Professional Documents
Culture Documents
Research Example 26 - Regression and Multiple Regression: Variable
Research Example 26 - Regression and Multiple Regression: Variable
themselves predict college success, as do SAT scores. Together, however, they predict better
than either one by itself (Sprinthall, 2000). As students, you might find the following Research
Example interesting. It is a good example of both simple, linear regression and multiple regres
sion and it might give you some insight into how the passion and motivation you may feel for
studying psychology can predict how engaged or overwhelmed you feel.
You might be wondering if motivation played some sort of role in students’ passion for study
ing. Maybe students who are more motivated also felt more passionate about studying. Well,
Stoeber et al., (2011) also controlled for motivation using multiple regression. Thus, passion and
motivation were predictor variables and engagement and burnout were criterion variables. Like
the linear regression analyses, they had to run separate multiple regression analyses – one for
each criterion variable.8 To measure motivation, students wrote down two personal goals they
wished to achieve when studying. They then rated on a 7‐point Likert scale their level of
agreement of the reason for each goal. For example, students rated to how much they agreed
that their goal was intrinsically motivated or extrinsically motivated, which was then coded by
Stoeber et al. as autonomous or controlled motivation, respectively. The results from the
multiple regression analyses were virtually identical to the linear regression analyses, meaning
that neither type of motivation factored into the relationships between passion, achievement,
and burnout demon
strated by the first set of regression analyses.
Stoeber et al.’s (2011) results were the first that demonstrated the relationships between pas
sion, engagement, and burnout in a college setting, and their work serves as an extension of
applied research in the field of industrial‐organizational psychology on employee burnout. The
external validity of the study should be considered, as the authors caution that their results
should not be generalized to all college students, in large part because they tested a lot more
women (n = 93) than men (n = 12). Such research certainly opens the door for more “what’s
next” type of thinking, including asking empirical questions about gender differences in passion
and moti vation, and whether passion and engagement are related to academic success. For you
as stu dents, harmonious passion is probably a healthier form of passion than obsessive passion.
Obsessive passion relates to more pressures to perform and burnout, which may sour you on
studying psychology. Harmonious passion for psychology can be better for you in terms of feel
ing more involved and in control your love of psychology.
One final point about a regression analysis is both procedural and ethical. In general, predic
tions should be made only for people who fall within the range of scores on which the correla
tion is based. For example, if a regression equation predicting college success is based on a
study using middle‐class suburban Caucasians with SAT scores ranging from 1000 to 1400,
then the equation should not be used to predict success for any future applicant not part of that
population.
SELF TEST
9.3
1. Consider the relationship between depression and exercise. Do you think it is a positive
or a negative correlation? Explain.
2. Suppose you wish to predict academic success in college by looking at high school
grades. Which is the criterion variable and which is the predictor variable?
3. How does simple, linear regression differ from multiple regression?
8
Stoeber et al. (2011) actually ran what is called a hierarchical linear regression model, which is a form of multiple regression
which allows the researcher to enter multiple predictors in a step-wise fashion, rather than adding all predictors into the model at
once. Technically, the hierarchical regression allows the research to statistically control for one predictor before another
predictor is entered into the model. In Stoeber et al.’s case, they controlled for the unique contributions of motivation before
testing the passion predictor.
282 NON‐EXPERIMENTAL DESIGN I: SURVEY METHODS
Directionality
If there is a correlation between two variables, A and B, it is possible that A is causing B to
occur (A → B), but it also could be that B is causing A to occur (B → A). That the causal
relation could occur in either direction is known as the directionality problem. The existence
of the correlation by itself does not allow one to decide about the direction of causality. For
instance, consider a study described in the New York Times in 2008, in which researchers
examined the research pro
ductivity and beer consumption of ornithologists in the Czech Republic. They found a negative
correlation: The more beer consumed by ornithologists, the less productive they were as scien
tists, a finding not likely to be taken well by those scientists who often claim that their best ideas
occur to them in bars or pubs. The Times article emphasized the interpretation that probably
occurred to you, that spending too much time drinking beer might cause the scientists to have
little time left for research. But one researcher, thinking in terms of directionality, suggested that
perhaps “those with poor publication records are drowning their sorrows” (For Scientists, 2008).
So it is conceivable that drinking lots of beer causes Czech ornithologists to fail in their publish
ing efforts (A → B), but it is also possible that failing to publish causes Czech ornithologists to
drink more beer (B → A). It is also worth noting the article also illustrated external validity, a
concept you learned about in Chapter 5. One non‐ornithologist critic of the study suggested that
perhaps the results were limited to scientists who studied birds. Another suggested the results
were limited individuals from the Czech Republic, which, the article claimed, has the highest
rate of beer consumption on Earth.
Research on the relationship between TV watching and children’s aggression typifies the
directionality problem. Some of these studies are correlational and take the following general
form. Some measure (variable A) of TV watching is made, the number of hours per week per
haps. For the same children, a second measure (variable B) of aggressive behavior is taken. It
might be a combined score of teacher ratings of the aggressiveness of those in the study.
Suppose this study yields a correlation of +.58, which is found to be significantly greater than
zero. What can be concluded?
