Essays, opinions, and professional judgments are welcome in this section of AJE.
Forum articles speak to and about the philosophical, ethical, and practical
dilemmas of our profession. Authors are invited to express their views to encourage
constructive dialogue centered on issues. To keep the dialogue substantive, other
articles motivated by previous Forum presentations should have independent titles
and themes. Items labeled “Comments on.. . ” and “Rejoinder to.. ” will not be
published in Forum-such responses are welcome and encouraged in the In
Response section ofAJE. Standard citationsand reference listsshould acknowledge
and identify earlier contributions and viewpoints.Manuscripts should not exceed 10
double-spaced typewrittenpages in length, unless the paper is invited by the editor.
The views expressed in this section are independent of and not necessarily shared
by AJE editors. By designs “Forums” are open to diverse views, in the hope that
such diversity willenhance professional dialogue.
The Qualitative Method of Impact Analysis
LAWRENCE
ABSTRACT
B. MOHR
Consider the qualitative approach to evaluation design (as opposed
to measurement) to be typified by a case study with a sample of just
one. Although there have certainly been elaborate and emphatic
defenses of the qualitative approach to program evaluation, such
defenses rarely attempt to qualify the approach explicitly and rigorously
as a method of impact analysis. The present paper makes that
attempt. The problem with seeking to advance a qualitative method
of impact analysis is that impact is a matter of causation and a nonquantitative
approach to design is apparently not well suited to the task
of establishing causal relations. The root of the difficulty is located in the counterfactual definition of
causality, which is our only broadly accepted formal definition of causality for social science. It is not,
however, the only definition we use informally. Another definition, labeled “physical causality,” is
widely used in practice and has recently been formalized. Physical causality can be applied to the
present problem. For example, it explains the persuasiveness of Striven’s “Modus Operandi” approach
Lawrence B. Mohr l SPP-Larch Hall, University of Michigan, Ann Arbor, MI 48109-1220; Tel: 313-763-4588; E-Mail:
lmohr@umich.edu.
American Journal of Evaluation, Vol. 20, No. 1, 1999, pp. 69-84. All rights of reproduction in any form reserved.
ISSN: 1098-2140 Copyright 0 1999 by American Evaluation Association.
69
70 AMElUCANJOURNAL
tailoredcasestudydesignwithasample
sizeof one in principleas strong a basis for making inferences program impact as
arandomized experiment. Crucial program evaluation finding
that people’s “operative reasons” for doing what they do are the physical actions.
it is shown that external validity using this qualitative approach would have exceptional strengths.
INTRODUCTION
Qualitative designs for program evaluation have become widely accepted and practiced
(Greene, 1994; Goddard & Powell, 1994). The idea has also generated a good deal of controversy
(summarized in part in Chen, 1990) wherein each side in the debate--qualitative vs.
quantitative-finds its own method to be clearly superior and the other to be woefully inadequate.
Many authorities take what they see as a middle ground, advocating that both
approaches may be used in concert (Reichardt & Cook, 1979; Chen, 1990; Greene & Caracelli,
1997). It has begun to be increasingly clear, however, that although both approaches
may well have something to contribute to program evaluation, their areas of contribution are
not identical. Although there is undoubtedly overlap between the two approaches in what they
seek to accomplish, they are not substitutes for one another.
Quantitative designs or approaches have been seen by their advocates as being especially
good for impact analysis. Qualitative designs, on the other hand, have been advocated primarily
in conjunction with other evaluative functions such as implementation analysis, process
analysis, community self-analysis, the empowerment of staff, and the interpretation and
understanding of experience. The functions in this latter list will not be reviewed here because
they are not germane to the present purpose, which focuses on impact analysis. Maxwell and
a small number of additional scholars (summarized in Maxwell, 1996) have begun to explore
a role for qualitative work in the assessing of impacts. More typical, perhaps, the defense of
the case study method by one of its most outspoken proponents in program evaluation does
not credit it with efficacy in the assessment of impacts and distinguishes it from quantitative
research in part on that basis (Stake, 1981). Of course, some proponents of the qualitative
approach suggest that it is better than the quantitative approach precisely because its particular
goals or functions have greater value than the goal of impact analysis (see, for example, Guba
& Lincoln, 1981), but that debate is also outside of the present scope.
Let us assume, recognizing that some might deny it, that there often is some value to
impact analysis in program evaluation. If one were concerned with impact analysis and if
there were a qualitative approach that did indeed have clear relevance for that function, then
it would be reasonable to bring it to light and discuss it. The purpose of this paper is to
advance the idea of one such qualitative approach and to provide an introductory treatment of
it. This idea must of necessity be viewed against a backdrop composed of concepts that, while
familiar to the evaluator, are here combined in a nontraditional picture that will help bring
clarity to the issue at hand.
CAUSATION
Impact analysis deals with causation. While some may take the position that program evaluation
as a discipline should rid itself of the preoccupation with causation (or equivalently here,
impactAnalysis 71
causality), let us also assume tentatively that causal analysis is suitable for at least some purposes
in program evaluation and therefore question how qualitative designs might contribute
to such analyses.
For the present discussion, it will be helpful to define the quantitative approach to impact
analysis as one that relies on the counterfactual definition of causality (a more defensible version
of this species of definition is called “factual causation” and will be elaborated below).
The counterfactual definition states that X was a cause of Y if and only if X and Y both
occurred and, in the circumstances, if X had not occurred, then neither would Y.It is the final
“if’ clause (“if X had not occurred.. .“) that we call “the counterfactual.” We will see momentarily
that this conceptualization leads naturally to what we usually think of as quantitative
research designs. It is also the source of the well known principle (whose validity will nevertheless
be disputed, below) that comparison is necessary to establish causality, the comparison
being between evidence for the “result” on the one hand, that is, the occurrence of Yafter X,
and evidence for the counterfactual claim on the other. Classically, we get our result from
observing a group to which a treatment has been applied and our estimate of the counterfactual
from a similar group (or a “before” measure) to which the treatment has not been applied.
