You are on page 1of 8

The Behavior Analyst 1999, 22, 109-116 No.

2 (Fall)

Statistical Inference in Behavior Analysis:


Experimental Control is Better
Michael Perone
West Virginia University
Statistical inference promises automatic, objective, reliable assessments of data, independent of the
skills or biases of the investigator, whereas the single-subject methods favored by behavior analysts
often are said to rely too much on the investigator's subjective impressions, particularly in the visual
analysis of data. In fact, conventional statistical methods are difficult to apply correctly, even by
experts, and the underlying logic of null-hypothesis testing has drawn criticism since its inception.
By comparison, single-subject methods foster direct, continuous interaction between investigator
and subject and development of strong forms of experimental control that obviate the need for
statistical inference. Treatment effects are demonstrated in experimental designs that incorporate
replication within and between subjects, and the visual analysis of data is adequate when integrated
into such designs. Thus, single-subject methods are ideal for shaping-and maintaining-the kind
of experimental practices that will ensure the continued success of behavior analysis.

Science is a social enterprise, and ties of human judgment, which is error-


the standards of scientific evidence are prone and subject to an assortment of
established by consensus. From this troubling biases.
perspective, the objective of research
design and data analysis is straightfor- Limitations of Statistical Inference
ward: to convince an audience of skep-
tical colleagues that a particular inter- Or so it seems. Unfortunately, statis-
pretation or inference is justified. The tics offer the investigator no panacea,
rules of statistical inference, set forth and no self-respecting statistician
in classic texts and promulgated in would claim otherwise. Despite the
mandatory graduate courses, provide central role played by null-hypothesis
an agreed-upon solution. By following statistical tests throughout the biologi-
these rules-and rejecting the null hy- cal, behavioral, and social sciences,
pothesis with a p value of less than fundamental problems have been rec-
.05-investigators assure their peers, ognized for some time (e.g., Bakan,
and themselves, of the significance of 1966; Lykken, 1968; Meehl, 1967). By
their findings. Statistics guide investi- 1970, the criticisms of statistical infer-
gators to inferences about their data ence had drawn enough attention from
that can be expressed in objective, psychologists to warrant a book pro-
quantitative terms. Indeed, the infer- vocatively entitled The Significance
ences seem to arise automatically from Test Controversy (Morrison & Henkel,
the application of the statistical for- 1970). But actual use of statistical
mula, as implied by the term most analysis has not changed much since
commonly used to describe the pro- then; null-hypothesis testing is as ro-
cess: statistical inference. When sci- bust as ever. Cautions may be decreed
entific inferences are produced by a by textbook authors and professors in
formula, the investigator is relieved of statistics courses, but when students
a burdensome responsibility, and sci- and investigators are confronted with
ence itself is protected from the frail- real research problems, they are be-
guiled by the reassuring directness of
statistical procedures, which offer sim-
Requests for reprints should be sent to Mi- ple rules for answering a host of prac-
chael Perone at the Department of Psychology,
West Virginia University, P.O. Box 6040, Mor- tical questions ("How many subjects
gantown, West Virginia 26506-6040 (E-mail: per cell?"). In return, textbooks and
mperone@wvu.edu). professors seem more than willing to
109
110 MICHAEL PERONE
offer simple recipes for cooking up the less well quantitatively trained psy-
answers ("Run Cohen's power analysis chologists?" (p. 796).
program and see what it says").
Browse through an assortment of sta- On the Search for Methodological
tistics texts, and you will find many Imperatives
with handy tables and flow-charts to
guide the reader to just the right test Eventually some investigators dis-
for the data at hand, putting the task of cover that there is no good cookbook
analyzing the results of an experiment for delicious servings of research de-
on the same level as looking up a tele- sign and data analysis. But too many
phone number. The appearance of still see design and analysis as obsta-
"point-and-click" software for statis- cles to good research rather than an in-
tical analysis has made matters worse. tegral part of it. If they have grants,
When asked by a puzzled associate ed- they hire statistical consultants. If they
itor to explain an unusual statistic in a are graduate students, they make sure
manuscript submitted for publication, a statistics professor is a member of
more than one author has responded by their dissertation committee. The pre-
providing the name and version num- vailing attitude is that the framing of
ber of the software package. research questions can proceed apart
The bottom line is this: Too much from the methods employed to answer
research design and data analysis is them.
performed without thinking. Thomp- The attraction to formulas and rules
son (1998) voiced his objection this is not confined to investigators who fa-
way: vor group-statistical approaches. Pro-
fessors who teach courses in single-
Most researchers mindlessly test [the null hy- subject research design are confronted
pothesis] because most statistical packages only by students of behavior analysis seek-
test such hypotheses. This ... does not require ing, for example, rules about the cri-
researchers to thoughtfully extrapolate expected teria used to decide that behavior has
results from the previous literature or from the-
ory. Instead, science becomes an automated, reached a steady state. Over the years
blind search for mindless tabular asterisks using many students have reported that they
thoughtless hypotheses. (p. 799) adopted the criterion recommended by
Sidman (1960) in his classic Tactics of
Although statistical analysis has its Scientific Research. But Sidman never
defenders (e.g., Dixon, 1998; Hagen, offered such a recommendation. In an-
1997, 1998; Wilcox, 1998), the criti- swer to the question "How does one
cisms of years past continue to cause select a steady-state criterion?" he ex-
trouble, and debate about statistical plained, "There is ... no rule to fol-
strengths and weaknesses is being re- low, for the criterion will depend upon
peated and expanded by a new gener- the phenomenon being investigated
ation of psychologists and statisticians and upon the level of experimental
(e.g., Cohen, 1994; McGrath, 1998; control that can be maintained" (p.
Tryon, 1998). The greatest strength of 258). On what basis, then, is one to
statistical inference-the automatic, decide? Sidman pointed to the inves-
objective, reliable assessment of data, tigator's "accumulated experience and
all independent of the skills or biases good experimental judgment" devel-
of the investigator-is a mirage. Re- oped in the course of "designing and
search summarized by Tryon indicates carrying out steady-state experiments"
that statistical tests are routinely mis- (p. 261).
interpreted by investigators publishing We are left with a dilemma: Group-
in our best journals, and even by stat- statistical methods incorporate tidy sets
isticians themselves. "How much more of rules, but the rules lead to less than
susceptible to misinterpretation," he satisfactory results, even in the hands
asks, "are the vast majority of other of veterans. Single-subject methods
EXPERIMENTAL CONTROL 111

