ERIC EJ964431: Data-Based Decision Making: The Impact of Data Variability, Training, and Context

2011·0.38 MB·English

by  ERIC

JOURNALOFAPPLIEDBEHAVIORANALYSIS 2011, 44, 767–780 NUMBER4 (WINTER2011)
DATA-BASED DECISION MAKING: THE IMPACT OF DATA
VARIABILITY, TRAINING, AND CONTEXT
NICHOLAS R. VANSELOW
NEWENGLANDCENTERFORCHILDREN
RACHEL THOMPSON
WESTERNNEWENGLANDCOLLEGE
AND
ALLEN KARSINA
NEWENGLANDCENTERFORCHILDREN
The current study examines agreement among individuals with varying expertise in behavior
analysisaboutthelengthofbaselinewhendatawerepresentedpointbypoint.Participantswere
askedtorespond tobaseline dataandtoindicate whentoterminate the baseline phase.When
onlyminimalinformationwasprovidedaboutthedataset,expertsandBoardCertifiedBehavior
Analystparticipantsgeneratedbaselinesofsimilarlengths,whereasnovicesdidnot.Agreement
wassimilaracrossparticipantswhenvariabilitywaslowbutdeterioratedasvariabilityinthedata
setincreased.Participantsgeneratedshorterbaselineswhenprovidedwithinformationregarding
theindependentordependentvariable.Implicationsfortrainingandtheuseofvisualinspection
arediscussed.
Keywords: data analysis, reversal design, scientific behavior, visual inspection
_______________________________________________________________________________
Appliedbehavioranalysisincorporatessingle- an ‘‘inconsistent treatment,’’ and an ‘‘irrevers-
subject design and visual inspection as the ible effect.’’ Editors of the Journal of Applied
primary means of identifying functional rela- Behavior Analysis (JABA) and the Journal of the
tions between independent and dependent Experimental Analysis of Behavior rated the
variables. Applied behavior analysts are charged demonstration of experimental control shown
with producing meaningful and important by each graph on a scale from 0 (low) to 100
changes in socially significant behavior in each (high), and poor agreement was obtained. The
individual they serve, and this task is facilitated meancorrelationbetweenraterswas.61,accord-
byvisualinspectionofsingle-subjectdata(Baer, ing to the Pearson product moment correlation.
1977; Skinner, 1956). In addition, the programmed differences in the
Despite the predominant use of visual data patterns accounted for only a very small
inspection in behavior analysis, its reliability proportion of the variance in the ratings. Other
has been questioned by studies that have shown
studies using similar procedures also demon-
poor agreement between scientists. For exam-
strated low agreement between visual inspectors
ple,DeProspero andCohen (1979)constructed
(Danov&Symons,2008;Furlong&Wampold,
graphs that depicted ABAB reversal designs
1982; Matyas & Greenwood, 1990; Park,
with data that reflected an ‘‘ideal’’ pattern,
Marascuilo, & Gaylord-Ross, 1990).
These studies question the utility of visual
Address correspondence to Rachel Thompson, Western inspectionasareliabledeterminantoftheeffect
New England College, Department of Psychology, 1215 of an independent variable. However, Kahng
WilbrahamRoad,Springfield,Massachusetts01119(e-mail:
etal.(2010)systematicallyreplicatedDeProspero
[email protected]).
