Experiments

Pre and Post condition

Classic experimental design

Random assignment to control and treatment

conditions
Why random assignment and control groups?

Classic experimental design

Random assignment helps with internal validity Some threats to internal validity: Experimenter/Subject expectation Mortality bias
Is there an attrition bias such that subjects later in the research

process are no longer representative of the larger initial group?

Selection bias
Without random assignment our treatment effects might be due to

age, gender etc. instead of treatments

Evaluation apprehension
Does the process of experimentation alter results that would occur

naturally?
Classic experimental design when done properly can help guard

against many threats to internal validity

Classic experimental design

Posttest only control group design:
Experimental Group Control Group

R R

X O1 O2

With random assignment, groups should be

largely equivalent such that we can assume the differences seen may be largely due to the treatment

Classic experimental design

Special problems involving control groups: Control awareness Is the control group aware it is a control group and is not receiving the experimental treatment? Compensatory equalization of treatments Experimenter compensating the control group's lack of the benefits of treatment by providing some other benefit for the control group Unintended treatments The Hawthorne effect (as it is understood though not actually shown by the original study) might be an example

Mixed design: prepost experiments

Back to our basic control/treatment setup A common use of mixed design includes a pre-test

post test situation in which the between groups factor includes a control and treatment condition Including a pretest allows:
A check on randomness Added statistical control Examination of within-subject change
2 ways to determine treatment effectiveness Overall treatment effect and in terms of change

Pre-test/Post-test

Random assignment
Observation for the two groups at time 1 Introduction of the treatment for the experimental group Observation of the two groups at time 2 Note change for the two groups

Mixed design
2x2

Between subjects factor of treatment

Within subjects factor of pre/post Example

treatment treatment treatment treatment treatment control control control control control

Pre 20 10 60 20 10 50 10 40 20 10

Post 70 50 90 60 50 20 10 30 50 10

SPSS output
Tests of Within-Subj ects Effects Measure: MEASURE_1 Source prepost prepost * treat Error(prepost) Type III Sum of Squares 1805.000 2205.000 1040.000 df 1 1 8 Mean Square 1805.000 2205.000 130.000 F 13.885 16.962 Sig. .006 .003 Partial Eta Squared .634 .680 Sphericity Assumed Sphericity Assumed Sphericity Assumed

Tests of Between-Subj ects Effects Measure: MEASURE_1 Transformed Variable: Average Source treat Error Type III Sum of Squares 1805.000 4240.000 df 1 8 Mean Square 1805.000 530.000 F 3.406 Sig. .102 Partial Eta Squared .299

Why are we not worried about sphericity here? No main effect for treatment (though close with

noticeable effect) Main effect for prepost (often not surprising) Interaction

Interaction
The interaction suggests
Estimated Marginal Means of MEASURE_1

Estimated Marginal Means

that those in the treatment are benefiting from it while those in the control are not improving due to the lack of the treatment

70

treat control treatment

60

50

40

30

20 Pre Post

factor1

Another approach: t-test

Note that if the interaction is the only thing of

interest, in this situation we could have provided those results with a simpler analysis Essentially the question regards the differences among treatment groups regarding the change from time 1 to time 2. t-test on the gain (difference) scores from pre to post

T-test vs. Mixed output

Tests of Within-Subj ects Effects Measure: MEASURE_1 Source prepost prepost * treat Error(prepost) Type III Sum of Squares 1805.000 2205.000 1040.000 df 1 1 8 Mean Square 1805.000 2205.000 130.000 F 13.885 16.962 Sig. .006 .003 Partial Eta Squared .634 .680 Sphericity Assumed Sphericity Assumed Sphericity Assumed

t2 = F

Independent Samples Test Levene's Test for Equality of Variances t-test for Equality of Means 95% Confidence Interval of the Difference Lower Upper -65.51672 -18.48328

F gain Equal variances assumed 2.246

Sig. .172

t -4.118

df 8

Sig. (2-tailed) .003

Mean Difference -42.00000

Std. Error Difference 10.19804

Another approach: ANCOVA

We could analyze this situation in yet another

way Analysis of covariance would provide a description of differences among treatment groups at post while controlling for individual differences at pre* Note how our research question now shifts to one in which our emphasis is in differences at time 2, rather than describing differences in the change from time1 to time 2

Special problems of before-after studies

Instrumentation change Variables are not measured in the same way in the before and after studies. A common way for this to occur is when the observer/raters, through experience, become more adept at measurement. History (intervening events) Events not part of the study intervene between the before and after studies and have an effect Maturation Invalid inferences may be made when the maturation of the subjects between the before and after studies has an effect (ex., the effect of experience), but maturation has not been included as an explicit variable in the study. Regression toward the mean If subjects are chosen because they are above or below the mean, one would expect they will be closer to the mean on re-measurement, regardless of the intervention. For instance, if subjects are sorted by skill and then administered a skill test, the high and low skill groups will probably be closer to the mean than expected. Test experience The before study impacts the after study in its own right, or multiple

