What Are the Potential Multiple-Case Study Designs (Types

3 and 4)?

The same case study may contain more than a single-case. When this occurs, the case

study has used a multiple-case study design, and such designs have increased in

frequency in recent years. A common example is a case study of a small group of public

versus private hospitals. Each hospital would be the subject of its own fieldwork, and the

multiple-case study would first cover each hospital as a single-case study before arriving

at findings and conclusions across the individual case studies.

Multiple- versus single-case designs.

In some fields, multiple-case studies have been considered a different methodology from

single-case studies. For example, both anthropology and political science have developed

one set of rationales for doing singlecase studies and a second set for doing what have

been considered “comparative” (or multiple-case) studies (see Eckstein, 1975; Lijphart,

1975).

This book, however, considers single- and multiple-case study designs to be variants

within the same methodological framework. No broad distinction is made between the

so-called classic (i.e., single) case study and multiple-case studies. The choice is

considered one of research design, with both being included as a part of case study

research.

Multiple-case study designs have distinct advantages and disadvantages in comparison

with single-case study designs. The evidence from multiple cases is often considered
more compelling, and the overall multiple-case study is therefore regarded as being more

robust (Herriott & Firestone, 1983). At the same time, the rationale for single-case

designs cannot usually be satisfied by the multiple cases. By definition, the unusual or

extreme case, the critical case, and the revelatory case all are likely to involve only

single-case studies. Moreover, the conduct of a multiple-case study can require extensive

resources and time beyond the means of a single student or independent research

investigator. Therefore, the decision to undertake a multiple-case study cannot be taken

lightly.

Selecting the multiple cases also raises a new set of questions. Here, a major insight is to

consider multiple-case studies as one would consider multiple experiments—that is, to

follow a “replication” design. This is far different from the misleading analogy that

incorrectly considers the multiple cases to be similar to the multiple respondents in a

survey (or to the multiple subjects within an experiment)—that is, to follow a “sampling”

design. The methodological differences between these two views are revealed by the

different rationales underlying the replication as opposed to sampling designs.

Replication, not sampling logic, for multiple-case studies.

The replication logic is directly analogous to that used in multiple experiments (see

Barlow, Nock, & Hersen, 2008). For example, upon uncovering a significant finding

from a single experiment, an ensuing and pressing priority would be to replicate this

finding by conducting a second, third, and even more experiments. Some of the

replications might attempt to duplicate the exact conditions of the original experiment.
Other replications might alter one or two experimental conditions considered challenges

to the original finding, to see whether the finding can still be duplicated. With both kinds

of replications, the original finding would be strengthened.

The design of multiple-case studies follows an analogous logic. Each case must be

carefully selected so that the individual case studies either (a) predict similar results (a

literal replication) or (b) predict contrasting results but for anticipatable reasons (a

theoretical replication). The ability to conduct 6 or 10 individual case studies, arranged

effectively within a multiple-case design, is analogous to the ability to conduct 6 to 10

experiments on related topics: A few case studies (2 or 3) might aim at being literal

replications, whereas a few other case studies (4 to 6) might be designed to pursue two

different patterns of theoretical replications. If all the individual case studies turn out as

predicted, these 6 to 10 cases, in the aggregate, would have provided

compelling support for the initial set of propositions pertaining to the overall multiple-

case study.6 If the

individual case studies are in some way contradictory, the initial propositions must be

revised and retested with another set of case studies. Again, this logic is similar to the

way researchers deal with conflicting experimental findings.

The logic underlying these replication procedures also should reflect some theoretical

interest, not just a prediction that two cases should simply be similar or different (e.g., in

a health care setting, see Dopson, Ferlie, Fitzgerald, & Locock, 2009). As another

example, consider the problem of advice-giving to city governments, on the part of

external expert groups. The typical experience is for an expert group to conduct some

research and then to present its advice in a report to a city agency. However, the common

outcome is for

such reports to receive little attention, much less to lead to any appropriate action. BOX

11 describes how a

multiple-case study addressed this issue.

Box 11 A Multiple-Case, Replication Design

Peter Szanton’s (1981) book, Not Well Advised, reviewed the experiences of numerous attempts by

university and nonuniversity research groups to advise city officials. The book is an excellent example of

a multiple-case replication design.

Szanton starts with eight case studies, showing how different university groups produced credible

research but nevertheless all failed to help city governments. The eight cases are sufficient “replications”

to convince the reader of a general phenomenon—the typical supposition being that the differences

between the academic and public policy cultures create an insurmountable communication barrier.

