Time Matters: On Theory and Method / Edition 2

by Andrew Abbott

Overview

What do variables really tell us? When exactly do inventions occur? Why do we always miss turning points as they transpire? When does what doesn't happen mean as much, if not more, than what does? Andrew Abbott considers these fascinating questions in Time Matters, a diverse series of essays that constitutes the most extensive analysis of temporality in social science today. Ranging from abstract theoretical reflection to pointed methodological critique, Abbott demonstrates the inevitably theoretical character of any methodology.

Time Matters focuses particularly on questions of time, events, and causality. Abbott grounds each essay in straightforward examinations of actual social scientific analyses. Throughout, he demonstrates the crucial assumptions we make about causes and events, about actors and interaction and about time and meaning every time we employ methods of social analysis, whether in academic disciplines, market research, public opinion polling, or even evaluation research. Turning current assumptions on their heads, Abbott not only outlines the theoretical orthodoxies of empirical social science, he sketches new alternatives, laying down foundations for a new body of social theory.

Product Details

ISBN-13:	9780226001036
Publisher:	University of Chicago Press
Publication date:	07/15/2001
Series:	Oriental Institute Publications
Edition description:	1
Pages:	296
Product dimensions:	6.00(w) x 9.00(h) x 0.90(d)

About the Author

Andrew Abbott is the Gustavus F. and Ann M. Swift Distinguished Service Professor of Sociology at the University of Chicago. He is author of Chaos of Disciplines, Department and Discipline, and The System of Professions.

Read an Excerpt

TIME MATTERS
ON THEORY AND METHOD

---

By Andrew Abbott
The University of Chicago Press
Copyright © 2001 The University of Chicago
All right reserved.
ISBN: 978-0-226-00103-6

---

Chapter One
Transcending General Linear Reality

A growing chasm divides sociological theory and sociological research. While the general linear model and other new techniques have reshaped empirical work, renewed acquaintance with the classics has transformed theory. These contradictory transformations have bred acrimony. Some have sought to reduce social statistics to the status of a substantive theory, while others accuse theorists of proliferating vague alternatives. The debate has been taken up by interactionists, macro theorists, and many others over the years. But the split did not assume its current proportions until the challenging and once-laborious mathematics of linear and characteristic equations became computerized. Quantitative work has since come to dominate central disciplinary journals, while theoretical and qualitative work has increasingly founded its own journals and/or chosen book form.

In this paper I identify one intellectual source for disagreement between theorists and empiricists. I shall argue that there is implicit in standard methods a "general linear reality" (GLR), a set of deep assumptions about how and why social events occur, and that these assumptions prevent the analysis of many problems interesting to theorists and empiricists alike. In addition to delineating these assumptions, I shall consider alternative methods relaxing them. The paper closes with a brief discussion of three alternative sets of methodological presuppositions about social reality. Through this analysis, I aim not to renew pointless controversies, for I believe the general linear model (GLM) is a formidable and effective method. But I argue that the model has come to influence our actual construing of social reality, blinding us to important phenomena that can be rediscovered only by diversifying our formal techniques.

I General Linear Reality

The phrase "general linear reality" denotes a way of thinking about how society works. This mentality arises through treating linear models as representations of the actual social world. This representational usage can be opposed to the more cautious use of linear models in which the analyst believes that some substantive causal process logically entails patterns of relations between variables, patterns which can then be tested by that model to discover whether the actual state of affairs is consistent with the substantive mechanism proposed. These two uses will be called the representational and entailment uses, respectively. The discussion of section II will outline precisely what theoretical assumptions are implicit in representational usage. To begin the analysis, however, we must first sketch the mathematics of the model.

The general linear model makes some particular variable dependent on a set of antecedent variables up to an error term:

y = Xb + u (1)

Lowercase letters here represent vectors and uppercase ones matrices. The row dimension of y, u, and X is the number of cases observed (m), while the column dimension of X and row dimension of b is the number (n) of antecedent variables. We can disregard the constant term without loss.

In formal terms, the model is a linear transformation from [R.sup.n] into [R.sup.1]. The transformation itself makes no assumptions about causality or direction; any column of X can be interchanged with y if the appropriate substitution in b is made. Using the transformation to represent social causality, however, assumes that y occurs "after" everything in X. In cross-sectional application, use of the model postulates a "causal time" that takes the place of actual time.

That the range of the linear transformation has but one dimension is a constraint imposed by problems of estimation. One can easily conceive a general-form GLM:

[X.sub.t] = [X.sub.t-1] B+U (2)

Here the index embeds the variables in actual time. Each succeeding value of each variable reflects a unique mix of all the antecedents. B becomes a square matrix of dimension n, and the full transformation is thus from [R.sup.n] into [R.sub.n]. This more general GLM underlies most panel studies, although the relevant coefficients can be estimated only by deleting on theoretical grounds some fraction of the dependence this model postulates. Loosely, this second model envisions the situation as a school of fish (the cases) swimming in some regular pattern (the transformation) through a multidimensional lake (the variable or attribute space).

