Statistics 196A, Education 401D                                  Spring 2021

    Multilevel Modeling Using R

Spring 2021 Remote Asynchronous Instruction

Instructor.  David Rogosa (usually)Sequoia 224,   rag{AT}stanford{DOT}edu   
Course web page:

Course Welcome and Logistics (first day stuff, posted in February, call it Week0)
      Lecture slides, week 0 (pdf)             Audio companion, week 0   
For recreation of in-classroom experience, linked below are youtube versions of the music I play
before starting lecture   and    after lecture concludes.      Some may wish to reverse that ordering.

                To see full course materials from Spring 2020 go here

From explorecourses
  STATS 196A (EDUC 401D): Multilevel Modeling Using R
Multilevel data analysis examples using R. 
Topics include: two-level nested data, growth curve modeling, generalized linear models for counts and categorical data, 
nonlinear models, three-level analyses. For more information, see course website:
Terms: Spr | Units: 1 | Grading: Satisfactory/Unsatisfactory  Remote: Asyncrhonous
Instructors: Rogosa, D. (PI) 
STATS 196A | 1 units | Class # 21356 | Section 01 | Grading: Satisfactory/No Credit | WKS |  
 Course Schedule, Remote Asynchronous
Four (2hr) classes (formerly located in  Sequoia 200)
 Weeks 1 - 4. 
    a. Introduction: Basic analyses for two-level nested data, normal models  (UK Exam data)
    b. Additional two-level (normal) models: experimental designs (Dyestuff), longitudinal data (growth curves, sleepstudy), 
        observational data (High School and Beyond) 
    c. Generalized linear mixed models for counts and categorical outcomes 
    d. Three-level analyses (nested data and longitudinal data)    
    e. Specialized applications (as time permits):  regression diagnostics, power calculations and design, ecological inference, survival analysis, 
        nonlinear functional forms, mediation analysis, propensity scores and matching, imputation, item response theory
 Week 5. Substitute for live "Student presentations of multilevel data analyses" by online submission of materials 
For the 1-unit enrollment in this short course, students are expected to engage in (i.e. consume) the four presentation class sessions.
   For the fifth session, formerly each student made a short (~10 min) presentation of a relevant data analysis they have conducted.
   In Spring 2021, as it was in Spring 2020, students instead submit the small project online (such as using rpubs or google drive). Refer to Week0 audio.

Core Sources
Many of the example presented in this short course are described in Examples from Multilevel Software Comparative Reviews  Douglas Bates.
        Code version of MlmSoftRev     R-package containing mlmRev data examples.     Bates talk on mlmRev      U Bristol documentation
     Additional examples in core package lme4
Another set of examples: lmer for SAS PROC MIXED Users Douglas Bates Department of Statistics University of Wisconsin Madison
   Data sets from SAS System for Mixed Models
Overviews and additional examples from Doug Bates:
lme4: Mixed-effects modeling with R  February 17, 2010 Springer (book chapters).
A merged updated version of Bates book lme4: Mixed-effects modeling with R  May 2020
R Journal intro Fitting linear mixed models in R Using the lme4 package Douglas Bates (pp.27-30)
Collection of all Doug Bates lme4 talks
   lme4 vignette: Douglas Bates Martin Machler Ben Bolker.   Fitting linear mixed-effects models using lme4, Journal of Statistical Software
    Technical topics: Mixed models in R using the lme4 package Part 4: Theory of linear mixed models
HSB and growth curve examples in  John Fox lme tutorial
Another nice lmer exposition with life sciences examples: Mixed-effects models, Remko Duursma, Jeff Powell  Hawkesbury Institute for the Environment, Western Sydney University. September 2016.  HIE Datasets

Current version of R is R version 4.0.4 released on 2021-02-15. For references and software: The R Project for Statistical Computing
   Berkeley mirror is no longer, choose a mirror from the main R page (first link, I use TN).