seem to offer no rules at all. What ner, 1956/1972, p. 113). Skinner was
guidance, then, is to be offered the stu- enthusiastic about the approach; he re-
dent embarking on a research career? ported that "the possibility of using
The answer may be found in three large groups of animals greatly im-
critical notions in the passages quoted proves upon (our) method ... since
from Sidman's (1960) book: experi- tests of significance are provided for
ence, experimental control, and judg- and properties of behavior not apparent
ment. These are recurring themes in in single cases may be more easily de-
Sidman's treatment of research tactics tected" (Skinner, 1956/1972, p. 113).
and, indeed, throughout the historical But Skinner's enthusiasm soon faded:
development of behavior analysis. Un- In actual practice that is not what happened....
derstanding the role they can play in You cannot easily make a change in the condi-
scientific research is the key to appre- tions of an experiment when twenty-four appa-
ciating why statistical inference has not ratuses have to be altered. Any gain in rigor is
been and need not become a major fac- more than matched by a loss in flexibility. We
tor in the experimental analysis of be- were forced to confine ourselves to processes
which could be studied with the baselines al-
havior. ready developed in earlier work. We could not
move on to the discovery of other processes or
Experience even to a more refined analysis of those we were
working with. No matter how significant might
Behavior analysts' interest in single- be the relations we actually demonstrated, our
statistical Leviathan had swum aground. (Skin-
subject as opposed to group-statistical ner, 1956/1972, pp. 113-114)
research may be regarded as the result
of an inductive process arising from in- Skinner, the consummate tinkerer,
tense interactions with data and various was quite willing to scout about for
practical considerations, rather than de- new ways to conduct experiments. He
ductions from a well-developed philos- rejected group-statistical methods not
ophy of science. The sophisticated because they collided with his radical
philosophical justification for single- behaviorist epistemology, but rather
subject research came later. because his experience revealed that
Behavior-analytic methods, of they insulated the investigator from the
course, derive from the work of Skin- behavior of the subject. The ongoing
ner, whose graduate training antedated interaction between experimenter and
the widespread adoption of the group- data that had characterized his earlier
statistical approach made possible by work-and led to his innovations in
Fisher (1925). Skinner's early research apparatus, measurement, and theory-
involved single-subject designs; most could not be sustained in group-statis-
of the experiments reported in his sem- tical research. Skinner returned to the
inal work, The Behavior of Organisms experimental analysis of individual be-
(Skinner, 1938), used only 4 rats. But havior, and directed his energies to de-
as large-group methods gained favor veloping stronger methods of experi-
within psychology in the late 1930s, mental control that would obviate the
Skinner, then an assistant professor at need for statistical inference.
the University of Minnesota, gave
them a try. He and Heron built a set of Experimental. Control
24 operant chambers and cumulative The tension between group-statisti-
recorders, interconnected so that the re- cal and single-subject methods is cre-
corders displayed mean performances ated by the relative roles played by ex-
for the entire group of 24 rats, as well perimental control in the two ap-
as subgroups of 12 and 6. Skinner said proaches. For Skinner and other advo-
that he and Heron "thus provided for cates of single-subject research,
the design of experiments according to group-statistical methods are ill suited
the principles of R. A. Fisher, which to the development of strong forms of
then were coming into vogue" (Skin- experimental control over behavior, in
112 MICHAEL PERONE
