Вы находитесь на странице: 1из 17

Student’s Companion Website for Assessment for Effective Teaching

The Regression Effect


Have you ever wondered why r is the symbol for the Pearson product-
moment coefficient of correlation? (Probably not, but let's consider it
anyway.) Why not c for correlation or P in honor of Karl Pearson?
Relationship, you say? Good guess, but it's wrong. Actually, r stands
for regression.

The regression effect is important and widely applicable. Yet it is


not well understood either among members of the public or among
educators. A command of this topic is extremely helpful in under-
standing our world and in being able to better predict the future.
Without knowledge of regression, people tend to err systematically in
certain kinds of predictions; this has great practical importance in a
wide range of fields, including educational assessment.

A Little History

Over a century ago, Sir Francis Galton noticed that very tall men didn’t
ordinarily have sons as tall as themselves. The sons of these tall
fathers were nearer the population mean in height. To be sure, most
of the sons were taller than average, but they were not typically as tall
as their fathers. Compared with their fathers, their stature had
regressed toward the mean.

Galton's contribution grew out of the work of his methodical cousin,


Charles Darwin. Recall that Darwin's seminal 1859 On the Origin of
Species posited that (a) species overproduce, (b) natural variation
occurs among the young, (c) the more fit (in their particular environ-
ment) tend to survive to reproduce, and (d) the characteristics of
those who reproduce tend to be transmitted to the next generation.

Galton focused much of his illustrious career on the topic that


biologists call natural variation; educators and psychologists more
often refer to it as individual differences. Sir Francis measured
individual differences in a wide variety of human traits. Having
measured variation among individuals, he then needed to quantify the
findings. This led him to mentor the young English statistician Karl
Pearson, who developed some very widely used descriptive statistics.
The scholarly measurement, study, and analysis of individual
differences constituted the great English contribution to the American
mental measurement movement.
Student’s Companion Website for Assessment for Effective Teaching

In the course of measuring most everything he could measure, Sir


Francis discovered regression in human stature. This led to his 1885
paper on regression toward the mean in human stature.1

The following examples illustrate how regression operates, how


erroneous conclusions are drawn when the phenomenon is not taken
into account, how ill-advised public policy decisions can result from
such faulty research, and how we all can better understand, predict,
and adjust to the world around us in light of Galton's insight.

Examples and Explanations of the Regression Effect

Father-Son Stature in Tall Sample2

Suppose a sample of tall men is selected on the basis of deviance


from the population mean and the following mean heights are found for
them and their adult sons.
• Mean for each generation's population = 178 cm.3
• Mean for the deviant sample of fathers = 190 cm.
• Mean for the adult sons of these fathers = 184 cm.

The mean deviation score (that is, the mean of the x scores) for the
fathers is +12 cm (190 cm - 178 cm). Similarly, the mean deviation
score for the sons is +6 cm (184 cm - 178 cm).
Half of the fathers' deviance was manifested in the sons. That is, half
of the fathers' mean deviation of +12 cm was maintained in the sons'
mean deviation of +6 cm. This .5 represents the regression
coefficient or Pearson r.

1
The title of Galton's paper was, "Regression Toward Mediocracy in Hereditary Stature."
His wording seems unfortunate. First, the word "mediocracy" is disparaging in tone; the
synonymous phrase, "the mean" serves better. Second, an ardent hereditarian, Sir
Francis unnecessarily went out on a speculative limb in using the word "hereditary."
Actually, no assumption concerning nature-nurture contribution or interaction is
necessary for considering and understanding the regression effect.

2
For human height examples, gender of parent and child must be specified because of sex
differences. We could, of course, just as well have used examples of father-daughter,
mother-son, or mother-daughter height. For subsequent examples, such as parent-
child IQ, gender need not be specified.

3
This assumes that the population mean and standard deviation remain constant across
the two generations. If the population were getting taller or if its variability were
changing, then we would have to adjust for this in our calculations. But to keep the
examples simple, let's assume that height, IQ, and other traits do not change from one
generation to the next.
Student’s Companion Website for Assessment for Effective Teaching

This provides us with a conceptualization of the meaning of the


Pearson r. It can be described as the nonregression of one variable to
the other. This idea led to the stastistic being named the “regression
coefficient.”

