Event History Analysis with Stata 2nd Edition

Hans-Peter Blossfeld download

https://textbookfull.com/product/event-history-analysis-with-
stata-2nd-edition-hans-peter-blossfeld/

Download full version ebook from https://textbookfull.com

We believe these products will be a great fit for you. Click
the link to download now, or visit textbookfull.com
to discover even more!

Education as a Lifelong Process The German National

Educational Panel Study NEPS Hans-Peter Blossfeld

https://textbookfull.com/product/education-as-a-lifelong-process-
the-german-national-educational-panel-study-neps-hans-peter-
blossfeld/

Applied Meta-Analysis with R and Stata (Chapman &

Hall/CRC Biostatistics Series) 2nd Edition Chen

https://textbookfull.com/product/applied-meta-analysis-with-r-
and-stata-chapman-hall-crc-biostatistics-series-2nd-edition-chen/

IUTAM A Short History 2nd Edition Peter Eberhard

https://textbookfull.com/product/iutam-a-short-history-2nd-
edition-peter-eberhard/

Sleep Wake Neurobiology and Pharmacology Hans-Peter

Landolt

https://textbookfull.com/product/sleep-wake-neurobiology-and-
pharmacology-hans-peter-landolt/
Functional Analysis in Clinical Treatment 2nd Edition
Peter Sturmey (Editor)

https://textbookfull.com/product/functional-analysis-in-clinical-
treatment-2nd-edition-peter-sturmey-editor/

On the Origin of Natural Constants Axiomatic Ideas with

References to the Measurable Reality Hans Peter Good

https://textbookfull.com/product/on-the-origin-of-natural-
constants-axiomatic-ideas-with-references-to-the-measurable-
reality-hans-peter-good/

Practical Microservices Build Event Driven

Architectures With Event Sourcing and CQRS 1st Edition
Ethan Garofolo

https://textbookfull.com/product/practical-microservices-build-
event-driven-architectures-with-event-sourcing-and-cqrs-1st-
edition-ethan-garofolo/

Derivatives and Internal Models: Modern Risk Management

Hans-Peter Deutsch

https://textbookfull.com/product/derivatives-and-internal-models-
modern-risk-management-hans-peter-deutsch/

Data Analysis with R 2nd Edition Tony Fischetti

https://textbookfull.com/product/data-analysis-with-r-2nd-
edition-tony-fischetti/
i

Event History Analysis with Stata

Nowadays, event history analysis can draw on a well-established set of

statistical tools for the description and causal analysis of event history
data. The second edition of Event History Analysis with Stata provides an
updated introduction to event history modeling, along with many
instructive Stata examples.
Using the latest Stata software, each of these practical examples develops a
research question, refers to useful substantive background information, gives a
short exposition of the underlying statistical concepts, describes the
organization of the input data and the application of the statistical Stata
procedures, and assists the reader in performing a substantive interpretation
of the obtained results. Emphasizing the strengths and limitations of event
history model techniques in each field of application, this book
demonstrates that event history models provide a useful approach with
which to uncover causal relationships or to map out a system of causal
relations. It demonstrates how long-term processes can be studied and how
changing context information on the micro, meso, and macro levels can be
integrated easily into a dynamic analysis of longitudinal data.
Event History Analysis with Stata is an invaluable resource for both novice
students and researchers who need an introductory textbook and
experienced researchers (from sociology, economics, political science,
pedagogy, psychology, or demography) who are looking for a practical
handbook for their research.

Hans-Peter Blossfeld is Professor of Sociology at the University of Bamberg

in Germany. He directed the European Research Council (ERC) funded
project ‘Education as a Lifelong Process – Comparing Educational
Trajectories in Modern Societies’ (eduLIFE) at the European University
Institute in Florence, Italy (2012–2016).

Götz Rohwer is Professor Emeritus of Methods of Social Research and

Statistics at the Ruhr-Universität Bochum in Germany.

Thorsten Schneider is Professor of Sociology with a special focus on

comparative analysis of contemporary societies at the University of Leipzig
in Germany since 2011.
iii

Event History Analysis

with Stata
Second Edition

Hans-Peter Blossfeld, Götz Rohwer

and Thorsten Schneider
With a Chapter on Sequence Analysis by Brendan Halpin
Second edition published 2019
by Routledge
2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
and by Routledge
52 Vanderbilt Avenue, New York, NY 10017
Routledge is an imprint of the Taylor & Francis Group, an informa business
© 2019 Hans-Peter Blossfeld, Götz Rohwer, Thorsten Schneider
The right of Hans-Peter Blossfeld, Götz Rohwer, and Thorsten Schneider
to be identified as authors of this work has been asserted by them in
accordance with sections 77 and 78 of the Copyright, Designs and Patents
Act 1988.
All rights reserved. No part of this book may be reprinted or reproduced or
utilised in any form or by any electronic, mechanical, or other means, now
known or hereafter invented, including photocopying and recording, or in
any information storage or retrieval system, without permission in writing
from the publishers.
Trademark notice: Product or corporate names may be trademarks or
registered trademarks, and are used only for identification and explanation
without intent to infringe.
First edition published by Routledge, 2007
British Library Cataloguing-in-Publication Data
A catalogue record for this book is available from the British Library
Library of Congress Cataloging-in-Publication Data
A catalog record for this book has been requested
ISBN: 978-1-138-07079-0 (hbk)
ISBN: 978-1-138-07085-1 (pbk)
ISBN: 978-0-429-26040-7 (ebk)
Typeset in Goudy
by Apex CoVantage, LLC
Visit the eResources: www.routledge.com/9781138070851
v

Contents

Preface viii
Acknowledgments xi

1 Introduction 1
1.1 Causal modeling and observation plans 3
1.1.1 Cross-sectional data 4
1.1.2 Panel data 11
1.1.3 Event history data 17
1.2 Event history analysis and causal modeling 19
1.2.1 Causal explanations 19
1.2.2 Transition rate models 28

2 Event history data structures 41

2.1 Basic terminology 41
2.2 Event history data organization 45
2.2.1 Using event history data files with Stata 49
2.2.2 Executing Stata with a do-file 51
2.2.3 Single episode data 54
2.2.4 Multiepisode data 57

3 Nonparametric descriptive methods 63

3.1 Life table method 63
3.2 Product-limit estimation 76
3.3 Comparing survivor functions 80

4 Exponential transition rate models 95

4.1 The basic exponential model 96
4.1.1 Maximum likelihood estimation 97
4.1.2 Models without covariates 99
4.1.3 Time-constant covariates 102
4.2 Models with multiple destinations 108
4.3 Models with multiple episodes 116
vi Contents vi
5 Piecewise constant exponential models 124
5.1 The basic model 124
5.2 Models without covariates 126
5.3 Models with proportional covariate effects 130
5.4 Models with period-specific effects 132

6 Exponential models with time-dependent covariates 137

6.1 Parallel and interdependent processes 137
6.2 Interdependent processes: the system approach 139
6.3 Interdependent processes: the causal approach 142
6.4 Episode splitting with qualitative covariates 144
6.5 Episode splitting with quantitative covariates 155
6.6 Application examples 160

7 Parametric models of time dependence 192

7.1 Interpretation of time dependence 193
7.2 Gompertz models 196
7.3 Weibull models 204
7.4 Log-logistic models 212
7.5 Log-normal models 218
7.6 Conclusion: Estimating time-dependent models with Stata 223

8 Methods to check parametric assumptions 224

8.1 Simple graphic methods 224
8.2 Pseudoresiduals 226

9 Semiparametric transition rate models 231

9.1 Partial likelihood estimation 232
9.2 Time-dependent covariates 236
9.3 The proportionality assumption 239
9.4 Stratification with covariates and for multiepisode data 243
9.5 Baseline rates and survivor functions 248
9.6 Application example 251

10 Problems of model specification 256

10.1 Unobserved heterogeneity 256
10.2 Models with a mixture distribution 261
10.2.1 Models with a gamma mixture 264
10.2.2 Exponential models with a gamma mixture 267
10.2.3 Weibull models with a gamma mixture 269
10.2.4 Random effects for multiepisode data 273
10.3 Discussion 276
vii Contents vii
11 Introduction to sequence analysis 282
BRENDAN HALPIN

