Neuropsychology Copyright 2007 by the American Psychological Association

2007, Vol. 21, No. 5, 611– 620 0894-4105/07/$12.00 DOI: 10.1037/0894-4105.21.5.611

Using Regression Equations Built From Summary Data in the

Neuropsychological Assessment of the Individual Case
John R. Crawford Paul H. Garthwaite
University of Aberdeen The Open University

Regression equations have many useful roles in neuropsychological assessment. This article is based on
the premise that there is a large reservoir of published data that could be used to build regression
equations; these equations could then be used to test a wide variety of hypotheses concerning the
functioning of individual cases. This resource is currently underused because (a) not all neuropsycholo-
gists are aware that equations can be built with only basic summary data for a sample and (b) the
computations involved are tedious and prone to error. To overcome these barriers, the authors set out the
steps required to build regression equations from sample summary statistics and the further steps required
to compute the associated statistics for drawing inferences concerning an individual case. The authors
also develop, describe, and make available computer programs that implement the methods. Although
caveats attach to the use of the methods, these need to be balanced against pragmatic considerations and
against the alternative of either entirely ignoring a pertinent data set or using it informally to provide a
clinical “guesstimate.”

Keywords: neuropsychological assessment, regression equations, single-case methods

Regression equations serve a number of useful functions in the predict memory scores at retest from baseline scores in a sample of
neuropsychological assessment of individuals (Chelune, 2003; temporal lobe epilepsy cases who had not undergone any surgical
Crawford, 2004; Crawford & Howell, 1998; Strauss, Sherman, & intervention in the intervening period. The equation was then used
Spreen, 2006; Temkin, Heaton, Grant, & Dikmen, 1999). For to assess the effects of surgery on memory functioning in further
example, they are widely used to estimate premorbid levels of individual patients.
ability in clinical populations by using psychological tests that are Regardless of whether the equation was built from a healthy or
relatively resistant to neurological or psychiatric dysfunction clinical sample, this approach simultaneously factors in the
(Crawford, 2004; Franzen, Burgess, & Smith-Seemiller, 1997; strength of correlation between scores at test and retest (the higher
O’Carroll, 1995). the correlation the smaller the expected discrepancies), the effects
Another common application of regression is in the assessment
of practice (typically scores will be higher on retest), and regres-
of change in neuropsychological functioning in the individual case.
sion to the mean (extreme scores on initial testing will, on average,
Here a regression equation can be built (usually by using healthy
be less extreme at retest).
participants) to predict a patient’s level of performance on a
Yet another common role of regression equations is as an
cognitive ability measure at retest from their score at initial testing.
alternative to the use of conventional normative data (Heaton &
An obtained retest score that is markedly lower than the predicted
Marcotte, 2000). For example, if (as is commonly the case) per-
score suggests cognitive deterioration (Crawford, 2004; Heaton &
formance on a neuropsychological test is affected by age, then age
Marcotte, 2000; Sherman et al., 2003; Temkin et al., 1999).
can be incorporated into a regression equation to obtain an indi-
Clinical samples can also profitably be used to build regression
vidual’s predicted score on the test. This provides what Zachary
equations for predicting retest scores. For example, Chelune,
and Gorsuch (1985) have termed continuous norms and can be
Naugle, Lüders, Sedlak, and Awad (1993) built an equation to
contrasted with the discrete norms formed by creating arbitrary age
bands. With the latter approach, the apparent relative standing of
an individual can change dramatically as they move from one age
John R. Crawford, School of Psychology, College of Life Sciences and band to another.
Medicine, King’s College, University of Aberdeen, Aberdeen, United It can be seen from the foregoing that regression equations
Kingdom, and Paul H. Garthwaite, Department of Statistics, The Open perform many useful roles in the neuropsychological assessment
University, Milton Keynes, United Kingdom. of individuals. However, we suggest that the potential of regres-
We are grateful to Sytse Knypstra of the Department of Econometrics, sion equations is far from being fully realized. There is a large
University of Groningen, The Netherlands, for providing an algorithm that reservoir of published data that could be used to build regression
finds the noncentrality parameter of a noncentral t distribution given a equations; these equations could then be used to test a wide variety
quantile, its associated probability, and the degrees of freedom. This
of hypotheses concerning the neuropsychological functioning of
algorithm is incorporated in the computer programs available at http://
www.abdn.ac.uk/~psy086/dept/regbuild.htm.
individual cases.
Correspondence concerning this article should be addressed to John R. For example, there are literally hundreds of published studies
Crawford, School of Psychology, College of Life Sciences and Medicine, that have examined performance at test and retest on a wide variety
King’s College, University of Aberdeen, Aberdeen AB24 3HN, United of commonly used neuropsychological tests. Many of these studies
Kingdom. E-mail: j.crawford@abdn.ac.uk have been tabulated by McCaffrey, Duff, and Westervelt (2000)

