Documente Academic
Documente Profesional
Documente Cultură
Overview
Role of ITS in the history of WSC
Two classes of ITS for WSCs
ITS
A series of observations on a dependent variable
over time
N = 100 observations is the desirable standard
N < 100 observations is still helpful, even with very few
observationsand by far the most common!
0.4
0.2
0
-0.2
Impact
Begins
-0.4
-0.6
Se
p86
Ja
n8
M 7
ay
-8
Se 7
p87
Ja
n8
M 8
ay
-8
Se 8
p88
Ja
n8
M 9
ay
-8
Se 9
p89
Ja
n9
M 0
ay
-9
Se 0
p90
Ja
n9
M 1
ay
-9
Se 1
p91
prenatal drinking
0.6
Label
Law
Date
Disadvantages
Typically very simple with only a few design elements (perhaps a
control group, little chance to introduce and remove treatment,
rarely even implemented with multiple baseline designs).
Usual problems with uncontrolled and unknown attrition and
treatment implementation
We have yet to find a really strong example in education
Disadvantages:
Of less general policy interest except in some circles (e.g., special
education) but
IES now allows them for both treatment development studies and for impact
studies under some conditions.
Increasing interest in medicine (e.g. CENT reporting standards).
Roifman et al (1987)
High-dose versus low-dose intravenous
immunoglobulin in hypogammaglobulinaemia and
chronic lung disease
12 patients in a longitudinal randomized cross-over
design. After one baseline (no IgG) observation:
Group A: 6 receive high dose for 6 sessions, then low dose
for 6 sessions.
Group B: 6 receive low dose for 6 sessions, then high dose
for 6 sessions
Analysis Strategy
To compare RE to SCD results, we analyze the
data two ways
As an N-of-1 Trial: Analyze Group B only as if it
were six single-case designs
As a RE: Analyze Time 6 data as a randomized
experiment comparing Group A and Group B.
Analyst blinding
I analyzed the RE
David Rindskopf analyzed SCD
Analytic Methods
The RCT is easy
usual regression (or ANOVA) to get group mean
difference and se.
We did run ANCOVA covarying pretest but results were
essentially the same.
Analysis: RCT
If we analyze as a randomized experiment
with the endpoint at the last observation
before the crossover (time 6):
Group A (M = 794.93, SD = 90.48)
Group B (M = 283.89, SD = 71.10)
MD = 511.05 (SE = 46.98) (t = 10.88, df = 10, p <
.001)
Analysis 2: SCD
If we analyze only Group B (6 cases) using a destimator1:
g = 4.59, V(g) = 1.43 (se = 1.196)
Close to RE estimate g = 5.80, V(g) = 1.98 (se = 1.41)
Example: PRT
Pivotal Response Training (PRT) for Childhood
Autism
This WSC method does a meta-analytic
comparison of results from SCDs to results from
an RE.
Meta-analytic WSCs have a long history but also
have significant flaws in that many unknown
variables may be confounded with the designs.
But those flaws may often be no more than in the
usual 3-arm nonrandomized WSC
Big difference is the latter usually has raw data but
meta-analysis does not. In the case of SCDs, however,
we do have the raw data (digitized).
Statistics
G
N
Valid
Missing
Mean
Median
Std. Deviation
Minimum
Maximum
54
0
1.311520
1.041726
1.2359393
-0.4032
5.4356
g
.25643452
.54197801
.63926057
.75588367
.99578116
1.3178189
1.3252908
1.6105048
1.6148902
1.6345153
2.5494302
2.5985178
2.5989373
3.5641752
vg
.06523894
.06563454
.02315301
.03473584
.17598106
.36416389
.3465313
.11272789
.1154149
.23100681
.26735438
.67319145
.78074928
.15104969
w
15.328268
15.235879
43.190927
28.788705
5.6824295
2.7460164
2.8857422
8.8709195
8.6643927
4.3288767
3.7403539
1.4854615
1.2808209
6.6203379
I2 = 85.14%
More on Michalopolous
They did not analyze the ITS as a time series.
Instead just substituted the comparison group
for the control group and analyzed like a
randomized experiment. This is not how an ITS
would be analyzed.
In addition, this has the usual confounds
between method and third variables.
But this perhaps can and should be done more
Does someone want to reanalyze Michalopolous?
Compare results from usual randomized estimate at
one time point to
Analyzing the randomized control as ITS
Analytic Issues
Because so many real examples are short ITS,
usual ARIMA modeling etc is not practical.