11.1 Defining distances 284

11.2 Doing sequence analysis in Stata 287
11.3 Unary summaries 289
11.4 Intersequence distance 291
11.5 What to do with sequence distances? 293
11.6 Optimal matching distance 296
11.7 Special topics 297
11.8 Conclusion 305

Appendix: exercises 308

References 320
About the authors 344
Index 347
Preface

Nowadays, event history analysis can draw on a well-established set of statis-

tical tools for the description and causal analysis of event history data. This
second edition of our book provides an updated introduction to event history
modeling along with many instructive Stata examples. Each of these practical
examples develops a research question, refers to useful substantive back-
ground information, gives a short exposition of the underlying statistical con-
cepts, describes the organization of the input data and the application of
Stata’s statistical procedures, and helps readers perform substantive interpre-
tation of the obtained results. The book also demonstrates the strength of
event history analysis based on many further research applications from
the economic and social science literature. Thus, the book is written for
both novices who need an introductory textbook and experienced research-
ers (from sociology, economics, political science, pedagogy, psychology, or
demography) who are looking for a practical handbook for their research.
Through a great number of examples and computer exercises, our book
might be particularly useful for undergraduate classes and graduate training.
We used the latest Stata software for our examples. Today, Stata is a widely
applied software package in economics and the social sciences that provides
many useful tools for data analysis, data management, and graphic presenta-
tion. For further information, we refer the reader to the Stata homepage.1
The data used in our examples are taken from the German Life History
Study (GLHS) but anonymized for data protection purposes. The GLHS
study was conducted by Karl Ulrich Mayer, as principal investigator at the
Max Planck Institute for Human Development and Education in Berlin.
The original data collection was funded by the German Research Foundation
(Deutsche Forschungsgemeinschaft, DFG). We would like to thank Professor
Mayer for his kind permission to use a sample of 600 job episodes from the
GLHS as a basis for our practical examples. This new edition of our book
additionally offers many exercises in the Appendix with data from the
German National Educational Panel Study (NEPS). The NEPS is a research
infrastructure project providing data on educational processes and trajectories
in Germany with several starting cohorts. One cohort focuses on adults and
ix Preface ix
has collected detailed retrospective information in the tradition of the
GLHS. From 2008 to 2013, the collection of NEPS data was funded by
the German Federal Ministry of Education and Research (Bundesministerium
für Bildung und Forschung, BMBF). Since 2014, the NEPS survey has been
carried out by the Leibniz Institute for Educational Trajectories (LIfBi) at the
University of Bamberg in cooperation with a Germany-wide network of edu-
cational experts (https://www.neps-data.de/tabid/2665/language/en-US). The
NEPS data used in our book are taken from a nonrepresentative subsample
restricted to West Germany that was additionally fully anonymized with
advanced strategies by the LIfBi. We would like to thank the LIfBi Board
of Directors for providing access to this subsample, and especially Tobias
Koberg from the LIfBi team, for data preparation and anonymization. The
GLHS and NEPS data files used throughout the book, the series of files con-
taining the Stata setups for the examples in the book, and the Appendix
can be downloaded from https://doi.org/10.5157/NEPS:Campus:1.0.0 or
http://home.uni-leipzig.de/~tschneid/. Thus, the reader is offered a unique
opportunity to easily run and modify all the application examples in the
book on a computer. In fact, we advise event history analysis beginners to
go through the application examples of the book step by step on their own
computers. Based on our own teaching experiences from many workshops
and introductory classes, going through the examples seems to be the most eff-
cient and straightforward way to become familiar with what are sometimes
complex analysis techniques.
Our book demonstrates that event history models provide a useful
approach with which to uncover causal relationships or to map out a
system of causal relations. Event history models are linked very naturally
to a causal understanding of social processes, because they allow us to
model a theoretically supposed change in future outcomes of a process in rela-
tion to its history and the (time-constant or changing) conditions of other
related processes. This book demonstrates how long-term processes can be
studied; how various forms of duration dependencies can be estimated with
nonparametric, parametric, and semiparametric models; and how parallel
and interdependent dynamic systems can be analyzed from a causal-analytical
point of view by applying the method of episode splitting. The book also
shows how changing context information on the micro, meso, and macro
levels can be integrated easily into a dynamic analysis of longitudinal data.
Finally, the book picks up the issues of unobserved heterogeneity of time-
constant and time-dependent omitted variables and makes suggestions on
how to deal with these sometimes difficult problems.
Depending on theoretical considerations about the process time axis and
the degree of measurement accuracy of event times, the literature often dis-
tinguishes between discrete-time and continuous-time event history models
(see, e.g., Tuma & Hannan, 1984; Blossfeld, Hamerle, & Mayer, 1989;
Mills, 2011). Discrete-time event history methods either (1) theoretically
presuppose a discrete process time axis (i.e., that the events can occur only
x Preface x
at fixed time intervals) or (2) assume a continuous process time axis (i.e., that
events can happen at any point in time, but that the measurement of these
event times could be achieved only in a sequence of rather crude time inter-
vals). For example, in some national panel studies, events such as job shifts
are recorded as happening only in yearly intervals. Various publications on
discrete-time models are available in the literature (see, e.g., Allison, 1982,
2014; Yamaguchi, 1991; Singer & Willet, 1993; Vermunt, 1997; Rohwer &
Pötter, 2007; Steele, 2008; Blossfeld & Blossfeld, 2015a, 2015b). Continuous-
time event history models assume that the timing of events is measured
exactly. How exactly the timing of an event actually has to be recorded is,
to a large extent, a substantive matter. For example, for long job careers,
monthly measurements of job mobility would seem to be suffcient. Of
course, at the end, all timing observations are discrete, but for continuous
models, they should aim to be recorded more precisely. More accurate mea-
surements of the timing of events lead to fewer tied observations in the
dataset and make it easier to reconstruct the causal order of events from dif-
ferent connected or interdependent processes. For example, retrospective life
history studies often collect data measured on the level of 1 month. It should
be noted here that, even if discrete- and continuous-time models use different
statistical techniques, in many concrete applications, the results of these
models are very similar, so that the substantive conclusions based on these dif-
ferent approaches are often quite the same (see Blossfeld & Blossfeld, 2015b).
The goal of this volume is to introduce the reader to the application of con-
tinuous-time event history models.
This updated edition of our book also adds a new chapter on sequence
analysis, written by Brendan Halpin. Sequence analysis is an approach that
treats the whole sequences of events as units, defining measures of similarity
and difference between pairs of sequences. Sequence analysis can be used
for different goals in the social sciences, including (1) performing a data-
driven classification of sequences of events, (2) assigning sequences to a
priori empirical or theoretical ideal types, (3) describing how the diversity
of sequences differs among groups, and (4) examining the determinants
of dyadic similarity (e.g., the event sequences of marital or nonmarital
partners).