611
612 REGRESSION EQUATIONS FROM SUMMARY DATA

and are classified by the test involved (e.g., verbal fluency, Wis- where rxy is the correlation between the predictor and criterion
consin Card Sorting Test, and so forth), the type of sample (e.g. variables and sX and sY are the standard deviations of these vari-
healthy elderly, traumatic brain injury [TBI] sample, etc.), the ables. The equation for the intercept a is
retest interval, and whether any intervention was applied in the
intervening period. a ⫽ Y៮ ⫺ bX៮ , (3)
Provided that studies such as those described above record the
means and standard deviations for the sample’s scores at test and where Y៮ is the mean score on the criterion or dependent variable
retest, and the correlation between these sets of scores, a regression and X៮ is the mean score on the predictor or independent variable.
equation can be built to predict retest scores from initial scores Having set out the steps to obtain a regression equation from
(these basic summary statistics are commonly reported). The equa- summary data, we now turn to the calculations required to draw
tion built with the summary statistics can then be used with an inferences concerning the discrepancies between an individual’s
individual to examine the discrepancy between the predicted score obtained score and the score predicted by such an equation. The
and the actual score obtained at retest. The statistics required to first step is to calculate the standard error of estimate (sY䡠X). The
draw inferences concerning this discrepancy can also be calculated standard error of estimate is a measure of the variability of obser-
from summary data alone; the only information required beyond vations about the regression line in the sample used to build the
that already noted is the size of the sample (essentially, it can be equation. The formula for the standard error of estimate when there
assumed that this will always be available). is only a single predictor variable is
The remainder of this article has two principal aims. The first is
to (a) set out the calculations required to build regression equations
from summary data and (b) set out the further calculations required
when applying these equations to draw inferences concerning an
冑
s Y䡠X ⫽ s Y 共1 ⫺ r XY
2
兲
N⫺1
N⫺2
, (4)

individual case. The second aim is to describe and make available where N is the size of the sample used to build the equation, and
computer programs that implement all the methods described. all other terms have been defined previously. Again it can be seen
Statistically minded neuropsychologists will be aware that re- that this statistic can be calculated when only summary data are
gression equations can be built by using summary data alone and available.
that the associated statistics required to apply such equations to an Having obtained sY䡠X, the next step is to calculate the standard
individual case can also be obtained without the sample raw data. error of a predicted score for a new case (denoted as sN⫹1;
However, on the basis of discussions at conferences, workshops, Crawford & Howell, 1998; Howell, 2002). When there is only one
and elsewhere, it is clear that many neuropsychologists are un- predictor variable, the required formula is
aware, or only vaguely aware, that such a possibility exists.

