Table 1.

Professor Ginsburg’s 3302 Review of Statistical Tests.

Knowing which analyses to use depends upon (1) the kind of numerical measurement
scale (2) number of samples (whether the research design has 2 or >2 treatment conditions)
and (3) whether samples are related (when all participants are measured under all
treatment conditions) or independent (different participants are each measured under only
one treatment condition).

STATISTICAL TESTS
Test Cause-Effect Test Relational
Inference for two or Inference for
> two samples two samples
Non-Interactive Types of Statistical Analyses
SAMPLE ▼ ▼
NUMBER → 1 Sample 2 Samples >2 samples

▼ ▼ ▼ ▼
/ \ / \
SAMPLE ▼ ▼ ▼ ▼
TYPE→ Independent Related Independent Related Independent

▼ ▼ ▼ ▼ ▼ ▼
MEASUREMENT
SCALE ▼ ▼ ▼ ▼ ▼
---------▼------------------------------------------------------------------------------------------------------------------
1 NOMINAL ↓ McNemar Chi square Cochran Q Chi square Contingency
▼ Coefficient
Binomial test –
(test for 50/50 odds)
-----------------------------------------------------------------------------------------------------------------------------
2 ORDINAL Wilcoxen Mann-Whitney Friedman Kruskal-Wallis
U

Spearman r

-----------------------------------------------------------------------------------------------------------------------------
3 INTERVAL/ t-test t-test One-way One-way Pearson r
4 RATIO repeated samples independent samples
(within-subjects) (between-subjects)
ANOVA ANOVA
Various Post-hoc Analyses (see SPSS ‘options’ )
 *Interactive Types of Analyses listed below all require interval/ratio numbers)

1. Factorial ANOVA test for Interaction Effects (2x2, 3x2. 4x3, ect.)
2. Covariant ANOVA (equates any pre-existing differences)
3. Linear Multiple Discriminant Regression Analysis or Path Analysis –

What Do Interactive Types of Analysis Show?

Why Are Interactive Analyses Typically Preferred?

I. Factorial ANOVA Designs and Interaction Effects –

A. What is an Interaction Effect? - An interaction occurs when the effect produced by one
manipulated, independent variable is different at each level of a second independent variable.

1. When may Interaction Effects occur? When there are two or > two treatments (independent
variables); When either a repeated (within-subjects) or independent (between subjects) samples
research design is used; and, when an interval/raio scale of measuring the dependent variable is
used.

2. This is called a FACTORIAL DESIGN.

a. Cells The factorial design and ANOVA statistical test are characterized by the number of levels
for each independent variable (treatment), described as ‘cells’. A factorial design with 2
treatments and 2 levels for each treatment is called a 2x2 factorial (2 x 2 = 4 cells). A design with 3
treatments having 4 levels for each would be a 3x4 factorial design (3x4 = 12 cells).

b. What are the advantages of performing research using factorial designs?

(1) INTERACTIONS - The investigator wants to know whether a significant interaction effect
occurs. Sometimes treatment A and treatment B will not have an effect when presented alone, but
will have an effect when combined with each other

(2) EFFICIENCY – it is more time and cost effective to examine several variables at once instead
of performing multiple studies on single independent variables.

(3) CONTROL – there is greater control over transient influences like room temperature,
different times of day when a unified factorial design research is conducted (3x4), rather than
conducting 12 separate studies.