doi:10.1901/jaba.2011.44-767 and Cohen (1979) and found much higher
767
768 NICHOLAS R. VANSELOW et al.
agreement for ratings made using the 100-point to make general comments about the data
scale (.93 compared to .61 in DeProspero & (Austin & Mawhinney), to characterize the
Cohen). Two methodological differences be- data in terms of structured categories (e.g.,
tween Kahng et al. and DeProspero and Cohen stability and trend; Mawhinney & Austin),
might be responsible for these discrepant results. and to identify the point at which in-
First, Kahng et al. used the same 100-point tervention was initiated. Compared with
scale as DeProspero and Cohen but also asked other research on visual inspection, these
participantswhetherthegraphsdid(‘‘yes’’)ordid two studies presented data in a manner that
not (‘‘no’’) demonstrate experimental control. was more similar to the daily decisions made
Second, participants included only editors of during the research process. These studies
JABA. Kahng et al. suggested that the increased also differed from previous research in that
agreementmaybeduepartlytothestandardized participants responded to published data sets,
trainingforappliedbehavioranalystssuchasthat whereas most other studies (e.g., DeProspero
& Cohen, 1979; Furlong & Wampold, 1982;
recommended by the Behavior Analyst Certifi-
Kahng et al., 2010; Matyas & Greenwood,
cationBoard(BACB).Replicationoftheseresults
1990) presented computer-generated graphs
and analyses of the variables that influence
that may not resemble data typically encoun-
agreement or disagreement among visual inspec-
tered by participants.
torswithdifferenttraininghistoriesarenecessary
Like Austin and Mawhinney (1999) and
toreconcileKahngetal.’srecentfindingswiththe
Mawhinney and Austin (1999), we presented
many studies that have shown poor agreement.
data point by point and used graphs that
Participants in the current study included three
includeddatafrompublishedstudies.However,
groups (experts, Board Certified Behavior Ana-
participants in our study were asked to make
lysts[BCBAs],andnovices)withvaryinglevelsof
decisions about whether or not to continue the
expertise, allowing an examination of agreement
baseline phase of the study and to provide
within and across these groups.
commentsabouttheir decisions. InStudy 1,we
The studies of agreement among scientists
examined agreement regarding baseline length
have focused on visual inspection of data sets
among participants with varying levels of
with at least two complete phases and agree-
expertise in behavior analysis. In Study 2, we
ment of experimental control. However, the
evaluated the effects of providing additional
final data set is a result of many decisions
information regarding the independent or de-
regardingwhethertocontinueorchangephases,
pendent variable on the lengths of participant-
and these decisions often occur before experi-
generated baselines.
mental control is demonstrated. Point-by-point
decisions are integral to deciding when to
STUDY 1
change phases, and small decisions may affect
the demonstration of experimental control in The purpose of Study 1 was to describe and
the final graph. We identified only two studies compare the length of baseline phases created
that attempted to capture point-by-point data by experts in applied behavior analysis, BCBAs,
analysis. and novices in applied behavior analysis when
Austin and Mawhinney (1999) and Ma- data were presented point by point.
whinney and Austin (1999) asked participants
to respond to data presented point by point. Method
Graphs were from currently or subsequently Participants. Participants in the expert group
published data sets with phase lines removed. had doctoral-level degrees in behavior analysis
Acrossthetwostudies,participantswereasked or a related field, had earned their BCBA (or
DATA-BASED DECISION MAKING 769
Board Certified Behavior Analyst–Doctoral) intellectual disabilities. Just prior to participating
certification, and were currently or previously in the study, the participants were assigned
Associate Editors or Editors for JABA. To Chapter 6 on visual inspection from the intro-
recruit participants for this group, the authors ductory behavior analysis textbook by Cooper,
contacted 13 experts via an e-mail with the Heron,andHeward(1987,pp.130–141).Seven-
experimental program attached. Ten experts teen participants from this group completed the
completed the experimental program and sent experimental program.
the file containing their responses back to the
authors. Procedure
The second group of participants did not Experimental program. A computer program
meet the expert criteria but were BCBAs. was designed to present a graph, one data
(Board Certified Assistant Behavior Analysts point at a time. The experimental program
or BCaBAs were excluded from this group.) utilized the Visual Basic for Applications pro-
BCBAs were recruited in two ways. First, we gramming language available with Microsoft
created a computerized spreadsheet from the Windows versionsofExcel 2003 and 2007.By
certificant registry on the BACB Web site using common spreadsheet software, we were
(http://www.bacb.com). The spreadsheet ran- able to send the experimental program in a
domly selected 20 BCBAs at a time. We small file attached to an e-mail. Participants’
contacted these people via e-mail to request data were recorded directly into the experi-
their participation in the study. After confir- mental program file that was sent back to the
mation, participants were sent the experimental experimenter.
program. Four of the 77 people randomly After the experimental program began, par-
selected from all of the BCBAs listed in the ticipants were provided with informed consent
certificantregistryrespondedtoaninitiale-mail followed by instructions for using the program.
