Professional Documents
Culture Documents
In this paper, we explore the rules that the diagnosis of a patient, or a person's
determine intuitive predictions and judg- future occupation. In a numerical case,
ments of confidence and contrast these the prediction is given in numerical form,
rules to the normative principles of statis- for example, the future value of a particular
tical prediction. Two classes of prediction stock or of a student's grade point average.
are discussed: category prediction and In making predictions and judgments
numerical prediction. In a categorical under uncertainty, people do not appear
case, the prediction is given in nominal to follow the calculus of chance or the
form, for example, the winner in an election, statistical theory of prediction. Instead,
they rely on a limited number of heuristics
1
Research for this study was supported by the which sometimes yield reasonable judg-
following grants: Grants MH 12972 and MH 21216 ments and sometimes lead to severe and
from the National Institute of Mental Health and systematic errors (Kahneman & Tversky,
Grant RR 05612 from the National Institute of
Health, U. S. Public Health Service; Grant GS 3250 1972; Tversky & Kahneman, 1971, 1973).
from the National Science Foundation. Computing The present paper is concerned with the
assistance was obtained from the Health Services role of one of these heuristics—representa-
Computing Facility, University ot California at tiveness—in intuitive predictions.
Los Angeles, sponsored by Grant MH 10822 from
the U. S. Public Health Service.
Given specific evidence (e.g., a person-
The authors thank Robyn Dawes, Lewis Gold- ality sketch), the outcomes under consider-
berg, and Paul Slovic for their comments. Sundra ation (e.g., occupations or levels of achieve-
Gregory and Richard Kleinknecht assisted in the ment) can be ordered by the degree to
preparation of the test material and the collection which they are representative of that evi-
of data.
J
Requests for reprints should be sent to Daniel dence. The thesis of this paper is that
Kahneman, Department of Psychology, Hebrew people predict by representativeness, that
University, Jerusalem, Israel. is, they select or order outcomes by the
237
238 DANIEL KAHNEMAN AND AMOS TVERSKY
TABLE 2
PRODUCT-MOMKNT CoKKKI.ATIONS OF MlCAN I.IKKI.IHOOD RANK WITH MlsAN S I M I L A R I T Y
R A N K A N D WITH BASIC RATK
manipulation of expected accuracy does no! produced under the low-accuracy instruc-
affect the pattern of predictions. The ex- tions were not significantly closer to the
perimental material consisted of five thumb- base-rate distribution than the orderings
nail personality sketches of ninth-grade produced under the high-accuracy instruc-
boys, allegedly written by a counselor on tions. A product-moment correlation was
the basis of an interview in the context of computed for each judge, between the
a longitudinal study. The design was the average rank he had assigned to each of the
same as in the Tom VV. study. For each nine outcomes (over the five descriptions)
description, subjects in one group (TV = 69) and the base rate. This correlation is an
ranked the nine fields of graduate special- overall measure of the degree to which the
ization (see Table 1) in terms of the simi- subject's predictions conform to the base-
larity of the boy described to their "image rate distribution. The averages of these
of the typical first-year graduate student individual correlations were .13 for subjects
in that field." Following the similarity in the high-accuracy group and .16 for
judgments, they estimated the base-rate subjects in the low-accuracy group. The
frequency of the nine areas of graduate difference docs not approach significance
specialization. These estimates were shown (/ = .42, df = 103). This pattern of judg-
in Table 1. The remaining subjects were ments violates the normative theory of
told that the five cases had been randomly prediction, according to which any de-
selected from among the participants in crease in expected accuracy should be
the original study who are now first-year accompanied by a shift of predictions to-
graduate students. One group, the high- ward the base rate.
accuracy group (N = 55), was told that Since the manipulation of expected ac-
"on the basis of such descriptions, students curacy had no effect on predictions, the
like yourself make correct predictions in two prediction groups were pooled. Sub-
about 55% of the cases." The low- sequent analyses were the same as in the
accuracy group (N = 50) was told that Tom W. study. For each description, two
students' predictions in this task are correlations were computed: (a) between
correct in about 27% of the cases. For mean likelihood rank and mean similarity
each description, the subjects ranked the rank and (b) between mean likelihood rank
nine fields according to "the likelihood that and mean base rate. These correlations
the person described is now a graduate are shown in Table 2, with the outcome
student in that field." For each descrip- judged most likely for each description.
tion, they also estimated the probability The correlations between prediction and
that their first choice was correct. similarity are consistently high. In con-
The manipulation of expected accuracy trast, there is no systematic relation be-
had a significant effect on these probability tween prediction and base rate: the correla-
judgments. The mean estimates were .70 tions vary widely depending on whether
and .56, respectively, for the high- and low- the most representative outcomes for each
accuracy group (/ = 3.72,/> < .001). Mow- description happen to be frequent or
ever, the orderings of the nine outcomes rare.