An alternative conceptualization was offered in the preceding


Extension where we used the formula for r that defined it as the mean
of products of corresponding z-scores.

Father-Son Height in Short Sample

Now assume we select a sample of very short men on the basis


of shortness. Suppose the mean height of our deviant sample is 160
cm and the mean height of the population is 178 cm. In this case let's
assume we are concerned with predicting the ultimate mean height of
their sons.

If we scurried off to the library to research the topic of father-son


correlation in stature, we would find that of the many studies done
since Galton's time, the typical finding is that the r is about .5. W e
can now use this figure to predict the sons' stature.

The fathers’ mean deviation score is -18 cm (160 cm - 178 cm). We


multiply this mean deviation score of the predictor variable (fathers'
mean height) by the regression coefficient of .5 to obtain the best
prediction for the criterion variable (sons' mean stature). The product
of .5 and –18 is –9. We therefore predict a mean stature of 169 cm
(178 cm - 9 cm) for the sons.

Father-Son Stature for Individuals

The same method is used in making predictions for individuals.


Suppose a particular man is 170 cm tall, and there is reason to predict
the adult stature of his future son (with no knowledge of the height of
the son’s mother). To do so, the father's deviation score of -8 cm
(170 cm - 178 cm) is multiplied by the .5 regression coefficient known
from previous research to obtain the most likely deviation score of the
son (-4 cm). The son's adult stature is thus predicted to be 4 cm
shorter than population mean—174 cm.

The method is the same for individual or group prediction. In either


case, we don’t, of course, expect a prediction to be exactly correct. In
the case of a prediction for the mean of a group of several hundred
father-son pairs, we expect the prediction to come within a
Student’s Companion Website for Assessment for Effective Teaching

centimeter or less of the mean height of the sons. In the case of a


prediction for an individual, we could miss by a meter! But in both
cases, the best prediction is made by multiplying the deviation score of
the predictor by the Pearson r between the two variables.4

Clearing up Confusion

Some readers will be troubled at this point by something along the


following lines.
If sons of tall men are shorter than their fathers and sons
of short men are taller than their fathers, then why isn't
the variable of stature becoming more homogeneous from one
generation to the next?
To paraphrase the concern,
The examples so far seem to indicate a convergence upon the
mean from both sides; the standard deviation seems to
decrease with each generation. Can this be?

No, the regression effect over several generations does not cause the
population to converge upon the mean. What has been seen is that tall
men don’t produce enough tall sons to replenish the ranks of the tall.
Likewise, short men don’t sire enough short sons to replace their
numbers in the next generation. But others do! The key point is that
the next generation of tall men doesn’t come exclusively from the
ranks of tall fathers; although tall fathers have more than their share
of the tall sons, some average fathers and even a few short fathers
have tall sons. Likewise, the next generation of short men doesn’t
derive solely from short fathers.

If there were a perfect correlation between father-son height, the


trait would be totally heritable through the male line. This would
preclude (a) any hereditary contributions from mothers and (b) any
environmental influence; both notions are, of course, nonsense.

Along the same lines, if there were perfect correlation between


parent-child IQ, it would mean that gifted parents would never produce
non-gifted children and that non-gifted parents would never produce
gifted children. This too is contrary to fact.

4
Note again that we are making the simplifying assumption that the standard deviation
remains the same in the two populations (generations in this case).
Student’s Companion Website for Assessment for Effective Teaching

Before moving on to the next examples, notice in passing that the


arithmetic of prediction in the above illustrations was with raw scores,
while that for the next several will be with derived scores. Either raw
scores or standard scores can appropriately be used because both
have (at least roughly) equal units of measure. Of course, we should
not use kinds of derived scores that lack equal units of measure (e.g.,
grade equivalents or percentile ranks) because it isn’t legitimate to
compute a mean, r, etc., with them.

The next several examples concern academic aptitude. It will be


convenient to use a kind of derived score known as the IQ which
historically , meant literally "intelligence quotient." However, modern
so-called IQs are standard scores that have a mean of 100 and a
standard deviation of 15 or 16. We will revisit them in Chapter 15.

