Featured news from SAS.


View as a web page
SAS The Power to Know logo

It’s back-to-school time in North Carolina, although that seems backwards here at SAS headquarters, as we’ve recently ended another successful summer of internships for both undergraduate and graduate students. Interns showed off their work in poster presentations in an exhibition or individual seminars. We saw some very impressive results! And speaking of students, in case the offspring of SAS employees didn’t have enough schoolwork ahead of them, SAS also provided a two-week class for these children who are in high school or college to teach them SAS® programming skills. The two students I spoke to—one heading to UNC and the other to Stanford—thought it was great fun.

We saw continued interest in SAS® University Edition at the SAS exhibition area at the Joint Statistical Meetings in Vancouver a few weeks ago, where it seemed as if the Educational Practice staff were constantly surrounded by attendees. We had oversized demand for several of our tutorials this year.

One of my favorite sessions at JSM was on how to evolve the teaching of statistics and data science to make them relevant in today’s world. Panelists covered K–12, undergraduate, and graduate education, and they discussed experiences with pilot programs that centered on getting data and software into the hands of students as quickly as possible.

The SAS fall regional conferences, which are coming up soon, provide a more intimate opportunity to learn about SAS from SAS and to interact with the staff. See Talks and Tutorials below for the list of conferences and tutorials we are giving.

Here’s to back-to-schooling!

Senior R&D Director, Statistical Applications


Technical Papers


Causal Mediation Analysis with the CAUSALMED Procedure

Important policy and health care decisions often depend on understanding the direct and indirect (mediated) effects of a treatment on an outcome. For example, does a youth program directly reduce juvenile delinquent behavior, or does it indirectly reduce delinquent behavior by changing the moral and social values of teenagers? Or, for example, is a particular gene directly responsible for causing lung cancer, or does it have an indirect (mediated) effect through its influence on smoking behavior? Causal mediation analysis deals with the mechanisms of causal treatment effects, and it estimates direct and indirect effects. A treatment variable is assumed to have causal effects on an outcome variable through two pathways: a direct pathway and a mediated (indirect) pathway through a mediator variable. This paper introduces the CAUSALMED procedure, new in SAS/STAT® 14.3, for estimating various causal mediation effects from observational data in a counterfactual framework. The paper also defines these causal mediation and related effects in terms of counterfactual outcomes and describes the assumptions that are required for unbiased estimation. Examples illustrate the ideas behind causal mediation analysis and the applications of the CAUSALMED procedure.


Tips and Techniques for Using the Random-Number Generators in SAS

SAS® 9.4M5 introduces new random-number generators (RNGs) and new subroutines that enable you to initialize, rewind, and use multiple random-number streams. This paper describes the new RNGs and provides tips and techniques for using random numbers effectively and efficiently in SAS. Applications of these techniques include statistical sampling, data simulation, Monte Carlo estimation, and random numbers for parallel computation.


Managing the Expense of Hyperparameter Autotuning

Determining the best values of machine learning algorithm hyperparameters for a specific data set can be a difficult and computationally expensive challenge. The recently released AUTOTUNE statement and autotune action set in SAS® Visual Data Mining and Machine Learning software automatically tune hyperparameters of modeling algorithms by using a parallel local search optimization framework to ease the challenges and expense of hyperparameter optimization. This implementation allows multiple hyperparameter configurations to be evaluated concurrently, even when data and model training must be distributed across computing resources because of the size of the data set.

With the ability to both distribute the training process and parallelize the tuning process, one challenge then becomes how to allocate the computing resources for the most efficient autotuning process. The best number of worker nodes for training a single model might not lead to the best resource usage for autotuning. To further reduce autotuning expense, early stopping of long-running hyperparameter configurations that have stagnated can free up resources for additional configurations. For big data, when the model training process is especially expensive, subsampling the data for training and validation can also reduce the tuning expense. This paper discusses the trade-offs that are associated with each of these performance-enhancing measures and demonstrates tuning results and efficiency gains for each.