Putting these two sets of observations together so that we have variance on the proposed
causal or treatment variable makes the whole amenable to statistical analysis. Design and
analysis together yield the classic quantitative package. Note that without data on the counterfactual,
a statistical technique such as regression or correlation analysis would not be applicable,
even if there were thousands of cases, because there would be no variance on the variable
denoting whether or to what extent the treatment was received.
Accordingly, we may provide a specialized definition of the term “qualitative,” as well as
“quantitative,” as they both apply to the idea of design (as opposed, for example, to measurement).
A design whose purpose is to determine impact will be considered qualitative if it
relies on something other than evidence for the counterfactual to make a causal inference. It is
qualitative first in the positive sense that it rests on demonstrating a quality-in the present
treatment the quality will be a physical connection in the natural world between the proposed
cause and effect-and second in the negative sense that it is not quantitative, that is, it does
not rely on a treatment variable, or on comparing what is with an estimate of what would otherwise
have been. A vast number of observations may be and usually are made in a qualitative
evaluation, but when it comes to the core of causal inference there is only one kind of observation
in the perspective of the counterfactual definition, that showing the co-occurence of X
and Y.Conversely, even if a study has only two empirical observations-one showing that the
proposed cause and effect occurred together and the other showing that when the cause did not
occur, then neither did the effect-it is clearly based on variables and counterfactual reasoning
and will therefore be considered quantitative.
These two approaches would seem to present an interesting and viable picture except for
one matter. The qualitative approach does not fulfill the requirements of the counterfactual
definition since it does not deal in variables and does not incorporate evidence on the no-treatment
condition, but, apparently, no other rigorous definition of causality is accepted in social
science (King et al., 1994). If it is indeed true that there is no other, then it would simply be
impossible to infer causation based on a qualitative design. Criticisms of the case study as
being based on a sample of only one case make exactly this point. On the other hand, there has
been at least one approach to causality in the literature on program evaluation that seems on
the surface to be both legitimate and to have nothing to do with counterfactuals or quantitative
analysis. It may be found in Striven (1976), where it is called the “modus operandi” method
72 AMERICANJOURNALOF EVALUATION,20(l), 1999
of demonstrating causality. An explicit alternative definition is not provided there, but the
method proceeds as follows:
Y has occurred and the task is to demonstrate that the treatment T caused Y. This is done
in part by the process of elimination. That is, there are several other possible causes of Y, such
as U, V, and W, and these must be eliminated as contenders in order to settle on T. The process
also depends critically on the idea that each possible cause has a known “signature,” and if the
signatures of U, V, and W can be shown to have been absent while that of T was present, then
it may be concluded that T was the cause. The signature may contain events of two types: (a)
a known causal chain of events by which T or U, etc., would lead to Y (George & McKeown,
1985); and (b) other events in addition to Y that are known to be associated with an active T,
or U, and so on (Campbell, 1979). The elimination of other possible causes is of course a
prominent tool in almost all causal inference but it is not enough by itself except in the rare
circumstance, such as in randomized experiments, when all other possible causes can be identified
or controlled. Even then, a reservation remains, as will be shown in a later section. In
short, we will see below that while the method of elimination can be extremely helpful in
causal inference from qualitative designs, we must rely more on demonstrating that the signature
of T itselfwas present and that this proves the causal point. For the present, it is important
to see, as Striven (1976) emphasizes, that the modus operandi scheme is extremely common
in everyday causal inference. It is not an arcane, academic conceptualization, but a method of
determining causality that is in constant use in the professions and in everyday life, such as in
diagnosing what is wrong with a car, or in medical diagnosis, detective work, cause-of-death
determination, and so forth. The auto mechanic can infer that a car is overheating because a
thermostat is stuck by feeling the temperature of certain hoses. The physician can determine
that a certain disease is at work by establishing the presence of two or three critical symptoms.
It is also important to emphasize that this common method is not at all the same as determining
causality by counterfactuals. Suppose that a woman has had a heart attack and has
died. Probably she would not have died at that time if she had not had the heart attack, but that
is not certain. In fact, suppose we could demonstrate that she would have died then anyway
from some other cause. Using the counterfactual definition, we would be forced to conclude
that the heart attack was not the cause of her death because we could not claim the counterfactual
(if no heart attack, then no death). But by examining the tissues and otherwise identifying
the signature of a heart attack, we might be able to establish the truth, namely, that she did
indeed die of a heart attack. Note that in making this determination we do not proceed by
bringing forth evidence for the counter-factual-what would have happened had she not had a
heart attack-which would in fact lead us completely astray in this case because she would
have died anyway. We do not make a dichotomous variable out of heart attack, the supposed
cause. We consider this to be a one-case study and determine causality by exposing the incontrovertible,
physical signature of the heart attack. Finally, it may well be that the signature is
known because of prior medical research and also that the prior research in question might
possibly have employed counterfactuals, but that does not mean that our causal inference rests
on the counterfactual definition. We are not doing that other research at present. We are
researching the death of this woman, just as we might research the outcome of a program
being evaluated, and we do not use counterfactuals to reach the particular causal conclusion
in the specific instance of causality that links this heart attack with this death. To drive the
point home, suppose that many impact studies have given us the knowledge that employment
can definitely decrease criminal recidivism, just as heart attacks can cause death. If we now
do a new study of recidivism and wish to explore a possible causal link with employment, we
Impact Analysis 73
have two distinct choices of approach-the quantitative and the qualitative. It does not matter
how the prior research was conducted. Both research options are real and independent in the
present.
Other research in social science and program evaluation has shown a role for qualitative
research in impact analysis, particularly Huberman and Miles (1985) and King et al. (1994).
When pressed, however, the core determination of causality in these treatments depends ultimately
on the counterfactual method. The best clue we have that a core determination may
indeed be qualitative lies in the concept of the signature as it emerges under the various labels
of “modus operandi” (Striven, 1976), “process tracing” (George & McKeown, 1985), and
“pattern matching” (Campbell, 1979). The possibility of a qualitative determination of causality,
then, has been very strongly suggested and what remains to consider is how this general
approach aspires to establish instances of causation without variance on a causal variable
(Maxwell, 1996). A totally different definition of causality must necessarily be involved.