Parent-Child IQ for High Sample

Suppose a group test of academic aptitude has been administered to a


thousand sophomores in a large high school. Now the test scores are
used to select a deviant sample having deviation IQs of 132 or above.5
Assume this yields a deviant sample of 20 high-scoring students whose
mean score is 140. Now we want to predict the mean IQ of the children
these 20 high-scoring people will produce (with yet unidentified mates).
Here are two questions to consider.
1 What else is needed in order to make this prediction?
2 What is the best prediction of the future mean IQ (say by
adolescence) of their children?

To predict the future IQ of the offsprings, the parent-child correlation


must be known. Many studies have correlated IQs of a parent with IQs
of children. By adolescence of the children, the average finding is an r
of about .5.6

5
This procedure would not be appropriate for identifying gifted students. Important
decisions should never be made on the basis of one datum alone, and rigid cutoff points
shouldn’t be used in defining giftedness (or most anything else). Moreover, in most
jurisdictions, classification would legally require use of an individually administered
test of scholastic aptitude.

6
From the reader’s point of view, this .5 is a given, not something that one should be
able to figure out. By the way, it is only coincidental that it equals the approximate
fraction of genes shared by parent-child pairs.
Student’s Companion Website for Assessment for Effective Teaching

Knowing that r = .5, what is the most likely mean IQ of the children of
this deviant sample? To obtain the predicted deviation score of the
children, the deviation score of the deviant sample of parents is
multiplied by the r. Thus, +40 (140 - 100) times .5 equals +20. This
expected mean deviation score of +20 is then added to the population
mean of 100 to obtain the answer of 120.

Of course there will be many exceptions. Some of the children of our


high IQ sample will surpass their parents, but on average they will be
lower. Lower, that is, than the deviant parents, yet higher than the
population mean.

A Bit of Fruitful Digression

By this point, some readers will have reasoned that prediction can be
improved if the IQs of both parents are used as predictors. Indeed it
can. Studies of the relationship between mean of parents' IQs and
those of adolescent offsprings report correlations of about .6.

To take an example, suppose parents have IQs of 110 and 130.


Predictive validity will be maximized if equal weights are assigned to
the parents' IQs.7 Thus, the mean value of 120 is used; this yields a
deviation score of +20 by which to multiple the .6 regression
coefficient. The resulting +12 is then added to the population mean of
100 to yield a predicted offspring IQ of 112.

The regression effect is bad news for the gifted. Gifted individuals
typically have children who are less bright than themselves. Along the
same lines, most famous athletes don’t have children as successful as
themselves. Celebrated composers, philosophers, and politicians are
likewise unlikely to produce children as prominent as themselves.
There are exceptions, but regression toward the mean tilts the odds
against it.

We suggest that the regression effect should be considered for


enrichment curriculum for students who are gifted. It would help them
to form realistic expectations and aspirations for their future children,
not to be disappointed in the likely event of regression, and not to
make their children feel inadequate because of it.

7
Equal weighting of the predictors is appropriate because parents have approximately
equal genetic and environmental impacts. In some problems, however, maximum
validity is attained by very unequal weighting of the predictors.
Student’s Companion Website for Assessment for Effective Teaching

Sibling-Sibling IQ for Low Individual

Suppose Jenny's IQ is 80, and we are interested in predicting the future


IQ of her baby brother. We again need to know the size of the
regression coefficient. Again it happens that the average research
finding yields a correlation of about .5 between sibling IQs by
adolescence. What, then, is the most likely IQ of Jenny's brother by
the time he gets to high school? See if you can get the answer before
reading on.

The answer is 90. It is obtained by finding Jenny's deviation score to


be -20 (her 80 minus the population mean of 100) and then multiplying
it by the r of .5. This yields a most likely deviation score for her
brother of -10. To this -10 the population mean is added to obtain his
most likely IQ—90.

Note again that we should expect much more error in individual


predictions than in group predictions. For a group of 20 or so people, a
prediction is not apt to err by more than a very few points. But for
Jenny's one brother, the prediction is likely to err by more than 8 or so
points.

The regression effect is good news for those who work with persons
having intellectual disabilities. The example of Jenny and her brother
could just as well have been Jenny and her child. Knowledge of this can
increase realistic optimism in some kinds of professional work.

Child-Parent IQ for High Individual

For the next example, think of your brightest friend with whose family
you are acquainted. Chances are that your friend will have children
who are less bright, will have siblings who are less bright, and will have
parents who are less bright.