The instructions stated that
and completed the experimental program. Due
to the low response rate with the first method, The goal for each graph is to continue adding data
we recruited additional participants by e- points to the baseline phase by clicking the
‘‘Continue Running Current Phase Button’’ until
mailing a variety of schools and service agencies
you have enough data points that you would begin
that employed BCBAs to ask if any of their the next phase (e.g., ‘‘treatment’’) if this were an
employees would be interested in participating. applied research project you were supervising or
conducting.
Throughthisrecruitingmethod,nineadditional
e-mails with the experimental program attached Before each graph was presented, a message
were sent, and six participants returned their box displayed the x- and y-axis labels, a
responses. A total of 10 BCBAs completed the reminder that this was the baseline phase, and
experimental program. the expected change in behavior for the next
The third group of participants included phase. For example, the message box indicated
novices in applied behavior analysis. All partici- thatthetargetbehaviorwasexpectedtodecrease
pants had been hired recently at a school for in the next phase if a reduction in behavior was
children and adults with intellectual disabilities expected in the phase after baseline (as in the
and were enrolled in an introductory course in case of a treatment for aggression). Participants
behavioranalysis.Theseparticipantsdidnothave could not view the graph until the ‘‘OK’’
previous training in the interpretation of graphs button in the message box was selected.
that depict single-subject experiments; however, After clicking ‘‘OK,’’ the first graph was
they were learning about behavior-analytic topics displayed with a single data point. An example
relevant to providing services to children with of Graph 5 with one data point is displayed in
770 NICHOLAS R. VANSELOW et al.
Figure 1.Graph 5 before (top) andafter(middleandbottom) datapoints wereadded.
DATA-BASED DECISION MAKING 771
Table 1 decision.Atthebottomofthedialogueboxwas
GraphCharacteristics a button labeled ‘‘continue.’’ If the participant
chose to continue with the current phase,
y-axis Data Variability Expected
clicking ‘‘continue’’ submitted the participant’s
Graph label points coefficient Trend change
responses,hidthedialoguebox,andshowedthe
1 rpm 6 .08 none decrease
2 rpm 4 .30 increasing decrease graph with another data point added. Figure 1
3 rpm 5 .51 none decrease (middle and bottom panels) displays Graph 5
4 rpm 11 .90 Upattern decrease
5 frequency 8 .88 none increase with data points added as if the participant had
clicked the ‘‘continue running current phase’’
Note. rpm 5 responses per minute. For Graph 5,
participantsinStudy1saw‘‘numberofcorrectresponses’’ button. If the participant chose to stop adding
andparticipantsinStudy2saw‘‘responses’’ontheyaxis. baseline sessions (i.e., the participant selected
the ‘‘stop baseline and start intervention’’
thetoppanelofFigure 1.Thegraphscontained
button), the current graph ended and the
axis labels that matched those in the previous
experimental program proceeded to the basic
information message box. In Study 1, the y axis
information about the next graph. Each subse-
waslabeled‘‘responsesperminute’’or‘‘number
quent graphfollowed thissamebasicprocedure.
ofcorrectresponses,’’andthetopographyofthe
Whenallofthegraphs(fivegraphsforStudy
responsewasnotspecified.Table 1providesthe
1) had been presented, the experimental pro-
specificlabelsusedforeachgraph.Therangeof
gram ended by thanking the participant and
y-axis values was determined by the default
providing instructions to send the completed
settings in Microsoft Excel. The x-axis labels
file back to the experimenter as an e-mail
were always ‘‘sessions.’’ The x axis always had
attachment. Clicking the ‘‘complete’’ button at
session numbers that corresponded to the data
thebottomofthismessagesavedthechangesto
points plus five session numbers (without data)
the file and closed the program.
beyondthecurrentdatapoint.Therewerethree
Selected published graphs. The graphs dis-
buttons on the screen in addition to the graph.
