Experimental, Quasi-experimental, and Single Subject Research
774/801 Sept 1, 2004
John Hattie & Tony Hunt
It is simple: There is no perfect experiment
in education
There is nearly always a trade off between the
Power to generalise - from sample to population- from items to the behaviour domain- from conditions in the study to all intended conditions
and the
Power to convince- there are many audiences
PG PC
Power to Generalise
How confident can be generalise from the study to all “similar” situations
Is the design replicable/reproducible/exchangeable?
Is the evidence/conclusions unique to this study?
Have the generalisations taken into account all possible competing views – plausible alternative rival explanations (PARE)
PG
PC
Power to convince
Who are we trying to convince If it is a colleague(s) then more situation
specificity may be convincing (kids/classrooms/schools like mine)
If it is the educational community, then situation needs to be less critical
PC
PG
Resolution: Linking Power
Experimental design consists of a series of links:– It is as strong as the weakest link– Each link influences the next link– Desirable to have equal strength– Does each link have explanatory power– Are conclusions credible to the intended audience
History
Stanley and Campbell (1963) Cook and Campbell (1979) Shadish, Cook & Campbell (2002)
Evidence based
All based on designing studies that can lead to explanation and claims of causality
Explanation and Cause
1. Cause and effect must be related
(e.g., self-concept & achievement)
1. There needs to be temporal order (cause before effect)
2. Need to rule out other explanations/ other Plausible Alternative Rival Explanations (PARE)
Campbell & Stanley (1963)
Pretest-Posttest Control Group Design
Pre Treatment Post
R O X O
R O O
Randomisation – aiming for representativeness
But can we randomise
No Child Left Behind Tennessee Class Size Study
Quasi-experimentation:
When you do not have so much control over allocation of treatment, conditions, sample
When you have non-equivalent groups
In quasi-experimentation, the researcher has to enumerate alternative explanations one by one, decide which are plausible, and then use logic, design, and measurement to assess whether each one is operating in a way that might explain any observed effect (Shadish, Cook & Campbell, 2002, p. 14)
Relates to the Popper notion of falsification: What evidence would you accept that you are wrong?
Examples of Quasi-experimental designs
Divorce LawsOzdowski, S.A. & Hattie, J.A. (1981).
The impact of divorce laws on divorce rate in Australia: A time series analysis. Australian Journal of Social Issues, 16, 3-17.
0
2
4
6
8
10
12
14
16
18
20
1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981
a. Time SeriesO1 O2 O3 O4 O5 X O6 O7 O8
O9
ABA designA B A
O1 O2 O3 X4 X5 X6 O7 O8 O9 Le Fevre, et.al. (2002). Adequate Decoders
-10
0
10
20
30
40
50
60
70
80
90
100
0 10 20 30 40
Sessions
School 3
-10
0
10
20
30
40
50
60
70
80
90
100
0 10 20 30 40
Generalization
BLine 1 BLine 2 Intervention FollowUp
Manitenance
Daily Mean
School 1
-10
0
10
20
30
40
50
60
70
80
90
100
0 40
School 2
Multilevel Design: Hierarchical Linear Modelling
Students within classes within schools
E.g., Tracking/Streaming
School 1 School 2…
Teacher 1 Teacher 2
Class 1 Class 2 Class 1 Class 2
Structural Equation Modelling
Nutrition
nutr138e5
.74nutr123e4
.78nutr120e3.84
Exercise
exer53
e6
.62exer54e7
.54exer55e8
.85
Self Care
selfc59
e14
.57
selfc58
e13
.79
selfc57
e12
.66
selfc56
e11
.68
exer131e9 .88
exer139e10
.76
nutr51e2 .68
nutr50e1
.84
.79
.69
.80
Structural Model for Physical Self
Social
jfriend e15.97
jlove e16
.92
Interactive
jintell e17
.93
jwork e18.89
jemot e19.92
jcontr e20
.94
jhumor e21.86
-.23
.54
.60
e22
e23
-.14.55
.67
Minimal requirements for Studies
Sampling– Items to behaviour domains– People to all possible people– Conditions to all possible conditions
Representative sampling via– Random sampling– Stratified random sampling
Variables
At the end of your study, can I say “Aha, so that is what you mean, now I am clear”
Open constructs NOT Definitions No such thing as immaculate perception Dependent - Manipulable Independent - Nonmanipulable
Dependability
How reliable/consistent/replicable are your measures/ observations
Validity = Interpretations
Validity - "an integrated evaluative judgement of the degree to which empirical evidence and theoretical rationales support the adequacy and appropriateness of inferences and actions based on test scores or other modes of assessment".
Not validity of a test, but validity of interpretations
Validity of your study …
Is related to having ruled out
Plausible Alternative Rival Explanations (PARE)
CONTROL CONTROL CONTROL
… some examples
1. PARE: Power
Is your study POWERFUL enough to detect the effect you are investigating
1. PARE: Power
Is your study POWERFUL enough to detect the effect you are investigating
Do chickens have lips?