To take a specific example, suppose your friend's IQ is 130. What is


the most likely value for a parent (or sibling or child)? To find the
answer, we multiply the friend's deviation score of +30 (130 - 100)
by the parent-child (or sibling-sibling or child-parent) correlation of .5
to obtain the most likely deviation score for the parent of +15. Thus,
we predict (expecting again to err to some considerable extent) that
the parent (or sibling or child) will have an IQ of 115 (100 + 15).
Student’s Companion Website for Assessment for Effective Teaching

Two Limiting Case Examples

To show the scope and limits of the regression effect, let’s now
consider some very extreme examples.

The Case of 0.0 Correlation. Suppose a university used its


admissions test scores to select its football squad. (Never mind that
the institution would have a new football coach and a new president the
next year.) Further suppose that the correlation between football skill
and the kind of achievement or aptitude assessed by the admissions
test is 0.00. Let’s suppose the university selected the top 1% of the
males on the basis of the admissions test scores. What skill in football
would we expect this “select” group to have? We multiply the mean
deviation score on the admissions test by the regression coefficient of
0.0. Whatever the mean deviation scores was, the regression would be
all the way to the mean.

That illustrates one limiting case. When there is zero correlation, then
regression will be to the mean. In the usual case where there is some
(either positive or negative) correlation, then regression is only
toward the mean.)

A Case of 1.0 Correlation. For another far-fetched example,


suppose two variables are perfectly correlated. You score a 20-word
spelling test for one point per word, then you convert these raw
scores into percent scores. Each pupil now has two scores—the raw
score and the percent score. Of course, they are perfectly
correlated. Now if you select the top person in the class on the basis
of raw score, you can be certain that he or she will be top in percent
score. Or, if you were to select the bottom on the basis of either
score, you can be certain that this person will also be at the bottom on
the other score.

This illustrates the other limiting case. When there is perfect


correlation, then regression does not occur.

Another Case of 1.0 Correlation. Finally, suppose a group of


students has taken an objective test and every student has answered
every question. Now you score the test both for number of items
answered right and for number answered wrong. In this case, the
correlation between the two variables is -1.00. Now if you were to
select the person who got the largest Right Score, you can be sure
that this person got the smallest Wrong Score. Or if you identified
Student’s Companion Website for Assessment for Effective Teaching

the person whose Wrong Score is at the class mean, you have also
identified the person whose Right Score is at the class mean. Of, if
you selected the person whose Right Score was second from the
bottom, you can be certain that this is the person whose Wrong score
is second from the top.

This example illustrates the case of perfect correlation. When there


is perfect correlation, then regression toward the mean does not
occur.

It also demonstrates that the regression effect works with negative


correlations in the same way as it works with positive ones—aside
from the fact that the predictions made from negatively correlated
variables are on the side of the mean opposite that of the predictor
scores.

Regression of IQ for Persons Selected for


Their Exceptionality

Now suppose there is a well-meaning researcher who is unacquainted


with the regression effect who hopes to raise the IQs of low-I Q
preschool children. This naive investigator proceeds to use an "IQ" test
as the only means by which to select a sample of 25 low-IQ three-year
olds. The plan is to provide these children with an experimental
intervention of an enriched environment.

Assume a .8 correlation between the IQs from the "intelligence" test


given at age 3 and again at age 4.

Suppose the 25 three-year-old children have a mean IQ of 70 and the


treatment lasts a year. In a research report, the investigator
indicates that the mean IQ of the children at age 4 is 80; therefore,
the intervention raised the IQ an impressive average of 10 IQ points!

Any research design for a study of this type should take the
regression effect into account. How much drift upwards toward the
mean would be predicted on the basis of the regression effect alone?
That is, if the intervention were worthless, what mean IQ would be
expected at age 4 for this sample selected because of negative
deviation from the mean? The answer is found by multiplying the .8 r
by the deviation score of -30 (70 - 100) to get the expected
Student’s Companion Website for Assessment for Effective Teaching

deviation of -24; this yields an expected IQ of 76 (100 - 24). Thus,


the regression effect accounts for over half of the 10 point gain!8

It is beyond the scope of this enrichment reading to treat all aspects


of a more suitable research design (e.g., use of control group and
sample size). We can, however, address one easy way of preventing
the regression effect from contaminating the findings. One can
compute how much regression is expected. Knowing this, the findings
can simply be "discounted" by that amount. Finding that a 6-point
"free ride" from the regression effect is expected, we would, if other
aspects of the research design were sound, herald only the remaining
4-point gain.