Szanton then provides five more case studies, in which nonuniversity groups also failed, concluding that

failure was therefore not necessarily inherent in the academic enterprise. Yet a third group of cases shows

how university groups have, in contrast, successfully and repeatedly advised sectors other than city

government, such as businesses and engineering firms. A final set of three cases shows that those few

groups able to help city government were concerned with implementation and not just with submitting a

research report containing new research-based ideas. The findings from all these case studies led to
Szanton’s major conclusion, which is that city governments may have peculiar needs in receiving advice

but then also putting it into practice.

Within each of the four groups of case studies, Szanton has illustrated the principle of literal replication.

Across the four groups, he has illustrated theoretical replication. This potent case study design can and

should be applied to many other topics.

The replication logic, whether applied to experiments or to case studies, must be

distinguished from the sampling logic commonly used in surveys. The sampling logic

requires an operational estimation of the entire universe or pool of potential respondents

and then a statistical procedure for selecting a specific subset of respondents to be

surveyed. The resulting data from the sample that is actually surveyed are assumed to

reflect the entire universe or pool, with inferential statistics used to establish the

confidence intervals for presuming the accuracy of this representation. The entire

procedure is commonly used when a researcher wishes to determine the prevalence or

frequency of a particular phenomenon.

Any application of this sampling logic to case study research would be misplaced. First,

case studies are not the best method for assessing the prevalence of phenomena. Second,

each individual case study would have to cover both the phenomenon of interest and its

context, yielding a large number of potentially relevant

variables (see Appendix B for a more detailed discussion). In turn, this would require an

impossibly large
sample of cases—too large to allow more than a superficial examination of any given

case.

Third, if a sampling logic had to be applied to all types of research, many important

topics could not be empirically investigated, such as the following problem: Your

investigation deals with the role of the presidency of the United States, and you are

interested in doing a multiple-case study of (a few) presidents to test your theory about

presidential leadership. However, the complexity of your topic means that your choice of

a small number of cases could not adequately represent all the 45 presidents since the

beginning of the Republic. Critics using a sampling logic might therefore deny the

acceptability of your study. In contrast, if you use a replication logic, a study is eminently

feasible.

The replication approach to multiple-case studies is illustrated in Figure 2.5. The figure

indicates that the

initial step in designing the study should preferably consist of theory development and

then shows that case selection and the definition of specific measures are important steps

in the design and data collection process. Each individual case becomes the subject of a

whole case study, in which convergent evidence is sought regarding the findings and

conclusions for the study; each case study’s conclusions are then considered to be the

information needing replication by the other individual case studies. Both the individual

case studies and the multiple-case results can and should be the focus of a summary

report. For each individual case study, the report should indicate how and why a

particular proposition was demonstrated (or not demonstrated). Across case studies, the
report should indicate the extent of the replication logic and why certain case studies

were predicted to have certain results, whereas other case studies, if any, were predicted

to have contrasting results.

An important part of Figure 2.5 is the dashed-line feedback loop. The loop represents the

situation where

important discovery occurs during the study of one of the individual cases (e.g., one of

the cases deviated unexpectedly from the original design). Such a discovery may require

you to reconsider one or more of the multiple-case study’s original theoretical

propositions. At this point, “redesign” should take place before proceeding further. Such

redesign might involve the selection of alternative cases or changes in the case study

protocol (see Chapter 3). Without such redesign, you risk being accused of distorting or

ignoring the

discovery, just to accommodate the original design. This condition leads quickly to a

further accusation—that you have been selective in reporting your data, to suit your

preconceived ideas (i.e., the original theoretical propositions).

Overall, Figure 2.5 depicts a different logic from that of a sampling design. The logic as

well as its contrast

with a sampling design may be difficult to follow and is worth extensive discussion with

colleagues before proceeding with any multiple-case study.

When using a multiple-case design, a further question you will encounter has to do with

the number of cases deemed necessary or sufficient for your study. However, because a
sampling logic should not be used, the typical criteria regarding the use of a power

analysis to determine the desired sample size (e.g., Lipsey, 1990) also are irrelevant.

Instead, you should think of the number of case replications—both literal and

theoretical— that you need or would like to have in your study.

Figure 2.5 Multiple-Case Study Procedure

Source: Cosmos Corporation.

Your judgment will be a discretionary, not formulaic, one. Such discretionary judgments

are not peculiar to case study research. They also occur in non–case study research, such

as in setting the criterion for defining a “significant effect” in experiments. Thus,

designating a “p < .05” or “p < .01” likelihood of detection, to set the confidence level for

accepting or rejecting the null hypothesis, is not based on any formula but is a matter of a

discretionary, judgmental choice. Note that when patient safety and well-being are at
stake, as in a clinical trial, investigators will usually not settle for a “p < .01” significance

level but may choose to attain a “p < .0001” or even more stringent level.

