Вы находитесь на странице: 1из 10

Quasi-Experimental Design

A quasi-experiment is one where the treatment variable is manipulated but the


groups are not equated prior to manipulation of the independent variable. I shall
discuss a few such designs here.
Pretest-Posttest Nonequivalent Groups Design
N O X O
This design looks a lot like the randomized pretest-posttest
design that I discussed earlier, but in this case the two groups have N O O
not been equated prior to treatment. Since one has not randomly
assigned subjects to groups, one cannot assume that the populations being compared
are equivalent on all things prior to the treatment, and accordingly internal validity is
threatened. When a post-treatment difference between groups is observed, one cannot
with great confidence attribute that effect to the treatment, since the groups may have
had pre-existing differences that caused the observed post-treatment difference.
In the language of Campbell and Stanley, the threat here is selection and all of
the various interactions involving selection. One can try to mitigate the problem by
assigning subjects to groups (or selecting intact groups) in ways that make it likely that
the groups do not differ greatly prior to the treatment, but one always worries about the
unknown variables on which the groups might differ and which might affect the criterion
variable.
Trochim has written several pages in which he presents hypothetical examples of
this design and then critiques them with respect to which threats to internal validity
seem of most concern and which seem unlikely. You should read those pages carefully
to get a feel for the sort of thinking that is involved when trying to determine how
suspicious one should be of conclusions drawn from research gathered through a
pretest-posttest nonequivalent groups design.
So, why would one ever choose to employ this design? Frankly, one should not,
if a better design (such as a randomized pretest-posttest design) is feasible, but
sometimes you cannot accomplish random assignment, especially when dealing with
human subjects (who don’t like to be told what treatments they are or are not going to
get) in field settings.

Double-Pretest Nonequivalent Groups Design


N O O X O
This modification of the pretest-posttest N O O O
nonequivalent groups design helps to control for a
Selection x Maturation interaction by including a second pretest. If the groups are
maturing at different rates, that difference may appear in the comparison between the
first pretest and the second pretest.
Regression-Discontinuity Design

Copyright 2003, Karl L. Wuensch - All rights reserved.

Research-8-QuasiExpDesign.doc
2

This design looks a lot like the pretest-posttest


nonequivalent groups design, but the groups are nonequivalent
C O X O
by choice. The ‘C’ the first column indicates that the subjects are C O O
assigned to groups based on their score on the covariate (the
pretest).
One usually starts by deciding what the treatment and criterion variables will be.
For example, I may decide that my treatment variable will involve an online tutorial
program in basic statistics and my criterion variable will be students’ performance in an
undergraduate statistics class. I shall offer the program to some students (and entice
them to use it) but not to other students. At the end of the semester I shall use scores
on the comprehensive final examination as the criterion variable.
So, what do I use as the pretest? I could administer the comprehensive final
examination twice, at the beginning of the semester (pretest) and the end of the
semester (posttest), but that might just reveal that none of my students know any
statistics before they take a statistics class. I might decide to use an alternative sort of
pretest, such as a test of statistics aptitude (based largely on verbal and logical
reasoning but with a little math thrown in too). I would want scores on this test to be
highly predictive of final examination performance -- that is students who do well on this
aptitude test (those with good verbal and logical skills, even if they don’t know any
statistics yet) will be highly likely to do well in the course, while those who door poorly
on the aptitude test will likely have great difficulty with the course.
Now, how am I going to assign subjects to groups? I could just announce the
availability of the tutorial program and let anybody use it. I would keep track of who
used it and who did not, and at the end of the semester I would compare those two
groups in terms of how they performed on the final examination. I would use the pretest
scores as a covariate in an ANCOV. Conducted in this fashion, this would be an
example of a pretest-posttest nonequivalent groups design. There almost certainly
would be pre-treatment differences between those who elected to use the tutorial
program and those who elected not to. While some researchers think that the use of
ANCOV enables one statistically to remove such a confound, that is not exactly true.
Interpretation of the results of an ANCOV where the experimental groups differ on the
covariate is very tricky. Even if the ANCOV could remove the confound involving pre-
treatment differences on the covariate, the groups probably differ on other important
characteristics that could affect the results.
So, I abandon the idea of using a pretest-posttest nonequivalent groups design.
I decide instead to use a randomized pretest-posttest design. I use a random number
generator to decide who gets to use the tutorial program and who does not. Not long
after the start of the semester I start to get angry phone calls from concerned parents
who are worried that their children will be at a disadvantage in the class because they
are in the deprived (control) group. They explain that being so deprived could lead to
them failing the class, or getting a grade that would keep them out of graduate school or
cost them a scholarship and so on. I am risking those students’ futures for my stinking
research and they are not going to stand for it. After a few of these parents (and/or the
students) call up my chair, my dean, the chancellor, the Board of Governors, their
legislative representatives, the local media, and so on, and their attorneys phone the
3