冑
Moreover, those neuropsychologists who know that summary
statistics are sufficient also know that, if the optimal statistical 1 共X 0 ⫺ X៮ 兲 2
s N⫹1 ⫽ s Y䡠X 1⫹ ⫹ 2 , (5)
methods are to be used, the calculations involved are time con- N s X共N ⫺ 1兲
suming, tedious, and prone to error. Therefore, there is the temp-
tation to either (a) not use regression equations in situations in where X0 is the individual’s score on the predictor, and the re-
which it would be helpful or (b) use computationally simple maining terms have been defined previously. Note that, once again,
methods that only approximate the results that would be obtained all the necessary quantities are available from summary statistics.
from the optimal methods. Alternatively, if the optimal methods The standard error of a predicted score for a new case (that is,
are used, there is the danger that clerical errors will unknowingly a case not in the sample used to build the equation) captures the
be made when carrying out the computations. The provision of a uncertainty associated with estimating a population slope from a
computer program that implements the optimal methods deals with sample. It can be seen that the sN⫹1 will increase in magnitude the
all of these problems. further the patient’s score on the predictor variable (X) is from the
mean; this is also a consequence of the uncertainty in estimating
Building a Regression Equation From Summary Data the slope. Any error in estimating ␤ will exert a more marked
effect on extreme values of X; to visualize this effect, imagine
A regression equation with one predictor variable (i.e., a biva- rotating the regression line around the mean of X (Crawford &
riate regression equation) takes the following form: Howell, 1998).
Next, consider
Ŷ ⫽ ␣ ⫹ ␤X, (1)
Y 0 ⫺ Ŷ
where Ŷ is the predicted score on the criterion variable, ␣ is the s N⫹1
, (6)
intercept (the value of Y when X takes the value of zero), ␤ is the
slope of the regression line, and X is the score on the predictor where Y0 is the individual’s obtained score and Ŷ is the individu-
variable. Let a and b denote the estimates of ␣ and ␤ that the al’s predicted score, to yield a standardized discrepancy between
complete sample data would give. Then the equation for calculat- the obtained and predicted score. Under the null hypothesis, that
ing the slope b with summary data from a sample is the discrepancy is an observation from the population sampled to
build the equation, this quantity has a t distribution on N – 2 df
sY (Crawford & Garthwaite, 2006a). Thus, for a specified level of
b ⫽ r xy , (2)
sX alpha (e.g., .05), one can test whether there is a statistically
 CRAWFORD AND GARTHWAITE 613