part because the group methods are un- When a monetary reinforcer was pre-
wieldy and in part because the nature sented (an occasional event given the
of statistical analysis reduces the in- intermittent nature of the schedule), the
vestigator's motivation to establish subject repositioned himself and
such control. The sensitivity of a sta- pressed a button on the console re-
tistical test is a direct function of the quired to collect the reinforcer, then re-
number of subjects, and weak control sumed the rocking motion. This topog-
can be tolerated if the number is large raphy was wholly compatible with the
enough. Averaging data across many monetary schedule, which involved
subjects can hide a multitude of sins: only the plunger response, but the in-
The experimental treatment may fail to vestigators worried that it would inter-
affect the behavior of some subjects, fere with the acquisition of the observ-
and may even lead to contrary effects ing response, because the observing
in others. As a consequence, statisti- keys would usually be out of the sub-
cally significant results based on large ject's reach. To block the chair-rocking
sample sizes are not persuasive. Given topography, the investigators replaced
a sufficiently large sample, statistical the chair with a wheeled stool. The
significance is assured. Meehl (1967) subject reacted by sitting on the floor,
pointed out that the only question is tying his bootlace to the plunger and
whether the direction of the statistical pulling the other end, and occasionally
difference will support the investiga- standing up to collect reinforcers. The
tor's hypothesis. Under these circum- new topography was no better than the
stances, the probability of support is a old one. Finally, the investigators
lofty .5-hardly a rigorous experimen- placed a limited hold on the collection
tal challenge. button: Once a monetary reinforcer
In single-subject research, by com- was earned, the subject had just 1 s to
parison, treatment effects are clarified get up and collect it before it was can-
not by increasing statistical sensitivity celed. This contingency was effective
but rather by improvements in experi- in moving the subject ornto the stool in
mental control. Individual differences front the console, with the collection
are not averaged into obscurity as sta- button and the observing keys within
tistical error, but instead are regarded easy reach. When the critical phase of
as revealing the limits of the control the experiment finally commenced, the
being exercised. subject acquired the observing re-
As a case in point, consider a situ- sponse and his data fell in line with
ation encountered in the course of an those of the other subjects.
experiment on "observing behavior" The close interaction between inves-
in adult humans (Perone & Baron, tigator and subject fostered by the sin-
1980). The main response was pulling gle-subject approach allowed a poten-
a plunger mounted underneath a table. tial disaster to be identified and avert-
On the table was a console with col- ed. The troublesome individual differ-
ored stimulus lamps and several re- ence was not relegated to a statistical
sponse keys. In the critical conditions, error term, but was eliminated by suit-
pressing the "observing" keys on the able adjustment in the experimental
console would turn on colored lights procedure. What would have happened
correlated with the schedules of mon- in a group experiment? Perhaps the ab-
etary reinforcement associated with the sence of a conditioned reinforcement
plunger response. During preliminary effect in the problem subject would
training, 1 subject adopted an unusual have been overlooked, if it did not ap-
response topography: He tied one end preciably affect the group mean. Or, if
of his bootlace to the plunger and the detected, the negative result might
other end to the leg of his chair, put his have been attributed to the regrettable
feet on the table, and executed the re- but inevitable appearance in the sample
sponse by rocking back and forth. of a recalcitrant subject whose person-
EXPERIMENTAL CONTROL 113