The working out of that definition, which is too lengthy to reproduce here in full, may be
found in Mohr, 1996. There it is argued that it is appropriate in social science to employ a dual
conceptualization of causality. Furthermore, there is abundant evidence that such a dual conceptualization
actually is employed. The two facets of the concept are labeled physical and
factual causality, respectively. The first is a relation between events in the natural world, as in
“The car skidded sideways and toppled the lamp post on the comer.” The second is a relation
between statements, truths, or facts (Strawson, 1985), as in, “The fact that the ex-prisoner had
steady employment caused the fact that he did not commit further crimes.”
Of the two, physical causation is more important in terms of the present inquiry, although
it is not necessarily the most commonly encountered in program evaluation and social science.
Its importance lies in the fact that (a) it is fundamental within the dual approach (it is itself part
of the definition of factual causation); and (b) it is the key to causal inference in qualitative
research. Physical causality is the common, mechanical or Newtonian sense of causation that
is familiar to all. Cause and effect are the relation between a force and a motion, respectively.
It is the causation we recognize in the lifting, smashing, pulling, pushing, or tearing of something.
The classic example in the Humean tradition is billiard balls on a table, but Hume
(1955) sought to exclude this idea as a definition. It is argued elsewhere (Strawson, 1985;
Mohr, 1996) that Hume put too much emphasis on the sense of vision and not enough on the
sense of touch in reaching this conclusion, but it would be a lengthy and unwarranted digression
to review the arguments here. We must proceed as though physical causation were not
only intuitively appealing but philosophically sound and refer the interested reader to the cited
works for a fuller treatment.
The counterfactual definition of causality, serviceable as it is for some purposes, is well
known to be plagued with serious philosophical problems (Ma&e, 1980; Mohr, 1996). Factual
causation is a variant achieved through the interjection of physical causality that preserves
the counterfactual definition in spirit while overcoming those problems. The definition
is as follows: X was a factual cause of Y if and only if X and Y both occurred and X occupied
a necessary slot in the physical causal scenario pertinent to Y. The interjection of physical
cause is clear. The spirit of the counterfactual definition is preserved in the idea of a necessary
slot: if something did not occupy that particular functional slot (and it happened in this case to
be X), Y would not have occurred. For our purposes, we may treat factual causation as though
it were the simpler, counterfactual definition without undue or misleading distortion. The two
facets, then, though related, are quite distinct. Physical causation rests on the idea of a direct
physical connection in the real world. Factual causation rests on a counterfactual claim and
74 AMERICAN OF EVALUATION, 1999
has no need to refer to a direct physical relation. (For example, in “Because the ex-prisoner
was employed, he committed no further crimes,” there is no direct physical connection referenced
between events in the natural world.)
REASONS AS CAUSES
We have seen that causation can be demonstrated through qualitative research when physical
rather than factual causality is the guiding conceptualization. That is why we needed no variables
or other quantitative paraphernalia to make an inference about the heart attack, the stuck
thermostat, or the diagnosis of the disease. There would seem to be a problem in terms of our
central purpose, however, in that the effects of interest in program evaluation and social science
are so frequently human behaviors. This is overwhelmingly true in the case of program
evaluation, in fact, where common outcomes of interest are behaviors such as self care in
nursing homes, criminal recidivism, juvenile delinquency, teen pregnancy, preventive health
behaviors, welfare-to-work, substance abuse, overcrowded housing, divorce, safe driving, and
safe sex. What are the physical causes of intentional human behaviors? They are at any rate
not so tangible or clearcut as hot and cold hoses, or sore throats and fevers. Many if not most
who ponder this issue seem to feel that mental states or acts such as decisions, choices, and
reasons-specially the latter-are the causes of behavior (Davidson, 1980). We will be
guided here by the finding, based in the physiological rather than the philosophical literature,
that reasons, in a special sense of the term, are indeed not only the causes but the physical
causes of intentional human behaviors (see Mohr, 1996, and works cited there).
Before elaborating the special sense of reasons referred to, let us review the more general
sense. Most definitions are similar and suggest two components: If you (1) want to accomplish
Y; and (2) believe that action X will do it, then you have a reason for undertaking action
X. For example, if you want to improve your agency’s productivity and believe that privatizing
some of the tasks will do it, then you have a reason for privatizing. So far, the general
sense of the concept of a reason does not seem very physical, since wanting and believing are
mental constructs. In addition, it is hard to see how reasons so construed can be the causes of
behavior. We all have a great many reasons for doing a variety of things at a given time but
we do not act on most of them at that time, and some of them are never implemented at all.
One might have a good reason, for example, to go to the market for eggs this evening and also
to replace a washer in a leaky faucet, but one might carry out only one of those activities, or
none. Moreover, a student may see, and even keenly feel that getting good grades is a reason
for studying long and hard, but never study very long or hard at all. What does it mean, then,
to say that reasons are causes when so many of the reasons that were definitely “there” at a
given time did not cause anything?
To answer, it is convenient to introduce the notion of “operative reason,” i.e., the reason
that actually operated to produce the behavior performed, and to stipulate that the operative
reason was different from all of its contemporary reasons in that it was the strongest. If the
person fixed the faucet, for example, instead of going to the market, it is because he wanted to
stop the drip more than he wanted the eggs. If the student studies little (given a belief in the
connection between studying and good grades), it is because she wants something else more,
perhaps being accepted as one of an “in” crowd with whom she regularly associates. And if
the Director is more interested in pleasing the public employees union than in productivity,
the tasks in question may not be privatized this year, or perhaps ever. In this way, in short, one
impact Analysis 75
of the many reasons that may be active at a given time becomes an operative reason and the
associated behavior ensues. As a footnote, we see that if a program aims at getting people to
behave in certain ways-showing up for work regularly and on time, for example-the program
must not only motivate them but motivate them strongly to carry this action out, and if
the evaluator wishes to relate the presence or absence of a successful result to the program, it
will be important to adduce evidence regarding the strength or weakness of the critical moti-
vation.