significant difference between the predicted score and the obtained individual or error in estimating the population regression equation
score, using either a one- or two-tailed test.1 from sample data). However, it should be borne in mind that
Opinions differ on the use of one- versus two-tailed tests in significance levels (whether they be one-, two-, or split-tailed) are
neuropsychological assessment. Our view is that one-tailed tests largely arbitrary conventions; the conclusion drawn when a pa-
are acceptable and appropriate if the neuropsychologist wishes to tient’s discrepancy is just above the significance level threshold
test a directional hypothesis, particularly when a discrepancy in the should be similar to the conclusion when it is just below that
opposite direction would not be the result of a clinical condition. threshold. Thus, we suggest that the neuropsychologist should be
When working with individual cases, power to reject the null primarily concerned with the more general issue of the degree of
hypothesis (in this context to reject the hypothesis that the patient’s abnormality of the patient’s discrepancy.
discrepancy is an observation from discrepancies in the reference Fortunately, an estimate of the abnormality of the discrepancy is
population) is often low (Crawford & Garthwaite, 2006b): Use of an inherent feature of the proposed method: The p value used to
a one-tailed test provides higher power than use of a two-tailed test significance is also a point estimate of the proportion of the
test. However, the logic of a one-tailed test is such that the
population with the same value on the predictor variable (i.e., X) as
direction of a difference must be specified a priori, and if a
the patient who would obtain a discrepancy more extreme than that
difference is in an unexpected or counterintuitive direction, then it
which was observed for the patient (Crawford & Garthwaite,
must be ignored. As such, unexpected findings may be clinically
2006a). As noted, the population referred to here is that sampled to
informative; some neuropsychologists may prefer to keep their
options open and use two-tailed tests. The programs that accom- build the regression equation; that is, if the equation was built by
pany this article (see below) provide both one- and two-tailed using healthy adults, then the population is the healthy adult
values, thereby allowing clinicians to choose which values to population. Alternatively if, for example, the equation was built in
attend to on the basis of their general preferences or the specific a sample of patients who had suffered a severe TBI 6 months
problem in hand. earlier, then the population is all patients with a severe TBI 6
Note that there is a potential “third way” with respect to this months postinjury.
issue. Many of us are used to thinking in terms of making a In Crawford and Garthwaite (2006a), the fact that the p value
dichotomous decision between a one- or two-tailed test. However, from the significance test also equals the estimated proportion of
an alternative approach is to use what has been termed a split- the population exhibiting a more extreme discrepancy than that of
tailed test (Harris, 1994). The rationale underlying the split-tailed the case was stated without proof. In the present article, we provide
test is that, having selected a Type I error rate (i.e., alpha) we a short formal proof in the Appendix. It is more convenient (and
regard as acceptable (5% by convention), we need not “spend” it more in line with convention) to multiply the p value referred to