university attorney about a lawsuit, I decide that maybe a randomized pretest-posttest


design is not feasible.
I might then decide to try a switching-replications design, where half the class
gets access to the tutorial program during the first but not the second half of the
semester, and the other half get access to it the second but not the first half of the
semester. I would, of course, measure performance in the class both on a midterm
exam and a final exam. This switch might, however, not satisfy my critics. Those who
have to wait until the second half of the semester might rightfully complain that by that
point they will already be hopelessly lost in the class without access to the tutorial --
after all, if you don’t have a good understanding of the material covered during the first
half the semester (things like means, variance, probability, and inference), you are not
likely to be able to understand much of what is taught during the second half of the
semester. Also, those who have access to the tutorial the first half of the semester are
likely to complain bitterly about having it taken away from them during the second half of
the semester. Maybe the switching-replications design is not feasible here either.
How can I possibly assign subjects to my treatment groups in a way that will not
upset so many people? One interesting possibility is to apply the treatment only to
those students who are most in need of it -- that is, provide the tutoring only to those
students whose scores on the pretest indicate that they will probably have a lot of
trouble with the course. There still might be some complaints. The more able students
may argue that they too could benefit from the tutorial and it is not fair that they be
denied just because they have higher aptitude -- but there is an American tradition of
favoring the underdogs, of giving “special education” to those with low aptitude but not
those with high aptitude, so social pressures would probably cause these whining
brainiacs to shut up. I might also get some complaints from the students who were
selected to get the tutorial -- “Why do I have to this extra work when those other
students don’t, it just isn’t fair!” Well, you know what they say about pleasing all the
people all of the time.
I should add that there might be circumstances in which I would want to give the
special treatment only to those who were most advantaged. For example, suppose that
my tutorial program was designed to be useful only for those who would quickly master
the basics of statistics and then be bored to death while I was repeating myself again
and again trying to get the rest of the class to understand. A special program that was
designed to capture and maintain those highly able students would best be applied only
to them -- but doing this might be considered “politically incorrect,” elitist.
You might now be thinking that I have lost my mind, deliberately producing
experimental groups that are nonequivalent on a covariate that is known to be highly
correlated with the criterion variable. How am I going to eliminate threats to internal
validity such as regression to the mean? As you will see, the answer to this question
comes from how the analysis is done, and it is not done by simply comparing the two
groups on the posttest.
Having chosen a treatment, a criterion variable, and a covariate (pretest), I now
must decide on the cutoff (on the covariate) for deciding who gets the treatment and
who does not. It might be reasonable to give the treatment to half of the students and
4

not to the other half, but I might not have enough resources to do so. For example,
staffing and equipment restrictions might allow me only to provide the treatment to the
students in the bottom 25% of the distribution of scores on the pretest.
To illustrate the use of the regression-discontinuity design, I have simulated data
from a hypothetical project employing that design. I started by simulating 38 pairs of
scores (pretest, posttest) randomly drawn from a population where
Post = 7 + 1.35∗Pre + error, ρ = .9, and σ error = 1.71 . These data are available at this
2

