The thesis deals with survival trees and their application to the analysis and prediction of innovation processes. The purpose of the conducted research was (1) to investigate the appropriateness of survival trees for innovation diffusion research by means of an application of the method to a real dataset, (2) to give an overview of the current state of research on the survival tree method, and (3) to compare survival trees to more established methods of event history analysis, such as hazard rate models, in order to discuss both advantages and disadvantages of survival trees vis-à-vis alternative approaches.
Table of Contents
1 Introduction
1.1 Context of Thesis
1.2 Contribution of Thesis
1.3 Structure & Internal Pattern of Thesis
2 Modelling Censored Event Data in the Context of Innovation Adoption- and Diffusion Theory
2.1 Analysing and Forecasting Innovation Diffusion by Dynamic Micro Models
2.2 General Concepts and Terminologies
2.3 Statistical Framework
2.4 Classical Methods for the Analysis of Event History Data
2.4.1 Non-Parametric Methods
2.4.2 Parametric Methods
2.4.3 Semi- Parametric Methods
3 Presentation and Analysis of the Survival Tree Method
3.1 Review of CARTTM
3.2 Principle Framework & Mechanics of Survival Trees
3.3 Splitting, Pruning, Tree-Selection & Alternative Proposals for Survival Trees
3.3.1 Splitting – Growing the Saturated Tree
3.3.2 Pruning – Generation of Optimal Subtree Sequence
3.3.3 Final Tree Selection
3.3.4 Alternative Approaches
3.4 Final Assessment
3.4.1 Assessment of the Splitting, Pruning and Tree Selection Proposals
3.4.2 Merits & Deficiencies of the Survival Tree Method
4 The use of Survival Trees to Forecast Innovation Diffusion
4.1 Applicability of Available Software
4.2 Data Description & Handling
4.3 Implementation
4.4 Results
4.5 Discussion
5 Summary
6 Appendix
6.1 Classification and Regression Tree for E-purchase Adoption
6.2 Cross Table for Sector and Country Coverage
6.3 Variable Description and Handling
6.4 R-Syntax for Survival Tree
6.5 Data Output for E-purchase Survival Tree
6.6 Saturated Survival Tree for E-purchase Adoption
6.7 Original Survival Tree for E-purchase
7 Indices
7.1 Index of Abbreviations
7.2 Index of Symbols
7.3 Index of Synonyms
7.4 Index of Tables
7.5 Index of Figures
8 Bibliography
Research Objectives and Key Topics
The main objective of this thesis is to examine the applicability and usefulness of survival trees as a novel method for analyzing and forecasting innovation diffusion. By extending the non-parametric classification and regression tree (CART) method to censored event history data, the research aims to bridge the gap between individual adoption decisions and aggregate diffusion patterns, ultimately providing a more robust tool for innovation marketing.
- Integration of micro-level adoption behavior into diffusion theory.
- Extension of CART methodologies to handle censored event history data.
- Comparative analysis of splitting, pruning, and selection criteria for survival trees.
- Practical implementation of survival trees using R for e-business adoption data.
- Evaluation of survival trees as a diagnostic and predictive tool for market targeting.
Excerpt from the Book
3.1 Review of CART
In contrast to classical regression and classification models the non-parametric CART method does not require a specified model structure. Rather than fitting a model to the sample data, a tree structure is generated by dividing the sample recursively into a number of groups, each division being chosen so as to maximize some measure of the difference in the response variable for the resulting binary subgroup.
In the first step the dataset is recursively split (i.e. partitioned) into subgroups until a so-called “saturated tree” is found. The root node of a tree contains the sample of subjects from which the tree is grown.
The criterion that each split has to comply with is that it splits each node into those two sub-groups that are most homogeneous in terms of predicting the binary dependent variable.
In the context of ADT, the sample is recursively split into groups that are most homogeneous in predicting whether a company adopted a certain technology or not. In other words, complete homogeneity means that a node contains either only adopters or non-adopters.
Predictor variables can be ordinal, (continuous) or nominal. The number of possible splits varies with the type of predictor variable. Usually, complete homogeneity is an ideal that is rarely realized. Thus, the numerical objective of partitioning is to make the contents of the nodes as homogeneous as possible. A number of methods have been proposed to assess the extent of node homogeneity for each node and to choose the best split. In general, splitting rules are based on node impurity measures such as the Gini index, the Bayes rule or the entropy function. Most frequently, the entropy function is used due to a number of desirable properties.
Summary of Chapters
1 Introduction: This chapter defines the research context, emphasizing the necessity of micro-level models in innovation diffusion, and outlines the thesis structure.
2 Modelling Censored Event Data in the Context of Innovation Adoption- and Diffusion Theory: It establishes the statistical framework of event history analysis, discussing how hazard models and censored data fit into the context of innovation adoption.
3 Presentation and Analysis of the Survival Tree Method: This chapter provides a comprehensive review of CART methodology and its extension into survival trees, detailing the processes of splitting, pruning, and final tree selection.
4 The use of Survival Trees to Forecast Innovation Diffusion: This section details the practical implementation of survival trees using R, applying them to real-world e-purchase data to evaluate their predictive power and usefulness for market segmentation.
5 Summary: This concluding chapter synthesizes the main findings, reiterating the value of survival trees as a supplemental tool for forecasting innovation diffusion and identifying areas for future research.
Keywords
Innovation Diffusion, Adoption Theory, Event History Data, Survival Analysis, Survival Trees, CART, Censored Data, Hazard Models, Recursive Partitioning, Micro-modelling, Market Segmentation, Predictor Variables, Machine Learning, Statistical Forecasting, Proportional Hazard Model
Frequently Asked Questions
What is the core focus of this research?
The thesis focuses on applying the survival tree methodology to innovation diffusion research to better capture individual heterogeneity and process dynamics.
What are the primary thematic areas?
The research intersects innovation marketing, survival analysis, event history modelling, and non-parametric statistical methods like CART.
What is the primary goal or research question?
The goal is to determine if survival trees can effectively bridge the gap between individual adoption decisions and aggregate diffusion patterns, offering better forecasting accuracy than traditional macro models.
What scientific methods are employed?
The work utilizes survival analysis techniques, specifically survival trees, and compares them with classical parametric and semi-parametric hazard models, supported by statistical software like R.
What is covered in the main section of the work?
The main part explains the theoretical background of survival trees, details the building blocks of the algorithm (splitting, pruning, selection), and demonstrates its application to e-purchase adoption data.
Which keywords best characterize this work?
Key terms include Innovation Diffusion, Survival Trees, Censored Data, Event History, and Predictive Modelling.
Why are survival trees considered better than traditional models for this specific dataset?
Unlike aggregate models, survival trees identify specific predictive subgroups and capture complex, non-linear interactions between variables, which provides deeper insights into which companies are likely to adopt an innovation first.
How does the author handle missing data in the survival tree model?
The model employs "surrogate splits," which allow the algorithm to classify observations with missing values based on alternative splitting variables that highly correlate with the best split variable.
- Arbeit zitieren
- Burkhard von Wangenheim (Autor:in), 2004, Survival trees - a new method in innovation theory, München, GRIN Verlag, https://www.grin.com/document/193797