by allocating it equally to both tails or solely to one tail. Rather, we above by 100 so that we have a point estimate of the percentage
can choose to allocate a greater proportion of it to the tail corre- (rather than proportion) of the population exhibiting a larger dis-
sponding to our directional hypothesis while still leaving a pro- crepancy. This latter index of abnormality is used in the examples
portion to cover the opposite tail. For example, we may decide to that follow and in the outputs from the computer programs that
allocate 80% to the former tail and 20% to the latter. The loss of accompany this article.
power for such a split-tailed test is only a little lower than that of
The above quantity is a point estimate of the abnormality of the
a one-tailed test (a p value less than .04, rather than less than .05,
discrepancy between an individual’s obtained and predicted score.
is required to be considered significant) but, unlike the latter test,
Recently, Crawford and Garthwaite (2006a), building on earlier
the researcher or clinician need not ignore a large difference in the
statistical results from Crawford and Garthwaite (2002), have
direction opposite to that expected, provided it is significant be-
yond the .01 level. provided a method of obtaining an interval estimate for this quan-
To our knowledge, split-tailed tests have not previously been tity. That is, the method provides 95% confidence limits on the
used or advocated for inferences concerning individual cases, but percentage of the population that would obtain a more extreme
they may be worthy of consideration. In practice they are easily discrepancy than that observed for the individual.
applied, that is, the user need simply attend to the one-tailed p The provision of these confidence limits is in keeping with the
value from the test. That is, if the difference is in the predicted contemporary emphasis in psychological assessment and statistics
direction and the one-tailed p is ⬍ .04, then the difference is on the utility of confidence limits (American Psychological Asso-
considered to be statistically significant at p ⬍ .05; if the difference ciation [APA], 2001; Wilkinson & APA Task Force on Statistical
is in the direction opposite to that predicted and the one-tailed p Inference, 1999). Confidence limits serve the useful general pur-
is ⬍ .01, then the difference would also be considered statistically pose of reminding the user that there is always uncertainty attached
significant at p ⬍ .05. Just as is the case with use of a one-tailed to an individual’s results (i.e., they counter any tendency to reify
test, it must be stressed that the user should take a principled the observed scores), but they also serve the specific purpose of
approach to such testing—the proportion of alpha allocated to each quantifying this uncertainty (Crawford & Garthwaite, 2002). The
tail should be decided before the data are observed and thereafter calculations involved in obtaining these limits involve noncentral
should be regarded as fixed.
Significance tests have a role to play in neuropsychological
assessment of the individual: When a discrepancy achieves statis- 1
Note that, as pointed out by a reviewer, this test can be construed in
tical significance, the neuropsychologist can be confident it is terms of the studentized deleted residual, which is equivalent to asking
unlikely to be a chance finding (in other words, it is unlikely that whether, if that individual had been included in the original sample, would
the observed discrepancy stems from random variation in an a separate dummy code parameter for the individual be necessary.
614 REGRESSION EQUATIONS FROM SUMMARY DATA