hot link: RegD0.txt. This data file is a plain text file. Each line contains data for one
subject. The first score is the letter ‘C’ or ‘T,’ indicating whether that subject was
assigned to the control group or the treatment group. Following a blank space (the
delimiter), the next score is the subject’s posttest score, and the next is the subjects’
pretest score. I assigned to the treatment group all subjects whose score on the pretest
was 6 or less. I defined the effect of the treatment to be exactly zero in the population.
If you remember how to use the statistical package you learned in your statistics
course, you can bring these data into that program and conduct a regression analysis
on them. Using all 38 data points, you would find that the estimated regression
parameters are: Post = 7.58 + 1.27∗Pre + error, r = .85, and MSE = 2.13. A plot of the
data with the regression line drawn in is shown in Figure 1 below. On this plot, data
points for subjects in the treatment group are plotted with the symbol ‘T’ and those for
subject in the control group with the symbol ‘C.’
If you were to conduct two regression analyses on these data, one for the
treatment group and one for the control group, you would find slightly different
regression lines, but those differences would be due totally to sampling error, because
the data for all subject in both groups were randomly sampled from exactly the same
population. I have estimated the separate regression lines and plotted them on the
scatter plot which is shown in Figure 2 below. For the treatment group the separate
regression estimates are Post = 8.09 + 1.17∗Pre + error, r = .62, and MSE = 2.13 and
for the control group Post = 6.33 + 1.43∗Pre + error, r = .72, and MSE = 2.29. Looking
at the plot, you should be able to see the difference between these two regression lines,
but they don’t look very different from one another.
5

Figure 1. Pooled Regression Line for Predicting Post from Pre


With No Treatment Effect.
6

Figure 2. Separate Regression Lines With No Treatment Effect.

Next I re-simulated the data, but with one change -- I built in a three point
treatment effect when defining the population for those in the treatment group. These
data are available at hot link RegD1.txt. I have estimated the separate regression lines
and plotted them on the scatter plot which is shown in Figure 3 below. For the
treatment group the separate regression estimates are Post = 11.27 + 1.07∗Pre + error,
r = .82, and MSE = 1.35 and for the control group Post = 7.90 + 1.18∗Pre + error, r = .
82, and MSE = 1.25. Looking at the plot, you can clearly see the difference in these two
regression lines. I extended the control group regression line into the treatment group
area to show what we would expect the regression line to look like for the treatment
group if there were no treatment effect.
7

Figure 3. Separate Regression Lines With a Three Point Treatment Effect.

The plot provides pretty convincing evidence that there is an effect of the
treatment. The regression line for the treatment group is clearly higher than what we
would expect it to be if the treatment were without effect. It would be hard to imagine
how one of the threats to internal validity that we have already discussed could have
created the observed discontinuity at exactly the cutoff point, but there is another threat
that must be considered. What if the true relationship between the criterion variable and
the pretest is curvilinear, but we have used a linear analysis? As illustrated in Chapter
11 of Trochim, this can lead one to conclude that there is a treatment effect when in fact
there is not.
8

Proxy-Pretest Design N O1 X O2
In this design one gathers the pretest information after the N O1 O2
experimental treatment has started. In other words, one finds an
archival proxy for the pretest. For example, suppose I ask the following question:
“Does completion of PSYC 2210 (experimental psychology) have an effect on a
student’s knowledge of statistics?” Ideally I would measure the students’ statistical
knowledge at the beginning of the semester, but suppose that the question did not
occur to me until the middle of the semester. I might decide to use as a proxy-pretest
students’ performance in their PSYC 2101 (statistics) class. My control group might
consist of a group of students taking some other class (not 2210). For each student I
would obtain a continuous measurement of the student’s performance in PSYC 2101
and, at the end of the semester, a continuous measurement of the student’s knowledge
of statistics. ANCOV would be used, with the proxy-pretest serving as a covariate.