Note
1 https://www.stata.com/
xi

Acknowledgments

We received support for our work from several sources and various places. In
particular, we received support from the project ‘Education as a Lifelong
Process – Comparing Educational Trajectories in Modern Societies’
(eduLIFE) supported by the European Research Council (ERC) and hosted
at the Otto Friedrich University of Bamberg, Germany, and the European
University Institute in Florence, Italy. We also were supported by the Depart-
ment of Social Sciences at the University of Bochum, Germany, and the
Institute of Sociology at the University of Leipzig, Germany. Finally, we
would like to thank Katrin Golsch, who was the third author of the first
edition of our book. Due to time constraints, she unfortunately has not
been able to continue her work on this new revised volume. We are therefore
very happy that Thorsten Schneider, an experienced expert in event history
modeling and statistical analysis with Stata, has supported the second edition
of our book. We would also like to thank Ulrich Pötter for helpful sugges-
tions; Jonathan Harrow for his rigorous proofreading and language editing
of the new manuscript; and Till Hovestadt, Mirjam Odile Nanko, and Fabi-
enne Wöhner, all from the University of Leipzig, for their final formatting
of the typescript in LaTeX. Finally, we also want to acknowledge that for
the production of the camera-ready copy for our book, we used Donald
Knuth’s typesetting program TeX in connection with Leslie Lamport’s
LaTeX and Tomas Rokicki’s DVIPS PostScript drivers.
Hans-Peter Blossfeld, Otto Friedrich University of Bamberg
Götz Rohwer, Ruhr University of Bochum
Thorsten Schneider, University of Leipzig
1

1 Introduction

Over the last three decades, social scientists have been collecting and analyz-
ing event history data with increasing frequency. This trend is not accidental,
nor does it reflect the prevailing fashion in survey research or statistical
analysis. Instead, it indicates a growing recognition among social scientists
that event history data are often the most appropriate empirical information
one can get on the substantive process under study.
James Coleman (1981, p. 6) characterized this kind of substantive process
in the following general way: (1) There is a collection of units (which may be
individuals, organizations, societies, or whatever), each moving among a
finite (usually small) number of states; (2) these changes (or events) may
occur at any point in time (i.e., they are not restricted to predetermined
points in time); and (3) there are time-constant and/or time-dependent
factors influencing the events.
Illustrative examples of this type of substantive process can be given for a
wide variety of social research fields: In labor market studies, workers move
between unemployment and employment,1 full-time and part-time work,2
or among various kinds of jobs;3 in social inequality studies, people become
home-owners over the course of their lives (life course);4 in demographic anal-
yses, men and women enter into consensual unions, marriages, or into father-/
motherhood or are getting a divorce;5 in sociological mobility studies, employees
shift through different occupations, social classes, or industries;6 in studies
of organizational ecology, firms, unions, or organizations are founded or closed
down;7 in political science research, governments break down, voluntary organi-
zations are founded, or countries go through a transition from one political
regime to another;8 in migration studies, people move between different
regions or countries;9 in marketing applications, consumers switch from one
brand to another or purchase the same brand again; in criminological studies,
prisoners are released and commit another criminal act after some time;
in communication analyses, interaction processes such as interpersonal and
small group processes are studied;10 in educational studies, students drop out
of school before completing their degrees, enter into a specific educational
track, or start a program of further education later in life;11 in analyses of
ethnic conflict, incidences of racial and ethnic confrontation, protest, riot,
2 Introduction 2
12
and attack are studied; in socialpsychological studies, aggressive responses
are analyzed;13 in psychological studies, human development processes are
studied;14 in psychiatric analyses, people may show signs of psychoses or neuro-
ses at a specific age;15 in social policy studies, entry to and exit from poverty,
transitions into retirement, or the changes in living conditions in old age
are analyzed;16 in medical and epidemiological applications, patients switch
between the states “healthy” and “diseased” or go through various phases of
an addiction career;17 and so on.
Technically speaking, in all of these diverse examples, units of analysis
occupy a discrete state in a theoretically meaningful state space, and transi-
tions between these states can occur virtually at any time.18 Given an event
history data set, the typical problem of the social scientist is to use appropri-
ate statistical methods for describing this process of change, to discover the
causal relationships among events, and to assess their importance.
This book was written to help the applied social scientist achieve these
goals. In this introductory chapter, we first discuss different observation
plans and their consequences for causal modeling. We also summarize the
fundamental concepts of event history analysis and show that the change
in the transition rate is a natural way to represent the causal effect in a sta-
tistical model. The remaining chapters are organized as follows:

• Chapter 2 describes event history data sets and their organization. It also
shows how to use such data sets with Stata.
• Chapter 3 discusses basic nonparametric methods used to describe event
history data, mainly the life table and the Kaplan-Meier (product-limit)
estimation methods as well as cumulative incidence functions, and finally
an appropriate nonparametric method for processes with competing risks.
• Chapter 4 deals with the basic exponential transition rate model.
Although this very simple model is almost never appropriate in practical
applications, it serves as an important starting point for all other tran-
sition rate models.
• Chapter 5 describes a simple generalization of the basic exponential
model, called the piecewise constant exponential model. In our view, this is
one of the most useful models for empirical research, and we devote a full
chapter to discussing it.
• Chapter 6 discusses time-dependent covariates. The examples are restricted
to exponential and piecewise exponential models, but the topic–and part
of the discussion–is far more general. In particular, we introduce the
problem of how to model parallel and interdependent processes.
• Chapter 7 introduces a variety of models with a parametrically specified
duration-dependent transition rate, in particular the Gompertz–Makeham,
Weibull, log-logistic, and log-normal models.
• Chapter 8 discusses the question of goodness-of-fit checks for parametric
transition rate models. In particular, the chapter describes simple graphical
checks based on transformed survivor functions and generalized residuals.
3 Introduction 3
• Chapter 9 introduces basic and advanced semiparametric transition rate
models based on an estimation approach proposed by D. R. Cox (1972).
The advanced part on Cox models describes fixed effects models for
multiepisode data.
• Chapter 10 discusses problems of model specification, in particular,
transition rate models with unobserved heterogeneity. The discussion on
specifying more than one error term on the same level is mainly critical,
and examples are restricted to using a gamma mixing distribution. This
chapter also provides a short introduction to transition rate models with
random coefficients in a multilevel framework.
• Chapter 11 by Brendan Halpin introduces sequence analysis as a com-
plementary alternative to event history analysis, with exploratory and
descriptive advantages. Sequence analysis focuses on longitudinal data
such as life course trajectories as wholes and calculates distances between
sequences. These distances can be used to create data-based typologies
using cluster analysis to calculate distance to theoretically or empirically
defined reference sequences, to compare trajectory diversity across vari-
ables such as cohort, to compare similarity of dyadic pairs of sequences
such as couples’ time use, and so on.
• The Appendix contains many exercises designed to help the reader gain
a deeper understanding of the (basic) concepts, more familiarity with
how to handle event history data and estimation commands in Stata, and
greater skill in estimating the results.

1.1 Causal modeling and observation plans

In event history modeling, design issues regarding the type of substantive
process are of crucial importance. It is assumed that the methods of data anal-
ysis (e.g., estimation and testing techniques) cannot depend on the particular
type of data alone (e.g., cross-sectional data, panel data, etc.) as has been the
case when applying more traditional statistical methodologies. Rather, the
characteristics of the specific kind of social process itself must “guide” both
the design of data collection and the way these data are analyzed and interpreted
(Coleman, 1973, 1981, 1990).
Different observation plans have been used to collect data generated by a
continuous-time, discrete-state substantive process (Coleman, 1981; Tuma &
Hannan, 1984). With regard to the extent of detail on the process of change,
one can distinguish between cross-sectional data, panel data, event count
data, event sequence data, and event history data.
In this book, we do not treat event count data (see, e.g., Hannan &
Freeman, 1989; Olzak, 1992; Andersen et al., 1993; Barron, 1993; Olzak &
Shanahan, 1996; Minkoff, 1997; Olzak & Olivier, 1998a, 1998b). These
simply record the number of different types of events for each unit (e.g.,
the number of upward, downward, or lateral moves in the employment
4 Introduction 4

State space a) Cross-sectional sample

Married •

Consensual union
Single

t2 Time t
State space b) Panel with 4 waves
Married • •

Consensual union •
Single

t1 t2 t3 t4 Time t
State space c) Event-oriented design
Married
••
••
•••
Consensual union
Single •••
••
•••
t4 Time t
Figure 1.1.1 Observation of an individual’s family career on the basis of a cross-
sectional survey, a panel study, and an event-history-oriented design.

career during a period of 10 years). Event sequence data (see, e.g., Rajulton,
1992; Abbott, 1995; Halpin & Chan, 1998) that document sequences of
states occupied by each unit are only used in Chapter 11. In this section,
we concentrate our discussion on cross-sectional and panel data as the main
standard sociological data types (Tuma & Hannan, 1984) and compare
these with event history data. For example, as shown in Figure 1.1.1, an indi-
vidual’s family career is observed in a cross-sectional survey, a panel survey,
and an event-oriented survey.