The most important aspect of the special sense of “reason” referred to is that, as used
here, a reason is an unaware physiological entity and not a mental construct. We all know that
we frequently do things intentionally, but without being aware of any particular reason. On the
other hand, we are often very much aware of particular reasons for doing things (which we
may or may not do). These conscious reasons or thoughts, it is argued, do not cause behavior,
although they may at times closely parallel the physiological elements that do. It is well
known that for an experience to be stored in and recallable from long-term memory, it must
be accompanied by a physical affect residue, a component that records the related affect, if
any, at the time of the experience. The “want” part of what we think of as a reason, although
we now speak of a “want” that is totally below the level of consciousness, is the positive or
negative affect residue attached to some object in memory. The strength of the reason is the
electrical or chemical strength of this affect residue. The “belief’ part of a reason is the memory
of having done some X to achieve Y or something similar in the past, or even of having
heard of accomplishing something like Y by doing X. The X is selected and activated when
the affect on the relevant Y becomes dominant by virtue of its strength. We see therefore, that
operative reasons as rendered here are the physical causes of undertaking action X. Finally, it
is important for the research implications to note that operative reasons, being unaware, are
not necessarily obvious, either to the subject or the investigator. They may frequently be
accompanied by a parallel reason or decision that is an aware, mental state, but they also may
not, and even if they are, it is the physiological “reason” and not the mental one that is causal.
Indeed, when we think sometimes that a certain conscious reason for our behavior was operating
we may well be wrong, as we might find out, perhaps, only by careful introspection or
with the help of a psychotherapist.
In sum, qualitative research to determine the causes of intentional human behaviors, such
as whether or not those behaviors were induced by a program being evaluated, must involve a
search for the operative reasons behind the behaviors. In most cases this would undoubtedly
involve obtaining information from the subjects whose behavior is at issue. Given that operative
reasons are unaware, however, that is not necessarily all that one must do. It might be necessary
to obtain information from many other people, from documents, and from the histories
of relevant events. These are methods, however, that are familiar in social science, especially
in such areas as history, anthropology, and area studies.
CAUSAL REASONING
We have seen that there are two types or conceptualizations of causation, the factual and the
physical. When research proceeds as sketched out just above, the causal inference will be
based on an argument that makes a case for the operative reason and connects it with the
behavior, perhaps with some reservations. The kind of argument involved in that case, as well
as that involved in previous qualitative examples such as establishing the connection between
76 AMERICAN JOURNAL OF EVALUATION, 20(l), 1999
of hoses and the overheating of the car, will be called “physical causal reasoning.”
When the causal inference is to be an inference of factual causation, a different sort
of research process is dictated, one that results in the ability to make the argument, again perhaps
with reservations, that factual causation has been established. The form of that particular
argument would be called “factual causal reasoning.”
Factual causal reasoning must necessarily rely on a counterfactual claim. A good deal of
energy and design ingenuity would be devoted to being able to argue persuasively that if X
had not occurred, neither would Y. This might actually be done within the confines of a single
case study, or when talking about a single event that is a minor part of a larger story. That is,
one might show that X and Y occurred and argue as best one can on the basis of knowledge of
past events, plus a careful and detailed examination of the context and circumstances surrounding
the present event, that if X had not occurred here, neither would Y. One might not,
that is, purposely make one or more empirical observations of a “control group”-a similar
subject and context in which X did not occur. Still, this would be factual causal reasoning and,
as the terms are used here, a quantitative design. Or, as featured prominently in the book by
Ring et al. (1994), one might study in depth one or two cases in which X was present and one
or two in which it was not. The authors tend to refer to such “small n” studies as qualitative
because of the depth of exploration and the manner of data collection and measurement, but
in design perspective, these are again quantitative. In most cases, however, factual causal reasoning
is based on designs and sample sizes for which statistical methods of analysis become
suitable. In that kind of case, it is important to remember from the counterfactual definition
that the control group or comparison group or “before” measure, for example, is not of interest
in its own right. It is not a matter of comparing what happens (present tense) when X does and
does not occur. Rather, the control or similar group is of interest only as a basis of estimating
what would have happened in the experimental group if X had not occurred there. That is the
essence of factual causal reasoning.
Take the evaluation of health education efforts, for example. There we have statistical
designs of various sorts in which investigators have concluded that knowing about certain risk
factors leads to (i.e., causes) certain health related behavior, such as nutritionally sound eating
practices (Edwards, 1985). Frequently, there is no thought that this knowledge about risk factors
might be a physical cause. Being exposed to information about proper nutrition, for example,
is not thought to have made the subjects eat properly, but rather we can see that
adequately motivated people (here is where the physical causation comes in) are much more
likely to eat properly if they indeed know what is proper than if they do not-accurate knowledge
is more auspicious than ignorance or misconception. In any case, the reasoning proceeds
by showing that those who received and absorbed the health education materials tended to eat
properly. Those who did not, however, or who did not understand it or remember it, tended
not to eat properly, thereby indicating that the knowledge in the first group was a necessary
condition. (Mohr, 1995, p. 267)
Physical causal reasoning, on the other hand, is focused on demonstrating persuasively
that a particular kind of physical connection took place between two events in the real world,
in essence, the connection between a force and a motion. There is some basis for considering
that it is this sort of reasoning, in contradistinction to factual causal reasoning, that gets at real
causes. Frequently, in other words, we are not satisfied with evidence from observational studies
or even randomized experiments because the real, that is, the physical causal mechanism
remains a mystery. The physical evidence may frequently be required to support the statistical,
and it will almost always contribute to a more convincing package if it can be obtained.
Impact Analysis 77
One excellent example is the long and persistent interest in the question of how, physiologically,
smoking operates to cause lung cancer. We have been bothered, in other words, by
the fact that the signature of smoking has remained obscure. In contrast, it took us less time to
become convinced of the effect of fluorocarbons on the ozone layer because we not only had
statistical time series relating the two, but an account of the chemical processes involved, as
well.