Pre-test sensitization
So what if exposure to the pretest automatically

influences posttest results in terms of how well the treatment will have its effect? Example:
Attitudes about human rights violations after exposure to a

documentary on the plight of Tibet

Pretests: questions about attitudes human rights violations

Initial Awareness State

More empathic response to the film

Scores on posttest that might reflect a greater treatment effect

Solomon 4-group design

A different design can allow us to look at the

effects of a pretest

Solomon 4-group design

Including a pretest can sensitize participants and

create a threat to construct validity. Combining the two basic designs creates the Solomon 4group design, which can determine if pretest sensitization is a problem:

R R

O O

If these two groups are different, pretest sensitization is an issue. Pre X Treatment interaction If these two groups are different, there is a testing effect in general.

R O
R O

O
O

Solomon 4-group design

Why not used so much? Requires more groups However, it has been show that this does not mean more subjects necessarily Even if overall N maintained with switch to S4, may have more power than a posttest only situation Not too many interested in pretest sensitization Regardless one should control for it when possible, just like wed control for other unwanted effects
Complexity of design and interpretation
Although understandable, as usual this is not a good

reason for not doing a particular type of analysis

Lack of understanding of how to analyze
How do we analyze it?

Solomon 4-group design

We could analyze the data in different ways

For example: One-way ANOVA on the four post-

test results
Treat all four groups as part of a 4 level factor Contrast treatment groups vs. non

This would not however allow for us to get a

sense of change/gain

Alternative approach (Braver & Braver)

2 x 2 Factorial design

with control/treat, pre/not as two between subjects factors

Test A: Is there an

interaction?
Significant interaction

would suggest pretest effect

Effect of treatment changes

depending on whether there is pretest exposure or not

Simple effects
Test B & C: simple effects

B: Treatment vs Control at

Prepresent C: Treatment vs Control at Preabsent

In other words, do we find that

the treatment works but only if pretest?

O2 > O4, O5 = O6

If so, terminate analysis

The treatment effects are due

to pretest

Simple effects
However, could there be a treatment effect in

spite of the pretest effect?

In other words, could the pretest merely be provide

an enhancement of the treatment Ex. Kaplan/Princeton Review class helps in addition to the effect of having taken the GRE before
If the other simple effect test C is significant

also (still assuming sig interaction) we could conclude that was the case

Non-significant interaction
If there is no interaction to begin with, check the

main effect of treatment (test D)

If sig, then treatment effect w/o pretest effect

However this is not the most powerful course of

action, and if not sig may not be indicative of no treatment effect because we would be disregarding the pre data (less power)

Non-significant interaction: alternatives to testing treatment main effect

Better would be to use analysis of covariance

that takes into account differences among individuals at pretest (Test E) T-test on gain/difference scores (Test F) Or mixed design (Test G)
Between groups factor of Treatment Within groups factor of Pre-Post

As mentioned, F and the interaction in G are

identical to one another However test E will more likely have additional power

Ancova
We can interpret the ANCOVA as allowing for

a test of the treatment after posttest scores have been adjusted for the pretest scores Basically boils down to:
What difference at post would we see if the

participants had scored the same at pre?

We are partialling out the effects of pre to

determine the effect of the treatment on posttest scores

In SPSS
The ancova (or

other tests) will only concern groups one and two as they are the only ones w/ pre-tests to serve as a covariate or produce difference scores for the mixed design/t-test approach

 If the Ancova results (or test F or G) show the

treatment to still have an effect, we can conclude that the treatment has some utility beyond whatever effects the pre-test has on the post-test If that test is not significant however, we may perform yet another test

 Test H

t-test comparing groups

3 and 4 (O5 vs.O6) Less power compared to others (only half the data and no pre info) but if it is significant despite the lack of power we can assume some treatment effect

Meta-analysis
Even if this test is not significant, Braver & Braver

(1988) suggest a meta-analytic technique that combines the results of the previous two tests (test E, F or G and that of H)
Note how each is done only with a portion of the data
More power from a consideration of all the data

Take the observed p-value from each test, convert to

a one-tailed z-score, add the two z-scores and divide by 2 (i.e. the number of z-scores involved) to give
zmeta

If that shows significance* then we can conclude a

treatment effect
Nowadays might want to use effect size r or d for the meta-

analysis (see Hunter and Schmidt) as there are obvious issues in using p-values One might also just examine the Cohens d for each (without analysis) and draw a conclusion from that

Problems with the meta-analytic technique for Solomon 4 group design

Note that the meta-analytic approach may not

always be the more powerful test depending on the data situation

Sawilosky and Markman (1990) show a case where

the other tests are sig meta not

Also, by only doing the meta in the face of non

significance we are forcing an inclusion criterion for the meta (selection bias)