A button above the top right corner of the played by the experimental program contained
graph was labeled ‘‘see graph information’’ and datafromstudiespreviouslypublishedinJABA.
displayed the message box with the basic First, graphs that met the following criteria
information about the previously described from issues of JABA from 1998 to 2008 were
graph when selected. enteredintoadatabase.Onlygraphsthatdepict
Below the graph were buttons labeled an ABA or ABAB reversal design with a single
‘‘continue running current phase’’ and ‘‘stop data path were selected. Graphs could contain
baseline and start intervention.’’ Clicking on additional phases (e.g., ABAC reversal design);
eitherof thesetwobuttonspresentedadialogue however,thereversaldesigncouldnotbepartof
box that asked if the participant was sure about another experimental design (e.g., a reversal for
this decision. Selecting ‘‘no’’ would return the ParticipantAinthetoppanelthatwasalsopart
participant to the graph with no changes. If the of a multiple baseline across Participants B and
participant selected ‘‘yes,’’ the graph disap- C). These criteria were developed to eliminate
peared from the screen and a dialogue box graphs that may have resulted from researcher
appeared. This box displayed two items with responses to features of the data that were not
space to reply to each. The first item asked why presented to participants in our study.
the participant chose to continue or stop the Fromthisdatabaseofgraphs,weselectedfive
current phase. The second item was a text box graphs that differed in trend and degree of
labeled ‘‘comments’’ in which the participant variability. The characteristics of each of graph
couldtypeadditionalinformationaboutthelast are described in Table 1. The five graphs were
772 NICHOLAS R. VANSELOW et al.
presented to each participant in semirandom (instead of generated uniquely for each partic-
order. Five possible sequences of graph presen- ipant) so that each participant saw the exact
tationwerecreated beforetheprogram wassent same data and decisions between participants
to the participants. To create the sequences, could be compared.
each of the graphs was presented first across the
Prior to the start of the study, computer-
five sequences, and the order of the remaining
generated graphs and the original graphs were
four graphs was randomized. Participants then
presented to four doctoral-level behavior analysts
were assigned randomly to receive one of the
who served as faculty for a PhD program
fivesequences;however,eachsequencewasused
in behavior analysis to determine whether the
once before a sequence was repeated.
computer-generatedgraphssimulatedtypicaldata
Computer-generated data points. A potential
patterns.Onepersonwhorated thesegraphsalso
limitation of using published data is that
participated as an expert in Study 1 (Participant
participants might choose to continue baseline
A-2). The other three did not participate in any
beyond the number of data points available in
otherpartofthestudy.Thebehavioranalystswere
the original graph. To address this problem,
providedwiththefiveoriginalgraphsandthefive
computer-generated data points were calculated
computer-generated graphs presented in a ran-
using the mean, range, and standard deviation
dom order. For each graph, the participant was
of the original baseline data. However, these
askedwhetherthegraphdisplayedpublisheddata
calculations were not basedon the entire graph.
or computer-generated data. A third option
Instead, the graph was divided into two to four
allowed the participant to respond that he or she
sections (consisting of at least three data points
did not know whether the data were obtained or
each) based on the number of data points,
computer generated. Three of the four partici-
and the mean, range, and standard deviation
pants responded that they did not know which
were calculated for each section separately.