t distributions and are complex, but the important point for present basic summary statistics are available for a suitable sample. This
purposes is that they can be calculated without requiring the raises the question of whether rounding errors contained in these
sample’s raw data. summary statistics pose a threat to the validity of the results
Confidence limits on the abnormality of an individual’s discrep- obtained. The main threat to precision comes from calculating the
ancy are implemented in the computer program that accompanies slope of the regression line from the reported correlation coeffi-
this article, and an example of their use is provided in a later cient for the sample (note however that the correlation also fea-
section. For full details of the derivation and evaluation of these tures in the formula for sY䡠X and hence also influences sN⫹1). This
limits, and of the calculations required to obtain them, see Craw- is because the correlation is intrinsically more influential in the
ford and Garthwaite (2006a). calculations but also because the other statistics required (i.e.,
Recovering a Correlation From Summary Data When It Is means and standard deviations) are normally reported with three or
Not Reported more significant digits.
In contrast, the correlation coefficient is normally reported with
In the foregoing section, it was assumed that the summary data only two significant digits, for example, .72. Thus, because of
for a sample included the correlation between the predictor and rounding error, the actual sample correlation coefficient in this
criterion variables. It might be thought that if the correlation is not example could be as low as .715 and would have its upper limit at
reported then all is lost, that is, the regression equation cannot be .725. Supposing, for convenience, that the standard deviations of X
built. However, it is possible to recover this correlation if the study and Y were both 1 (so that the slope is identical to the correlation
has reported the results of a paired-samples t test to compare mean coefficient), then the slope calculated from the sample raw data
scores on the two variables. For example, in a test–retest study,
could be as low as .715 or as high as .725 rather than .72.
even if the correlation between test and retest was not reported, it
At the margins, this rounding error may make a difference to the
will often be the case that a t test was conducted to test whether
results obtained for an individual. If there were grounds for con-
there had been a significant decline or improvement over the
cern then, given that the range of error introduced by rounding is
intervening period. In this situation, the correlation can be recov-
known, the calculations could be repeated substituting the upper
ered by using simple algebra.
and lower limits on the value of the correlation coefficient. The
The formula for a paired-samples t test is
results could then be compared with those obtained originally to
X៮ ⫺ Y៮ examine whether they make a difference to the substantive con-

冑
t⫽ , (7) clusions.
s ⫹ s Y2 ⫺ 2s Xs Yr XY
2
X
Admittedly, this is not a particularly appealing prospect were the
N calculations to be done by hand or calculator, but with the com-
where all terms have previously been defined. Rearranging this puter program developed to accompany this article, it would take
formula to solve for rXY gives seconds to run this check (i.e., by substituting the upper or lower
limit on the correlation attributable to rounding error for the

r XY ⫽
s X2 ⫹ s Y2 ⫺ N 冉 冊
X៮ ⫺ Y៮
t
2

, (8)
correlation as reported in the relevant data field). Before leaving
this topic, it should be borne in mind that errors arising from a lack
2s Xs Y of precision are not limited to equations built from summary data.
Such errors can also arise when using published regression equa-
and the correlation can then be used to obtain the regression tions if the slope and standard error of estimate and so on are
equation and its associated statistics as set out in the previous reported with a limited number of significant digits.
section. Note that this formula should not be confused with for-
mulas used to convert a t value to a point biserial correlation; the
latter provide an index of effect size and represent the correlation Computer Programs Implementing These Methods
between occasion (before vs. after) or group membership (e.g., A computer program for PCs is available that implements the
control vs. clinical) and test scores. methods covered in the present article. The program (regbuild.exe)
The principal computer program that accompanies this article prompts the user for the sample means and standard deviations of
(regbuild.exe) builds the regression equation directly from the the predictor and criterion variables, the correlation between these
correlation coefficient. However, a companion program (regbuild_ variables, and the N for the sample.
t.exe) is available to build the regression equation from a t value. The output consists of the regression equation, that is, the slope
If a study has used an analysis of variance rather than a t test to b and intercept a, as well as the standard error of estimate. These
compare the two means, it is still possible to build an equation outputs are followed by the results obtained from analyzing the
either by hand or by using this latter program. F ⫽ t2, so all that individual case’s data. These consist of the following: the individ-
need be done is to take the square root of F and enter it into ual’s predicted score; the standardized discrepancy between the
Formula 8 in place of t. predicted and obtained scores; the results of the significance test
The Precision of Regression Equations and Their (one- and two-tailed probabilities are provided); the point estimate
Associated Statistics When Computed From Summary of the percentage of the population that would obtain a larger
discrepancy; and (by default) the 95% confidence limits on this
Data
percentage (alternatively, a one-sided upper or lower 95% limit
As noted, the main purpose of this article is to set out a can be requested). The results can be viewed on screen, printed, or
procedure for building and using regression equations when only saved to a file. There is also the option of entering user notes (e.g.,
 CRAWFORD AND GARTHWAITE 615

