Online Program

This workshop will cover the use of Stata to perform meta-analysis (MA), a statistical technique for combining the results from several similar studies. The course will provide a brief introduction to MA and will demonstrate how to perform MA in Stata 16. Stata’s new meta command offers full support for MA—from computing various effect sizes and producing basic meta-analytic summary and forest plots to accounting for between-study heterogeneity and potential publication bias. A number of case studies demonstrating how to conduct an MA within Stata will be provided. These examples will focus on the interpretation of MA under various models, meta-regression and its postestimation features, subgroup analysis, small-study effect and publication bias, and various types of forest, funnel, and other plots. No prior knowledge of Stata is required, but basic familiarity with MA will prove useful.

Outline & Objectives

Outline
This workshop is geared toward researchers wanting to perform MA and those who already
know about MA and wish to learn how to do it using Stata.
1. Brief overview of MA
2. Data setup and effect sizes using meta set and meta esize
• Effect sizes for binary data
• Effect sizes for continuous data
• Generic (precomputed) effect sizes
3. MA models
• Random-effects model (seven estimation methods)
• Fixed-effects model (Mantel–Haenszel and inverse-variance methods)
• Common-effect model (Mantel–Haenszel and inverse-variance methods)
4. Graphical and numerical MA summary using meta summarize and meta forestplot
• Standard MA
• Subgroup MA with one or many grouping variables
• Cumulative MA with and without stratification
1
5. Meta-regression
• Continuous and categorical moderators
• Fixed-effects and random-effects regression
• Multiplicative and additive residual heterogeneity
• Knapp–Hartung standard-error adjustment
• Postestimation features: prediction, bubble plots, etc.
6. Small-study effects and publication bias
• Standard and contour-enhanced funnel plots
• Traditional and random-effects versions of tests for funnel-plot asymmetry or
small-study effects
• Nonparametric trim-and-fill method

Performance objectives
Participants of this workshop will walk away with the following knowledge:
• A brief overview of MA as a statistical procedure
• How to declare and compute effect sizes using meta set and meta esize
• How to summarize the meta-analytic results via meta sumamrize and meta forestplot
• How to interpret the results under different MA models
• How to address the problem of heterogeneity
• How to perform meta-regression using meta regress
• How to assess the validity of the MA against the threat of publication bias
• How to test for funnel-plot asymmetry using meta bias
• How to conduct a trim-and-fill analysis using meta trimfill
• How to differentiate between various reasons behind funnel-plot asymmetry

This presentation will provide methods and formulas and demonstrate how to perform MA
with real data. Participants who bring their own laptop will be able to interactively follow
along provided they have Stata 16 installed and a working Internet connection for down-
loading datasets from http://www.stata-press.com. However, interactive participation is
not required. The notes will provide sufficient information to reproduce all analyses at the
attendees’ convenience.

About the Instructor

Houssein Assaad is a Senior Statistician and Software Developer at StataCorp LLC and
the primary developer of Stata’s MA suite. Houssein has a PhD in statistics from the
University of Texas at Dallas. He is a former research assistant professor at Texas A&M
University, where his research focused on longitudinal and functional data analysis.

Relevance to Conference Goals

This demonstration will provide researchers with the tools to use MA in real-world applica-
tions. Participants will learn about MA as a statistical procedure and how to perform the
steps of MA in Stata.

SAS/STAT® 15.1 includes PROC BGLIMM, a new, high-performance, sampling-based procedure that provides full Bayesian inference for generalized linear mixed models (GLMMs). PROC BGLIMM models data from the exponential family distributions that have correlations or nonconstant variability; uses syntax similar to that of the MIXED and GLIMMIX procedures (the CLASS, MODEL, RANDOM, REPEATED, and ESTIMATE statements); deploys optimal sampling algorithms that are parallelized for performance; handles multilevel nested and non-nested random-effects models; and fits models to multivariate or longitudinal data with repeated measurements. PROC BGLIMM provides convenient access, with improved performance, to Bayesian analysis of complex mixed models that you could previously perform with the MCMC procedure. This workshop starts with a general discussion of Bayesian GLMM, then presents the important features of PROC BGLIMM, showing you how to use it for estimation, inference, and prediction.

Outline & Objectives

OUTLINE
1. Overview of Bayesian GLMM
2. Syntax and options of PROC BGLIMM
3. Demonstration of PROC BGLIMM through examples
3.1 Simple normal regression
3.2 Logistic regression with random intercepts
3.3 Normal regression with repeated measurements
3.4 Non-nested logistic random-effects model with prediction
3.5 Poisson regression with random effects
3.6 Repeated growth measurements with internal difference

Target Audience
This presentation is intended for a broad audience of statisticians who are interested in Bayesian inference for generalized linear mixed models. It would be helpful for attendees to have a basic understanding of normal regression analysis, generalized linear mixed models, and Bayesian methods, but it is not required.