data were published or computer generated. The
To generate a data point, the program first
fourth participant correctly identified only five
semirandomly selected one of the sections, with
ofthe10graphs,whichwaschancelevelrespond-
the last section being selected more often than
ing. These results suggested that the computer-
the other sections. Next, a number that fell
generated data could not be distinguished from
between a range that was 20% larger than the
thepublisheddata,increasingourconfidencethat
original range was generated randomly. For
the computer program generated data with
example, if the original range was between 10
patterns similar to data collected in the course of
and20,adatapointbetween8and22couldbe
applied research.
generated. The slightly larger range was includ-
ed because, visually, computer-generated data Calibration and testing. Extensive testing and
points in graphs created without this increase calibrationoccurredbeforethecomputerprogram
appeared at more predictable values and more was sent to the participants. The first author
often within a much smaller range than the generatedlistswiththenumberofdatapointsfor
original graphs. The computer program was each graph and responses to open-ended ques-
designed to generate data points that would tions.Then,hecreatedgraphswithintheprogram
preservethemeanandstandarddeviationofthe usingthoselistsofdatapointsandresponses.The
current graph. Using this type of calculation, datarecordedbythecomputerprogramandthose
thepatterns apparent in the originalgraph were from the lists were compared. Any calculations
preserved as the graph continued past the data in the computer program were checked against
presented in the original published graph. Each calculations completed manually by the author.
graphwasextendedto300datapointsusingthe After this stage was completed, the authors sent
program described above and was pregenerated the programs to colleagues with instructions to
DATA-BASED DECISION MAKING 773
Figure2.DatafromtheexpertparticipantsinStudy1.Thedifferencefromthemeanwascalculatedbysubtractingthe
mean numberofadded points averaged for agraph fromthe numberofdata points generated byasingle expert.
complete the program and provide comments regarding agreement easier among graphs by
about their user experience. These testers also providing a standard reference point from
were instructed to try to cause the program to which to analyze the differences in agreement
fail. Information collected from the initial testers for each of the graphs. The number and range
was used to improve the program and solve of data points are provided in some figures to
any problems the participants might experience. show differences in the graphs created by
Finally, a second group of testers completed the different participant groups. Using difference
updated program. The first author checked the from the mean in all graphs might show the
data collected throughout the process to ensure levelofagreementamonggroupsbutwouldnot
that it was complete and calculations were provide information about different baseline
accurate. lengths created by the different participants.
Data analysis. The two primary dependent
measures were the number and range of data Results and Discussion
points added to each graph and the difference InStudy1,thedecisionsmadeinresponseto
from the mean number of data points. Differ- data presented point by point were described
ence from the mean was calculated by subtract- for three different participant groups. Figure 2
ing the mean number of data points averaged displays the results for the participants in the
across all participants for a graph from the expert group. This graph presents the number
number of data points generated by a single of data points each participant generated for
participant. Data points concentrated around each graph as a difference from the mean. For
zeroindicatethatmostoftheparticipantsadded Graphs 1 and 2, which contained data paths
a similar number of data points to the graph. with little variability, the experts generated
More disagreement among participants is ap- graphs with very similar numbers of data
parent when the data points vary further points. Nine of the 10 participants generated
from zero. This analysis makes comparisons graphs within one data point of each other.
774 NICHOLAS R. VANSELOW et al.
Participant D-1 added approximately two more studies varied from .1 to .25 (DeProspero &
data points than most of the other participants Cohen, 1979; Kahng et al., 2010). In the
for Graphs 1 and 2. Graph 1 contained a data current study, the variability coefficient varied
path with no trend, whereas the data path from.08to.90.Participantscreatedbaselinesof
in Graph 2 had a steadily increasing trend similar lengths for graphs with variability
(Table 1). Disagreement among the expert coefficients less than .25 in the current study.