to keep a record of the source of the summary data or further Table 1

details of the sample or individual case); these are reproduced in Illustrative Examples of the Use of Summary Data From
the output from the program. Published Studies to Draw Inferences Concerning an Individual
For convenience, the summary statistics for the sample used to Patient
build the equation are saved to a file and reloaded when the
Variable Study 1 Study 2 Study 3 Study 4
program is rerun. Therefore, when the program is used with a
subsequent individual, the required data for the new case can be Predictor Age FAS SF Time 1 SF Time 1
entered, and results obtained, in a few seconds. The program has Predictor M 63.80 36.60 44.3 22.3
the option of clearing the sample data to allow the user to build a Predictor SD 8.42 12.50 12.8 11.5
new equation if required. As noted previously, a further program, SF M 41.30 43.4 48.1 16.4
SF SD 13.20 13.14 13.6 11.8
regbuild_t.exe, is designed for use when a study fails to report the Correlation (r) ⫺0.58 0.67 0.76 —
correlation between predictor and criterion variable but does report Paired t value — — — 4.79
a t value obtained from performing a paired-samples t test to Sample size (N) 160 120 45 50
compare the means of the two variables. The functionality, anal-
Slope (b) ⫺0.909 0.708 0.808 0.740
ysis, and output from this program are identical to regbuild.exe Intercept (a) 99.31 17.51 12.33 ⫺0.094
except that it also reports the estimated correlation between the sY·X 10.787 10.110 8.941 8.264
predictor and criterion variables obtained with Formula 8. A com- sN⫹1 10.821 9.872 9.367 8.346
piled version of these programs can be downloaded (as a zip file)
Note. For clarity, a blank line separates those statistics required to gen-
from the following Web address: http://www.abdn.ac.uk/⬃psy086/ erate the equations and those statistics calculated from them. Unlike the
dept/regbuild.htm other statistics for the regression equation, sN⫹1 is not a fixed quantity as
it is partly determined by the extremity of an individual’s score on the
predictor variable. Studies 1– 4 are hypothetical, and statistics for the
Examples of the Use of the Methods and Accompanying resultant regression equations were built with either the formulas in the text
Programs or accompanying computer programs. FAS ⫽ a letter fluency test; SF ⫽
semantic fluency.
In this section, we illustrate some ways in which the methods
and accompanying computer programs can harness summary data
from published studies to assist the neuropsychologist to draw elderly participants. Finally, Study 4 was another longitudinal
inferences concerning the cognitive status of individual patients. study that included the semantic fluency test but was concerned
Suppose that a neuropsychologist has seen a 65-year-old patient with cognitive change in a sample of patients with early AD.
with suspected early Alzheimer’s disease (AD) and has adminis- Suppose, as is the case for many neuropsychological instru-
tered a semantic (category) fluency test at the initial assessment ments, that the normative data for the elderly on the semantic
and again after 5 months; further suppose that an initial letter fluency test are modest. These conventional normative data could
fluency test (e.g., FAS) was also administered at the first assess- be supplemented by using the data from Study 1 to build an
ment. The patient’s score on the semantic fluency test at initial equation for prediction of individuals’ semantic fluency scores
testing was 21, and the score at retest was 11; the patient’s FAS from their age. If the predicted score is substantially higher than
score at initial testing was 25. the patient’s obtained score, this suggests impaired performance.
Table 1 sets out details of four hypothetical studies: For each This is an example of the use of regression to provide continuous
study, it lists the summary data required to build regression equa- norms (Zachary & Gorsuch, 1985) as referred to in the Introduc-
tions and to calculate the associated statistics for drawing infer- tion.
ences concerning an individual case. The resultant regression Entering the summary statistics for the sample in Study 1 (i.e.,
equations and their associated statistics, calculated by using either the means and standard deviations for age and semantic fluency
the formulas presented in the text or by using the accompanying together with the correlation between these variables) into Equa-
computer programs, are also presented in Table 1. Although the tions 2 and 3, the slope b and intercept a of the regression equation
accompanying computer programs are designed to be intuitive, the are calculated as – 0.909 and 99.31. Thus, on the basis of the
provision of the summary data in Table 1 and the worked examples patient’s age, his predicted semantic fluency score is 40.21. With
below allows users to run these examples themselves. This will Equations 4 and 5, the standard error of estimate for this equation
help users become familiar with the mechanics of the process prior (sY䡠X) is 10.787, and the standard error for an additional individual
to using the methods with other data. (sN⫹1) is 10.821. The difference between these two quantities is
Study 1 was a study conducted on a sample of healthy partici- modest in this example because the patient’s value on the predictor
pants (age range 50 to 80 years) on the effects of aging on variable (i.e., their age) is not extreme relative to the sample mean
neuropsychological test performance; in the course of this study, and also because the sample used to build the equation is large; it
the correlation between age and performance on the semantic will be appreciated that this will not always be so.
fluency test (–.58) was reported. It can be seen that age exerts a The raw discrepancy between the patient’s obtained semantic
substantial effect on performance on the semantic fluency task. fluency score of 21 and predicted score of 40.21 is –19.21. Divid-
Study 2 was also a study of cognitive aging in healthy elderly ing this discrepancy by sN⫹1 (i.e., applying Formula 6) yields a
participants and included among its results the correlation between standardized difference of –1.775. Under the null hypothesis, this
the semantic fluency test and the FAS test. Study 3 was a longi- standardized difference is distributed as t on N – 2 ⫽ 158 df (in this
tudinal (test–retest) study of neuropsychological tests, including case, the null hypothesis is that the individual’s discrepancy is an
the semantic fluency test; again the sample consisted of healthy observation from the population of discrepancies found in the
616 REGRESSION EQUATIONS FROM SUMMARY DATA