To use such a model to actually represent social reality one must map the processes of social life onto the algebra of linear transformations. This connection makes assumptions about social life: not the statistical assumptions required to estimate the equations, but philosophical assumptions about how the social world works. Such representational use assumes that the social world consists of fixed entities (the units of analysis) that have attributes (the variables). These attributes interact, in causal or actual time, to create outcomes, themselves measurable as attributes of the fixed entities. The variable attributes have only one causal meaning (one pattern of effects) in a given study, although of course different studies make similar attributes mean different things. An attribute's causal meaning cannot depend on the entity's location in the attribute space (its context), since the linear transformation is the same throughout that space. For similar reasons, the past path of an entity through the attribute space (its history) can have no influence on its future path, nor can the causal importance of an attribute change from one entity to the next. All must obey the same transformation.

There are, of course, ways of relaxing some of these assumptions within standard methods, all of them at substantial cost in interpretability. But it is striking how absolutely these assumptions contradict those of the major theoretical traditions of sociology. Symbolic interactionism rejects the assumption of fixed entities and makes the meaning of a given occurrence depend on its location-within an interaction, within an actor's biography, within a sequence of events. Both the Marxian and Weberian traditions deny explicitly that a given property of a social actor has one and only one set of causal implications. Marx's dialectical causality makes events produce an opposite as well as a direct outcome, while Weber and the various hermeneutic schools treat attributes as infinitely nuanced and ambiguous. Marx, Weber, and work deriving from them in historical sociology all approach social causality in terms of stories, rather than in terms of variable attributes. To be sure, Marx and Weber discuss variable attributes in some of their purely conceptual writing, but their most currently influential works are complex stories in which attributes interact in unique ways-the Protestant Ethic, the General Economic History, the Eighteenth Brumaire, and even much of Capital.

The contrast between these assumptions and those of GLR suggests that theorists may reject empirical sociology because of the philosophical approach implicit in representational use of the GLM. In the rest of this paper, I shall consider the assumptions of that use, drawing examples from work by some of the best exponents of the GLM. For each assumption, I will discuss its nature, the attempts made to relax it within standard methods, and the types of alternative methods extant or possible. My focus throughout on the problems with GLR and the potentialities of its alternatives does not imply any derogation of its very great successes, and in particular any derogation of the studies I use as examples. But by exploring the theoretical limits of the GLM, I hope to suggest new lines of development in empirical sociology.

II The Fundamental Assumptions

A. Fixed Entities with Attributes

A central assumption of the GLM is that the world consists of entities with attributes. Entities are fixed; attributes can change. In practice, standard empirical work overwhelmingly concerns biological individuals, governmental units, and other entities considered to be "stable" by common cultural definitions. The GLM is less often applied to social groups like occupations, professions, and social movements whose members and social boundaries are continually changing.

The entities/attributes model for reality can best be understood by contrasting it with its most common alternative, the central-subject/event model. A historical narrative is organized around a central subject. This central subject may be a sequence of events (the coming of the Second World War), a transformation of an entity or set of entities into a new one (the making of the English working class), or indeed a simple entity (Britain between the wars). The central subject includes or endures a number of events, which may be large or small, directly relevant or tangential, specific or vague. Delineating a central subject and the relevant events-the task of colligation-is the fundamental problematic of classical historiography.

Precisely the same phenomena are organized by the entities/attributes and central-subject/event approaches, but in different ways. Consider the problem of the spread of the multidivisional form (MDF) among American firms as analyzed by Fligstein. There is a set of entities-the firms-which at any given moment have fairly clear boundaries. Firms can be thought of as having properties-size, rate of asset increase, domination by certain kinds of individuals, business strategies. We can imagine generalizing across the "cases" in terms of these "variables" and asking about the relation of the variables to the use of MDF. Yet we could also think about the history of a given "area" of firms, say the utilities area. We will see some entities in that area disappear through merger, others appear through internal differentiation and separation. Firm sizes will fluctuate through this appearance and disappearance as well as through variation in continuous entities. Some dominant individuals will control certain firms continuously, while other leaders will move from one firm to another through the mergers and divisions. Strategies will come and go, shaped by interfirm contagion and by period events like the depression. The histories of individual firms will be seen to follow unique paths shaped by the contingencies of their environments. In such a view, what GLR saw as variables describing entities become events occurring to central subjects.

This example shows a profound difficulty with the fixed entities approach; it ignores entity change through birth, death, amalgamation, and division. One way the MDF can arrive is through merger; yet merger removes entities from the sample and replaces them with new ones. It is not merely a strategy, but an event changing the sample frame. The social science of demography does indeed deal with appearance and disappearance of entities, and demographic models are now being applied to organizations in the work of the Stanford school of organizational ecologists. Yet the event history models so applied are essentially simple GLMs treating rates of change (usually of organizational death) as dependent variables and using a log-linear group of independent variables to predict them. Entities are grouped in synthetic cohorts and existence becomes yet another variable attribute to be predicted. Moreover, while such demographic methods address the appearance/disappearance problem, they do not address the merger/ division problem in any formal way.