Improvement in Reading

Suppose a school district establishes a new program to remediate


reading deficiencies among sixth graders. The results of Form L of a
recently administered general reading comprehension test are used to
identify the poorest 2% of the district's first-semester, sixth-grader
readers. The mean z-score of the resulting 95 pupils identified turns
out to be -2.50. Form L z-scores serve double duty; they are also
used as the pretest or benchmark against which progress is measured
during an intensive one-semester remedial program for these poor
readers. At the end of the program, Form M of the same reading test
is administered to the pupils.

The Form L z-scores are based on a national cross-section of first-


semester sixth graders, while Form M z-scores are based on a national
sample of second-semester sixth graders. Consequently, normal
growth won’t show up in the scores because the reference groups also
differed by normal growth. Hence, we don’t need to attend to that
issue.

8
Accounting for the other 4 points of gain is more problematic. Since we know the
investigator was so unsophisticated as to fail to account for the regression effect, we
might worry that there may have been some tilting of the intervention toward the test--
some teaching to the test. This could easily account for the other 4 points of gain.
Another possibility, of course, is that the treatment really does have merit and that it
did contribute to the gain. Another is that the outcome was, in part, a fluke of chance;
this is especially likely, of course, with small samples. Yet another viable hypothesis is
a practice effect that raised the scores on the second occasion. Although all of this is
speculative, it is not speculative to say that the researcher lacked competence and that it
is therefore hazardous to assume that all is well with all aspects of the research design.
Student’s Companion Website for Assessment for Effective Teaching

The test manual reports that for a typical sample, the parallel-form
correlation over a one-semester interval was .84. Now the amount of
"free ride" from the regression effect can be calculated. How much
apparent gain is to be expected from the regression effect? Try to
compute it before reading on.

The -2.50 mean z-score of the deviant group is multiplied by the .84 r
to yield an expected mean z-score of -2.10 if there is no real gain.
That is, the regression effect alone should raise the scores about .4 z-
score units. The interpretation of results should "discount" the
findings by this amount.

Here is an alternative research design for preventing the regression


effect from contaminating findings: Separate instruments are used
for selecting participants and for obtaining pretest data. Hence,
one device is used to select participants because of their deviance;
let’s call it the selection test. Then another test is administered
before the treatment is started; its scores are used as the pretest.
These pretest scores will show regression toward the mean from
scores of the selection test.

The regression has now been “flushed out” before the comparison of
pretest and posttest scores. If the researcher refrains from any
further selection (or de-selection) on the basis of the second test's
results (i.e., the results of the pretest), then the difference between
the pretest and the posttest is not contaminated by the regression
effect. That is, if there is no further selection between pretest and
posttest, then there will be no further regression between them.

To paraphrase, regression will occur between the selection test and


the pretest. The same amount of regression will occur between the
selection test and the posttest. Hence, the pretest and postest are
stable with regard to the regression effect.

This brings us to a key principle. Regression is expected when


and only when a sample is selected because of its deviance.
In each example, a sample was selected because of its exception-ality
on a given measure. In such cases, regression between the selection
measure and any other variable is to be expected. That is why
separating the functions of the selection variable and the premeasure
avoids regression between premeasure and postmeasure. If no
selection is performed by use of the premeasure, then no regression is
to be expected between it and any other variable.
Student’s Companion Website for Assessment for Effective Teaching

Student Achievement Status in a Course

Suppose a teacher administers the first social studies test of the year
to a tenth-grade class. The performance of 5 of the 25 students is
reported below.
• Abe received the highest score.
• Bertha obtained a high-average score.
• Maxie got an average score.
• Olivia received a low-average score.
• Rob got the lowest score.

Assuming we know nothing else about these or other students, whose


relative status is most likely to decrease on the next test? The best
bet is Abe. Whose status is most likely to go up? Rob’s is.