ON THE PSYCHOLOGY OF PREDICTION 241
Here again, considerations of base rate group). Subjects in another group (the
were neglected. In the statistical theory, high-engineer, H group; N = 86) were
one is allowed to ignore the base rate only given identical instructions except for the
when one expects to be infallible. In all prior probabilities: they were told that the
other cases, an appropriate compromise set from which the descriptions had been
must be found between the ordering sug- drawn consisted of 70 engineers and 30
gested by the description and the ordering lawyers. All subjects were presented with
of the base rates. It is hardly believable the same five descriptions. One of the
that a cursory description of a fourteen- descriptions follows:
year-old child based on a single interview Jack is a 45-year-old man. He is married and has
could justify the degree of infallibility im- four children. He is generally conservative, careful,
plied by the predictions of our subjects. and ambitious. He shows no interest in political
Following the five personality descrip- and social issues and spends most of his free time on
his many hobbies which include home carpentry,
tions, the subjects were given an additional sailing, and mathematical puzzles.
problem: The probability that Jack is one of the 30 engineers
in the sample of 100 is %.
About Don you will be told nothing except that he
participated in the original study and is now a first- Following the five descriptions, the subjects
year graduate student. Please indicate your order- encountered the null description:
ing and report your confidence for this case as well.
Suppose now that you are given no information
For Don the correlation between mean whatsoever about an individual chosen at random
likelihood rank and estimated base rate from the sample.
was .74. Thus, the knowledge of base The probability that this man is one of the 30
engineers in the sample of 100 is .%.
rates, which was not applied when a de-
scription was given, was utilized when no In both the high-engineer and low-engi-
specific evidence was available. neer groups, half of the subjects were asked
to evaluate, for each description, the prob-
Prior versus Individuating Evidence ability that the person described was an
engineer (as in the example above), while
The next study provides a more stringent the other subjects evaluated, for each de-
test of the hypothesis that intuitive pre- scription, the probability that the person
dictions are dominated by representative- described was a lawyer. This manipulation
ness and are relatively insensitive to prior had no effect. The median probabilities
probabilities. In this study, the prior assigned to the outcomes engineer and
probabilities were made exceptionally sali- lawyer in the two different forms added to
ent and compatible with the response about 100% for each description. Con-
mode. Subjects were presented with the sequently, the data for the two forms were
following cover story: pooled, and the results are presented in
A panel of psychologists have interviewed and ad- terms of the outcome engineer.
ministered personality tests to 30 engineers and 70 The design of this experiment permits
lawyers, all successful in their respective fields. On the calculation of the normatively appro-
the basis of this information, thumbnail descriptions
of the 30 engineers and 70 lawyers have been written. priate pattern of judgments. The deriva-
You will find on your forms five descriptions, chosen tion relies on Bayes' formula, in odds form.
at random from the 100 available descriptions. Let 0 denote the odds that a particular
For each description, please indicate your prob- description belongs to an engineer rather
ability that the person described is an engineer,
on a scale from 0 to 100. than to a lawyer. According to Bayes'
The same task has been performed by a panel of rule, O = Q-R, where Q denotes the prior
experts, who were highly accurate in assigning odds that a randomly selected description
probabilities to the various descriptions. You will belongs to an engineer rather than to a
be paid a bonus to the extent that your estimates lawyer; and R is the likelihood ratio for a
come close to those of the expert panel.
particular description, that is, the ratio of
These instructions were given to a group the probability that a person randomly
of 85 subjects (the low-engineer, or L drawn from a population of engineers will
242 DANIEL KAHNEMAN AND AMOS TVKRSKY
with the alternative outcomes, and judg- reflect predictive accuracy. When pre-
ments were controlled, we suggest, by the dictive accuracy is perfect, one predicts
degree to which the descriptions appeared the criterion value that will actually occur.
