Skip to main content
Fachgebiete
Automobil + Motoren Bauwesen + Immobilien Business IT + Informatik Elektrotechnik + Elektronik Energie + Nachhaltigkeit Finance + Banking Management + Führung Marketing + Vertrieb Maschinenbau + Werkstoffe Versicherung + Risiko
DE
EN
Themenseiten
Organisationspsychologie Projektmanagement Marketing Smart Manufacturing
Bücher
Zeitschriften
Weitere Formate
Podcasts Webinare Technik Webinare Wirtschaft Kongresse Veranstaltungskalender Awards
Jetzt Einzelzugang starten
Zugang für Unternehmen
Referenzkunden
MTZ-Fachkongress

Antriebe und Energiesysteme von morgen


Austausch von Automobil- und Energiewirtschaft

Am 14./15. Mai geht es in Chemnitz um die Mobilitätswende durch Elektro- und Wasserstofffahrzeuge.
Erweiterte Suche
Anmelden
nach oben

2021 | Buch

Introduction Kapitel lesen Erstes Kapitel lesen

New Statistics for Design Researchers

A Bayesian Workflow in Tidy R

verfasst von: Dr. Martin Schmettow

Verlag: Springer International Publishing

Buchreihe : Human–Computer Interaction Series

Enthalten in: Springer Professional "Wirtschaft+Technik" , Springer Professional "Technik"

Einloggen, um Zugang zu erhalten
insite
SUCHEN

Über dieses Buch

Design Research uses scientific methods to evaluate designs and build design theories. This book starts with recognizable questions in Design Research, such as A/B testing, how users learn to operate a device and why computer-generated faces are eerie. Using a broad range of examples, efficient research designs are presented together with statistical models and many visualizations.

With the tidy R approach, producing publication-ready statistical reports is straight-forward and even non-programmers can learn this in just one day. Hundreds of illustrations, tables, simulations and models are presented with full R code and data included.

Using Bayesian linear models, multi-level models and generalized linear models, an extensive statistical framework is introduced, covering a huge variety of research situations and yet, building on only a handful of basic concepts. Unique solutions to recurring problems are presented, such as psychometric multi-level models, beta regression for rating scales and ExGaussian regression for response times. A “think-first” approach is promoted for model building, as much as the quantitative interpretation of results, stimulating readers to think about data generating processes, as well as rational decision making.

New Statistics for Design Researchers: A Bayesian Workflow in Tidy R targets scientists, industrial researchers and students in a range of disciplines, such as Human Factors, Applied Psychology, Communication Science, Industrial Design, Computer Science and Social Robotics. Statistical concepts are introduced in a problem-oriented way and with minimal formalism. Included primers on R and Bayesian statistics provide entry point for all backgrounds. A dedicated chapter on model criticism and comparison is a valuable addition for the seasoned scientist.

Inhaltsverzeichnis

Frontmatter

Preparations

Frontmatter
Chapter 1. Introduction
Abstract
In this chapter, I will loosely define the terms of the title: design research, New Statistics and tidy R. Then I give some possible routes through the book, as well as ideas on how to use it in the classroom.
Martin Schmettow
Chapter 2. Getting Started with R
Abstract
In this book, we will be using the statistical computing environment R. R at its core is a programming language that specializes in statistics and data analysis. R comes with a complete set of standard packages that cover common routines in statistical analysis. However, the classic standard packages are known to be incomplete, inconsistent and difficult to master. In the past few years, Hadley Wickham has started an initiative known as the Tidyverse. Under this new dogma, a fast-growing collection of tidy packages emerged, which share a set of data engineering principles. This makes Tidyverse packages interoperable and easy to learn. This chapter introduces the reader to tidy R.
Martin Schmettow
Chapter 3. Elements of Bayesian Statistics
Abstract
The aim of scientific research is to avoid the pitfalls of our minds and act as rational as possible by translating our theory into a formal model of reality, gathering evidence in an unbiased way and weigh evidence against noise in a controlled way. This chapter introduces the Bayesian approach to statistical modeling from the ground up and illustrates its use for rational decision-making. Probability is derived from set theory and relative frequencies before we dive into Bayes Bayesian thinking. The elements and practical details of statistical modeling are introduced by its two components: the structural part, which typically carries the research question or theory, followed by a rather deep account of the second component of statistical models: the random part.
Martin Schmettow

