Lecture 19: Program Evaluation and Quasi-Experiments

1
Lecture 19 Program Evaluation and
Quasi-Experiments
2
Correlation and Causality

• Fact 1 Correlation does not imply causality.
• Fact 2 Causality implies correlation.
• A correlational study can provide evidence
  AGAINST a causal hypothesis. The problem is that
  a correlational study (by itself) does not give
  us much confidence when making affirmative causal
  inferences.

3
Program Evaluation
4
What is Program Evaluation?

• The use of social science methods to
  systematically investigate the effectiveness of
  social intervention programs.
• Focus on summative evaluation. Formative
  evaluation is where the goal is to help improve
  existing programs rather than making
  effectiveness judgments.
• Senator D. P. Moynihan (1927-2003) If there is
  any empirical law that is emerging from the past
  decade of widespread evaluation activity, it is
  that the expected value for any measured effect
  of a social program is zero.

5
Reforms as Experiments (Campbell, 1969)

• Problem Specific programs are often
  advocated as though they were certain to be
  successful.
• Campbells Solution Hard-Headed Evaluation of
  Effects. Focus on the Importance of the Problem.
  Demand Empirical Evidence that Programs Work.
• Political Stance This is a serious problem. We
  propose to initiate Policy A on a trial basis.
  If after five years there has been no significant
  improvement, we will shift to Policy B (p. 410)

6
Surefire Paths to Success (p. 428)

• What to do if you want to see that your program
  works? Some ideasrely on testimonials and
  capitalize on regression artifacts
• Human courtesy and gratitude being what it is,
  the most dependable means of assuring a favorable
  evaluation is to use voluntary testimonials for
  those who have had the treatment (p. 426)

7
Regression Toward the Mean

• Extreme Scores at one time are not likely to be
  as extreme on a second testing.
• Why? Two sets of scores are never perfectly
  correlated.
• Take those 25 people who scored 65 or worse on
  Exam 1. What was their average gain from Exam 1
  to Exam 2? 4.10 points. What about those 13
  people who scored 95 or better? What was their
  average difference? A loss of 2.34 points.

8
Regression to the Mean - 2

• This true story illustrates a saddening aspect
  of the human condition. We normally reinforce
  others when their behavior is good and punish
  them when their behavior is bad. By regression
  alone, therefore, they are most likely to improve
  after being punished and most likely to
  deteriorate after being rewarded. Consequently,
  we are exposed to a lifetime schedule in which we
  are most often rewarded for punishing others, and
  punished for rewarding. (Kahneman Tversky,
  1973, p. 251).

9
How could you use this phenomenon to create a
positive impression of a program?
10
Death, Taxes, and Regression to the Mean

• Regression to the mean is as inevitable as death
  and taxes. Academic performance, emotional
  well-being, medical diagnosis, investment return,
  athletic feats, motion picture sales, and any
  other variable you can think of all exhibit
  regression toward the mean. (Reichardt, 1999, p.
  ix)

11
Applied Research
12
Example

• Palmgreen, P., Donohew, L., Lorch, E. P., Hoyle,
  R. H., Stephenson, M. T. (2001). Television
  campaigns and adolescent marijuana use Tests of
  sensation seeking targeting. American Journal of
  Public Health, 91, 292-296.

13
Palmgreen et al. (2001)

• Objective To evaluate the effectiveness of one
  television media campaign designed to reduce
  marijuana use among an at-risk group, high
  sensation seekers.
• Sensation seeking is a trait associated with the
  need for novel and intense stimuli and the
  willingness to take risks to obtain such stimuli.
• Design 32-month interrupted time series design.
• Logic Before and After Comparisons. Cross-site
  comparisons.

14
Data Collection Basics

• Duration Beginning 8 months before the first
  campaign and lasting 8 months after the last
  campaign.
• Sites Two Counties in Southern States.
• Each month draw a random sample of 100 public
  school students in each county
• Measures Sensation Seeking and 30-day use of
  ATOD.

15
(No Transcript)
16
Interrupted Time Series Design

• Extension of the pretest-posttest design
• A stronger argument can be made to eliminate
  maturation, testing, and history effects
• Can also be used with multiple groups with and
  without the treatment

Group 1
O2
O3
O1
O5
O6
O4
X1
17
Replicated Interrupted Time Series Design 2 (p.
323)
Group 1
O2
O3
O1
O5
O6
O4
X1
Group 2
O8
O9
O7
X2
O13
O11
O10

• Both groups are exposed to treatment but at
  different times.
• Quite strong from the point of view of internal
  validity
• Why?

18
What makes this a quasi-experimental design?

• One or more independent variables are being
  manipulated but participants are not randomly
  assigned to conditions.
• What is the textbook threat in this case?
• How plausible is this threat?
• Quasi-experiments, however, rely heavily on
  researcher judgments about assumptions,
  especially on the fuzzy but indispensable concept
  of plausibility. (Shadish et al., 2002, p. 484)

19
Quasi-Experimental Designs
20
Notation

• X IV, a treatment, or putative (supposed)
  cause
• O DV, an observation, or putative effect
• Notation used in Campbell and Stanley (1966)
• Recall Threats Selection, Maturation, History,
  Instrumentation, Attrition

21
One-Shot Case Study

• Single Group Studied Once
• As been pointed out (e.g., Boring, 1954
  Stouffer, 1949) such studies have such a total
  absence of control as to be of almost no
  scientific value (Campbell Stanley, 1966, p.
  6). Basic to scientific evidence is the process
  of comparison. There is no point of comparison
  here. Misplaced precision.

X
O1
22
Static-Group Comparisons

• The dashed line indicates a lack of random
  assignment
• Selection is a serious threat to internal
  validity
• Temporal precedence is often hard to establish

Group 1
X
O1
Group 1
X1
O1
OR
Group 2
X2
O2
Group 2
Not X
O2
Group 3
X3
O3
23
Pretest-Posttest Nonequivalent control group
design

• Can help evaluate the extent to which selection
  is a threat to validity
• Temporal precedence is clear

Group 1
X1
O2
O1
Group 1
X1
O2
O1
OR
Group 2
X2
Group 2
O4
O3
O4
O3
Group 3
X3
O6
O5