ality leads to sabotaging experimental of their discipline, theoretical predilec-


goals. Of course, nothing about group tions, or epistemological convictions.
designs prevents the kind of corrective The adoption of group-statistical
action taken in this case. But the ready methods does not eliminate the need
acceptance of individual differences for an investigator's sound judgment,
and other forms of "error variance" nor does the adoption of single-subject
seems more amenable to the theory methods guarantee it. The two kinds of
and practice of group-statistical re- methods do, however, place different
search than to single-subject research. judgmental burdens on the investigator.
As noted elsewhere (Perone, 1991), And because of the relative rarity of
a successful experimental science is single-subject methods, the burdens of
one that exerts high degrees of control that tradition are often misunderstood.
over its subject matter. The ability to Perhaps the greatest misunderstanding
control variables that affect behavior is revolves around the so-called "visual
prerequisite to the study of steady analysis" of data.
states. Thus, because single-subject de- When it comes to analyzing experi-
signs require investigators to seek strict mental results, the difference between
levels of control, their adoption en- group-statistical and single-subject
courages the development of an exper- methods is sometimes characterized
imental science of behavior. along these lines: In group research, in-
ferences about causal relations between
Judgment independent and dependent variables
are guided by precise, sophisticated
Although some discussions of statistical tests free of subjectivity and
group-statistical methods may suggest bias. In single-subject research, inves-
otherwise, human judgment is an un- tigators stumble along with only a sim-
avoidable component of the scientific ple graph of the results to inspect un-
enterprise. Investigators must exercise aided, leaving their causal inferences
their best judgment repeatedly over the susceptible to all manner of idiosyn-
course of a research project. At the out- cratic influences. Again, the absence of
set they must decide what line of in- codified rules for conducting the visual
vestigation is likely to make a contri- analysis is seen as the culprit. Kazdin
bution to knowledge. Then they must (1982) expressed the problem this way:
devise appropriate experimental de-
signs and procedures, often balancing Perhaps the major issue pertains to the lack of
concrete decision rules for determining whether
competing interests based on conve- a particular demonstration shows or fails to
nience, economy, and the availability show a reliable effect. The process of visual in-
of apparatus and personnel. They must spection would seem to permit, if not actively
puzzle over the measures to employ, encourage, subjectivity and inconsistency in the
analyses to conduct, which results are evaluation of intervention effects. (p. 239)
worth reporting, and the implications Research is available to bolster this
of the results for contemporary theo- criticism. Investigators given session-
retical debate. They must decide how by-session graphs of concocted behav-
methods, results, and arguments should ioral data and asked to judge the pres-
be conveyed to the scientific commu- ence of treatment effects may disagree
nity in the form of grant applications, with one another, be swayed by seem-
publications, and professional presen- ingly minor details of the graphic pre-
tations. Sometimes they must decide sentation, overlook small but reliable
whether a negative outcome should effects, or see effects when they are ab-
spur a reappraisal of one's experimen- sent (DeProspero & Cohen, 1979;
tal strategy or abandonment of a cher- Knapp, 1983; Matyas & Greenwood,
ished theoretical position. All these 1990; but see Parsonson & Baer, 1992,
judgments and more are a matter of for a more appreciative account of vi-
routine for active scientists regardless sual analysis).
114 MICHAEL PERONE