A recent text (potentially) provides more infrastructure for this short course, but sadly it has many shortcomings.
This text has free access at Stanford via
     Multilevel Modeling Using R        Chapman & Hall/CRC Statistics in the Social and Behavioral Sciences Published:June 23, 2014 by CRC Press      Journal of Statistical Software Book Review      Book website, including data

               2021 Remote Asynchronous Content.
Updated from 2020 content each instructional week, starting Second Week of Spring Quarter

Week 1 April 7 2021
2021 Lecture slides, week 1 (pdf)
Audio companion, week 1
parta   partb   partc
1. Introductory Example. Nested data, two-levels. Goldstein Exam Data.
    Exam {mlmRev}  UK HS data.   Subset: Coed schools exam data, mixed schools, cleaned
    a. Introductory descriptive approaches for gender gap analysis (Smart First Year Student analyses using lmList, additional plots).
    b. Various lmer analyses for gender gap.
        Gender gap data analysis:    scanned class handout          Rogosa R-session     basic plots    models used         2020 isSingular fix           
     c. Residual plots, add-on regression diagnostics: packages   HLMdiaginfluence.ME
           Rogosa session with Exam data (week 1) (ascii)         resulting plots
     d. more P-values, tests add-ons to lmer.
       afex package with Exam data ggaplmer2
       Faraway text addendums: Inferential Methods for Linear Mixed Models
     e. Plus     Plots for random, fixed effects.

2. Matrix Formulation for Mixed Effects Models (growth curves and nested data).

Week 2 April 14 2021
2021 Lecture slides, week 2 (pdf)
Audio companion, week 2
parta   partb   partc   partd
1. Recap Introductory Example. Nested data, two-levels. Goldstein Exam Data.
     merMod objects from lmer
    Add-on package merTools     merTools vignette
prediction with lmer : predict with lme4   predictInterval from merTools  prediction vignette
Rogosa session.   plots

2. Common/canonical two-level examples (measured outcome)
A. Growth Curve models and analysis. Bates Sleepstudy example (week2 Stat222).
Chap. 4 Bates book          [more Doug Bates Slides (pdf pages 8-28) ]
     Sleepstudy class handout, pdf scan     Sleepstudy, 2018 clean ascii         Individual plots (frame-by-frame)       Plot of straight-line fits         
Reduced/constrained models: growth curve example
B. Two-level nested (normal) data recap.   Brief overview of HSB (High School and Beyond) analysis (from Stat209): plots and model.
      a full single Bryk dataset (longform)              (abbreviated) Rogosa R-session  Bryk data        plots, Rogosa R-session
Caution froma prior year: side-by-side boxplot creation and lmList subset issue
     A nice teaching document from Indiana that does HSB data with every known statistical package (including lmer)

3. Growth Curve Modeling exercise, Brain Volume Data Analysis. analyses from "Variation in longitudinal trajectories of regional brain volumes of healthy men and women (ages 10 to 85 years) measured with atlas-based parcellation of MRI"     cartoon plot of Lateral Ventricles data;     actual data plot of Lateral Ventricles data;    development of lmer (mixed effect) growth models

4. Data from designed experiments.(basics).
     a. Dyestuff data, Bates book, Chapter 1 (sec 1.2, 1.3)      Rogosa Dyestuff session
     b. Penicillin data (also Pastes, ratbrain), Bates book, Chapter 2.
         From Doug Bates presentation       Rogosa R-session
     Random effects anova recap (see Bates book Chap1,    Chap2).

Week 3 April 21 2021
2021 Lecture slides, week 3 (pdf)
Audio companion, week 3
parta   partb   

         Main topic:   Generalized Linear Mixed Models: counts and proportions.
    nice overview: Generalized Linear Mixed Models   from Encyclopedia of Statistics in Behavioral Science.   
    Github GLMM Faq (Ben Bolker)