1.1.1 Cross-sectional data

Let us first discuss cross-sectional observations. In the social sciences, this is
the most common form of data used to assess sociological hypotheses. The
family history of the individual in Figure 1.1.1 is represented in a cross-
sectional study by one single point in time: her or his marital state at the
time of interview. Thus, a cross-sectional sample is only a “snapshot” of
the substantive process being studied. Normally, the point in time when
researchers take that “picture” is not determined by hypotheses about
the dynamics of the substantive process itself, but by external considerations
such as acquiring research funds, finding an appropriate institute to conduct
the survey, and so on.
5 Introduction 5
Coleman (1981) has demonstrated that one must be cautious when
drawing inferences about the effects of explanatory variables in logit
models on the basis of cross-sectional data because, implicitly or explicitly,
social researchers have to assume that the substantive process under study
is in some kind of statistical equilibrium. Statistical equilibrium, steady-state,
or stability of the process mean that, although individuals (or any other unit
of analysis) may change their states over time, the state probabilities are
fairly trendless or stable. Therefore, an equilibrium of the process requires
that the inflows to and the outflows from each of the discrete states are, to
a large extent, equal over time. Only under such time-stationary conditions
is it possible to interpret the estimates of logit and log-linear analyses, as dem-
onstrated by Coleman (1981).
Even if the assumption of a steady state is justified in a particular applica-
tion, the effect of a causal variable in a logit and/or log-linear model should
not be taken as evidence that it has a particular effect on the substantive
process (Coleman, 1981; Tuma & Hannan, 1984). This effect can have an
ambiguous interpretation for the process under study, because causal variables
often influence the inflows to and the outflows from each of the discrete
states in different ways. For example, it is well known that people with
higher educational attainment have a lower probability of becoming poor
(e.g., receive social assistance), but at the same time educational attainment
obviously has no significant effect on the probability of escaping poverty (see,
e.g., Leisering & Leibfried, 1998; Leisering & Walker, 1998; Zwick, 1998).
This means that the causal variable educational attainment influences the
poverty process in a specific way: It decreases the likelihood of inflows into
poverty, but it has no impact on the likelihood of outflows from poverty.
Given that the poverty process is in a steady state, a logit and/or log-linear
analysis of cross-sectional data tells us only the difference in these two direc-
tional effects on the poverty process (Coleman, 1981). In other words, cross-
sectional logit and/or log-linear models can show only the net effect of the
causal variables on the steady state distribution and that can be misleading,
as the following example demonstrates.
Consider that we are studying a process with two states (“being unem-
ployed” and “being employed”) that are in equilibrium (i.e., the unemploy-
ment rate is trendless over time), and let us further assume that the
covariate “educational attainment” increases the probability of movement
from unemployment to employment (UE ! E) and increases the probability
of movement from employment to unemployment (E ! UE) for each indi-
vidual. In a cross-sectional logistic regression analysis using the probability of
being employed as the dependent variable, the estimated coefficient for “edu-
cational attainment” tells us only the net effect of both directional effects.
Therefore, if the positive effect of educational attainment on UE ! E
offsets the positive effect on E ! UE, the net effect of “educational attain-
ment” in the logistic regression on the steady-state probability will be
about zero and not significant. That is, a zero effect of a covariate in a
6 Introduction 6
cross-sectional logistic regression analysis could mean two very different
things: that there is no effect at all of the respective covariate on UE ! E
and on E ! UE or that the directional effects on UE ! E and on E !
UE offset each other. Thus, a nonsignificant effect in the cross-sectional
logit model should not be taken as evidence that a variable is irrelevant to
the process, but only that it has no net effect on the equilibrium distribution
(Tuma & Hannan, 1984).
Similarly, if the net effect of “educational attainment” in a cross-sectional
logistic regression on the probability of becoming employed is positive, then,
in principle, the following four different interpretations are possible: (1) that
the positive effect on UE ! E is greater than the positive effect on E ! UE,
(2) that the negative effect on UE ! E is smaller than the negative effect on
E ! UE, (3) that there is only a positive effect on UE ! E and no effect on
E ! UE, and (4) that there is no effect on UE ! E and only a negative effect
on E ! UE. Conversely, for negative effects in the cross-sectional logistic
regression, the four interpretations have to be reversed.
If there is no equilibrium in the process, however, cross-sectional
coefficients may not only be ambiguous but also present a completely
misleading picture. In a study on unemployment incidence, Rosenthal
(1991), for example, demonstrated how confusing cross-sectional estimates
can be if the proportion of people being unemployed increases or decreases
in a specific region and if the process of change is, therefore, not in
equilibrium.
In the social sciences, one can expect that stability is very rare. For
example, life history studies (Blossfeld, 1989, 1995; Mayer, 1990; Blossfeld &
Hakim, 1997; Blossfeld & Drobnič, 2001; Blossfeld & Müller, 2002/2003;
Blossfeld & Timm, 2003) show that change across age, cohort, and histor-
ical period is an enduring and important feature in all domains of modern
individuals’ lives (Mayer & Tuma, 1990); organizational studies demon-
strate that most social organizations seem to follow a program of growth
and not of stability (Carroll & Hannan, 2000); and most modern societies
reveal an accelerating rate of change in almost all of their subsystems (cf.
the rapid changes in family systems, job structures, educational systems,
etc.; see Mayer, 1990; Heinz, 1991a, 1991b, 1992; Huinink et al., 1995;
Blossfeld & Hakim, 1997; Leisering & Leibfried, 1998; Blossfeld &
Drobnič, 2001; Blossfeld & Müller, 2002/2003; Blossfeld & Timm, 2003).
But even in areas considered to be fairly stable, one must ask the crucial
methodological question: To what extent is the process under study close
to an equilibrium (Tuma & Hannan, 1984)? This question can be answered
only if longitudinal data are applied because only this type of data can indi-
cate whether a steady state actually exists or how long it will take until a
system returns to a new equilibrium after some external upheaval.
Beyond the crucial assumption of process stability, cross-sectional data
have several other inferential limitations with regard to causal modeling.
We want to address at least some of the more important problems here.19
7 Introduction 7
Direction of causality. In only a few situations, the direction of causality can
be established on the basis of cross-sectional data (Davies, 1987). For example,
consider the strong positive association between parental socioeconomic char-
acteristics and educational attainment of sons and daughters, controlling for
other important influences (Shavit & Blossfeld, 1993; Erikson & Jonsson,
1996; Blossfeld et al., 2014, 2016a, 2016b, 2017). A convincing interpretation
of this effect might be that being born into a middle-class family increases the
likelihood of attaining a university degree, because one is unable to think of
any other plausible explanation for the statistical association. However, such
recursive relationships in which all the causal linkages run “one way” and
have no “feedback” effects, are rare in social science research. For example,
there is very often an association between the age of the youngest child and
female labor force participation in modern industrialized societies (Blossfeld
& Hakim, 1997; Blossfeld, Drobnič, & Rohwer, 1998; Blossfeld & Drobnič,
2001). The common interpretation is that there is a one-way causality with
young children tending to keep mothers at home. However, it is quite possible
that the lack of jobs encourages women to enter into marriage and mother-
hood, suggesting a reversed relationship (Davies, 1987).
The ambiguity of causation seems to be particularly important for the mod-
eling of the relationship between attitudes and behavior. There are two interest-
ing aspects of this relationship: There is a direct effect in which behavior
affects attitudes, and there is a “feedback” process in which attitudes
change behavior (Davies, 1987).20 The well-known disputes among sociolo-
gists as to whether value change engenders change in social behavior or
whether structural change in behavior leads to changing values of individuals
often originate from the fact that cross-sectional surveys are used that can
assess only the net association of these two processes.
Various strengths of reciprocal effects. Connected with the inability
to establish the direction of causality in cross-sectional surveys is the drawback
that these data cannot be used to discover the different strengths of reciprocal
effects. For example, many demographic studies have shown that first marriage
and first motherhood are closely interrelated (Blossfeld & Huinink, 1991;
Blossfeld et al., 1999; Blossfeld & Mills, 2001). To understand what has
been happening to family formation in modern societies, it might be of inter-
est to know not only the effect of marriage on birth rates, but also the effect of
pregnancy or first birth on getting married (Blossfeld & Huinink, 1991; Bloss-
feld, 1995; Blossfeld et al., 1999; Blossfeld & Mills, 2001; Mills & Trovato,
2001), and, perhaps, how these effects have changed over historical time
(Manting, 1994, 1996).
Observational data. Most sociological research is based on nonexperimental
observations of social processes, and these processes are highly selective. For
example, in a study examining the influence of type of school (private vs.
public) on test performance among students, Lieberson (1985) distinguished
at least three types of nonrandom processes: (1) self-selectivity, in which the
8 Introduction 8
units of analysis sort themselves out by choice (e.g., specific students choose
specific types of schools); (2) selective assignment by the independent variable
itself, that determines, say, what members of a population are exposed to spe-
cific levels of the independent variable (e.g., schools select their students
based on their past achievement); and (3) selectivity due to forces exogenous
to variables under consideration (socioeconomic background, ethnicity,
gender, previous school career, changes in intelligence over time, etc.).
Many of these sources are not only not observed but also effectively unmeasur-
able. Of course, no longitudinal study will be able to overcome all the prob-
lems in identifying these various effects; however, cross-sectional data offer
the worst of all opportunities to disentangle the effects of the causal
factors of interest on the outcome from other forces operating at the same
time, because these data are least informative about the process of change.
Cross-sectional analysis therefore requires a particularly careful justification,
and the substantive interpretation of results must always be qualified appro-
priately (Davies, 1987; Pickles & Davies, 1989).