Analogously, designating the number of replications depends upon the certainty you want

to have about your multiple-case results. For example, you may want to settle for two or

three literal replications when your theory is straightforward and the issue at hand does

not demand an excessive degree of certainty. However, if your theory is subtle or if you

want a higher degree of certainty, you may press for five, six, or more replications.

In deciding upon the number of replications, an important consideration also is related to

your sense of the strength and importance of rival explanations. The stronger the rivals,

the more additional cases you might want, each case showing a different but predicted

result when some rival explanation had been taken into account. For example, your

original hypothesis might be that summer reading programs improve students’ reading

scores, and you already might have shown this result through two to three programs

whose case studies served as literal replications. A rival explanation might be that parents

also work more closely with their children during the summer and that this circumstance

can account for the improved reading scores. You would then find another case, with

parent participation but no summer reading program, and in this theoretical replication,

you would predict that the scores would not improve. Having two such theoretical

replications would provide even greater support for your findings.

Rationale for multiple-case designs.

In short, the rationale for multiple-case designs derives directly from your understanding

of literal and

theoretical replications (refer again to BOX 11). The simplest multiple-case design would

be the selection of

two or more cases that are believed to be literal replications, such as a set of case studies

with exemplary outcomes in relation to some evaluation question, such as “how and why

a particular intervention has been implemented smoothly.” Selecting such cases requires

prior knowledge of the outcomes, with the multiplecase inquiry focusing on how and why

the exemplary outcomes might have occurred and hoping for literal (or

direct) replications of these conditions from case to case.7

More complicated multiple-case designs would likely result from the number and types

of theoretical replications you might want to cover. For example, investigators have used

a “two-tail” design in which cases from both extremes (of some important theoretical

condition, such as extremely good and extremely bad outcomes) have been deliberately

chosen. Multiple-case rationales also can derive from the prior hypothesizing of different

types of conditions and the desire to have subgroups of cases covering each type. These

and other similar designs are more complicated because the study should still have at

least two individual cases within each of the subgroups, so that the theoretical

replications across subgroups are complemented by literal replications within each

subgroup.

Multiple-case studies: Holistic or embedded.

The fact that a design calls for multiple-case studies does not eliminate the variation

identified earlier with single-case studies: Each individual case study may still be holistic

or contain embedded subunits. In other

words, a multiple-case study may consist of multiple holistic cases (see Figure 2.4, Type

3) or of multiple embedded cases (see Figure 2.4, Type 4). The difference between these

two variants depends upon the type of

phenomenon being studied and your research questions. In an embedded multiple-case

design, a study even may call for the conduct of a survey at each case study site.

For instance, suppose a study is concerned with the impact of the training curriculum

adopted by different nursing schools. Each nursing school may be the topic of a case

study, with the theoretical framework dictating that nine such schools be included as case

studies, three to replicate a direct result (literal replication) and six others to deal with

contrasting conditions (theoretical replications).

For all nine schools, an embedded design is used because surveys of the students (or,

alternatively, examination of students’ archival records) are needed to address research

questions about the performance of the schools. However, the results of each survey will

not be pooled across schools. Rather, the survey results will be part of the findings for the

individual case study of each nursing school. The results may be highly quantitative and

even involve statistical tests, focusing on the attitudes and behavior of individual

students, and the data will be used along with information about the school to interpret

the success and operations with the training curriculum at that particular school. If, in

contrast, the survey data are pooled across schools, a replication design is no longer being
used. In fact, the study has now become a mixed-methods study (see discussion of mixed-

methods designs at the end of this chapter), the collective survey providing one set of

evidence and the nine case studies providing a separate set. Such a turn of events would

create a pressing need to discard the original multiple-case design. The newly designed

mixed-methods study would require a complete redefinition of the main unit of analysis

and entail extensive revisions to the original theories and propositions of interest.

Summary.

This section has dealt with situations in which the same investigation calls for multiple

cases and their ensuing case studies. These types of designs are becoming more

prevalent, but they are more expensive and timeconsuming to conduct.

Any use of multiple-case designs should follow a replication, not a sampling, logic, and a

researcher must choose each case carefully. The cases should serve in a manner similar to

multiple experiments, with similar results (a literal replication) or contrasting results (a

theoretical replication) predicted explicitly at the outset of the investigation.

The individual cases within a multiple-case study design may be either holistic or

embedded. When an embedded design is used, each individual case study may in fact

include the collection and analysis of quantitative data, including the use of surveys

within each case study