representative of these stereotypes. In When uncertainty is maximal, a fixed value
other problems, the outcomes are more is predicted in all cases. (In category pre-
naturally viewed as segments of a dimen- diction, one predicts the most frequent
sion. Suppose, for example, that one is category. In numerical prediction, one
asked to judge the probability that each of predicts the mean, the mode, the median,
several students will receive a fellowship. or some other value depending on the loss
In this problem, there are no well-delineated function.) Thus, the variability of predic-
stereotypes of recipients and nonrecipients tions is equal to the variability of the
of fellowships. Rather, it is natural to criterion when predictive accuracy is per-
regard the outcome (i.e., obtaining a. fellow- fect, and the variability of predictions is
ship) as determined by a cutoff point along zero when predictive accuracy is zero.
the dimension of academic achievement or With intermediate predictive accuracy, the
ability. Prior probabilities, that is, the variability of predictions takes an inter-
percentage of fellowships in the relevant mediate value, that is, predictions are re-
group, could be used to define the outcomes gressive with respect to the criterion.
by locating the cutoff point. Consequently, Thus, the greater the uncertainty, the
they are not likely to be ignored. In addi- smaller the variability of predictions. Pre-
tion, we would expect extreme prior prob- dictions by representativeness do not follow
abilities to have some effect even in the this rule. It was shown in the previous
presence of clear stereotypes of the out- section that people did not regress toward
comes. A precise delineation of the condi- more frequent categories when expected
tions under which prior information is used accuracy of predictions was reduced. The
or discarded awaits further investigation. present section demonstrates an analo-
One of the basic principles of statistical gous failure in the context of numerical
prediction is that prior probability, which prediction.
summarizes what we knew about the prob-
lem before receiving independent specific Prediction of Outcomes versus Evaluation of
evidence, remains relevant even after such Inputs
evidence is obtained. Bayes' rule trans- Suppose one is told that a college fresh-
lates this qualitative principle into a multi- man has been described by a counselor as
plicative relation between prior odds and intelligent, self-confident, well-read, hard-
the likelihood ratio. Our subjects, how- working, and inquisitive. Consider two
ever, failed to integrate prior probability types of questions that might be asked
with specific evidence. When exposed to about this description:
a description, however scanty or suspect,
of Tom W. or of Dick (the engineer/ (a) Evaluation: How does this description impress
you with respect to academic ability? What per-
lawyer), they apparently felt that the dis- centage of descriptions of freshmen do you believe
tribution of occupations in his group was would impress you more? (b) Prediction: What is
no longer relevant. The failure to appre- your estimate of the grade point average that this
ciate the relevance of prior probability in student will obtain? What is the percentage of
the presence of specific evidence is perhaps freshmen who obtain a higher grade point average?
one of the most significant departures of There is an important difference between
intuition from the normative theory of the two questions. In the first, you evalu-
prediction. ate the input; in the second, you predict an
outcome. Since there is surely greater un-
NUMERICAL PREDICTION certainty about the second question than
A fundamental rule of the normative about the first, your prediction should be
theory of prediction is that the variability more regressive than your evaluation.
of predictions, over a set of cases, should That is, the percentage you give as a pre-
244 DANIEL KAHNEMAN AND AMOS TVKRSKY
within the data of each subject. A com- between-subjects and within-subject com-
parison of these values indicated no signifi- parisons. Although the judges were un-
cant differences in variability between the doubtedly aware of the multitude of factors
evaluation and the prediction groups, that intervene between a single trial lesson
within the range of values under study. In and teaching competence five years later,
the adjectives study, the average standard this knowledge did not cause their predic-
deviation was 25.7 for the evaluation group tions to be more regressive than their
(N = 38) and 24.0 for the prediction group evaluations.
(N = 36) (/ = 1.25, df = 72, ns). In the
reports study, the average standard devia- Prediction versus Translation
tion was 22.2 for the evaluation group The previous studies showed that pre-
(N = 37) and 21.4 for the prediction group dictions of a variable are not regressive
(N = 63) (t = .75, df = 98, ns). In both when compared to evaluations of the inputs
studies the prediction and the evaluation in terms of that variable. In the following
groups produced equally extreme judg- study, we show that there are situations in
ments, although the former predicted a which predictions of a variable (academic
remote objective criterion on the basis of achievement) are no more regressive than
sketchy interview information, while the a mere translation of that variable from
latter merely evaluated the impression ob- one scale to another. The grade point
tained from each description. In the statis- average was chosen as the outcome variable,
tical theory of prediction, the observed because its correlates and distributional
equivalence between prediction and evalua- properties are well known to the subject
tion would be justified only if predictive population.