Optimization Modeling with Python and SAS® Viya®

Python has become a popular programming language for both data analytics and mathematical optimization. With SAS Viya and its Python interface, Python programmers can use the state-of-the-art optimization solvers that SAS provides. This paper demonstrates an approach for Python programmers to naturally model their optimization problems, solve them by using SAS® Optimization solver actions, and view and interact with the results. The common tools for using the optimization solvers in SAS for these purposes are the OPTMODEL and IML procedures, but programmers more familiar with Python might find this alternative approach easier to grasp.


Regime-Switching Models: Capturing Structural Changes in Time Series

Stock market conditions, government policy changes, or even weather patterns can be regarded as stochastic processes that are driven by unobserved regimes. A powerful tool to explore these behavioral patterns is the regime-switching model (RSM) that is offered in the HMM procedure and the associated action in SAS® Econometrics software. This model, which is widely used in finance, economics, science, and engineering, has two characteristics: it allows different parameter values for different regimes, and it models the transition probabilities between regimes. These characteristics enable it to fully capture the structural changes in the time series. This paper uses two examples to illustrate how you can use RSMs to better understand the regime patterns in your data and improve your economic analysis. The first example demonstrates how regime-switching autoregression (RS-AR) models help you characterize the volatility and dynamics of stock returns. The second example examines the relationship and movement between the Japanese yen and the Thai baht by using regime-switching regression (RS-REG) models.


Technical Highlights

Video camera icon

Video Prize  

Filmmakers Brett Wujek and Patrick Koch, also known as Principal Data Scientist Wujek and Principal Machine Learning Developer Koch, were awarded the Audience Appreciation Award for Best Promotional Video based on their autotuning paper for the KDD conference in London this past week. This is the machine learning and data mining conference, with over 3,500 attendees. For some fun, watch their short piece on the need for autotuning.  

And you thought YouTube was only for learning how to put your kid’s bike together!


Blog Icon

Graphically Speaking

Distinguished Research Statistician Developer Warren Kuhfeld discusses the SGPLOT procedure, BY groups, and SG annotation.


Tech Support Points Out

Tech Support Points Out icon

Usage Note 60335: Choice of Continuous Response Distribution in Log-Linked GLMs

There are several response distributions available when you are fitting generalized linear models (GLMs) in procedures such as PROC GENMOD and PROC GLIMMIX. The distributions include the normal, Poisson, gamma, inverse Gaussian, Tweedie, and others. One important aspect of how these distributions differ is in the relationship between the mean and the variance. When it comes to selecting a distribution to use when modeling a response, choosing a distribution that matches the observed mean-variance relationship is important. This note discusses and illustrates some tools for finding a suitable response distribution if it is not already known, including graphical and analytical methods.


Talks and Tutorial


WUSS 2018

Date: Sept. 5–7, 2018

Location: Sacramento, CA

Introduction to Mixed Models — Gordon Brown



MWSUG 2018

Date: Sept. 30 – Oct. 2, 2018

Location: Indianapolis, IN

Advanced ODS Graphics Examples — Warren Kuhfeld



SESUG 2018

Date:  Oct. 14–17, 2018

Location: St. Pete Beach, FL

Modeling Longitudinal Categorical Response Data — Maura Stokes



SCSUG 2018

Date: Nov. 4–7, 2018

Location: Austin, TX

Causal Analysis with Observational Data: Methods and Applications


Quick Links


SAS® Statistics and Operations Research News

Don't miss important updates from SAS! Please add sas.com as a domain in your safe sender list.
Want to get more out of your relationship with SAS? Update your SAS profile

About this e-mail:
If you do not wish to receive future SAS® Statistics and Operations Research News editions, please unsubscribe.
SAS places great value in fair information practices and in connection with the management of our contact database, we would like to remind you that SAS Institute Inc. may use your personal contact details for marketing purposes, as stated in its privacy policy. To contact SAS via postal mail: SAS, SAS Campus Drive, Cary NC 27513 USA. ATTN: Legal Division/Privacy Manager.

SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS Institute Inc. in the USA and other countries. ® indicates USA registration. Other brand and product names are trademarks of their respective companies. Copyright © SAS Institute Inc. All rights reserved.