Previous history. One aspect of observational data deserves special attention

in the social sciences. Life courses of individuals (and other units of analysis
such as organizations, etc.) involve complex and cumulative time-related
layers of selectivity (Mayer & Müller, 1986; Mayer & Schöpflin, 1989;
Mayer & Tuma, 1990; Mayer, 1991; Huinink et al., 1995; Mayer & Baltes,
1996). Therefore, there is a strong likelihood that specific individuals have
been entering a specific origin state (see, e.g., the discussion in Blossfeld et
al., 1999 on consensual unions). In particular, life-course research has shown
that the past is an indispensable factor in understanding the present (Buch-
mann, 1989; Mayer, 1990; Heinz, 1991a, 1991b, 1992; Allmendinger, 1994;
Huinink et al., 1995; Weymann, 1995; Weymann & Heinz, 1996). Cross-
sectional analysis may be performed with some proxy-variables and with
assumptions about both the causal order and the interdependencies between
the various explanatory variables. However, it is often not possible to appropri-
ately trace back the time-related selective processes operating in the previous
history, because these data are simply not available. Thus, the normal control
approaches in cross-sectional statistical techniques will rarely succeed in isolat-
ing the influence of some specific causal force (Lieberson, 1985).

Age and cohort effects. Cross-sectional data cannot be used to distinguish

between age and cohort effects (Tuma & Hannan, 1984; Davies, 1987).
However, in many social science applications it is of substantive importance
to know whether the behavior of people (e.g., their tendency to vote for a spe-
cific party) differs because they belong to different age groups or because
they are members of different birth cohorts (Blossfeld, 1986, 1989; Mayer &
Huinink, 1990).

Historical settings. Cross-sectional data are also not able to take into account
that processes emerge in particular historical settings. For example, in addition
9 Introduction 9
to individual resources (age, education, labor force experience, etc.), there are
at least two ways in which a changing labor market structure affects career
opportunities. The first is that people start their careers in different structural
contexts. It has often been assumed that these specific historical conditions
at the point of entry into the labor market have an impact on people’s subse-
quent careers. This kind of influence is generally called a cohort effect (Glenn,
1977). The second way that a changing labor market structure influences career
opportunities is that it improves or worsens the career prospects of all people
within the labor market at a given time (Blossfeld, 1986). For example, in a
favorable economic situation with low unemployment, there will be a relatively
wide range of opportunities. This kind of influence is generally called a period
effect (Mason & Fienberg, 1985). With longitudinal data, Blossfeld (1986) has
shown that life-course, cohort, and period effects can be identified based on
substantively developed measures of these concepts (see also Rodgers, 1982,
as well as Becker & Blossfeld, 2017) and that these effects represent central
mechanisms of career mobility that must be distinguished in any valid account.

Multiple clocks, historical eras, and point-in-time events. From a theoreti-

cal or conceptual point of view, multiple clocks, historical eras, and point-in-
time events very often influence the substantive process being studied (Mayer
& Tuma, 1990). For example, in demographic studies of divorce, it is impor-
tant to consider types of clocks, such as age of respondent, time of cohabita-
tion, duration of marriage, ages of children, different phases in the state of
the business cycle, or changes in national (divorce) laws, as important (Bloss-
feld et al., 1995; Blossfeld & Müller, 2002/2003). With respect to cross-sec-
tional data, such relationships can hardly be studied without making strong
untestable assumptions.

Contextual processes at multiple levels. Social scientists are very often inter-
ested in the influences of contextual processes at multiple aggregation levels
(Huinink, 1989). Contextual process effects refer to situations in which
changes in the group contexts themselves influence the dependent variable.
For example, the career mobility of an individual may be conceptualized as
being dependent on changes in resources at the individual level (e.g., social
background, educational attainment, experience, etc.); the success of the
firm in which she or he is employed (e.g., expansion or contraction of the
organization) at the intermediate level; and changes in the business cycle at
the macro level (Blossfeld, 1986; DiPrete, 1993; Becker & Blossfeld, 2017).
Cross-sectional data do not provide an adequate opportunity to study such
influences at different levels (Mayer & Tuma, 1990).