participants increased as the variability in the This supports the evidence that agreement
graphs increased. Participants created graphs among visual inspectors may be higher than
withawiderrangeinthenumberofdatapoints previously reported when variability is within
for Graphs 3, 4, and 5. For Graph 5, the ranges previously studied (Kahng et al.,
Participant C-1 continued baseline until 41 2010). However, as variability in the data
data points were presented. increased beyond .30, agreement began to
Similar results were obtained for the BCBA decrease. The variability of Graphs 3, 4, and 5
group; however, the results for the novices was .51, .90, and .88, respectively. At this level
differed from those of the experts and BCBAs. of variability, there was a wider range of data
Themeannumberofbaselinedatapointsforall points added to the graph before participants
graphs across the three participant groups is began the next phase.
displayed in Figure 3. For Graphs 1, 2, and 3, Participants also provided an explanation for
the results obtained from the BCBA group their decision. The two types of comments that
closely matched those from the expert group. occurred most often after participants were
For Graphs 4 and 5, the mean number of data asked why they continued baseline was that
points in the BCBA graphs differed slightly either one data point was insufficient to change
from the experts, although the general pattern phases (six experts and nine BCBAs) or that
acrossgraphswassimilar(thevariabilityandthe three data points are required for a baseline
mean number of data points in each graph (four experts and two BCBAs). Although some
increased as the variability in the data path participants changed phases at three data points
increased). For novice participants, the mean for graphs with stable data paths, some
number of data points remained approximately participants chose to exceed this minimum by
thesameforeachofthefivegraphs.Inaddition, one or two data points. We also provided the
the variability among participants decreased as participants with the opportunity to provide
the data paths in the graphs became more any additional comments. In this section, three
variable; an opposite relation was observed with experts stated that their decisions might have
experts and BCBAs. Based on these data, it been different if more information about the
appears that, in isolation, rules for visual type of treatment that would follow was
inspection provided in textbooks may be provided. These experts might have reduced
insufficient to produce expert performance. It the number of data points in baseline if the
seems likely that additional experience with treatment was one that was likely to produce a
data-based decision making accounts for the strong effect. Three experts and three BCBAs
differences among these groups. commented that their decisions may have been
To compare the current results with previous affectedbyinformationaboutthetopographyof
studies, a variability coefficient for each of the theresponse(e.g.,severeself-injuriousbehavior).
five graphs was calculated by dividing the Although the response topography never was
standard deviation of a set of data by the mean provided, one participant in the BCBA group
(Kahng et al., 2010). The variability coefficient stoppedthegraphafteronedatapointandmade
ofthedata presentedto participantsin previous commentssuggestingthatshewouldevaluatethe
DATA-BASED DECISION MAKING 775
Figure 3. The mean number of data points for each graph from all groups in Study 1. The bars represent the
minimumandmaximum numberof datapoints foreach group.
safety of conducting sessions with such a high- the problem behavior and the likely effects of
rate behavior or attempt different analyses. the subsequent treatment. These two categories
were selected due to comments from the
participants in Study 1 and suggestions in
STUDY 2
previous research (e.g., DeProspero & Cohen,
ParticipantsinStudy1andinpreviousstudies 1979; Kahng et al., 2010; Lerman et al.,2010).
(Danov & Symons, 2008; DeProspero & In addition, Cooper, Heron, and Heward
Cohen, 1979; Kahng et al., 2010; Matyas & (2007) specifically addressed these two situa-
Greenwood,1990)wereaskedtomakedecisions tions as an exception to the ‘‘rule of the-more-
about data with minimal information. Com- data-points-the-better’’ (p. 150).
ments collected from our participants suggested
that their decisions about the length of baseline Method
would be influenced by information about the Participants.TheparticipantsinStudy2were
type of treatment planned and the form or students studying behavior analysis in doctoral
severity of the targeted behavior. In addition programs across the United States. Students
to these variables, studies have suggested that were contacted either directly via e-mail or
decisions might be influenced by research through their academic advisers. The experi-
previously published in the area, the potential mental program was sent to 19 students, and
significanceoftheresults(DeProspero&Cohen, eight participants in this group completed both
1979), or the characteristics of the participants Sessions1and2.Oftheeightparticipants,four
(Kahng et al., 2010; Lerman et al., 2010). wereBCBAs,onewasaBCaBA,andthreewere
The purpose of Study 2 was to evaluate the not certified.