To elaborate in the context of more standard program evaluation, consider that we might
be trying to learn whether to attribute a lower fatality rate from motor vehicle accidents to a
well-publicized police crackdown on speeding in one community. Application of the beforeafter
or the interrupted time series design to the fatality rate is one option. Comparison to a
community that is similar and that has had no crackdown is another. Both of these are in the
classic, quantitative or factual-causal mode and both are weakened by the standard threats to
the validity of quasi-experiments. An alternative to be pursued instead, or perhaps better yet
in addition, would lead us to confine our attention to the one community and the one time
period of concern-a case study insofar as this particular aspect of the design is concerned.
We might use a survey to determine not only whether and to what extent but how the citizenry
became aware of the existence of the crackdown. That is, we would not try to show that they
would never have heard of the crackdown if not for the media used by the project (factual
causal reasoning), but rather to document that they actually did read about it here, or did see
it on TV there, or were told about it by a police officer, or by a friend who claimed to have
seen the TV spot, and so forth-all physical processes. Furthermore, we would be interested
in understanding the behavior itself-not only whether the citizens consciously drove more
slowly, but why-was it a heightened sense of the dangers of speeding, for example, or a
heightened subjective probability of getting caught? Note that the quantification involved in
both of these links in the project, given the survey involved, would be just a matter of counting
citizens to get an idea of the prevalence of the causal processes uncovered. The quantification
would not in any way be part of a design to permit causal inference. The design is unmistakably
qualitative. No evidence for the counterfactual would be brought forward because the
survey would not contain any. On the contrary, whether one citizen is surveyed or a hundred,
the design is critically dependent on persuasively establishing the occurrence of the two physical
causal processes in the individuals concerned-learning about the crackdown because of
project activities and modifying behavior because of strong motivations sparked by the
project’s messages.
Another brief example might be helpful in suggesting the kinds of roles available for
physical causal reasoning. In the kind of research exemplified by eating practices and knowledge
of good nutrition in the quoted paragraph above, the efforts to increase knowledge would
necessarily be evaluated by factual causal reasoning alone. The nutritional knowledge simply
was not a physical cause. As suggested, however, there would have to be separate motivations,
or reasons, impelling the sound dietary behavior apparently advocated. A more thorough
program would seek to inculcate these, as well. An impact analysis of the program might
then proceed either by factual or by physical causal reasoning or both. There is no reason, in
other words, why the two forms of logic or research design cannot both be brought to bear (a)
within the same evaluation, as in evaluating the efforts to impart knowledge by factual and
motivation by physical causal reasoning; or (b) on the same outcome, as long as we are concerned
with a physical cause, as in employing two different designs to determine whether or
not people changed their eating habits because of the program’s efforts to persuade them of
the benefits.
78 AMERICANJOURNALOF EVALUATION,20(l), 1999
The purpose of this paper is in large part to identify the idea of physical causal reasoning
and to suggest its functions. The details of successfully implementing such reasoning in evaluation
and other research are beyond the scope of our efforts here and no doubt will only be
accumulated over many trial projects. Something about the basic elements of implementation,
however, rises out of the definitional logic of physical causation itself. This conceptualization
of causality has to do with the relation between a force and a motion. With regard to human
behavior, these components are bound up in the operative reason and the behavior itself,
respectively, as in the relation between fear of getting caught by the police and keeping within
the speed limit during the crackdown. To establish this relation persuasively, four elements
are apparently important. One must show first that the person or group studied must have had
the reason to which we want to attribute causality- sometimes whether they themselves claim
it or not. Second, it is important to make the case that this reason was very strong relative to
other considerations that could have been active at the time. Third, it is generally crucial to
show by various other manifestations that it was indeed this reason that was operating-tracing
out the “signature” of the reason in the style of the modus operandi approach. And last, it
will generally be helpful to establish that it is proper to interpretthe actors’ behavior as intentionally
keeping below the speed limit, for example, or following sound nutritional practice,
and not just as driving at a certain speed or eating certain foods, as would be common in quantitative
research.
It would be a digression to discuss at length the various other perspectives on causality
and causal reasoning that have been suggested in relation to qualitative research, such as constructivism,
neorealism, and so forth. It may be well, however, to emphasize one point, and
that is that what is offered here are defznitionsof causation such as must normally support all
such perspectives. If one takes a constructivist view of “causation,” therefore, or emphasizes
“webs of causation” rather than single, linear, cause and effect dyads, or aims with the neorealist
at discovering “causal mechanisms” rather than causes, one can always be challenged to
specify what one actually means by the root term “causation” or “causal” in one’s own particular
context. Exactly as was the case with the modus operandi method, these other perspectives
do not obviate the need to conceptualize causality, nor do they automatically provide the
required conceptual deftition in themselves. They tend to use the term “cause” as though it
were previously defined. It is suggested here that the response to this challenge to define can
always be rendered in terms of one of the two conceptualizations proposed above. It is therefore
suggested further that the two types of causal reasoning, physical and factual, will in that
sense always be pertinent, whatever the philosophical framework, although they may not
always be important enough to an author’s point to warrant attention. Finally, although qualitative
research is firmly established as a scholarly approach in social science and program
evaluation, still, consideration of the notion of causality seems to be ruled by the counterfactual
definition, which, from the standpoint of design, is basically quantitative in orientation.
Thus, the possibility of thinking in physical causal terms should open up notably wider prospects
for thought, particularly within the qualitative regime.
INTERNAL AND EXTERNAL VALIDITY IN THE QUALITATIVE APPROACH
Let us fust accept the term internalvalidityto mean the validity of a causal inference based
on the research, as was the case when the term was first used by Campbell and Stanley (1966).
In this light, the concept has not been very relevant for qualitative evaluation in the past
impact Analysis 79
because qualitative evaluation has not been much preoccupied with causality. Given that the
present paper has been vitally concerned with the marriage of “qualitative” and “causal,”
however, internal validity does become germane as an issue.