Classical demography also provides preliminary models for the other major problem with treating entities as fixed, the fact that names often stay the same while the things they denote become different. This problem is most evident in the situation of exchange between aggregate entities.

Consider the attempt of Simpson and others to estimate the ability of occupations to recruit and retain cohorts of workers. The entities analyzed are occupations, characterized by the attributes of (1) strength, skill, and educational requirements, (2) product markets, industrial dispersion, and sex-specific growth, (3) earnings and earning growth potential, and (4) unionization or licensure. The dependent variable is an occupation's relative retention of a twenty-year age cohort, measured by the ratio of the odds of a cohort member's being in that occupation in the base year to those odds twenty years later, suitably standardized for death, relative occupational growth, and so on. Four twenty-year time-frames are analyzed, starting in 1920, 1930, 1940, and 1950.

There are two central problems with this daring design. First, the occupations themselves do not denote a constant body of work or activities. Simpson et al. have addressed this by excluding groups for which census classifications are not commensurate throughout the period. But this rules out, for example, the occupations reshaped by technology-a substantial fraction of the occupational structure, and a fraction that may in fact be determining what happens to the rest. Yet even those remaining in the sample changed drastically. Accounting, for example, began this period as a solo profession doing public auditing and ended it as a bureaucratized one doing nearly as much work in taxes and corporate planning as in auditing. The name stayed the same; the thing it denoted did not.

Second, the original cohort members present in an occupation after twenty years are not necessarily the same individuals who were in it at the outset. Evans and Laumann have shown that even the professions have extraordinarily high turnover and that they continue to recruit until well into middle age. Thus, the individuals aggregated under the labels are not necessarily the same individuals at one time as at another. Retention is confused with migration. Moreover, the cohort barriers are so wide that as each cohort ages twenty years, some individuals go from the start of their careers to their career midpoints, while others go from midpoint to near retirement. The cohorts-themselves presumed entities like the occupations-are thus no more coherent entities than are the occupations themselves.

One might handle such problems by disaggregation. But this is the counsel of despair. Both occupation and cohort do have some sort of reality, some sort of causal power. To disaggregate and model the occupations as properties of individuals would forfeit any sense of occupations' reality as structures. The classic answer to such multilevel problems is ecological regression. 10 But to assign coherent group-level terms to individuals-as is standard ecological practice-is completely impossible. The individuals don't stay in the same aggregates over time, and the aggregates themselves change-both by migration of their members and by change in emergent properties like type of work. These transformations make ecological parameters meaningless.

Some writers have noted the possibility of combining demographic and attribute methods to deal with such problems. In such methods, underlying demographic dynamics provide members-with their own attributes-to an emergent level of aggregates, which in turn have their own attributes. Event history methods to some extent so mix demographic and attribute models. On the theoretical side, a number of writers have argued that iterative processes of interaction between micro-level units in fact provide the structure that is macrostructure. Thus, there are a variety of preliminary attempts to address these issues, but clearly much work-both theoretical work on the formal structure of central-subject/event approaches and mathematical work on how to realize them-is required to develop this area further.

B. Monotonic Causal Flow

Between the various attributes of entities that it analyzes, GLR assumes that causality flows either from big to small (from the contextual to the specific) or between attributes of equivalent "size." Cause can never flow from small to large, from the arbitrary to the general, from the minor event to the major development. This assumption has several constituent parts.

The assumption of monotonic causal flow begins with the assumption of "constant relevance." A given cause is equally relevant at all times because the linear transformation, in most models, doesn't change over time periods (because the reestimation required is impractical). Of course, the B matrix of the general-form GLM can change, but GLR practitioners seldom take the position, common in historical writing, that "at time t, x was important, while later, the conjuncture of things made y more important." That kind of thinking-in which B is mostly zeroes and the nonzero elements differ from iteration to iteration-is not common. The first constituent of the monotonic causal flow assumption is thus the assumption, not necessary but nearly universal, of constant relevance.

(Continues...)

---

Excerpted from TIME MATTERS by Andrew Abbott Copyright © 2001 by The University of Chicago. Excerpted by permission.
All rights reserved. No part of this excerpt may be reproduced or reprinted without permission in writing from the publisher.
Excerpts are provided by Dial-A-Book Inc. solely for the personal use of visitors to this web site.

---

Table of Contents

Acknowledgments
Prologue An Autobiographical Introduction
Part One Methods and Assumptions
1 Transcending General Linear Reality
2 Seven Types of Ambiguity
3 The Causal Devolution
Part Two Time and Method
4 What Do Cases Do?
5 Conceptions of Time and Events in Social Science Methods
6 From Causes to Events
Part Three Time and Social Structure
7 Temporality and Process in Social Life
8 On the Concept of Turning Point
9 Things of Boundaries
Epilogue Time Matters
References
Index