Separate Pre-Post Samples Design


N O
In this design the sample of subjects that you use for the
pretest is different from the sample of subjects that you use for N X O
the posttest. There are several variations of this design. N O
Suppose that I want to evaluate the effect of a tutorial program
in my statistics class. My colleague Suzie Q and I each taught N O
statistics both this semester and the previous semester and we
each gave our students a standardized test of statistics achievement at the end of the
semester (as part of a departmental evaluation of the course). This semester I shall
make the tutorial program available to all of the students in my class, but my colleague
will not. Again, both teachers administer the statistics achievement test at the end of
the semester.
Look at the design notation in the box above. Note that there are four
nonequivalent groups. The first line represents my students last semester, when I did
not make the tutorial program available. The second line represents my students this
semester, when the tutorial program was made available. The third line represents my
colleagues’ students last semester, and the fourth line represents my colleague’s
students this semester. The potential for selection problems is clearly large with this
design.

Nonequivalent Groups Switching Replications Design


Recall our earlier (Chapter 7) discussion of
the randomized groups switching replications
N O X O O
design. The quasi-experimental version of this N O O X O
design differs in that the comparison groups are not
equated by randomization. For example, when evaluating the effect of my experimental
statistics tutorial, I could make it available to one class of students during only the first
half the semester and to the other class during only the second half of the semester.
This might reduce, somewhat, complaints about not getting the special treatment right
away, unless the one class learns what is going on in the other class.
9

Nonequivalent Dependent Variables Design O1 X O1


In this design you have only one group of subjects but
you have two or more dependent variables. In the most
N O2 O2
simple variation you have one dependent variable that you
expect to be affected by the experimental treatment and a second dependent variable
that you do not expect to be affected by the experimental treatment. The second
dependent variable serves as a control variable. You want your control variable to be
one that is similar enough to the primary variable that it should be affected in the same
way by history, maturation, and other threats to a single group pretest-posttest design,
but different enough from the primary variable that it will not also be affected by the
experimental treatment.
As you are aware, students in an experimental psychology class are expected to
have mastered basic statistics prior to enrolling in experimental psychology. However,
some students need some remedial instruction in statistics. Suppose that I decide to
make an online statistics remediation program available to my students in experimental
psychology. To avoid complaints about who gets the program and who does not, or
who gets it first, I decide to make the remediation program available to all of the
students. The primary dependent variable is performance on a test of statistics
knowledge, administered at the beginning and then again at the end of the semester. I
substitute a control variable for a control group. I decide to use a test of vocabulary as
the control variable, thinking that history and maturation should affect performance on
such a test in the same way that they would affect the primary dependent variable.
Conclusions drawn from research conducted with this design may be more
impressive if there are multiple control variables and an a priori prediction regarding to
what extent each of the control variables is likely to be affected by the treatment. For
example, suppose that my control variables include tests of logical thinking, verbal
reasoning, arithmetic skills, vocabulary, and artistic expression. I might be able to
predict, on theoretical grounds, that the treatment should affect the primary dependent
variable (statistics knowledge) more than any of the control variables, but among the
control variables logical thinking and verbal reasoning should be affected more than
arithmetic skills, arithmetic skills more than vocabulary, and vocabulary more than
artistic expression. If that pattern of results is obtained, my results would be more
impressive than they would be with just a single control variable.

Regression Point Displacement Design N(n = 1) O X O


In this design you have only one experimental unit N O O
(subject) in the treatment group, but several (or many) in
the control group. For example, suppose that you want to evaluate the effectiveness of
a novel economic development plan that is being tried in your home town. Every year
the state provides statistics on an assortment of economic variables, such as
unemployment rate, for every county and most cities (excluding small municipalities like
Ayden or Winterville). Included in this database are 20 cities that are comparable (in
size and ?) to your home town. For each city you obtain the unemployment rate for the
year before you start the novel economic development plan and the year two years after
you start that plan. You plot the posttest versus the pretest data, and draw in a
10

regression line for predicting posttest from pretest in the control group only, as in Figure
4 below. In the plot, I have plotted the posttest data in the vertical dimension and the
pretest data in the horizontal dimension. The data point for your home town is plotted
with an ‘X’ instead of an ‘O.’ Evidence of the effectiveness of the program is the
displacement of the ‘X’ point from where it would be expected to fall given the
regression line obtained for the control group and your home town’s pretest score.

Figure 4. Economic Development Plan Lowers Unemployment in Dingleville, NC.

Copyright 2003, Karl L. Wuensch - All rights reserved.

Вам также может понравиться