The rejoinder is that criticism of vi- study of behavior the instrument is the
sual analysis is based on a profound experiment. To be valid, a single-sub-
misunderstanding. Indeed, the very ject experiment must show that behav-
term visual analysis-and the research ioral states can be replicated at will in
into it-does not adequately represent different subjects and at different times
the process as it occurs in actual re- within the same subject. Replication
search. Perhaps the problem can be thus establishes the investigator's suc-
traced to the comparison with statisti- cess in identifying and controlling rel-
cal analysis. Statistical tests are con- evant variables and confirms the ade-
ducted after an experiment is complet- quacy of the stability criteria that guide
ed and the results are in. At that point, the investigator's decisions about the
the investigator is left to sift though the attainment of steady states (see Baron
data and seek evidence that an effect & Perone, 1998, and Perone, 1991, for
was brought about by the experimental detailed discussion of the validity of
manipulations. Critics of visual analy- single-subject experiments). In this
sis seem to believe that it is merely an connection, it is noteworthy that the
unsophisticated version of the same previously cited research questioning
process: After the experiment the in- the adequacy of visual analysis does
vestigator draws a graph of the results not address the role of replication
and decides about the influence of the across subjects, nor does it express
independent variable. But in practice doubt about the conclusions of actual
no single-subject experiment is con- single-subject research.
ducted in such a fashion. Visual anal-
ysis is an ongoing activity throughout Conclusions
the experiment; indeed, it is an integral
part of the experimental analysis and The question prompting this essay is
as such it cannot be separated from the the role inferential statistics should
methods employed to collect the data play in behavior analysis. Ever since
in the graphs. group-statistical methods gained favor
The point may be clarified by re- in psychology, behavior analysis has
stating it with a more appropriate em- drawn criticism for its devotion to sin-
phasis: Experimental analysis is an in- gle-subject methods. This essay has
tegral part of visual analysis. By this tried to show that the criticisms are
account, it is a mistake to suggest that based on an exalted and erroneous
investigators in the single-subject tra- view of the power of statistical infer-
dition prefer the visual inspection of ence, one that regards statistical tests
graphs over statistical analysis. What is as a set of tried and true rules that re-
preferred is an experimental analysis liably and inevitably guide investiga-
so thorough, so powerful in its control tors to objective answers for their ex-
over the subject matter of interest, that perimental questions. In practice, how-
cause-effect relations are plain to see. ever, statistical inference is not so sim-
The experiment may be regarded as ple. The rules, such as they are, have
any other scientific instrument, such as proven difficult to apply, even in the
a microscope, whose resolution is hands of statisticians, and the under-
painstakingly refined until the object of lying logic of null-hypothesis testing
study comes into clear focus. The be- has drawn fire since its popularization
havior analyst does not rely on unaided by Fisher nearly 75 years ago. Para-
senses to see causal relations in behav- doxically, the criticism most often lev-
ior any more than the biologist relies eled against single-subject methods-
on the naked eye to see subcellular ob- that they do not ensure consistent out-
jects. The adequacy of visual analysis comes across investigators-seems to
depends on, and can be no greater than, apply equally to group-statistical meth-
the adequacy of the instrument aiding ods.
the investigator's vision, and in the Tests of statistical inference may
EXPERIMENTAL CONTROL 115