Duration dependence. Another problem with cross-sectional data is that they

are inherently ambiguous when it comes to their interpretation at the level of the unit
of observation. Suppose we know that 30.6% of employed women in West
Germany were working part-time in 1970 (Blossfeld & Rohwer, 1997). At
the one extreme, this might be interpreted as implying that each employed
10 Introduction 10
woman had a 30.6% chance of being employed part-time in that year. But at
the other extreme, one could infer that 30.6% of the employed women always
worked part-time and 69.4% were full-timers only. In other words, cross-
sectional data do not convey information about the time women spent in
these different employment forms. They are therefore open to quite different sub-
stantive interpretations (Heckman & Willis, 1977; Flinn & Heckman, 1982;
Blossfeld & Hakim, 1997; Blossfeld & Drobnič, 2001). In the first case, each
woman would be expected to move back and forth between part-time and
full-time employment. In the second, there is no mobility between the two
types of work, and the estimated percentages describe the proportions of
two completely different groups of employed women. From an analytical
point of view, it is therefore important to have data about durations in a
state. Also, repeated cross-sectional analyses using comparable samples of
the same population (e.g., a series of microcensuses or cross-sectional
surveys) can show only net change, but not the flow of individuals.
Variability in state dependencies. In many situations, cross-sectional data
are problematic because the rate of change is strongly state dependent and
entries into and exits from these states are highly variable over time (e.g., over
the life course and historical period or across cohorts). For example, it is
well known that the roles of wife and mother (the latter in particular)
have been central in women’s lives. Therefore, the family cycle concept has fre-
quently been used in sociology to describe significant changes in the circum-
stances that affect the availability of women for paid work outside the home.
The basic idea is that there is a set of ordered stages defined primarily by var-
iations in family composition and size that could be described with cross-
sectional data. However, this view often leads to the tendency to assume
that what happens to different women at one point in time in various
phases in the family cycle is similar to the pattern that women experience
when they make these transitions in different historical times (which has
been called the life course fallacy). Moreover, there is the well-known
problem that individuals and families often fail to conform to the assumption
of a single progression through a given number of stages in a predetermined
order (see, e.g., Blossfeld & Hakim, 1997; Blossfeld & Drobnič, 2001).
There may be at least three reasons for this (Murphy, 1991): (1) The chronol-
ogy of the timing of events may not conform to the ideal model (e.g., child-
bearing may start before marriage); (2) many stages are not reached (e.g.,
by persons who never marry); and (3) the full set of stages may be truncated
by events such as death or marital breakdown. Such complex constellations
between the family cycle and women’s labor-force participation could hardly
be described or studied meaningfully on the basis of cross-sectional data.
Changes in outcomes. Cross-sectional models very often have a tendency
to overpredict change and consistently overestimate the importance of explanatory
variables (Davies, 1987). The reason for this phenomenon is that these data
analyses cannot be based on how changes in explanatory variables engender
11 Introduction 11
changes in outcomes. They are concerned only with how levels of explanatory
variables “explain” an outcome at a specific point in time. However, if an
outcome at time t (e.g., choice of mode of travel to work in June) is depen-
dent on a previous outcome (e.g., established choice of mode of travel to
work), and if both outcomes are positively influenced in the same way by
an explanatory variable (e.g., merits of public transport), then the effect of
the explanatory variable will reflect both the true positive influence of the
explanatory variable on the outcome at time t and a positive spurious
element due to that variable acting as a proxy for the omitted earlier
outcome (established mode of travel to work). Hence, a cross-sectional anal-
ysis of the travel to work choice (e.g., public vs. private transport) would tend
to overpredict the effect of policy changes (e.g., fare increases or faster buses)
because there is a strong behavioral inertia (Davies, 1987).
In sum, all these examples show that cross-sectional data have many severe
inferential limitations for social scientists. Therefore, it is not surprising that
causal conclusions based on cross-sectional data have often been altered radically
after the processes have been studied with longitudinal data (Lieberson, 1985).
Longitudinal studies also have a much greater power than cross-sectional
ones, both in estimating bias from missing data and in the means for correct-
ing it. This is because longitudinal studies often have data from previous
points in time, thus enabling the characteristics of nonresponders or lost
units to be assessed with some precision. It is noteworthy that almost all
the substantive knowledge concerning the biases associated with missing
data, which all studies must seek to minimize, is derived from longitudinal
research (Medical Research Council, 1992).
Although longitudinal data are no panacea, they are obviously more effec-
tive in causal analysis and have fewer inferential limitations (Mayer, 1990;
Mayer & Tuma, 1990; Magnusson, Bergmann, & Törestad, 1991; Clogg &
Arminger, 1993; Arminger, Clogg, & Sobel, 1995; Blossfeld, 1995; Blossfeld
& Hakim, 1997; Carroll & Hannan, 2000; Blossfeld & Drobnič, 2001).
They are indispensable for the study of processes over the life course (of all
types of units) and how these relate to historical change. Therefore, research
designs aimed at a causal understanding of social processes need to be based
on longitudinal data at the level of the units of analysis.

1.1.2 Panel data

The temporal data most often available to sociologists are panel data. In panel
studies, the same persons or units are reinterviewed or observed at a series of
discrete points in time (Chamberlain, 1984; Hsiao, 1986; Arminger & Müller,
1990; Engel & Reinecke, 1994; Blossfeld et al., 2016c). Figure 1.1.1 shows a
four-wave panel in which the family career of the respondent was observed
at four different points in time. This means that there is only information
on states of the units at predetermined survey points, but the course of the
events between the survey points remains unknown.
12 Introduction 12
Panel data normally contain more information than cross-sectional data
but involve well-known distortions created by the method itself (see, e.g.,
Hunt, 1985; Magnusson & Bergmann, 1990).
Panel bias. Respondents often answer the same questions differently in the
second and later waves than they did the first time because they are less
inhibited, or they have mulled over or discussed the issues between question-
ing dates.
Modification of processes. Panel studies tend to influence the very phenom-
ena they seek to observe—this sometimes changes the natural history of the
processes being observed.
Attrition of the sample. In panel studies the composition of the sample nor-
mally diminishes selectively over time. Usually, these processes are particu-
larly strong during the first panel waves and then gradually diminish.
Therefore, what researchers observe in a long-term panel may not provide
a good picture of what has actually happened to the process under study.
Nonresponses and missing data. In a cross-sectional analysis, one can afford
to throw out a small number of cases with nonresponses and missing data, but
in a long-term panel study, throwing out incomplete cases at each round of
observation can eventually leave a severely pruned sample that will have very
different characteristics from the original one.
Fallacy of cohort centrism. Very often, panel studies focus on members of a
specific cohort (cf., e.g., the British National Child Development Study). In
other words, these panels study respondents who were born in, grew up in,
and have lived in a particular historical period. This creates a danger that
researchers might assume that what happens to a particular group of
people over time reveals general principles of the life course (fallacy of
cohort centrism). Many events may simply be specific to that generation.
Fallacy of period centrism. Many panel studies include just a few waves and
cover only a short period of historical time. At the time of these observations,
special conditions may have existed (e.g., high unemployment) and this can
result in individuals responding differently than they would under different
historical conditions (fallacy of historical period).
Confounded age, period, and cohort effects. In any long-term panel study in
sociology, three causal factors—individual’s age, cohort, and historical period
effect—are confounded (cf. the Panel Study of Income Dynamics). Method-
ological techniques are needed to unconfound these three factors and reveal
the role of each. As discussed in more detail later, panel data do have some
specific problems unconfounding the three major factors. However, in order
to gain scientific insights into the interplay of processes governing
life courses from birth to death, they appear to be a better approach than
applying cross-sections. But a mechanical and atheoretical cohort analysis is
13 Introduction 13
a useless exercise, and statistical innovations alone will not solve the age-
period-cohort problem (Blossfeld, 1986; Mayer & Huinink, 1990).
Most of the previously mentioned difficulties concerning panel studies
can be overcome with better data collection methods, more sophisticated sta-
tistical procedures, or more panel waves. However, panel data also lead to a
series of deficiencies with respect to the estimation of transition rates (Tuma
& Hannan, 1984): First, there is the problem of “embeddability,” meaning
that there may be difficulties in embedding a matrix of observed transition
probabilities within a continuous-time Markov process (Singer & Spilerman,
1976a); second, another problem may be that there may be no unique matrix
of transition rates describing the data (Singer & Spilerman, 1976b); and
third, there is the drawback that the matrix of transition probabilities may
be very sensitive to sampling and measurement error (Tuma & Hannan,
1984). Multiple waves with irregular spacing or shorter intervals between
waves can reduce these problems. However, as Hannan and Tuma (1979)
have noted, this only means that the more panel and event history data
resemble each other, the less problematic modeling becomes.
Lazarsfeld (1948, 1972) was among the first sociologists to propose panel
analysis of discrete variables. In particular, he wanted to find a solution to
the problem of ambiguity in causation. He suggested that, if one wants to
know whether a variable X induces change in another variable Y, or
whether Y induces change in X, observations of X and Y at two points in
time would be necessary. Lazarsfeld applied this method to dichotomous var-
iables whose time-related structure he analyzed in a resulting sixteen-fold
table. Later on, Goodman (1973) applied log-linear analysis to such tables.
For many years, such a cross-lagged panel analysis for qualitative and quan-
titative variables (Campbell & Stanley, 1963; Shingles, 1976) was considered
to be a powerful quasi-experimental method for making causal inferences. It
was also extended to multiwave-multivariable panels to study more complex
path models with structural-equation modeling (Jöreskog & Sörbom, 1993).
However, it appears that the strength of the panel design for causal inference
was hugely exaggerated (Davis, 1978). Causal inferences in panel approaches
are much more complicated than has been generally realized. There are
several reasons for this.