effects of information about the target behavior Procedure. The experimental program from
or the independent variable on the decisions of Study 1 was used for Study 2 and functioned
individualparticipants.Weexaminedtheeffects similarly, with a few exceptions. Only the two
of information about the form and severity of graphs with the highest variability coefficients
776 NICHOLAS R. VANSELOW et al.
Table2 each graph before it was presented; however, in
Graph Orderin Study2 Study2,themessageboxalsoincludedfictitious
client initials, and different information was
Session1 Session2
provided according to the no-information,
Graph Informationtype Graph Informationtype
strong treatment, and severe self-injury condi-
5 none 4 none tions described below. The client initials were
4 severeself-injury 5 severeself-injury
generated by selecting two consecutive letters of
5 strongtreatment 4 strongtreatment
4 strongtreatment 5 strongtreatment the alphabet randomly (e.g., PQ or LM) and
5 severeself-injury 4 severeself-injury
were added to imply that each graph displayed
4 none 5 none
a different set of data. The order of graphs
and conditions varied across sessions, but all
were included (Graphs 4 and 5). These graphs
participants experienced the same sequence of
were chosen because responses from Study 1
graphs and conditions.
showed poor agreement with these graphs. The
No information. The information presented
variability in responses in Study 1 suggested
to the participants was similar to that in Study
that these graphs would provide a baseline that
1, except that for Graph 5 the expected
might be sensitive to information manipula-
direction of change was ‘‘decrease’’ instead of
tions. However, Graph 5 was modified slightly
‘‘increase’’ and the y-axis label was ‘‘responses’’
to prevent information in the graph from
instead of ‘‘number of correct responses.’’ The
conflicting with the additional information
participant was informed of the participant’s
provided. First, the y-axis label in Graph 5
initials, the axis labels, and the expected
was changed from ‘‘number of correct respons-
direction of change.
es’’to‘‘responses.’’Second,theexpectedchange
Strong treatment. In addition to the basic
in direction was altered from ‘‘increase’’ to
information, an additional line in the message
‘‘decrease’’ because one type of information
boxstatedthatthetreatmentimplementedafter
involved labeling the target response as severe
baseline was expected to have a strong effect on
self-injury. This modification seemed reason-
behavior.Nodefinitionofstrongtreatmentwas
able, given the relatively high level of behavior
provided.
depicted in Graph 5 (see Figure 1).
Severe self-injury. In addition to the basic
Participants completed two separate visual
information, an additional line in the message
inspection sessions, with the experimental
program for the second session sent to box stated that the target behavior was severe
participants after submission of their responses self-injury. No additional information about
fromthefirstsession.Tocreategraphsforthese the form of behavior was provided.
two sessions, the data from Graphs 4 and 5
were used to generate 300 points per graph. Results and Discussion
Computer-generated Points 150 to 300 were Figure 4 displays the main effects of Study 2.
presented as needed in Session 1; the published For both graphs, additional information resulted
data and, if necessary, computer-generated inashorterbaseline,onaverage,comparedtono
Points 1 to 149 were presented in Session 2. information.Theaveragelengthofbaselineacross
Each of the two graphs was presented three participants changed even with the addition of
times in a session (once for each of the three veryminimalinformation.Althoughthebaselines
conditions), fora totalof six graphs persession. decreased in length and agreement increased
Table 2 displays the order of graphs for each slightlyforthestrongtreatmentcondition,much
session. Similar to Study 1, the experimental greaterandconsistentdecreasesacrossparticipants
program presented the basic information about occurred in the severe self-injury condition.