Problems
Braver and Braver acknowledge that the meta-

analytic technique should be conducted regardless of the outcomes of the previous tests
If test A & D nonsig, do all steps on the right side

However they note that the example Sawilosky used

had a slightly negative correlation b/t pre and post for one setup, and an almost negligible positive corr in the other, and only one mean was significantly different from the others
Probably not a likely scenario

Since their discussion the Braver and Braver

approach has been shown to be useful in the applied setting, but there still may be concerns regarding type I error rate Gist: be cautious in interpretation, but feel free to use if suspect pre-test effects

MCs summary/take
1. Do all the tests on the right side if test A

and D nonsig
If there is a treatment effect but not a pretest effect,

the meta-analysis is more powerful for moderate and large sample sizes
With small sample sizes the classical ANCOVA is slightly

more powerful

As the ANCOVA makes use of pretest scores, it is noticeably more powerful than the meta-analysis, whereas the t test is only slightly more powerful than the meta-analysis.

When a pretest either augments or diminishes the

effectiveness of the treatment, the ANCOVA or t test is typically more powerful than the meta-analysis.
2. Perhaps apply an FDR correction to the

analyses conducted on the right side to control for type I error rate 3. Focus on effect size to aid your

More things to think about in experimental design

The relationship of reliability and power

Treatment effect not the same for everyone

Some benefit more than others

Sounds like no big deal (or even obvious), but all

of these designs discussed assume equal effect of treatment for individuals

Reliability
What is reliability? Often thought of as consistency, but this is more of a

by-product of reliability
Not to mention that you could have perfectly consistent scores

lacking variability (i.e. constants) for which one could not obtain measures of reliability

Reliability really refers to a measures ability to

capture an individuals true score, to distinguish accurately one person from another on some measure

It is the correlation of scores on some measure with

their true scores regarding that construct

Classical True Score Theory

Each subjects score is true score + error of

measurement Obsvar = Truevar + Errorvar

Reliability = Truevar/ Obsvar = 1 Errorvar/

Obsvar

Reliability and power

Reliability = Truevar/ Obsvar = 1 Errorvar/ Obsvar

If observed variance goes up, power will decrease However if observed variance goes up, we dont know automatically

what happens to reliability Obsvar = Truevar + Errorvar

If it is error variance that is causing the increase in observed

variance, reliability will decrease*

Reliability goes down, Power goes down

If it is true variance that is causing the increase in observed

variance, reliability will increase Reliability goes up, Power goes down
The point is that psychometric properties of the variables play an

important, and not altogether obvious role in how we will interpret results, and not having a reliable measure is a recipe for disaster.

Error in Anova
Typical breakdown in a between groups

design SStot = SSb/t + SSe Variation due to treatment and random variation (error) The F statistic is a ratio of these variances F = MSb/MSe

Error in Anova
Classical True Score Theory Each subjects score = true score + error of measurement MSe can thus be further partitioned Variation due to true differences on scores between subjects and error of measurement (unreliability) MSe = MSer + MSes MSer regards measurement error MSes systematic differences between

individuals

MSes comes has two sources Individual differences Treatment differences

Subject by treatment interaction

Error in Anova
The reliability of the measure will determine

the extent to which the two sources of variability (MSer or MSes) contribute to the overall MSe If Reliability = 1.00, MSer = 0
Error term is a reflection only of systematic

individual differences
If Reliability = 0.00, MSes = 0 Error term is a reflection of measurement error only MSer = (1-Rel)MSe MSes = (Rel)MSe

 We can test to see if systematic variation is

significantly larger than variation due to error of measurement

MSes (Rel)( MSe ) (Rel) F MSer (1 Rel)( MSe ) (1 Rel)

df (n 1, n 1)

 With a reliable measure, the bulk of MSe will be

attributable to systematic individual differences However with strong main effects/interactions, we might see sig F for this test even though the contribution to model is not very much Calculate an effect size (eta-squared)
SSes/SStotal Lyons and Howard suggest (based on Cohens rules of thumb)

that < .33 would suggest that further investigation may not be necessary

How much of the variability seen in our data is due to

systematic variation outside of the main effects?

Subjects responding differently to the treatment

Summary
Gist: discerning the true nature of treatment effects,

e.g. for clinical outcomes, is not easy, and not accomplished just because one has done an experiment and seen a statistically significant effect Small though significant effects with not so reliable measures would not be reason to go with any particular treatment as most of the variance is due poor measures and subjects that do not respond similarly to that treatment

One reason to perhaps suspect individual differences due to

the treatment would be heterogeneity of variance For example, lots of variability in treatment group, not so much in control

Even with larger effects and reliable measures, a

noticeable amount of the unaccounted for variance may be due to subjects responding differently to the treatment Methods for dealing with the problem are outlined in Bryk and Raudenbush (hierarchical linear modeling), but one strategy may be to single out suspected

Resources
Zimmerman & Williams (1986)

Bryk & Raudenbush (1988)

Lyons & Howard (1991)