There are two components to the practice of establishing internal validity in quantitative
research, and the same two will hold for qualitative research, as well. The first has to do with
showing the fulfillment of the governing idea of causation through the design and the operations
of the research. In the quantitative mode, this primarily means arranging to produce evidence
for the counterfactual (since establishing that X and Y both occurred is generally not
much of a challenge)-as for example by the device of a comparison group. In the qualitative
mode, it means arranging to produce evidence for a certain sort of physical relation between
X and Y. In both cases, however, the evidence will never be unassailable, so another important
component of establishing internal validity becomes the attempt to rule out certain other,
plausible causes. Even in the randomized experiment, the evidence for the counterfactual is
only probabilistic, and not certain, so it is common to achieve greater certainty on at least
some dimensions by introducing control variables. In both approaches, therefore, ruling out
alternative causes occupies a prominent place in the quest for internal validity, but the deductive
logic of the process of elimination does not obviate the role of the factual causal argument
in the quantitative approach, nor does it eliminate the role of the physical causal argument in
the qualitative approach.
In this light, we see that the elimination of other causes in the modus operandi method is
not meant to be the same as establishing evidence for the counterfactual, which would be mixing
the two forms of reasoning, and indeed it is not the same. First of all, one could in no way
claim to prove the counterfactual by that method unless one eliminated all other causes, and it
will always be impossible to make that claim legitimately. Second, even if we knew that nothing
else caused Y, we would not necessarily know that nothing else would have caused Y.
This is illustrated by our prior example of the woman and the heart attack. Recall that we
assumed for the sake of argument that she would have died then anyway, even if she had not
had the heart attack. Suppose that the woman had been found dead in her car, which had
crashed into a tree. The question is whether she had a heart attack and died, crashing into the
tree while the death was in process, or if she died of a concussion as a result of crashing into
the tree. We might be able to tell by careful autopsy that she died of a heart attack, so that all
other causes are categorically eliminated, but it may in fact be true that if she had not actually
died of the heart attack, she would surely have died of the concussion or other injuries. To
eliminate all other possible causes, therefore, as by finding the incontrovertible signature of
the heart attack, does not establish the truth of the counterfactual-here, she actually would
have died anyway. In quantitative research, the process of elimination is meant rather to
strengthen a basic claim for the counterfactual that will generally have weaknesses. In qualitative
research, it is meant to strengthen the basic claim of a certain physical connection-for
example, in striving to rule out excess greenhouse gases resulting from forces independent of
industrial and other human activities. In fact, it will by no means always be necessary to resort
to the effort to identify and eliminate other plausible causes. The design-based claim for the
counterfactual or the physical relation may suffice. Just as in a randomized experiment, and
even in some quasi-experiments, we might frequently consider the case for the counterfactual
to be strong without bothering to try to identify other plausible causes and enter them as control
variables, so in a qualitative study we might consider the case for the physical relation to
be strong, or obvious, without bothering to try to identify other plausible causes and show the
absence of their signatures. For example, if we show that certain drivers became aware of the
80 AMERICANJOURNALOF EVALUATION,20(l), 1999
police crackdown because they got tickets and the officers told them about it, we do not necessarily
have to speculate on how those particular drivers might have become aware of the
crackdown otherwise, but did not.
Nevertheless, many will say that internal validity in the qualitative approach must be
inherently weak. Usually, the claim that a certain reason was the operative reason for a behavior,
for example, must in the end be speculative. In that approach, there is nothing like the laws
of probability to give credence to an inference. This reservation is valid and important, but it
needs to be put in perspective. The difficulty is relative. Other approaches have similar prob-.
lems, and weak internal validity for a method has never been a good reason to reject it categorically.
In particular, one cannot make an airtight case on the basis of factual causal
reasoning, either. There is always a necessary residue of subjective assessment. In randomized
experiments, the threat of contamination (impure treatment delivery or inadequate control of
the environment) can never be extinguished altogether, nor can the reservation that the statistical
causal inference carries a probability of error equal to the significance level. At the 5%
level, we will make errors about 5% of the time. In quasi-experimental designs, which are
extremely well represented in evaluation journals, the threat of selection bias is similarly persistent
and is in fact notoriously troublesome. Furthermore, ex-post-facto or observational
studies are well known to be pathetic in terms of internal validity, yet they are in constant use,
probably because there is often much to be learned from them nevertheless. On the other hand,
just as we have imperfect but substantial confidence at times in the conclusions of experiments,
quasi-experiments, and observational studies, we may often have substantial confidence
in the case made for a physical cause through the modus operandi method or other
physical causal reasoning, whether the subject be the overheating of an engine, the penetration
of a message by means of the mass media or a policeman, the controlling desire to avoid
speeding tickets or, let us say, the controlling desire to avoid the inevitable future hassles with
bureaucrats by immediately correcting violations noted on a safety inspection. Reasons and
other physical causes that are not obvious may nevertheless frequently be demonstrated with
considerable satisfaction and edification.
This is not to say that further constructive thought regarding internal validity is either
futile or unnecessary in the context of qualitative designs for impact analysis. On the contrary,
much should be done and there is in fact evidence that the issues are not intractable. For example,
as will be reviewed momentarily, Lincoln and Guba (1986) have summarized much that
has been done to conceptualize for qualitative research what would seem to be analogous to
issues of validity in quantitative research. Much of the collective conceptualization has more
to do with qualitative approaches to measurement or description rather than design, but some
concepts, credibility for example, are clearly applicable to the present discussion and make a
stimulating contribution. The whole idea is both new enough and challenging enough, however,
that we should not expect definitive analyses to have been produced already. Finally,
Mohr (1999) has suggested the criterion of a “consummate causal understanding” that should
also be helpful in developing methods of achieving and testing for internal validity in the qualitative
approach to impact analysis. A consummate causal understanding of the past behavior
investigated is achieved when all pertinent challenges and questions regarding that past
behavior and one’s explanation of it are answerable from the data presented and their logical
extensions.
