Temporal Change and the Process of European Union Decision-Making
| Author | Christopher Zorn |
| Date | 01 December 2007 |
| DOI | 10.1177/1465116507082815 |
| Published date | 01 December 2007 |
| Subject Matter | Articles |
European Union Politics
Temporal Change and the
DOI: 10.1177/1465116507082815
Volume 8 (4): 567–576
Process of European Union
Copyright© 2007
SAGE Publications
Decision-Making
Los Angeles, London, New Delhi
and Singapore
Christopher Zorn
Pennsylvania State University, USA
Introduction
Survival analysis – also variously referred to as ‘event history analysis’,
‘duration analysis’ and ‘reliability analysis’ – is an increasingly widespread
approach to the quantitative study of political phenomena. Models for
survival data can be traced to early work in biostatistics, and before that to
life tables developed in the actuarial sciences; the latter are traceable at least
to 18th-century work by Bernoulli, D’Alembert and their contemporaries.1 Yet
it was not until the 20th century that statisticians, particularly those interested
in engineering, operations research and medicine, began developing methods
for analyzing time-to-event data. The development of such models took a
two-pronged path, with scholars in engineering and other physical sciences
gravitating toward (and thus developing) parametric models, while those in
epidemiology and medicine focused to a far greater extent on Cox’s (1972)
semi-parametric approach.2
Since the initial application of survival models in the 1970s, their use in
the social and behavioral sciences has grown tremendously. That growth can
be traced to a number of related causes. Certainly increases in computing
power and the widespread availability of software to estimate such models
have played a role, as have the greater availability, reliability and extent of
time-to-event data in the social sciences. Equally important, however, has been
a renewed substantive focus on social and political processes, and in particular
on how preferences and institutions interact to shape decision-making.
The articles on EU decision-making in this and the forthcoming issue
of European Union Politics reflect that renewed focus, taking as their
phenomenon of interest the speed with which the EU passes legislation. The
5 6 7
5 6 8
European Union Politics 8(4)
authors offer several improvements – theoretical, methodological and
empirical – on existing studies, many of which revolve around the import-
ance of the ‘proportional hazards’ assumption common to such models.
Here, I shall address some aspects of those improvements and discuss how
analysts might integrate these insights with theories to develop and test more
complete models of EU decision-making.
Time-varying covariates
Whereas early studies of event histories in political science tended to adopt
parametric approaches, more recent work has increasingly moved toward
Cox’s (1972) proportional hazards model as the model of choice. As noted
above, that model was developed in the context of biostatistics and life-table
analysis in epidemiology. In such fields, the paradigmatic study is the double-
blind clinical trial: subjects with some medical condition are randomly
selected into treatment and control (placebo) groups, the treatment is admin-
istered, and the subjects are then followed and the time to the event of interest
(e.g. recurrence of the condition) is recorded. Such trials contemplate a rela-
tively simple model, where the hazard of the event of interest is a function
of a single, binary covariate:
h t
( ) = f {β0 + β X
i
1
i }.
(1)
In this model, X denotes the treatment indicator, coded 1 for treated indi-
viduals and 0 for those in the placebo group; the usual expectation is that
β1 as written here, the model in equation (1) is a general one, encompassing both
parametric and semi-parametric alternatives. Nonetheless, equation (1)
implies (among other things) that the value of the covariate X for a given
subject i is fixed in the study, i.e. either subjects receive the treatment (placebo)
only once, or those selected to receive the treatment (placebo) continue to
receive it over the life of the study.
The simplest extension of this model is the two-period crossover design
(Jones and Kenward, 2003: chapter 2). In this design, the follow-up is divided
into two periods: period 1 corresponds to the design above, whereas in period
2 the sample of subjects initially receiving the treatment is ‘switched’ to the
placebo, and vice versa.3 Crossover designs thus represent the most basic
form of ‘time-varying covariate’: each subject now has both X = 0 and
X = 1, with the value of X varying across the two periods. We can re-express
equation (1) to capture this variation by writing:
h t
( ) = f {β0 + β X
i
1
it },
(2)
Zorn
Temporal Change and the Process of European Union Decision-Making
5 6 9
where the new subscript t on X denotes that the value of the covariate now
depends both on the identity of the subject and on the time of measurement.
Importantly, the specification in (2) retains a key assumption of (1): that
the effect of the treatment on the hazard is invariant across the two periods.
Put differently, (2) assumes that the period in which the treatment is admin-
...
To continue reading
Request your trial