A "common sense," understanding of these predictions can come from


considering the question: “What probably happens to cause an Abe to
be top or a Rob to be bottom?” Often several things conspire to cause
an unusually high performance. For example:
• The student probably knows more than average (not necessarily the
most in the class, but more than average).
• The student probably had good luck in happening to know the items
that appear on the test and having the test fortuitously avoid his
or her weak areas.
• The student may well have had some good luck in guessing on
some questions.
• The student may have had unusually harmonious home environment
the week before.
• The student probably did not have any unusually distressing events
(e.g., an argument with a good friend) happen right before the test.

Now which of these kinds of things are likely to recur at the time of
the next test? Superior knowledge, but not luck. Abe is likely to be
well above average in his knowledge in the next unit. But the good luck
he may have had isn’t any more likely to revisit him than to visit
anyone else. Since he is the most likely to have had good luck, he is
the one with the greatest chance of declining on the next test.

Of the five, Bertha is the next most likely to decline. The best bet is
that Rob would improve the most and that Olivia would improve the
Student’s Companion Website for Assessment for Effective Teaching

next most. The only person for whom we would predict stability would
be Maxie; since she did not deviate from the mean, there can be no
tendency to move toward it.

A word of caution may be in order about not overreacting to the


regression effect. If asked to “put your money” on one of the 25
students in the class for top score on the next text, Abe would surely
be the best single bet. His chances are better than those of any other
one person's. The reason his chances aren't too good is that there are
multiple opportunities for others to outscore him. The odds are less
than even that any other one student will overtake Abe, but when
there are several who have only slightly less than even odds of passing
him, the odds are good that at least one of them will.

Improvement During Counseling

Suppose the director of a university counseling center wants to


investigate the effect of counseling on the general life satisfaction of
self-referred undergraduates undergoing emotional counseling. A
research design might involve administering a self-report inventory to
self-referred counselees at the beginning of counseling and again
after, say, three months.

This situation is ripe for contamination by the regression effect if it


isn’t taken into account. At the common sense level, one might ask:
“When do people most often self-refer for personal counseling?”
Often at an all-time emotional low; almost always at a relative low.
Since life satisfaction is not a perfectly stable attribute, which way
are things more likely to go when one is experiencing a low? Up!
Therefore, adequate research designs must take the regression effect
into account.

Implications for Research

It has been well over a century since Sir Francis Galton described the
regression effect, and it seems reasonable to expect competent
researchers to be cognizant of the phenomenon and to allow for it in
designing their investigations. Otherwise, the artifact can seriously
contaminate research findings and interpretations.

In research with low-performing samples, failure to recognize the


impact of regression tends to yield findings that make treatments look
more successful than they really are. Programs for low
Student’s Companion Website for Assessment for Effective Teaching

populations--e.g., basic skills programs or intervention for high-risk


populations--are vulnerable to having their effectiveness
overestimated. If regression, or drift, upward toward the mean goes
unrecognized, then squanderous public policy decisions can be made to
duplicate programs that appear to be highly successful only because of
the free ride they received from the regression artifact.

Inversely, in research concerning high-performance samples, failure to


take regression into account tends to produce findings that make
interventions appear less successful than they really are.
Investigations of programs for high populations--e.g., high academic
aptitude, exceptional artistic abilities, extraordinary creativity, and
exceptional athletic achievement--are vulnerable to under-estimating
program effect. If regression, or movement, downward toward the
mean, is not taken into account, it can lead to tragic public policy
decisions to abandon good programs.

A Potpourri of Other Examples

Attentiveness to the regression effect is appropriate for a broad


variety of situations. Following are a few examples both from
education and from other fields.
• If you finally get a perfect score for the first time in a course quiz,
it is realistic to expect to drop on the next quiz. This is true not
only from the perspective of regression toward the group mean, but
even more from the perspective of regressing toward your o w n
mean.
• If you find this the dullest of all topics to date, then you have a basis
to be optimistic that the next one will be less boring.
• An elementary school principal carefully plots the percentage of
children who reach a mastery threshold on each objective on a state
minimum competence test. The objectives on which the school
performed the lowest are identified for special attention the
following year. If no action were taken, we would expect the school
to improve on those objectives next year. Now if some hard work is
devoted to them, the principal who is not conversant with the
regression effect will be prone to overestimate the impact of the
efforts.
• If Mary comes home from her softball game proclaiming, "This was
my best game ever!" then one can reasonably predict that her
performance will decline in the next game.
Student’s Companion Website for Assessment for Effective Teaching