Time until the effect starts to occur. It is important to realize that the role of
time in causal explanations does not just lie in specifying a temporal order in
which the effect follows the cause in time. It also implies that there is a tem-
poral interval between the cause and its impact (Kelly & McGrath, 1988). In
other words, if the cause has to precede the effect in time, it takes some finite
amount of time for the cause to produce that effect. The time interval may be
very short or very long, but it can never be zero or infinity (Kelly &
McGrath, 1988). Some effects take place almost instantaneously. For
example, if the effect occurs at microsecond intervals, then the process
must be observed within these small time units to uncover causal relations.
14 Introduction 14
However, some effects may occur in a time interval too small to be measured
by any given methods, so that cause and effect seem to occur at the same point
in time. Apparent simultaneity is often the case in those social science appli-
cations in which basic observation intervals are relatively crude (e.g., days,
months, or even years), such as, for example, yearly data about first marriage
and first childbirth (Blossfeld, Manting, & Rohwer, 1993). For these parallel
processes, the events “first marriage” and “first childbirth” may be interdepen-
dent, but whether these two events are observed simultaneously or succes-
sively depends on the degree of temporal refinement of the scale used in
making the observations. Other effects need a long time until they start to
occur. Thus, there is a delay or lag between cause and effect that must be spec-
ified in an appropriate causal analysis. However, in most current sociological
theories and interpretations of research findings, this interval is left unspec-
ified. In nearly all cases, reasearchers assume at least implicitly, that the effect
takes place almost immediately and is then constant (Figure 1.1.2b). Of
course, if this is the case, then there seems to be no need for theoretical state-
ments about the time course of the causal effect. A single measurement of the
effect at some point in time after a cause has been imposed might be sufficient
to catch it (see Figure 1.1.2a). However, if there is a reason to assume that
there is a lag between cause and effect, then a single measurement of the
outcome is inadequate for describing the process (see Figure 1.1.2b), and
the interpretation based on a single measurement of the outcome will then
be a function of the point in time chosen to measure the effect (the substan-
tial conclusion based on p3 or p4 in the figure would obviously be different).
Therefore, a restrictive assumption of panel designs is that either cause and
effect occur almost simultaneously or the interval between observations is
of approximately the same length as the true causal lag. The greater the dis-
crepancy, the greater the likelihood that the panel analysis will fail to dis-
cover the true causal process. Hence, as expressed by Davis (1978), if one
does not know the causal lag exactly, panel analysis is of little use when it
comes to establishing the causal direction or time sequencing of causal
effects. Unfortunately, we rarely have enough theoretically grounded argu-
ments about the structure of a social process to specify the lags precisely.

Temporal shapes of the unfolding effect. In addition to the issue of the

length of the delay between the timing of the cause and the beginning of
the unfolding of the effect, it is also possible that the way the effect develops
in time takes different shapes. Although the problem of time lags is widely rec-
ognized in social science literature, considerations on the temporal shape of
the effect are quite rare (Kelly & McGrath, 1988). In fact, social scientists
seem to be quite ignorant of the fact that causal effects could be highly
time-dependent.
Figure 1.1.2 illustrates several possible shapes these effects may trace over
time. In Figure 1.1.2a, an almost simultaneous change in the effect occurs
that is then maintained. In Figure 1.1.2b, the effect occurs with some
15 Introduction 15

Effect a) Effect occurs almost immediately and is then

of x on y time-constant.

0
p1 tx p2 p3 p4 Time t
Effect b) Effect occurs with a certain time lag and is then
of x on y time-constant.

0
p1 tx p2 p3 p4 Time t
Effect c) Effect occurs almost immediately and then
of x on y increases continuously.

0
p1 tx p2 p3 p4 Time t
Effect d) Effect occurs almost immediately, rises monotonically
of x on y at first, then declines, and finally disappears.

0
p1 tx p2 p3 p4 Time t
Effect e) Effect occurs almost immediately and oscillates
of x on y over time.

0
p1 tx p2 p3 p4 Time t
Figure 1.1.2 Different temporal lags and shapes of how a change in a variable x,
occurring at point in time tx, effects a change in a variable y.

lengthy time lag and is then time-invariant. In Figure 1.1.2c, the effect starts
almost immediately and then gradually increases. In Figure 1.1.2d, an almost
simultaneous increase reaches a maximum after some time before then decreas-
ing. Finally, in Figure 1.1.2e, a cyclical effect pattern over time is described.
If the effect increases or decreases monotonically or linearly, oscillates
in cycles, or shows any other complicated time-related pattern, then the
strength of the observed effect in a panel study will depend on the timing
of the panel waves. A panel design might be particularly problematic if
16 Introduction 16
there are nonmonotonic cycles of the effect, because totally opposite conclu-
sions about the effects of the explanatory variable can be arrived at depend-
ing on whether the panel places measurement points at a peak or at an ebb in
the curve (see Figures 1.1.2d and 1.1.2e).
Reciprocal effects with different time paths. In cases of reciprocal causality,
additional problems will arise in panel studies if the time structure of the
effects of X1 on X2 and of X2 on X1 differ in terms of lags and shapes. In
these situations, a panel design might turn out to be completely useless for
those wishing to detect such time-related recursive relationships.
Observational data and timing of measurement of explanatory variables.
Most sociological research is based on observational data, meaning that
manipulation of the timing of the independent variables is generally not
possible. For example, if a researcher is going to study the effects of job
mobility on marriage behavior, it is impossible to force respondents to
change their jobs, say, at the time of the first panel wave. Thus, the longer
the interval between panel waves, the more uncertainty there will be regard-
ing the exact point in time when an individual moved to another job and
therefore regarding the point we evaluate in the time path of the effect
(Coleman, 1981). The situation may be even more problematic if changes
in independent variables are repeatable and several changes are possible
between two successive panel waves, as might be the case with job exits
observed in yearly intervals (cf. Sandefur & Tuma, 1987). In such panel
studies, even the causal order of explanatory and dependent events may
become ambiguous.
Observational data and the timing of control variables. Observational panel
studies take place in natural settings and therefore offer little control over
changes in other important variables and their timing. If these influences
change arbitrarily and have time-related effect patterns, then panel studies
are useless in disentangling the effect of interest from time-dependent
effects of other parallel exogenous processes.
Continuous changes of explanatory and control variables. In observational
studies, explanatory and control variables may not only change stepwise from
one state to another, but can often change continuously over time. For
example, individuals continuously age; constantly acquire general labor
force experience or job-specific human capital if they are employed (Blossfeld
& Huinink, 1991); are exposed to continuously changing historical condi-
tions (Blossfeld, 1986); are steadily changing their social relationships in
marital or consensual unions (Blossfeld et al., 1995; Blossfeld et al., 1999);
and so on. Even in cases when these continuous changes are not connected
with lags or time-related effect patterns, panel data reveal deficiencies in
their ability to detect time dependence in substantive processes. This is
why panel analyses are often unable to appropriately identify age, period,
and cohort effects (Blossfeld, 1986).
Other documents randomly have
different content
When swells in their little tootsies,
Little tootsies, suffer pain,
Unto him they bring their footsies,
Footsies, to put right again;

For they say, sir, "None but he, sir,

He, sir, understands the toe."
Earls and Dukes wait every day, sir,
Every day, sir, in a row.

This the history of Farley,

Doctor Farley, sir, M.D.,
Others—in the street of Harley—
Others like him there may be.

There's a moral to this story,

To this story, if you're wise:
If you'd win both wealth and glory,
Wealth and glory—Specialize.
 XXXVIII.
THAT OF JEREMIAH SCOLES, MISER.
I sing of joys, and junketings,
Of holly, and of such-like things;
I sing of merry mistletoe,
And,—pardon me,—I sing also
Of Jeremiah Scoles.
I sing of Mister Scoles because
So singular a man he was,
And had so very strange a way
Of celebrating Christmas Day—
Unlike all other souls.

Myself, I am a cheerful man,

Enjoying life as best I can.
At Christmas-time I love to see
The flow of mirth and jollity
About the festive board;
I love to dance, I try to sing;
On enemies, like anything,
At Christmas-time I heap hot coals,
But not so Jeremiah Scoles—
He loves a miser's hoard.