accuracy were perfect, a condition which Three groups of subjects participated in
could not conceivably be met in these the experiment. Subjects in all groups
studies. predicted the grade point average of 10
Further evidence for the equivalence of hypothetical students on the basis of a
evaluation and prediction was obtained in single percentile score obtained by each of
a master's thesis by Beyth (1972). She these students. The same set of percentile
presented three groups of subjects with scores was presented to all groups, but the
seven paragraphs, each describing the per- three groups received different interpreta-
formance of a student-teacher during a tions of the input variable as follows.
particular practice lesson. The subjects 1. Percentile grade point average. The
were students in a statistics course at the subjects in Group 1 (N = 32) were told
Hebrew University. They were told that that "for each of several students you will
the descriptions had been drawn from be given a percentile score representing
among the files of 100 elementary school his academic achievements in the freshman
teachers who, five years earlier, had com- year, and you will be asked to give your
pleted their teacher training program. Sub- best guess about his grade point average
jects in an evaluation group were asked for that year." It was explained to the
to evaluate the quality of the lesson de- subjects that "a percentile score of 65, for
scribed in the paragraph, in percentile example, means that the grade point
scores relative to the stated population. average achieved by this student is better
Subjects in a prediction group were asked than that achieved by 65% of his class, etc."
to predict in percentile scores the current 2. Mental concentration. The subjects
standing of each teacher, that is, his overall in Group 2 (N = 37) were told that "the
competence five years after the description test of mental concentration measures one's
was written. An evaluation-prediction ability to concentrate and to extract all the
group performed both tasks. As in the information conveyed by complex messages.
studies described above, the differences It was found that students with high grade
between evaluation and prediction were not point averages tend to score high on the
significant. This result held in both the mental concentration test and vice versa.
246 DANIEL KAHNEMAN AND AMOS TVERSKY
TABLE 3
AVERAGES OF INDIVIDUAL PREDICTION STATISTICS FOR THE THREE GROUPS AND RESULTS OF
PLANNED COMPARISONS BETWEEN GROUPS 1 AND 2, AND BETWEEN GROUPS 2 AND 3
Group
Statistic 1. PercenLile
grade point 1 vs. 2 2. Mental 2 vs. 3 3. Sense of
concentration humor
Mean predicted
grade point average 2.27 ns 2.3S .05 2.46
SD of predictions .91 ns .87 .01 .69
Slope of regression .030 ns .029 .01 .022
r .97 ns .95 ns .94
that is most representative of the input. they encountered conformed to these ex-
We propose that the degree of confidence pectations. (For half of the subjects the
one has in a prediction reflects the degree labels of the correlated and the uncorre-
to which the selected outcome is more repre- lated pairs of tests were reversed). Sub-
sentative of the input than are other out- jects were told that "all tests were found
comes. A major determinant of represen- equally successful in predicting college per-
tativeness in the context of numerical pre- formance." In this situation, of course, a
diction with multiattribute inputs (e.g., higher predictive accuracy can be achieved
score profiles) is the consistency, or co- with the uncorrelated than with the corre-
herence, of the input. The more consistent lated pair of tests. As expected, however,
the input, the more representative the subjects were more confident in predicting
predicted score will appear and the greater from the correlated tests, over the en-
the confidence in that prediction. For tire range of predicted scores (t = 4.80,
example, people predict an overall B aver- df = 129, p < .001). That is, they were
age with more confidence on the basis of B more confident in a context of inferior
grades in two separate introductory courses predictive validity.
than on the basis of an A and a C. Indeed, Another finding observed in many pre-
internal variability or inconsistency of the diction studies, including our own, is that
input has been found to decrease confidence confidence is a J-shaped function of the pre-
in predictions (Slovic, 1966). dicted level of performance (see Johnson,
The intuition that consistent profiles 1972). Subjects predict outstandingly high
allow greater predictability than incon- achievement with very high confidence, and
sistent profiles is compelling. It is worth they have more confidence in the predic-
noting, however, that this belief is in- tion of utter failure than of mediocre
compatible with the commonly applied performance. As we saw earlier, intuitive
multivariate model of prediction (i.e., the predictions are often insufficiently regres-
normal linear model) in which expected sive. The discrepancies between predic-
predictive accuracy is independent of tions and outcomes, therefore, are largest
within-profile variability. at the extremes. The J-shaped confidence
Consistent profiles will typically be en- function entails that subjects are most
countered when the judge predicts from confident in predictions that are most likely
highly correlated scores. Inconsistent pro- to be off the mark.