Models

Frontmatter
Chapter 4. Basic Linear Models
Abstract
Linear models answer the question of how one quantitative outcome, say ToT, decreases or increases, when a condition changes. This chapter introduces the three basic ways how such conditions can enter a linear model. The most basic LM, the grand mean model, does not account for any conditions and produces just a single estimate: the grand mean in the population. In linear regression models, a metric predictor (e.g. age) is linked to the outcome by a linear function, with a slope and an intercept. A very common type of question is how an outcome changes by a set of discrete conditions, such as two different designs. Factorial model uses a mechanism called dummy variables to make categorical variables fit into the linear term. With ordinal factor models, dummy variables are put to a good use.
Martin Schmettow
Chapter 5. Multi-predictor Models
Abstract
Design researchers are often collecting data under a variety of conditions, each of which qualifies as a predictor in its own right. The advantage of the linear terms is that they can be combined, which allows estimating the simultaneous influence of multiple predictors in one model. We will start with models, where all predictors act mutually independent on the outcome. In reality, it often happens that predictors are not independent, but act conditional upon each other. Conditional effects models can adjust a model for common violations of linearity, saturation and amplification, but can also uncover relevant structures, or test theories. The chapter ends with polynomial models, which allow to fit non-linear relationships between metric predictors and outcome.
Martin Schmettow
Chapter 6. Multilevel Models
Abstract
Multi-level linear models introduce a special type of categorial variable, the random factor, which applies when the factor levels can be seen as members of a population, such as participants in a sample. Multi-level models allows to simultaneously produce estimates on population level and on participant level. That makes multi-level models interesting for a variety of applications, three of which are covered in this chapter. First, we will see how multi-level models render diversity of users, which is expressed as the random factor variance. Taken this to an extreme, participant-level coefficients can answer the question, whether a design-related impact factor is universal, in that it can be observed for every participant. Finally, multi-level models are very well-suited to handle psychometric, or design-o-metric, situations, where some of the populations are non-human.
Martin Schmettow
Chapter 7. Generalized Linear Models
Abstract
The preceding chapters were all about specifying an appropriate (and often sophisticated) predictor term. In this chapter, we will give the outcome variables their due respect. First, the assumptions of Gaussian linear model are reviewed and generally discarded. Next, the framework of Generalized Linear Models is explained from ground up. After that is established, I will introduce a good dozen of model families, organized by types of measures. Next to some commonly known families, such as Poisson or Logistic regression models, this chapter will cover outcome variables for which good defaults have been lacking, such as rating scale responses and ToT measures. For RT and ToT data, I will suggest exponentially-modified Gaussian models and for (quasi) continuous rating scales, I will introduce a rather novel approach, Beta regression. The chapter closes with a look beyond GLMs. Distributional models allow to link predictors to the mean, but also other properties of the outcome distribution. I will show how a distributional model covers differences in individual answer styles to rating scales.
Martin Schmettow
Chapter 8. Working with Models
Abstract
In Chaps. 4 and 6, we have seen a marvelous variety of models spawning from just two basic principles, linear combination of multiple effects and Gaussian distribution. Chapter 7 further expanded the variety of models, letting us choose response distributions that sit snug on the outcome variable. This chapter is dedicated to methods that assess how snug a model is.
Martin Schmettow
Backmatter
Metadaten
Titel
New Statistics for Design Researchers
verfasst von
Dr. Martin Schmettow
Copyright-Jahr
2021
Electronic ISBN
978-3-030-46380-9
Print ISBN
978-3-030-46379-3
DOI
https://doi.org/10.1007/978-3-030-46380-9

Neuer Inhalt

    Bildnachweise