LEARNING OUTCOMES
(a) Performance objectives
By attending this presentation, participants will improve their knowledge of generalized linear mixed models and Bayesian methods, and they will be able to use the BGLIMM procedure in SAS/STAT software to conduct Bayesian analyses.
(b) Content and instructional methods
The presentation will alternate between the use of slides and software demonstrations. Handouts given to attendees will cover both.

About the Instructor

Amy Shi is a senior research statistician developer in the Advanced Analytics Division at SAS Institute Inc. She received a Ph.D. in biostatistics from the University of North Carolina at Chapel Hill. She joined SAS in 2010, and her work involves implementation of Bayesian methods in software. Amy’s main responsibility is developing and enhancing SAS’ Bayesian capability, with a focus on generalized linear mixed models, discrete choice models, and multilevel hierarchical settings.

Relevance to Conference Goals

Data is abundant in modern society, and a raft of statistical and machine learning algorithms have been developed to assist researchers, managers and lay-people to understand what inferences can be made from their data, and what decisions would best progress their goal. And yet, the current \p-value crisis in science" is evidence that even in the scientific community, access to these sophisticated algorithms is not owing through to many researchers, particularly those who do not have dedicated statistical or data science support.

The AutoStat Institute was founded by a group of academics and consultants who believe that this issue is, in a large part, due to the need to code in programs like R or Python to gain access to these algorithms. While the operability of these and similar platforms continues to improve, there are many potential users of data that will never have the skill set, time, interest or level of exposure required to become au fait with these packages. As a result, many users are excluded from realising the potential of the Big Data World by virtue of a coding barrier. AutoStat solves this problem by offering its users a modern feel GUI environment for sophisticated statistical analysis that aims to provide academics, students, business and interested people access to scalable modern algorithms and visualizations in a code free environment.

Outline & Objectives

This 2-hour workshop will focus on the user experience and provide practical demonstrations of both businessand research projects from the practical implementations of

Data management: Making new variables with the calculator tool, various methods for easy
data splitting test / train, merging datasets and much more.

Visualisations: Easy exploratory plots through to sophisticated layering approaches for publication and presentation quality output.

Model Building: A range of machine learning and statistical models (both frequentist and
Bayesian approaches)

Results and Inference: Standard outputs and tools to create users own inference metrics

Team work: Project sharing and collaboration from early stage data management, modeling
and report writing,

Tutorials and other help facilities to enable the user to get full benefit from their data analysis.
We will then illustrate how the software can enhance the research or business output using real case studies and implementing the following tools:

Pipeline construction for ease of updating results as new data becomes available via easy point,
click and record;

Dashboard building for effective deployment to end users and broadening the reach of your
research;

Document builders that are available in AutoStat with a range of templates that can be
customized by the user.

About the Instructor

Dr Clair Alston-Knox is a Senior Statistician with Predictive Analytics Group (Melbourne Australia). She had been an research and academic statistician since 1992, with a number of biometric and statistical consulting positions in government and universities. She joined Predictive Analytics and the AutoStat Institute in 2018 because her teaching, consulting, advising and ethics committee roles were frequently frustrated by researchers who were very capable of understanding the objective and benefits of statistical or machine learning approaches, but did not have the resources to learn the required platform to enable next level analysis.

Dr Theo Gazos is the Managing Director of Predictive Analytics Group. Theo has over 25 years of experience building economic and econometric models that isolate and quantify the impact of changing market dynamics (domestic and international), competition effects and government policy on private and government sector organisations. Theo is passionate about bringing the power of statistics and machine learning to all levels within organisations, and has used his years of experience to develop an interface and user ow within AutoStat R that makes this objective achievable.

Relevance to Conference Goals

Communication, collaboration and career development
AutoStat is an ideal environment for sophisticated statistical analysis, such as Bayesian models with stochastic search variable selection. The report building, collaboration and visualization feature all assist users in communicating outcomes.

Data Modeling and Analysis, Data Science and Big Data
AutoStat will help different users of big data in many different ways. For example, the point and click nature of AutoStat will allow data analysts to perform sophisticated machine learning and produce the standard results by default without needing to implement code, decide on the most appropriate libraries or construct their own visualisations. The Bayesian models provided in AutoStat R are highly optimised and scalable to big data. Default settings have been based on the latest research in the area of each model, are well documented and are prominently displayed so that users are aware of their settings (and can easily change them).

Software, Programming and Data Visualization
AutoStat provides modern graphics using drop and drag, with many customisable styles and the option of layering within charts. Users can produce high quality graphics without the need to code.