files, on the other hand, are more frequent The foregoing analysis shows that the
when the intercorrelations are low. Be- factors which enhance confidence, for
cause confidence increases with consistency, example, consistency and extremity, are
confidence will generally be high when the often negatively correlated with predictive
input variables are highly correlated. How- accuracy. Thus, people are prone to ex-
ever, given input variables of stated va- perience much confidence in highly fallible
lidity, the multiple correlation with the judgments, a phenomenon that may be
criterion is inversely related to the correla- termed the illusion of validity. Like other
tions among the inputs. Thus, a para- perceptual and judgmental errors, the
doxical situation arises where high inter- illusion of validity often persists even when
correlations among inputs increase con- its illusory character is recognized. When
fidence and decrease validity. interviewing a candidate, for example,
To demonstrate this effect, we required many of us have experienced great con-
subjects to predict grade point average on fidence in our prediction of his future
the basis of two pairs of aptitude tests. performance, despite our knowledge that
Subjects were told that one pair of tests interviews are notoriously fallible.
(creative thinking and symbolic ability)
was highly correlated, while the other pair INTUITIONS ABOUT REGRESSION
of tests (mental flexibility and systematic Regression effects are all about us. In
reasoning) was not correlated. The scores our experience, most outstanding fathers
250 DANIEL KAHNEMAN AND AMOS TVERSKY
have somewhat disappointing sons, brilliant intervals that were symmetric around 140,
wives have duller husbands, the ill-adjusted failing to express any expectation of re-
tend to adjust and the fortunate are even- gression. Of the remaining 35 subjects, 24
tually striken by ill luck. In spite of these stated regressive confidence intervals and 11
encounters, people do not acquire a proper stated counterregressive intervals. Thus,
notion of regression. First, they do not most subjects ignored the effects of un-
expect regression in many situations where reliability in the input and predicted as if
it is bound to occur. Second, as any teacher the value of 140 was a true score. The
of statistics will attest, a proper notion of tendency to predict as if the input informa-
regression is extremely difficult to acquire. tion were error free has been observed
Third, when people observe regression, repeatedly in this paper.
they typically invent spurious dynamic The occurrence of regression is some-
explanations for it. times recognized, either because we discover
What is it that makes the concept of regression effects in our own observations
regression counterintuitive and difficult or because we are explicitly told that re-
to acquire and apply? We suggest that a gression has occurred. When recognized, a
major source of difficulty is that regression regression effect is typically regarded as a
effects typically violate the intuition that systematic change that requires substan-
the predicted outcome should be maximally tive explanation. Indeed, many spurious
representative of the input information. 5 explanations of regression effects have
To illustrate the persistence of nonre- been offered in the social sciences.6 Dy-
gressive intuitions despite considerable ex- namic principles have been invoked to ex-
posure to statistics, we presented the fol- plain why businesses which did excep-
lowing problem to our sample of graduate tionally well at one point in time tend to
students in psychology: deteriorate subsequently and why training
A problem of testing. A randomly selected indi-
in interpreting facial expressions is beneficial
vidual has obtained a score of 140 on a standard IQ to trainees who scored poorly on a pretest
test. Suppose than an IQ score is the sum of a and detrimental to those who did best.
"true" score and a random error of measurement Some of these explanations might not have
which is normally distributed. been offered, had the authors realized that
Please give your best guess about the 95% upper
and lower confidence bounds for the true [Q of this given two variables of equal variances, the
person. That is, give a high estimate such that you following two statements arc logically
are 95% sure that the true IQ score is, in fact, equivalent: (a) Y is regressive with respect
lower than that estimate, and a low estimate such to X; (b) the correlation between Y and
that you are 95% sure that the true score is in fact X is less than unity. Explaining regres-
higher.
sion, therefore, is tantamount to explaining
In this problem, the respondents were why a correlation is less than unity.