One possibility, of course, is that watching large amounts of TV inevitably exposes the child
to a great deal of violence, and we know children learn by observation; thus, it would follow that
a large dose of TV watching causes children to become aggressive—that is, A → B. Even if a
regression analysis is conducted and watching TV (A) is a significant predictor of aggression
(B), this still doesn’t mean that watching TV causes aggression. Why? Because the regression
itself is based on a correlation, and the research design is non‐experimental. But could causality
be work
ing in the reverse direction? Could it be that children who are already aggressive for some other
Interpreting Correlational Results 283
reason simply like to watch more TV than their nonaggressive peers? Knowing that much of
television involves violent programming, perhaps aggressive children choose to watch more of
the things that really interest them. In short, perhaps being aggressive causes children to watch
more TV—that is, B → A.
Choosing the correct causal direction is not possible based on an existing correlation.
However, the directionality problem can be addressed to some extent. The approach derives
from the crite ria for determining causality first described in Chapter 1. As you recall, research
psychologists are generally satisfied with attributing causality between A and B when they
occur together with some regularity, when A precedes B in time, when A causing B makes
sense in relation to some theory, and when other explanations for their co‐occurrence can be
ruled out.
For the TV and aggressiveness study, all we have is A and B occurring together and the fact
that A causing B makes some sense from what is known about observational learning theory
(remember Bandura’s Bobo doll study from Chapter 5—Box 5.2?). However, using a procedure
called a cross‐lagged panel correlation, it is possible to increase one’s confidence about direc
tionality. In essence, this procedure investigates correlations between variables at several points
in time. Hence, it is a type of longitudinal design, adding the causal element of A preceding B.
The following Research Example illustrates the procedure.
FIGURE 9.5
.21 −.05 .31 Results of a cross‐lagged panel study of
9
Actually, the study began with another purpose: to see if parental child rearing practices led to aggressiveness in children. But
in a questionnaire for parents, Eron et al. included “filler” questions, items designed to disguise the study’s true purpose. One
asked parents to list their child’s three favorite TV programs. When looking at the data, Eron et al. were surprised to see a
relation ship emerge: Children who preferred violent programs tended to be more aggressive. The entire focus of the study then
changed (Huesman & Dubow, 2008). You might recognize this as an example of how serendipity (Chapter 3) can alter the
direction of research.
284 NON‐EXPERIMENTAL DESIGN I: SURVEY METHODS
Of special interest are the diagonal or cross‐lagged correlations because they measure the
relationships between two main variables but separated in time. If third‐grade aggressiveness
caused a later preference for watching violent TV (B → A), then we would expect a fair‐sized
correlation between aggressiveness at time 1 and preference at time 2; in fact, the correlation is
virtually zero (+.01). On the other hand, if an early preference for viewing violent TV programs
led to a later pattern of aggressiveness (A → B), then the correlation between preference at time
1 and aggressiveness at time 2 should be substantial. As you can see, this correlation is + .31,
not terribly large but significant. Based on this finding, as well as on other indications in their
study, Eron and his colleagues concluded that an early preference for watching violent TV is at
least partially the cause of later aggressiveness.
Cross‐lagged panel correlations must be interpreted cautiously, however (Rogosa, 1980). For
one thing, if you examine the overall pattern of correlations in Figure 9.8, you will notice the cor
relation of +.31 may be partially accounted for by the correlations of +.21 and +.38—that is,
rather than a direct path leading from 3rd‐grade preference to 13th‐grade aggression, perhaps
the path is an indirect result of the relationship between preference for violent TV and
aggression in the 3rd grade and between the two measures of aggression. A child scoring high
on preference for violent TV in 3rd grade might also be aggressive in 3rd grade and still be
aggressive (or even more so) in 13th grade. Alternatively, it could be that aggressiveness in third
grade produced both (a) a prefer ence for watching violent TV in third grade and (b) later
aggressiveness. Thus, cross‐lagged panel correlations help with the directionality dilemma, but
problems of interpretation remain. More generally, interpretation difficulties take the form of
the third variable problem.
Third Variables
The June 4, 2000, issue of the New York Times Magazine, contained a playful article entitled
“Greens Peace” (Plotz, 2000). In it the author addressed the weighty issue of why some
countries seem to be always at war, while others remain relatively peaceful. His answer was
golf: Countries where a substantial portion of the population plays golf are less belligerent than
countries without golf. As Plotz put it:
Every peaceful European nation loves golf. But Russia, at war in Chechnya, doesn’t hit
the links. Non‐golf Greece and non‐golf Turkey have long warred over non‐golf Cyprus.
The former Yugoslavia has fragmented into five states. Only peaceful Slovenia swings
the sticks. Do India or Pakistan golf? Of course not. Algerians shoot one another;
Moroccans next door shoot par. (p. 32)
Although the slogan “make par, not war” (p. 37) might have merit, I think you can see the
absurdity of the argument that golf causes peace. And it is only slightly more likely that the
reverse is true—that peace causes golf. Rather, if there really is a correlation between peace and
golf on a national level, and Plotz doesn’t present a Pearson’s r, of course, its existence is an
exaggerated example of what researchers call the third variable problem. Because
correlational research may not attempt to control extraneous variables directly, these variables
often provide an explanation for the correlation found—that is, rather than A causing B or B
causing A, an unknown third variable, C, might be causing both A and B to happen. C is an
uncontrolled third variable (or variables—it is often the case that more than one uncontrolled
variable lies behind a correlation). Can you think of third variables that could produce the
alleged golf‐peace correla tion? Economic prosperity perhaps? Highly prosperous countries
might be more likely to be peaceful and also have more time for leisure, including golf.
The relationship between watching violent TV programming and children’s aggressiveness
provides a clear example of the third variable problem. As we’ve already seen, it is possible that