• If this year's Christmas temperature is the coldest in recorded


history, then it is prudent to expect a warmer Christmas next year.
• If you are the most depressed today that you have ever been, then
things will probably get better tomorrow.
• If today's New York Stock Exchange volume is extraordinarily low,
then it will probably increase tomorrow.
• A top race horse is valuable as a stud. His chances of siring the
next champion are better than are those of a less celebrated horse,
yet his chances are not all that good. Most of his progeny will be
good, but not as good as their champion father.
• The prize milk-producing cow at the fair will probably not mother the
prize winner in the next generation.
• The world's best laying hen is not likely to produce the next record
breaker, although her chances are better than those of any other
single biddy.

In spite of the last three examples, selective breeding is not futile


because there is some parent-offspring correlation. Although
movement toward the mean occurs, movement is not all the way to the
mean. Regression is manifest in each generation only because the
breeder (or nature) persists in each generation in selecting for
breeding the stock with the most desirable (or adaptive)
characteristics. And this brings us full circle back to Darwin’s
explanation of the origin of species.

Summary

Regression toward the mean appears in a remarkably wide variety of


educational, biological, social, economic, and everyday-life situa-ions.
Whenever a deviant sample of individuals or events is singled
out for study because of deviance, movement in the
direction of the mean on any other variable in the universe
will likely be observed.

Two things control the degree of regression:


• First, the more deviant the sample, the stronger the "pull" toward
the mean.
• Second, the less the correlation between the selection variable
and the variable of interest, the greater the regression toward
the mean.
Student’s Companion Website for Assessment for Effective Teaching

An understanding of the regression effect can do much to (a) enable


the design of research that is not contaminated by the phenomenon,
(b) facilitate the critical evaluation of research, (c) enhance
comprehension of events, and (d) enable more accurate prediction of
the future.

Thinking Cap Exercise

1 Josh grows a certain kind of melon. This year he was delighted to


produce the state’s largest specimen. He carefully kept the seeds
of this prize-winning melon. He would best expect the seeds to
produce melons that are
A larger than the “champ.”
B about as large as the “champ.”
C smaller than the “champ,” but larger than average.
D average in size.
E smaller than average.

2 Jane is shorter than any other female on either side of her family.
Her daughters are most apt to be
A shorter than she.
B about equal to her stature.
C taller than she, but shorter than most female members of her
family.
D as tall as most female members of her family.
E taller than most female members of her family.

3 Suppose that the first two tests in a course are correlated +0.7.
Jason receives a z-score of -2.0 on a the first test. What z-
scores would we best predict that he would make on the next
quiz? (Ignore affective factors.)
A +.0
B .0
C -.7
D -1.0
E -1.4

4 Suppose arithmetic test scores and reading test scores among


sixth graders correlate .6. Compared with a national sample of
other sixth graders, Jean’s arithmetic test z-score is about one
standard deviation above average. We would best expect her
reading test z-score to be ____.
Student’s Companion Website for Assessment for Effective Teaching

5 Suppose that in a certain country (a) longevity between the


length of life of mothers and daughters correlates .4, (b) the means
and standard deviation of life spans is the same for both
generations, and (c) the mean age at death in both generations is
60. Anita’s mother lived to be 110. You would best predict Anita’s
age at death to be _____.

Key to Thinking Cap Exercise

Item Key Comment


1 C We expect regression toward the mean. Not all the
way to the mean and certainly not through the mean
to the other side.

2 C Here too, we anticipate regression—i.e., drift—


toward the mean. Not all the way, just toward it.

3 E The variables are correlated .7. Hence we multiply


the z-score on the first (-2.0) by the regression
coefficient (.7) to find the answer (-1.4).

4 +.6 The variables are correlated .6. Jean’s status on


one of the variables is z = +1.0. To predict the
her status on the other variable, we multiply these
two values and obtain +.6.

5 80 Anita’s mother lived 50 years longer than average


(110-60). We multiply that 50 years by the
r of .4 and obtain 20 years as the amount above
average to be predicted for Anita.

Вам также может понравиться