(4) GENERALITY – in nature, 2 or more variables may often occur simultaneously, thus creating
a more natural effect than when single factors are studied sequentially.
c. How are Interaction Effects identified? The results of a factorial design are analyzed by
Analysis of Variance, ANOVA. ANOVA describes MAIN EFFECTS for each separate treatment
and INTERACTION EFFECTS when the treatments are combined with each other.

d. Examples: The Marilyn Monroe Effect - When taken alone, it takes relatively high amounts of
alcohol or barbiturates to cause sudden cardiac arrest. When combined with ‘sleeping pills’
much less alcohol is required to produce death.

e. How can interaction effects be readily identified in tables or figures? In tables, interaction
effects can be recognized when the mean recorded in one cell appears very different than the
means shown in the other cells. In figures, interaction effects are recognized when the lines
depicting dependent measures of the different levels of treatments appear to intersect. When the
lines appear by-and-large parallel, there are no interaction effects.

f. Post-hoc Analyses - Post hoc analyses (meaning after data has been collected) are performed
when the overall ANOVA shows some significant main effect(s) and/ or any interaction effect. The
investigator may perform appropriate post-hoc tests to localize the specific cells in the factorial
design producing the statistically significant main effects, and those combined cells responsible for
interaction effects. For example in a 3x3 ANOVA, 3 (levels of a drug, 1,2,3) x 3 (levels of
environmental stress, low, moderate, high), when the ANOVA shows significant and interaction
effects, which specific level of the drug produced main effects on the outcome measure? Which
specific level of stress? Which level of drug interacted with which level of stress to produce an
interaction? Sometimes these specific effects are apparent from examining a table or figure. Often
they are not readily apparent, which is when post-hoc analyses are very useful for understanding
what happened in a research investigation. The choice of post-hoc analysis depends upon the type
of primary design.

II. Analysis of Co-variance - Another common statistical technique is called the analysis of co-
variance. COANOVA should be used to equate different treatment conditions when some factor
makes the two or more groups unequal before any treatment was introduced. You can only use
COANOVA when a measurement is made (often a pre-test) has been administered and you know
that the participants in the different conditions had preexisting differences in some salient
measured variable. You may want to covary to hold a variable constant and equate that variable
for all the participants. Suppose that you were studying how a particular therapy improved
stroke patients’ recovery of motor skills. You create groups assigned to control or therapy a or
Therapy B conditions, but in your pretest measures you discover that participants in the control
group already had a 5-point IQ score advantage compared to the ones in therapy groups. You feel
that a person’s pre-IQ may influence the outcome. So, after data is collected, your initial ANOVA
may show that there were no differences between the therapy & control group. A Covariate
ANOVA will factor the pre-IQ difference to allow ‘fair and equal’ comparisons when the groups
are initially unequal in some relevant dimension. With IQ covaried, there may be a significant
effect of one or both of the therapies with a COANOVA.

III. Linear Multiple Dicriminant ANOVA and other ‘Path Analyses’ may be used with advanced
research designs with multiple ‘predictor variables’*. You may wish to determine which cluster of
factors has the greatest predictive values – for example, suppose you have identified 4 clusters of
measured factors that may predict a childhood injury (e.g., style of parenting, child’s perceptual-
motor skills, child’s temperament, home safety index, etc). With a path analysis, one can
determine how predictor variables may interact and which predictor variable carries the greatest
weight (e.g., home safety accounts for the greatest variance of the dependent variable (measured
injury rates, followed by child’s temperament, motor coordination, & parenting styles).

IV. Eta 2 - is a post-hoc measure showing how much total variance between conditions accrues
from the treatment, apart from any random ‘error variance’. Eta 2 is a another post-hoc test
(meaning after initial analysis is performed) used to determine the extent to which the
independent (or predictor) variables studied account for the amount of variance between the
treatment (or predictor) conditions. A high * Eta 2 score means that most of the effects were due
to the actual variables used to predict the outcome differences. Low * Eta 2 means that the
variables didn’t account for much of the total variability between the treatment conditions. A
researcher could obtain a statistically significant ANOVA effect, but a low Eta 2 means that the
treatment (or predictor) variables themselves did not account for very much of the variation of
outcome scores. Low Eta 2 means that some other factors that were not considered and measured
accounted for more of the between condition differences (variance) than the actual treatment or
predictor variables. High Eta 2 means that the specific variables under study did account for most
of the variance in the study.