2. PARE: Chance
Did the effect/conclusion occur by chance
E.g., That two means are the same – the
hypothesis of no difference
Setting a rejection level, say =.05
3. PARE: Type II errors
Type I errors – Rejecting a claim when it is true (=.05)
Type II errors – Accepting a claim when it is false (e.g., chickens do not have lips, if it is indeed true)
4. PARE Reliability of your measures
If the reliability is low, then the scores “wobble” and no guarantee you will get same results using these instruments (tests, observations, interviews, etc.)
Was the treatment “consistent” in the various classes/implementations?
5. PARE: Was the treatment implemented?
Degree of implementation The Hong Kong Practical Science Study
(Cheung, Hattie, & Bucat, 1997)
6. PARE: Maturation
Showing change may not be enough as kids improve anyway (e.g., by maturation)
Method to measure change = Effect-sizesPost-Pre/spread = Effect-size
X2 – X1 sddiff
e.g., Before = 12, After = 15, spread = 6 15-12 = .5
6
0
5000
10000
15000
20000
25000
< -1
.01 -.98-.9
3-.8
8-.8
3-.7
8-.7
3-.6
8-.6
3-.5
8-.5
3-.4
8-.4
3-.3
8-.3
3-.2
8-.2
3-.1
8-.1
3-.0
8-.0
3 0.0
7.1
2.1
7.2
2.2
7.3
2.3
7.4
2.4
7.5
2.5
7.6
2.6
7.7
2.7
7.8
2.8
7.9
2.9
71.
021.
071.
121.
171.
221.
271.
321.
371.
421.
471.
521.
571.
621.
671.
721.
771.
821.
871.
921.
97>2
.0
Distribution of effects
Zero achievement
Average effect
0
5000
10000
15000
20000
25000
< -1
.01 -.98-.9
3-.8
8-.8
3-.7
8-.7
3-.6
8-.6
3-.5
8-.5
3-.4
8-.4
3-.3
8-.3
3-.2
8-.2
3-.1
8-.1
3-.0
8-.0
3 0.0
7.1
2.1
7.2
2.2
7.3
2.3
7.4
2.4
7.5
2.5
7.6
2.6
7.7
2.7
7.8
2.8
7.9
2.9
71.
021.
071.
121.
171.
221.
271.
321.
371.
421.
471.
521.
571.
621.
671.
721.
771.
821.
871.
921.
97>2
.0
Distribution of effects
Maturation
The disasters …
71 programmed instruction 801 .14
72 finances 1634 .14
73 problem based learning 41 .12
74 diet 255 .12
75 gender (female-male) 9020 .09
76 inductive teaching 570 .06
77 team teaching 41 .06
78 ability grouping 3355 .05
79 class size 2559 .05
80 open vs. traditional 3426 -.01
81 summer vacation 269 -.06
82 retention 3626 -.17
83 transfer of school 354 -.26
84 disruptive students 1511 -.78
The also rans …56 metacognitive intervention 921 .29
57 math programs 3326 .27
58 audio-visual 2699 .26
59 gifted programs 47 .25
60 coaching 1076 .24
61 behavior objectives 157 .24
62 calculators 238 .24
63 mainstreaming 1641 .21
64 questioning 493 .20
65 learning hierarchies 168 .19
66 attitude to math 1122 .19
67 desegregation 1590 .18
68 play 129 .16
69 television 4337 .15
Almost there …
42 tutoring 136 .35
43 activity-based programs 674 .35
44 remedial programs 1438 .35
45 classroom climate 2726 .35
46 social skills training 5472 .35
47 time 1680 .34
48 CAI 18231 .32
49 inquiry based teaching 2740 .32
50 preschool 242 .32
51 whole language 198 .31
52 within class grouping 2359 .31
53 testing 1463 .31
54 problem solving 1141 .30
55 background 692 .30
In the middle …
29 parent involvement 2597 .46
30 bilingual programs 1501 .46
31 adjunct aids 659 .45
32 concept mapping 18 .45
33 advance organizers 2106 .44
34 hypermedia instruction 317 .44
35 socio economic status 1657 .44
36 perceptual-motor skills 7592 .42
37 individualised instruction 5948 .42
38 homework 568 .41
39 competitive learning 144 .41
40 simulations 972 .37
41 expectations 912 .36
Worth having …
14 self-assessment 152 .54
15 mastery learning 1933 .53
16 creativity programs 2340 .52
17 interactive video 1152 .52
18 psycho-linguistics 4404 .51
19 goals 959 .51
20 peer influence 366 .50
21 early intervention 30971 .49
22 outdoor education 294 .49
23 science 4124 .49
24 inservice ed 18644 .48
25 acceleration 371 .47
26 motivation 2196 .47
The MAJOR Influences …
Influence # effects Mean
1 Direct instruction 1925 .93
2 Reciprocal teaching 52 .86
3 Feedback 13209 .81
4 Cognitive strategy training 7649 .80
5 Classroom behaviour 361 .71
6 Prior achievement 2094 .71
7 Phonological awareness 2630 .70
8 Home encouragement 25706 .69
9 Piagetian programs 786 .63
10 Cooperative learning 1153 .59
11 Reading programs 14945 .58
12 Quality of teaching 808 .55
13 Study skills 3224 .54
Identifying that what matters
Percentage of Achievement Variance
Students
Teachers
Home
PeersSchools Principal
7. PARE: Testing
People become test wise and/or may respond different when under test conditions
White space and testing in asTTle Testwiseness
Test of Objective EvidenceEach of the questions in the following set has a logical or
"best" answer from its corresponding multiple choice answer set. Please record your eight answers.