Key analytic issues:
What is the metric for the comparison of ITS and
RE?
Dealing with trend
Dealing with autocorrelation
Analytic Issues:
When the Metric is the Same
To compare RE to ITS, the same effect estimate has to be
measured in the same metric
Not a problem
In longitudinal randomized crossover designs
In three arm studies in which participants are randomized to all
methods and treated identically and simultaneously (e.g.,
Shadish et al., 2008, 2011)
In three-arm studies like Michalopolous that are all part of one
large study
Or over multiple studies if they just happened to use the same
outcome variable.
Analytic Issues:
How to Deal with Trend in the ITS
The issue is to prevent the presence of linear
or nonlinear trend from causing a spurious
difference between RE and ITS estimate.
6
5
Baseline
Series1
Treatment
2
1
0
1
10
Trend:
Outcome Already in Common Metric
If you do not need to convert to a common metric (e.g., outcome in
RE and ITS is identical). Two options:
Model trend using ordinary regression (Huitema 2011) or multilevel
models (Van den Nortgate & Onghena; Kyse, Rindskopf & Shadish).
But may produce two estimates:
Main effect of treatment
Interaction of trend with treatment
Trend:
Outcomes in Different Metrics
E.g., in the PRT meta-analysis where each
study had the same construct but different
measures:
Detrend the data using methods previously
described
Then compute HPS d
Or wait till HPS d with trend adjustment is ready in
a few years.
Introduction to GAM
Like a parametric regression, but replacing some or all
of the parametric predictors with smoothed
nonparametric predictors. E.g.,
Parametric:
Yt = 0+ 1Xt + 2zt + 3[Xt (n1 + 1)]zt + t.
Fit indices
Borderline linear
parametric trend
Tmt effect is signif.
GAM Conclusions
About our three questions:
A trend by treatment might be present
If so, it may be highly nonlinear
But the treatment effect is robust to trend compared
to the usual GLM.
Typically computed with a standard YuleWalker estimator on the residuals from four
n 1
parameter regression:
yt yt 1
rj t 1n
yt 2
t 1
Variability in Autocorrelations
Raw autocorrelations can be quite variable:
(1
2
j
) (t j 3)
F
D
N
J
L
A
F
D
N
J
L
A
K
H
B
E
K
H
B
E
I
0.20
C
C
N
A
D
K
JFI
L
B
E
H
M
G
C
FI
N
D
JL
A
K
E
H
B
M
G
C
N
D
JFI
L
A
K
H
E
B
M
G
I
F
N
D
J
L
A
K
H
E
B
M
G
F
N
D
J
L
A
K
H
B
E
M
G
0.10
0.15
C
A
B
D
E
G
H
K
M
N
JFI
L
F
D
N
J
L
A
K
H
B
E
0.05
G
M
0.2
M
-0.8
-0.2
0.0
Conditional Mean
-0.4
0.25
-0.6
0.30
0.0
0.4
0.35
0.005
0.013
0.030
0.065
0.140
Tau
0.303
0.668
1.558
4.157
A= (Intercept) B= 1 C= 2 D= 3 E= 4 F= 5 G= 6 H= 7 I= 8 J= 9 K= 10 L= 11 M= 12 N= 13
0.4
0.3
0.35
0.2
0.0
0.1
Conditional Mean
0.20
A
B
C
E
B
C
E
B
C
E
B
C
E
B
C
E
B
C
E
0.399
0.676
1.189
2.288
-0.3
0.05
B
C
E
-0.1
A
B
C
E
0.10
0.15
D
D
A
B
C
E
-0.2
0.25
0.0
0.30
0.025
0.048
0.085
0.143
0.239
Tau
A= (Intercept) B= 1 C= 2 D= 3 E= 4
Logistical Issue
How to get the data for the time series.
Sometimes it is available in an archive etc.
Sometimes it has to be digitized from graphs
Questions to Ask
The one area where we still need studies on
the main effect question of can ITS = RE?
Design variations to also examine:
Does it help to add a nonequivalent control?
Does it help to add a nonequivalent DV?
What about variations in ITS design?
Ordinary ITS with one intervention at one time
Multiple baseline designs with staggered
implementation of intervention over time
Over cases
Over measures within one case
Conclusion
WSCs of ITS and RE badly needed if ITS is
going to regain the credibility it once had
(assuming it does, in fact, give a good answer).
Challenging to design and analyze, but much
progress has been made already, especially in
SCDs and N-of-1 Trials.
Questions?