Status and Intergroup Perception

Seven Deadly Sins of
Dyadic Data Analysis
David A. Kenny
February 14, 2013
Many of the sins to be described were
common because we did not have the proper
dyadic tools.
Now that we have those tools, we need to
avoid the following practices.
Moreover, several of the practices are
acceptable IF proper empirical justification is
It Would Help to Know
Basic definitions in dyadic data analysis
It would also help to know about the actorpartner interdependence model (APIM) and
1. Assuming that dyad members are
distinguishable without testing whether
such an assumption is warranted
Most common and problematic case: Separate
analyses for men and women in heterosexual
Perform a test of distinguishability.
Even when that test is significant, it does not
mean that all parameters differ across the
distinguishing variable.
Remember: Parsimony is valued in science and
that estimating a single model results in greater
precision and power.
2. Assuming that a statistically significant effect
for one member of the dyad but not significant
for the other implies that the effect varies
across members
Statistical significance of each parameter
tests whether the parameter differs from
A path of .3 may be significant for men
and the path of .1 is not be significant for
women. However, it still may be that .3
does not differ significantly from .1.
We need to test the interaction with the
distinguishing variable (e.g., gender) to draw
this conclusion.
3. Summing or averaging the scores
for two members of the dyad without
any empirical justification
It is acceptable to sum or average the dyad members’
scores if they are highly correlated though sometimes it
might be better to use a common fate model.
To sum predictors, it must be the case that the actor
effect = partner effect:
Y1 = aX1 + pX2 + E1
If a = p then
Y1 = 2a(X1+X2)/2 + E1
It is acceptable to sum the outcome if the ALL the
predictors are between-dyads because a between-dyads
predictor can explain only between dyads variation.
4. Using discrepancy score as a
predictor without controlling for the
main effects
Discrepancy scores or lX1 – X2l are really
interactions and need to control for the main
effects before considering whether an interaction
(e.g., similarity) has an effect.
Also a difference score or X1 – X2 is confounded
with (or are a linear combination of) actor and
partner! You can use if a difference score as
predictor if actor and partner effects are of equal
and opposite signs (a + p = 0).
5. Assuming that one member causes
the other without realizing that the
other has influence
Sometimes called pseudounilaterality
Human Development: Parent to child
Social Psychology: Confederates
Health Research: Having only the caretaker
determine the outcomes of the patient
One needs to allow bidirectional influence either
using the actor-partner interdependence or the
mutual influence model.
6. Not including partner
characteristics as predictors
Researchers sometimes say that their theory
does not make predictions about partner
effects, and so they exclude them.
If dyad might be interdependent, analysts
should at least examine partner effects
before dropping them. If there are partner
effects, the actor effects may be biased!
If no partner effects, they can be dropped.
Less commonly, actor characteristics are
7. Treating individual as the unit
of analysis and ignoring dyad
Becoming less common but still occurs
Is permissible if the data are
independent on the outcome variable.
Still include partner characteristics
as possible predictors.
Although the focus is on poor data analysis,
I have tried to show what needs to be done
Many of these practices are acceptable IF
first preliminary analyses are conducted.
Readings: Chapter 15, pages 421-424 in Dyadic
Data Analysis by Kenny, Kashy, and Cook
Thanks to Deborah Kashy for help in preparing
these slides!

similar documents