1. The purpose of the cluss in 2 Trassig is true when furmpaling is to removeA clump trasses the vonA cluss-prags B the viskal flans, if the viskal is B tremails donwil or zortilC cloughs C the belgo frulsD pluomots D dissels lisk easily
3 The sigia frequently overfesks the 4. The fribbled breg will minter best trelsum because with an
A all sigias are mellious A derstB all sigias are always votial B morstC the trelsum is usually tarious C sortarD no trelsa are feskable D ignu
Test of Objective Evidence, Part II
5 The reasons for tristal doss are 6 Which of the following is/are always present when trossels
are being gruven?
A the sabs foped and the doths tinzed A rint and vostB the dredges roted with the crets B vostC few rakobs were accepted in sluth C shum and vostD most of the polats were thonced D vost and plone
7 The mintering function of the ignu is most 8effectively carried out in connection with
A a razma toi AB the groshing stantol BC the fribbled breg CD a frailly sush D
8. PARE Statistical Regression
When taking extreme groups the means tend to move to the middle.
Why do the tallest fathers have shorter sons, and the shortest fathers have taller sons?
…. Regression to the Mean
Special Education (e.g., Sesame Street) Effective schools Gifted education
9. PARE Response rates
The returns of questionnaires/tests/interviews
should be high
What is typical?
Meta-analyses of Response Rates
Typical return is 50%
Three major factors:
1. Salience (77% vs 42%)
2. Number of follow ups (halve each time)
3. Lack of clutter/ orderliness
Not length (ave 7 pages, 72 questions), colour,
10. Change scores
The difference between post-pre scores
Problems
1 Unreliable
2 Are you measuring same thing both times
3 Regression to the mean
11 PARE: Experimenter effects
Hawthorne effect: Because we know we are in an experiment this alters our responses
Hans the Horse Pygmalion in the classroom Christine Rubie’s thesis Stanley Milgrim’s experiment
12. PARE: Restriction of range
When you choose/focus on a narrow range of abilities (etc.) this can be misleading
Picture …
0
2
4
6
8
10
12
14
16
18
20
0 5 10 15 20 25
0
2
4
6
8
10
12
14
16
18
20
0 5 10 15 20 25
13. PARE: Specification of target and accessible sample/population
Most experiments are highly local but have general aspirations
Often, there are two groups you are generalising to: e.g., all secondary students in NZ, and to all secondary students you have access -- from which to sample
14: Interactions
The model of individual differences indicates that we should modify our teaching methods to allow for individual differences in the class
0
5
10
15
20
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Competitive Learning
Cooperative Learning
Girls Boys
The art of research design is to devise experiments to identify the explanation and cause of effects – by
Maximising the chance that the conclusions are defensible and
Minimising the PAREs
Such that you have
Power to Generalise and
Power to Convince
Unobtrusive measures
Which painting do most people watch? Friendship in cities Racism in suburbs/cities
Statistical Methods to assist …
Correlation Analysis of variance (anova) Cross-tabulation
Comparing means: Magnitude and Chance
Magnitude Effect-sizes Chance Analysis of Variance
Well-being
What are the differences in levels of WELL-BEING among males and females, and between Australia and New Zealand
Country * GENDER Cross tabulationCount GENDER
MALE FEMALE Total
Country New Zealand 516 644 1160Australia 421 694 1115
Total 937 1338 2275
Australia Mn sd Effect-size
Male 45.7 10.6
Female 46.2 10.6 0.04
Total 46.0 10.6
New Zealand
Male 53.6 7.5
Female 54.3 7.4 0.08
54.0 7.5
NZ - Australia 0.89
anova
Source df MS F p
Country 1 35211.9 416.71 <.001
Gender 1 151.8 1.80 .180
Country * Gender 1 6.1 0.07 .787
Error 2271 84.5