The issue of extend vdidity must also be considered, but the first observation to be
made, it would seem, is that it is important to keep the two types of validity distinct. Because
a sample size of just one case appears to threaten both types, there may be a tendency to
ImpactAnalysis 81
blend them together when critiquing the case study as a design. “Let us accept,” the critic
might suggest, “that causality can be established with a sample of one, but is it not better to
have a large sample? Is it not better to observe many causal instances rather than just one or
two? The one or two might have been unique.” Perhaps so. We proceed now to consider
exactly that issue. But it is necessary to recognize at the outset that it is an issue of external
validity and not internal. If causation is established then causation is established. The question
whether it is better to establish it in more subjects rather than fewer is purely an issue of
generalizability.
It is well known that external validity is critical (Cronbach 1982). There is hardly a point
in finding out for sure that T caused Y if we have no interest in applying the finding either to
the same subjects in a subsequent time period or to other subjects. The strongest case for internal
validity, however, guarantees nothing in regard to external validity. True, if the observed
subjects were randomly sampled from a larger population we can generalize the results to that
population, but (a) that pertains to internally valid physical as well as factual causal results;
and (b) such a generalization is still limited to that particular larger population and, even more
restrictive, to the past time period in which the subjects were observed. In particular, there is
no basis for claiming that a result is theoretically valuable just because one can generalize it
to a certain population that has been randomly sampled.
Since internal validity is in no sense an aid to establishing external validity, it follows that
factual causal reasoning is no more auspicious for external validity than physical causal reasoning.
Both are oriented toward causality, not generalizability. As noted previously, for
example, if we used old or impressionistic evidence to reason for one individual subject that
if Thad not occurred, then neither would Y,the mode would definitely be one of factual causal
reasoning (quantitative reasoning, that is, even in the single case), but it is hard to see that this
would give us any particular basis for generalizing the results to other subjects. We see, therefore,
that it is not the design that leads us to prefer quantitative research over case studies as a
basis for generalization, but the mere fact of the number of subjects involved, whether people,
schools, cities, or countries. It will usually be more comfortable to generalize when many subjects
have been causally affected by a treatment than when only one or two have been
affected. The one or two might be unusual, even unique. Of course, the many subjects
observed in a large sample may also be quite different from others to which one would like to
generalize, and in fact they generally are, as is evidenced by the fact that one can almost
always find interacting variables of interest that qualify a relationship. There is therefore some
advantage in numbers, but it should not be oversold. Conditions change over time. One subculture
is different from another. One school is even different from another within the same
district, so that one may very well not observe the same regression coefficients as one replicates
the analysis.
At the same time, qualitative studies have their own advantage with respect to external
validity that has long been recognized. It is that the study in depth should enable one to
understand the process, and there may be no factor as important as this for generalizability
in social science. The more thoroughly we understand the causal process by which a treatment
has affected an outcome in one case, the better the position we will be in to know
when a similar outcome will result in another case (cf. “extrapolation” in Cronbach, 1982;
see also Lincoln & Guba, 1986). Thus, the quantitative approach generally has an advantage
in numbers and the qualitative approach generally has an advantage in understanding.
It is unfortunately difficult to capture both simultaneously because they tend to work
against each other in practical terms. There is no reason why quantitative studies cannot
82 AMERICANJOURNAL OF EVALUATION,20(l), 1999
strive for depth, but limited resources will constrain the effort. Similarly, physical causal
reasoning can be carried out on a large sample, as the example of the survey in connection
with the crackdown on speeding was meant in part to show, but again resources constrain
depth when numbers are large. We can conclude regarding external validity that it is well to
be sensitive to both advantages-numbers and understanding-regardless of the approach
one takes, but the primary point to be made in the present context, perhaps, is that by no
means is the qualitative approach to impact analysis inherently inferior in terms of external
validity.
In this section, physical causal reasoning has been scrutinized in terms of internal and
external validity. One might well ask, in addition, how it would fare on criteria of appraisal
that have been more specifically devised for qualitative evaluation research. This area does
not have as long a history as internal and external validity and is not as well developed. We
may briefly consider, however, the overview by Lincoln and Guba (1986) of some of the
major suggestions and ideas. They are summarized into the two umbrella categories of “trustworthiness”
and “authenticity.”
It appears that some of these ideas, at least, are applicable to physical causal reasoning
and that the latter can be expected to score high on such criteria in a well planned and executed
study. Under trustworthiness, in particular, the category of transferability is suggested as a
parallel criterion to external validity. As indicated above in the discussion of “understanding
the process,” the criterion of the capability of the transfer of results from an observed application
to a potential one is a strong form of external validity and, perhaps, the strongest form that
can be achieved in social science, where universal laws are dubious. Similarly, the whole idea
of “audits” that has been developed in connection with qualitative research is fully applicable
to physical causal reasoning and would seem to be an effective way to insure against any
unavowed and unwelcome intrusions of personal values and biases. Finally, when applied to
qualitative design, Lincoln and Guba’s category of “credibility” would seem to have almost
the same meaning as internal validity in more traditional contexts. The issues that are important
to consider in this regard were summarized earlier in this section under the rubric of internal
validity.
In the category of “authenticity,” Lincoln and Guba (1986) discuss a number of criteria
that are not meant to be parallel to the criteria traditionally applied to quantitative evaluation
but grow rather out of the specific characteristics of qualitative evaluation. These criteria, such
as fairness, educative authentication, and catalytic authentication are less applicable to the
approach outlined here. The reason is that they are largely bound up with alternative goals for
evaluation, beyond impact analysis. In some dimensions they should be important for all evaluation,
including quantitative, as for example the importance of obtaining fully informed consent
from human subjects and informants or of considering the values of all stakeholders
rather than just a few when assessing the merit or worth of a program. For the most part, however,
they are not only pertinent primarily to goals and functions of evaluation other than
impact analysis, but at times they articulate those very goals. For example, according to Lincoln
and Guba, “It is.. .essential that they come to appreciate.. .the constructions that are made
by others and to understand how those constructions are rooted in the different value systems
of those others” (p. 81). Authenticity, in sum, seems less applicable to the physical causal
approach in qualitative evaluation than trustworthiness, but that is not surprising when the
central role of causation is considered.