1.        Dichotomous outcomes, glmer analysis examples.
    a. Respiratory clinical trial from HSAUR.     lmList does logistic, introducing glmer      lmList, glmer for respiration data (placebo group)
    b. Contraception (Bangladesh) use from Bates review         Rogosa R-session     glmer model slide
    c. Test scores (pass/fail outcome) from Ch 8, Multilevel Modeling Using R.    Rogosa R-session
2.     Count outcome, GLMM poisson models.
        a.   Count data: Contagious bovine pleuropneumonia, data(cbpp) in lme4.
          Rogosa R-session     herd plots
        b. Factorial design, Count outcome. From HIE Sydney. EucFACE ground cover data   Rogosa glmer session.
        c.  Another count data example from mlmRev package,
      data(Mmmec): Malignant Melanoma Mortality in the European Community associated with the impact of UV radiation exposure.
        Rogosa glmer session       more Rogosa Mmmec session
glmer.nb  Fits a generalized linear mixed-effects model (GLMM) for the negative binomial family, building on glmer, and initializing via from MASS.
3.    hglm -- a different package for fitting hierarchical generalized linear models.
R Journal December 2010.    manual    vignette

Addendum. A little more on Overdispersion in Generalized Linear Mixed Models
         pdf slides    companion audio     Rogosa R-session
Resources:  Basic Overdispersion Overview: Overdispersion, and how to deal with it in R
Overdispersion in Github GLMM Faq (Ben Bolker)
R-package RVAideMemoire    function overdisp.glmer  man page 1   2

Week 4 April 28 2021
2021 Lecture slides, week 4 (pdf)
Audio companion, week 4
parta   partb   
1. Three-level (and above) lmer examples.
Measured outcomes
  a. Achieve data from Multilevel Modeling Using R book.    Rogosa R-session
  b. example from mlmRev package, data(Chem97): Scores on A-level Chemistry in 1997.    Rogosa R-session
Count outcome.
  c. data(Mmmec): Malignant Melanoma Mortality in the European Community associated with the impact of UV radiation exposure.      Rogosa 3-level session

2. Missing Data and Imputation Methods for Multilevel Data Analysis
Vignette:  Analyzing Imputed Data with Multilevel Models and merTools
          Rogosa R-session for vignette
Missing data wide-form imputation: mice multiple regression example, nhanes data in package mice     R-session using mice package
Missing data background. Multiple Imputation. Nhanes data example (mice primer) in van Buuren S and Groothuis-Oudshoorn K (2011). mice: Multivariate Imputation by Chained Equations in R. Journal of Statistical Software, 45(3), 1-67.
See also    Flexible Imputation of Missing Data. Stef van Buuren Chapman and Hall/CRC 2012. Chapter 9, Longitudinal Data Sec 3.8 Multilevel data. He is the originator of mice      book extras
     R resources.  Multivariate Analysis Task View, Missing data section, esp packages mice
  New package: hmi: hierarchical multiple imputation,   vignette

Addendum.  Study Design (power) for Mixed Models
Most activity and resources are for longitudinal designs (c.f Week2, topic 2A): How often and How many subjects?
In my Stat222  course longitudinal experimental design is taken up in Week 5, Lecture topic 4.
Audio for the topic starts at partc that week, with the lecture content for design starting at pdf page 93.
Resources for that Stat222, Week5 content  Power Calculations for Longitudinal Group Comparsions.
   R-package longpower Vignettes found by "browseVignettes(package = "longpower")" .    Functions in MBESS package--ss.power.pcm.
   R-package: powerlmm
   Background pubs:  Power for linear models of longitudinal data with applications to Alzheimer's Disease Phase II study design Michael C. Donohue, Steven D. Edland, Anthony C. Gamst
Sample Size Planning for Longitudinal Models: Accuracy in Parameter Estimation for Polynomial Change Parameters Ken Kelley Notre Dame Joseph R. Rausch Psychological Methods 2011
Additional Resources:         basic R analogues, power.t.test   power.anova.test
Power calculations and Design packages;
   Package pamm Title Power Analysis for Random Effects in Mixed Models
   Package simr Title Power Analysis for Generalised Linear Mixed Models by Simulation.
    Package MultiRR Title Bias, Precision, and Power for Multi-Level Random Regression
more powerlmm
intro:   Introducing 'powerlmm' an R package for power calculations for longitudinal multilevel models
vignette:   Power Analysis for Two-level Longitudinal Models with Missing Data
more simulation approaches to mixed-model power calculations:
Power for Multilevel Analysis    Power Analysis in R for Multilevel Models    Simulation approaches for mixed-model power calculations