have their place in psychology, and son, single-subject methods put inves-
perhaps even in behavior analysis. But tigator and subject into repeated con-
there is no room for the unthinking tact, and force the investigator to iden-
methodological orthodoxy that often tify and control variables relevant to
accompanies statistical inference. Per- the object of study. Thus, the methods
haps the trouble started when Camp- are ideal for shaping-and maintain-
bell and Stanley (1963) proclaimed ing-the kind of experimental practic-
that the only "true experiment" is one es that will ensure the continued suc-
with random assignment of subjects to cess of behavior analysis.
treatment groups. Campbell and Stan-
ley directed their monograph to field REFERENCES
researchers in education, and it seems
unlikely that they intended to dismiss Bakan, D. (1966). The test of significance in
single-subject experiments (or, for that psychological research. Psychological Bulle-
matter, virtually all natural science be- tin, 66, 423-437.
fore 1925) as invalid. But by parroting Baron, A., & Perone, M. (1998). Experimental
design and analysis in the laboratory study of
Campbell and Stanley's monograph human operant behavior. In K. A. Lattal & M.
with insufficient thought or circum- Perone (Eds.), Handbook of research methods
spection, several generations of text- in human operant behavior (pp. 45-91). New
books on psychological research meth- York: Plenum.
ods have surely had that unfortunate Campbell, D. T, & Stanley, J. C. (1963). Ex-
perimental and quasi-experimental designs
effect. for research. Chicago: Rand McNally.
Whatever methods are adopted by Cohen, J. (1994). The earth is round (p < .05).
behavior analysts, let us ask that they American Psychologist, 49, 997-1003.
be adopted thoughtfully. The cookbook DeProspero, A., & Cohen, S. (1979). Inconsis-
recipes sometimes associated with sta- tent visual analysis of intrasubject data. Jour-
nal of Applied Behavior Analysis, 12, 573-
tistical inference are easy to criticize, 579.
but more thoughtful statistical appli- Dixon, P. (1998). Why scientists value p values.
cations may be welcome. In the same Psychonomic Bulletin & Review, 5, 390-396.
vein, it must be recognized that the de- Fisher, R. A. (1925). Statistical methods for re-
mand for cookbooks is not altogether search workers. Edinburgh, UK: Oliver &
Boyd.
absent from the behavior-analytic com- Hagen, R. L. (1997). In praise of the null hy-
munity. Sidman (1960), as he wrote his pothesis statistical test. American Psycholo-
Tactics, was perhaps the first to feel the gist, 52, 15-24.
demand. His response was to stead- Hagen, R. L. (1998). A further look at wrong
fastly refuse to offer any recipes. In- reasons to abandon statistical testing. Ameri-
can Psychologist, 53, 801-803.
stead, he asked his readers to think an- Kazdin, A. E. (1982). Single-case research de-
alytically about their research ques- signs: Methods for clinical and applied set-
tions, to explore new procedures, and tings. New York: Oxford University Press.
to learn from experience-in short, to Knapp, T. (1983). Behavior analysts' visual ap-
develop good experimental judgment. praisal of behavior change in graphic display.
Behavioral Assessment, 5, 155-164.
The present view, derived from the Lykken, D. T (1968). Statistical significance in
insights and advice offered by Sidman psychological research. Psychological Bulle-
and Skinner, is that in a science of be- tin, 70, 151-159.
havior good judgment is shaped by in- Matyas, T. A., & Greenwood, K. M. (1990). Vi-
tensive interplay between investigator sual analysis of single-case time series: Ef-
fects of variability, serial dependence, and
and subject in the course of experi- magnitude of intervention effects. Journal of
mental analysis. Group-statistical Applied Behavior Analysis, 23, 341-35 1.
methods seem ill suited to the task, McGrath, R. E. (1998). Significance testing: Is
tending to insulate the investigator there something better? American Psycholo-
from the immediate results of experi- gist, 53, 796-797.
Meehl, P. E. (1967). Theory testing in psychol-
mental operations and reducing the ogy and physics: A methodological paradox.
motivation for seeking and exercising Philosophy of Science, 34, 103-115. (Reprint-
strong forms of control. By compari- ed in D. E. Morrison & R. E. Henkel, Eds.,
116 MICHAEL PERONE
The significance test controversy, pp. 252- correlated with extinction or increased effort.
266. Chicago: Aldine, 1970) Journal of the Experimental Analysis of Be-
Morrison, D. E., & Henkel, R. E. (Eds.). (1970). havior, 34, 239-261.
The significance test controversy. Chicago: Sidman, M. (1960). Tactics of scientific re-
Aldine. search. New York: Basic Books.
Parsonson, B. S., & Baer, D. M. (1992). The Skinner, B. F (1938). The behavior of organ-
visual analysis of data, and current research isms. New York: Appleton-Century.
into the stimuli controlling it. In T. R. Skinner, B. F (1972). A case history in scien-
Kratochwill & J. R. Levin (Eds.), Single-case tific method. In B. F Skinner (Ed.), Cumula-
research design and analysis: New directions tive record (3rd ed., pp. 101-124). New York:
for psychology and education (pp. 15-40). Appleton-Century-Crofts. (Original work
Hillsdale, NJ: Erlbaum. published 1956)
Perone, M. (1991). Experimental design in the Thompson, B. (1998). In praise of brilliance:
Where that praise really belongs. American
analysis of free-operant behavior. In I. H. Iver- Psychologist, 53, 799-800.
sen & K. A. Lattal (Eds.), Techniques in the Tryon, W. W. (1998). The inscrutable null hy-
behavioral and neural sciences: Vol. 6. Ex- pothesis. American Psychologist, 53, 796.
perimental analysis of behavior, Part I (pp. Wilcox, R. R. (1998). How many discoveries
135-171). Amsterdam: Elsevier. have been lost by ignoring modem statistical
Perone, M., & Baron, A. (1980). Reinforcement methods? American Psychologist, 53, 300-
of human observing behavior by a stimulus 314.

You might also like