I chanced one year, on Christmas Day,

To call upon him, just to say
That we'd be very pleased to see
Him, if he'd care to come to tea.
I found him quite alone.
He sat before a fireless grate;
The room looked bare and desolate,
And he, unkempt, in dressing-gown,
Received me with an angry frown,
And spoke in surly tone.
"Ha! what d'ye want?" said he to me
And eyed me most suspiciously.
I laughed and gave a hearty smack
Upon the grumpy fellow's back,
And cried: "Come home with me.
We'll treat you well. There's lots of fun—"
But ere I scarcely had begun
He cut me short. "Pooh! folly! stuff!
See here; I've fun—quite fun enough!"
He laughed, but mirthlessly.

Before him on the table lay

Gold, silver, coppers, in array;
Some empty bottles; stacks of bills;
Some boxes for containing pills—
And that was all. Said he:
"This gold is what I haven't spent
In presents; and the silver's meant
To show what could be wasted in—
Pah!—Christmas boxes. 'Tis a sin
I don't encourage—no, not me?

"The coppers—little boys, no doubt,

Would like 'em—they may go without;
While these long bills I should have had
From tradesmen, had I been so mad
As to have bought the things
They represent for Christmas cheer;
These bottles and pill-boxes here
Show what I will not have to take,
Because I'll have no stomach-ache
That over-eating brings.
"And thus I spend my Christmas Day,
Thinking what silly fools are they
Who spend so much in solid cash
On so much sentimental trash.
And now, good-day to you!"
He showed me out, he banged the door,
And I was—where I was before.

***

I really think, upon my word,

His line of reasoning's most absurd.
No doubt you think so, too?
 XXXIX.
THAT OF THE HIGH-SOULED YOUTH.

A year or so ago, you know,

I had a friend, at Pimlico,
For want of better name called Joe
(This name is not his right 'un).
He was a sweet, poetic youth,
Romantic, gallant, and in sooth
Might well be called, in very truth
An "Admirable Crichton."

And oh! it grieved him sore to see

The lack,—these times,—of chivalry.
He'd now and then confide to me
His views upon the matter.
"Good, never now is done by stealth!"
He'd say, "Men ruin mind, and health
In sordid scramble after wealth;
And talk,—is idle chatter."
"That simple virtue, Modesty,
Alas! it now appears to be
A valueless commodity,
Though once men prized it highly."
He went on thus,—like anything,
Until I heard, one day last Spring,
That he intended marrying
The daughter of old Riley.

I knew the Riley girls, and thought

"Now this has turned out as it ought.
Joe is a reg'lar right good sort
To marry 'Cinderella.'"
The younger one, (thus called by me)
A sweet good girl as e'er might be
Was poor; the elder—rich was she—
Her name was Arabella.

An Aunt had left her lots of gold,

While 'Cinderella'—so I'm told,—
She left entirely in the cold
Without a single shilling.
The elder one,—though plain to see,—
Of suitors had some two, or three;
Poor Cinderella, nobody
To marry her seemed willing.

Until the noble high-souled Joe—

That Errant-knight of Pimlico—
Came forth, the world at large to show
That he at least knew better.
In spirit I before him bowed,
"To know a man like that I'm proud
And happy!" I remarked aloud,
And sent to him this letter.
"ARABELLA."
"Dear Joe;—Wealth as you say's a trap
Gold is but dross,—not worth a rap—
How very like you—dear old chap!—
To marry 'Cinderella.'"
***
He wrote:—"I must expostulate,
I'm not a FOOL at any rate!
Of course I've chosen as a mate
The RICH one, Arabella!"
 XL.
THAT OF MR. JUSTICE DEAR'S LITTLE JOKE AND
THE
UNFORTUNATE MAN WHO COULD NOT SEE IT.
Again of Mr. Justice Dear
My harmless numbers flowing,
Shall tell a story somewhat queer
About His Worship, showing,
How sensitive the legal wit.
It is. There is no doubt of it.

Before good Justice Dear one day

A man—for some small matter,
Was hailed, and, in his own sly way
(The former, not the latter)
Made,—and I thought the Court would choke,—
An unpremeditated joke.

The prosecuting Counsel roared,

The Jury giggled madly,
Only the Prisoner looked bored,
He took it rather sadly.
"Why don't you laugh?" the Usher said,
The Prisoner, he shook his head.

"I cannot see," said he, "that's flat—

A fact that's most annoying,—
What everyone is laughing at,
And seemingly enjoying."
This strange remark, it reached his ear
And irritated Justice Dear.

"When I am pleased to make a joke

That's not the way to treat it."
Thus, warningly, his Worship spoke,
"Now listen! I'll repeat it."
He did. He said it o'er and o'er.
At least a dozen times or more.
"Excuse me, sir," the Prisoner said,
"At what may you be driving?"
Good Justice Dear turned very red,
"This joke of my contriving,
If you don't see it, Sir, you ought;
If not—well—'tis contempt of Court."

The Counsel then explained it, but

Quite failed the point to show him;
The Usher muttered "Tut-tut-tut!"
The Jury whispered "Blow him!"
Then several people wrote it down.
The Prisoner still wore a frown.

"Am I supposed to laugh at that?

Why? I can't see the reason."
It was too much. His Lordship sat
Aghast. "'Tis almost treason!
That unpremeditated joke before
Has never failed to raise a roar.

"Defective in morality,
Must be that man misguided,
Who fails its brilliancy to see."
His Lordship then decided
To send the man,—despite his tears,—
To servitude, for twenty years.
 XLI.
THAT OF THE LADIES OF ASCENSION ISLAND.

On the Island of Ascension

There are only ladies ten,
The remaining population
Being officers or men.
"Dear me!" I hear you saying,
"How united they must be!"
But in this you'd be mistaken,
As you'll very quickly see.

For each lady on the Island

Thinks she ought to take the lead
In social matters, and on this
They're not at all agreed.
And Mrs. Smith's told Mrs. Brown
She thinks her most absurd,
While others cut each other dead
And don't exchange a word.
This state of thing's been going on
They tell me year by year,
And the husbands have grown tired of it
As we should do I fear;
For connubial felicity
Is doomed, if all our lives
Are spent in listening to the faults
Of other people's wives.

Quite recently a steamer called

For cinnamon and spice,
And her Captain and the officers
Were asked for their advice.
They gave it promptly. It was this—
"'Twere better you agreed,
In social matters, just to let
The eldest lady lead."
They tried it. But—good gracious!
They are worse off than before,
For every lady in the place
Is firm upon that score.
Impossible it is that age
Shall be the final test,
For every one insists that she
Is younger than the rest!
 XLII.
THAT OF THE ARTICULATING SKELETON.

There was a worthy Doctor once

Who unlike Mother Hubbard
Had many bones (a skeleton)
Shut up within a cupboard.

One night the worthy Doctor dreamt,

(He'd been up rather late)
His articulated skeleton
Did thus articulate:—
"Come! Doctor, come! confess that you're a fraud
A very specious humbug and a sham.
Though meek as any lamb.
Don't glare at me! I'll tell it not abroad
But merely in your ears alone applaud
The wily artifice of pill and dram.

"You know as well as I do, you don't mean,

One half the things you tell 'our patient.' No!
Why, I can clearly show,
That Mrs. Gobbles' ailments are but spleen,
('Tis quite the simplest cause that e'er was seen)
And yet what crack-jaw names you now bestow.

"Because, forsooth, the longer you can prey

Upon her pocket, that doth please you best,
So, Doctor, you protest
'The case is serious,' from day to day,
'And it must run its course,' you gravely say
With wisest head-shake and a look distressed.

"And then those pills! Absurd you know to try

To gammon me with bolluses of bread;
While Aqua P. I've said,
Often, is good (if nothing else be nigh)
To drink when thirsty and our throats are dry,
But not for medicine—though coloured red.

"So, Doctor, when we're by ourselves alone,

Don't try to put on 'side' with me, good lack,
For I can surely track
Full many a 'fatal case' you'd fain disown.
And I can tell aright why you should groan
When harmless ducks in passing cry 'Quack! Quack!'

***
The Doctor woke. "Dear me!" said he,
"This skeleton's too wise
For me." He therefore packed it up,
And sent it off to Guy's.
 XLIII.
THAT OF YE LOVE-PHILTRE: AN OLD-ENGLISH
LEGEND.
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

textbookfull.com