Você está na página 1de 27

Using BetweenSubjects &

Within-Subjects
Experimental Design

Think about this


Imagine an experiment testing different types
of treatment on depression.
Drug, Therapy, Combination of drug & therapy,
Control (i.e., no treatment is given)

12 Participants with depression are randomly


assigned to a condition, are treated for 12
weeks, and then we test their levels of
depression.
Here are their scores:

*Higher scores indicate higher levels of depression*


Why do scores within a condition differ?
People are different
Measurement error

Both of these causes of differences are random.


They are due to chance.

Why do scores between conditions differ?


People are different
Measurement error
Both of these causes of differences are random.

AND, differences between the treatments/


treatment effects

Error variance ___________________


__________________________________
_________________________________
To reduce error variance:
Hold extraneous variables constant
Increase the effectiveness of IV (stronger
manipulation)
Random assignment to randomize error
across groups

Experimental Design
Between-subjects design ___________
___________________________________
________________________________

AKA: independent-samples
Can have 2 or more groups
The example at the beginning of this lecture is an
example of a between-subjects design. Different
participants were in each of the four groups.

Example: Which is better


Coke or Pepsi?

Two independent
groups provide
ratings.
All members of
each group rate
one of the
beverages.
Ratings of Coke
vs. Pepsi are
compared.

Experimental Design
Within-subjects design _____________
__________________________________
________________________________
_

AKA: repeated-measures
Can have 2 or more levels

Example:
Coke or

Which is better
Pepsi?

Each person rates both.


Ratings of Coke vs. Pepsi
are compared.

Comparison of Between
vs. Within
Between-subjects designs are more
conservative
there will not be any contamination between groups
For example, if you are doing a Coke vs. Pepsi taste-test
and your participants drink the Coke first and then the
Pepsi (in a within-subjects design) the results of the Pepsi
rating could be contaminated by a Coke aftertaste.

If groups are different from the beginning, could


confound results
There could be something that is different about
your two (or three, or four, etc.) groups that could
have an effect on your results (other than the IV
manipulation)

Comparison of Between
vs. Within
You need fewer participants with a withinsubjects design
You recycle participants through the levels
of your IV, so you do not need as many

Within-subjects designs are statistically


more powerful
The variance associated with having
different people in the different levels of the
IV is eliminated

Assigning Participants
to Groups

We want our groups to start out equal


We want the same average intelligence, same
number of men and women in each group, same
levels of education, etc.

Assigning Participants
to Groups
How do we maintain equivalency of groups?
There are a few different ways:
1. Randomization ignoring the characteristics of
the participants when assigning them to
conditions
Randomly assigning to groups

2. Matching participants are matched on some


variable assumed to be correlated with the
dependent variable
Then randomly assigned to conditions

Example of Matching:

Example:
Here we might
choose age
as the variable
to match our
participants.
We match the
participants on
age, then we
would
randomly
assign each
member of the
pair to one of
the groups.

Pair 5
Pair 3
Pair 2

Pair 1

Pair 4

A more realistic
example
10 people matched on
GPA
P1: 3.24
P2: 3.91
P3: 2.71
P4: 2.05
P5: 2.62
P6: 2.45
P7: 3.85
P8: 3.12
P9: 2.91
P10: 2.21

Put GPAs in order


P4: 2.05
Pair 1
P10: 2.21
Pair 2
P6: 2.45
P5: 2.62
P3: 2.71
Pair 3
P9: 2.91
P8: 3.12
Pair 4
P1: 3.24
P7: 3.85
Pair 5
P2: 3.91
Randomly assign from each pair

This helps us be more certain that our two groups will end up with equivalent
average GPAs.

To do matching we will:
Match pairs
Randomly assign each member of the pair
to one of the groups
However,
Matching has problems
You cannot match for everything!

Participant Loss
Mechanical subject loss occurs when a
participant fails to complete the
experiment because of equipment failure
or experimenter error
Not too big of a concern

Participant Loss

Selective subject loss participants are lost:


1. differentially across the conditions of the
experiment (more in one condition than the others)
2. A characteristic of the participant is responsible for
the loss
3. The characteristic is related to the DV

Destroys equivalency - subject attrition in


one group makes groups unequal.

Imagine that you are interested in determining what


type of exercise program is most effective for helping
obese individuals lose weight: a slow & steady
approach or a quick & intense approach. One group is
assigned to a condition where they are supposed to go
for a walk for 45 minutes, 6 days a week and another
group is assigned to a condition where they must
engage in heavy weightlifting and intensive cardio
exercises for 45 minutes, 3 days a week. Because of
the intensive nature of the second exercise program,
the individuals may be less likely to want to pursue the
experiment than those individuals who only have to go
for a walk each day. Therefore, more of those
individuals might drop out of the experiment, leading to
selective subject loss.

Controlling for selective subject loss


Pretest and screen out people who will potentially
quit
This can be quite costly
In our example, you could try to screen out people who have
dropped out of exercise programs in the past.

Pretest participants, randomly assign to one


condition
If a person from the experimental group quits, then drop a
comparably-scored person from the control group
In our example, you could pretest the participants for weight
at the beginning. Then, if a person drops out of the intensive
exercise group, a participant with a comparable weight in the
other group is dropped. (*The dropped person would
continue with the experiment, but their data would not be
included in the analyses.)

Comparing Between vs.


Within Again
Within-subjects designs:
Are more efficient fewer participants are
needed
Differences caused by manipulation are not
due to different participants
Risks:
Carryover effects ____________________
_____________________________________
___________________________________

Avoiding Carryover
Effects
Randomization of order of procedures
The order of the conditions are randomly
determined for each participant

Avoiding Carryover
Effects
Counterbalancing any technique used
to vary systematically the order of
conditions in an experiment to distribute
the effects of time of testing so they are
not confounded with conditions
We counter the effects of potential
confounding variables by balancing them
over the periods when treatments are
administered

Think about this


We are conducting a within-subjects design where the
participants are going to learn two lists of words and
have two memory tests. While learning one of the lists
of words, the participants are asked to think about how
many vowels are in each of the words (this is the low
meaningfulness condition). While learning the other list
of words, the participants are asked to think about how
that word relates to their life (the high meaningfulness
condition). Each participant will listen to a list of words,
take the memory test (recall), listen to the second list of
words, take another memory test (recall). In order to
eliminate carryover effects, we must counterbalance the
order of conditions: (see next slide)

COUNTERBALANCING
R=
randomly
assigned to
Order 1 or 2

Half of our participants will do the low condition first, half will do the
high condition first. This way, if there is a carryover effect (e.g., being in
the high condition first makes them think the same way in the low
condition) the effect will be minimized.

Complete counterbalancing
All possible treatment orders are used
2 treatments: A B; B A
3 treatments: A B C; A C B; B C
A; B A C; C A B; C B A
Four would have 24
Five would have 120
Quickly becomes impractical

Incomplete counterbalancing
Latin-square design form of
counterbalancing where each condition
occurs equally often during each time
period of the experiment

Você também pode gostar