Impact Analysis 83
CONCLUSION
Physical causation as a way of explaining the world seems to come naturally. There is ample
room in the practice of program evaluation for using it to demonstrate the causal nature of one
or more links on an outcome line. In particular, in-depth study and interpretation are well
suited to discovering the true or operative reason behind.a behavior. In this introductory treatment,
the purpose has not been to show how qualitative evaluation is done, but rather to show
how awareness of an alternative causal epistemology may be coupled with current knowledge
of qualitative methods of data collection and measurement to fashion a sound, independent
approach to the analysis of impacts.
ACKNOWLEDGMENTS
The text of this paper is based in part on Mohr (1995), Chapter 11. A prior version of the paper
itself was delivered at the annual meeting of the American Evaluation Association, Vancouver,
November l-5, 1995. I wish to express my appreciation to the editor, to his editorial assistant,
to Jennifer Greene, and to three anonymous reviewers for helpful comments and
criticisms on an earlier draft-to which, no doubt, I have not done full justice.
REFERENCES
Campbell,D. T., & Stanley, J. C. (1966).Experimental and quasi-experimental designs for research.
Chicago, IL: RandMcNally.
Chen, H. T. (1990). Theory-driven evaluations. Newbury Park, CA: Sage.
Cook, T. D., & Reichardt, C. S. (Eds.). Qualitative and quantitative methods in evaluation research (pp.
49-67). Beverly Hills, CA: Sage.
Cronbach, L. J. (1982). Designing evaluations of educational and social programs. San Francisco, CA:
Jossey-Bass.
Davidson, D. (1980). Essays on actions and events. Oxford: Clarendon Press.
Edwards, P. K., Acock, A. C., &Johnston, R. L. (1985). Nutrition behavior change: Outcomes of an educational
approach. Evaluation Review 9(4), 441-460.
George, A. L., & McKeown, T. J. (1985). Case studies and theories of organizational decision making.
In L. S. Sproull & P. D. Larkey (Eds.), Advances in information processing in organizations. R.
F. Coulam & R. A. Smith (Eds.), Research on public organizations. Vol. 2. (pp. 2l-58). Greenwich,
CT: JAI Press.
Goddard, A., & Powell, J. (1994). Using naturalistic and economic evaluation to assist service planning.
Evaluation Review, l&4), 472-492.
Greene, J. C. (1994). Qualitative program evaluation: Practice and promise. In N. K. Denzin & Y. S.
Lincoln @is.), Handbook ofqualitative research (pp. 530-544). Thousand Oaks, CA: Sage.
Greene, J. C., & Caracelli, V. J. (Eds.). (1997). Advances in mixed-method evaluation: The challenges
and benefits of integrating diverse paradigms. New Directions for Program Evaluation, No. 74.
San Francisco, CA: Jossey-Bass.
Guba, E. G., & Lincoln, Y. S. (1981). EfSective evaluation: Improving the usefulness of evaluation
results through responsive and naturalistic approaches. San Francisco, CA: Jossey-Bass.
Huberman, A. M., & Miles, M. B. (1985). Assessing local causality in qualitative research. In D. N. Berg
& K. K. Smith (Eds.), The selfin social inquiry (pp. 351-381). Newbury Park, CA: Sage.
84 AMERICAN JOURNAL OF EVALUATION, 20(l), 1999
Hume, D. (1955). An inquiry concerning human understanding. C. W. Hendell (Ed.). New York, NY:
Liberal Arts Press.
King, G., Keohane, R. O., & Verba, S. (1994). Designing social inquiry: Scientific inference in qualitative
research. Princeton, NJ: Princeton University Press.
Lincoln, Y. S., & Guba, E. G. (1986). But is it rigorous? Trustworthiness and authenticity in naturalistic
evaluation. In D. D. Williams (Ed.), Naturalistic evaluation (pp. 73-84). New Directions for Program
Evaluation, No. 30. San Francisco, CA: Jossey-Bass.
Ma&e, J. L. (1980). The cement of the universe: A study of causation. Oxford: Clarendon.
Maxwell, J.A. (1996). Using qualitative research to develop causal explanations. Working Papers. Cambridge,
MA: Harvard project on schooling and children.
Mohr, L. B. (1995). Impact analysis for program evaluation (second ed.). Newbury Park, CA: Sage.
Mohr, L. B. (1996). The causes of human behavior: Implications for theory and method in the social sciences.
AM Arbor, MI: University of Michigan Press.
Mohr, L. B. (1999). One hundred theories of organizational change: The good, the bad, and the ugly. In
H. G. Frederickson & J. M. Johnston (Eds.), Public administration in a time of turbulence: The
management of reform, reinvention, and innovation. Birmingham, AL: University of Alabama
Press, forthcoming.
Reichardt, C. S., & Cook, T. D. (1979). Beyond qualitative versus quantitative methods. In T. D. Cook
& C. S. Reichardt (Eds.), Qualitative and quantitative methods in evaluation research (pp. 7-32).
Beverly Hills, CA: Sage.
Striven, M. (1976). Maximizing the power of causal investigations: The modus operandi method. In G.
V. Glass (Ed.), Evaluation studies review annual, Vol. 1 (pp. 101-l 18). Beverly Hills, CA: Sage
Publications.
Shadish, W. R., Jr., Cook, T. D., & Leviton, L. C. (1991). Foundations ofprogram evaluation: Theories
of practice. Newbury Park, CA: Sage.
Stake, R. E. (1981). Case study methodology: An epistemological advocacy. In W. Welch. (Ed.), Case
study methodology in educational evaluation. Minneapolis, MN: Minnesota Research and Evaluation
Center.
Strawson, P. F. (1985). Causation and explanation. In B. Vermazen & M. B. Hmtikka @is.), Essays on
Davidson: Actions and events (pp. 155136). Oxford: Clarendon.