told to regard the observed score as the As a final illustration of how difficult it
sum of a "true" score and an error com- is to recognize and properly interpret re-
ponent. Since the observed score is con- gression, consider the following question
siderably higher than the population mean, which was put to our sample of graduate
it is more likely than not that the error students. The problem described actually
component is positive and that this indi- arose in the experience of one of the authors.
vidual will obtain a somewhat lower score A problem of training. The instructors in a flight
on subsequent tests. The majority of sub- school adopted a policy of consistent positive rein-
jects (73 of 108), however, stated confidence forcement recommended by psychologists. They
6 verbally reinforced each successful execution of a
The expectation that every significant particle of flight maneuver. After some experience with this
behavior is highly representative of the actor's training approach, the instructors claimed that
personality may explain why laymen and psychol- contrary to psychological doctrine, high praise for
ogists alike are perenially surprised by the negligible
6
correlations among seemingly interchangeable mea- For enlightening discussions of regression falla-
sures of honesty, of risk taking, of agressiou, and of cies in research, see, for example, Campbell (1969)
dependency (Mischel, 1968). and Wallis and Roberts (1956).
ON THE PSYCHOLOGY OF PREDICTION 251
good execution of complex maneuvers typically re- and sons. Evidently, statistical training
sults in a decrement of performance on the next try. alone does not change fundamental intui-
What should the psychologist say in response?
tions about uncertainty.
Regression is inevitable in flight maneu-
vers because performance is not perfectly REFERENCES
reliable and progress between successive
BEYTH, R. Man as an intiutive statistician: On
maneuvers is slow. Hence, pilots who did erroneous intuitions concerning the description
exceptionally well on one trial are likely to and prediction of events. Unpublished master's
deteriorate on the next, regardless of the thesis (in Hebrew), The Hebrew University,
instructors' reaction to the initial success. Jerusalem, 1972.
CAMPBELL, D. T. Reforms as experiments. Ameri-
The experienced flight instructors actually can Psychologist, 1969, 24, 409-429.
discovered the regression but attributed it GRAY, C. W. Predicting with intuitive correlations.
to the detrimental effect of positive rein- Psychonomic Science, 1968, 11, 41-42.
forcement. This true story illustrates a GUILFORD, J. P. Psychometric methods. New York:
saddening aspect of the human condition. McGraw-Hill, 1954.
JOHNSON, D. M. Systematic introduction to the
We normally reinforce others when their psychology of thinking. New York: Harper &
behavior is good and punish them when Row, 1972.
their behavior is bad. By regression alone, KAHNEMAN, D., & TVERSKY, A. Subjective prob-
therefore, they are most likely to improve ability: A judgment of representativeness. Cog-
after being punished and most likely to nitive Psychology, 1972, 3, 430-454.
MISCHEL, W. Personality and assessment. New
deteriorate after being rewarded. Conse- York: Wiley, 1968.
quently, we are exposed to a lifetime Sl.ovic, P. Cue consistency and cue utilization in
schedule in which we are most often re- judgment. American Journal of Psychology, 1966,
warded for punishing others, and punished 79, 427-434.
SLOVIC, P., & LICHTENSTEIN, S. Comparison of
for rewarding. Bayesian and regression approaches to the study
Not one of the graduate students who of information processing in judgment. Organ-
answered this question suggested that re- izational Behavior and Hitman Performance, 1971,
gression could cause the problem. Instead, 6, 649-744.
they proposed that verbal reinforcements STEVENS, S. S., & GREENBAOM, H. B. Regression
effect in psychophysical judgment. Perception &
might be ineffective for pilots or that they Psychophysics, 1966, 1, 439-446.
could lead to overconfidence. Some stu- TVEKSKY, A., & KAHNKMAN, D. Belief in the law
dents even doubted the validity of the of small numbers. Psychological Bulletin, 1971,
instructors' impressions and discussed pos- 76, 105-110.
sible sources of bias in their perception of TVERSKY, A., & KAHNEMAN, D. Availability: A
heuristic for judging frequency and probability.
the situation. These respondents had Cognitive Psychology, 1973, in press.
undoubtedly been exposed to a thorough WALLIS, W. A., & ROBERTS, H. V. Statistics: A
treatment of statistical regression. Never- new approach. New York: The Free Press, 1956.
theless, they failed to recognize an instance WOODWORTH, R. S. Experimental psychology. New
York: Holt, 1938.
of regression when it was not couched in
the familiar terms of the height of fathers (Received November 8, 1972)