low), with all other xs held constant? In general, dierence can depend on x Average over distribution of x in the dataYou cant just use a central value of x
Compute APE for each input variable x Multilevel factors are categorical input variables
Andrew Gelman
Fitting and understanding multilevel models
Eectiveness Ubiquity Way of life
Building and tting models Displaying and summarizing inferences Conclusions
Average predictive eects
What is E (y | x1 = high) E(y | x1 = low), with all other xs held constant? In general, dierence can depend on x Average o