healthy elderly). Evaluating this t value reveals that the patient’s icant at the .05 level, the discrepancy is nevertheless fairly unusual
obtained score is significantly below the score predicted from her and is consistent with a differential deficit in semantic versus
or his baseline score ( p ⫽ .0389, one-tailed). initial letter fluency.
The point estimate of the abnormality of this discrepancy (i.e., Note that a good case could be made for the use of a two- rather
the percentage of the population that would be expected to exhibit than one-tailed test in this situation. That is, a case may turn out to
a discrepancy larger than that observed) is thus 3.89%. The ac- have a discrepancy favoring semantic fluency over initial letter
companying 95% confidence interval on the percentage of the fluency (a pattern that is liable to be relatively uncommon in AD).
population that would exhibit a larger discrepancy than would the Had this occurred in the present case, then the logic of hypothesis
patient ranges from 2.12% to 6.32%. Thus in summary, there is a testing would have precluded testing for the significance of this
large and statistically significant discrepancy between the patient’s difference. The two-tailed p value in this example is .152.
predicted and obtained scores. This size of discrepancy is esti- This foregoing analysis also provides a good opportunity to
discuss the weight given to null hypothesis significance tests in
mated to be unusual in the healthy elderly population and is
neuropsychology. When a difference does achieve statistical sig-
consistent with impaired performance on the semantic fluency
nificance, the neuropsychologist can be particularly confident that
task.
a problem has been uncovered. However, as noted, we suggest that
Moving on to Study 2, in neuropsychological assessment much
the principal focus should be on the abnormality of the patient’s
emphasis is placed on the use of intraindividual comparison stan- discrepancy rather than whether it falls below or above the cusp for
dards when attempting to detect acquired impairments (Crawford, conventional statistical significance. In this example, the discrep-
2004; Lezak, Howieson, Loring, Hannay, & Fischer, 2004). Com- ancy between initial and semantic fluency tests is fairly uncom-
parison of semantic and initial letter fluency performance provides mon. Thus such evidence should be given weight when arriving at
a good example of such an approach as (a) scores vary widely as a formulation, particularly when it is consistent with other indica-
a function of an individuals’ premorbid verbal ability and thus tions of impairment, such as here, where the patient’s semantic
there are limits to the usefulness of normative comparison stan- fluency score was also lower than expected given his age.
dards (Crawford, Moore, & Cameron, 1992), and (b) the two tasks Turning to Study 3: Neuropsychologists commonly have to
are highly correlated in the general adult population (Henry & attempt to detect change in cognitive functioning in the individual
Crawford, 2004). Thus, if an individual exhibits a large discrep- case, whether this be to gauge recovery following a stroke or TBI,
ancy between these two tasks, this suggests an acquired impair- to monitor the positive or negative effects of intervention, or to
ment on the more poorly performed task. detect decline in degenerative conditions. In the case of AD, serial
In this specific example, there is an additional consideration: assessment plays a particularly important role in diagnosis because
There is much evidence that semantic fluency performance is more the results of testing from a single time period will often be
severely disrupted by AD than is initial letter fluency. For exam- equivocal (Morris, 2004). When test data from two occasions are
ple, a random effects meta-analysis of a large number of studies of to be compared, regression provides a useful means of drawing
semantic and initial letter fluency in AD patients versus healthy inferences concerning change: The neuropsychologist simply
control participants revealed very large effects for semantic flu- needs to find test–retest data for the measures used in an appro-
ency coupled with more modest effects on initial letter fluency priate sample retested over an interval similar to that of the patient.
(Henry, Crawford, & Phillips, 2004). That is, the semantic fluency Study 3 is a hypothetical test–retest study in which a sample of
deficits qualified as differential deficits relative to initial letter healthy elderly participants (N ⫽ 45) were tested on the semantic
fluency. On the basis of such evidence, a discrepancy in favor of fluency test and retested after 6 months (this is a slightly longer
initial letter fluency over semantic fluency would be consistent interval than that for the case but sufficiently close to justify use of
with an Alzheimer’s process. the data).
One means of examining whether this pattern is observed in the The regression equation and associated statistics obtained from
individual case is to use a healthy sample to build an equation to Study 3 are presented in Table 1. With this equation, the patient’s
predict semantic fluency from initial letter fluency and to compare predicted semantic fluency score at retest is 29.29 (based on his
the individual’s predicted and obtained scores. The regression initial score of 21) compared with his obtained retest score of 11.
equation and associated statistics built with the hypothetical data Dividing the raw discrepancy between the obtained score and
from Study 2 are presented in Table 1. With this equation, the predicted score (–18.29) by sN⫹1 yields a standardized difference
patient’s predicted semantic fluency score is 35.23 (based on his of –1.95. Evaluating this t value reveals that the patient’s retest
initial letter fluency score of 25) compared with his obtained score score is significantly below the score predicted from his score on
of 21. Dividing the raw discrepancy between the obtained score first testing ( p ⫽ .0287, one-tailed). The point estimate of the
and predicted score (–14.23) by sN⫹1 yields a standardized differ- abnormality of this discrepancy is thus 2.87%, and the 95% con-
ence of –1.442. Evaluating this t value reveals that the patient’s fidence interval is from 0.27% to 9.88%. In conclusion, the anal-
obtained score is not significantly below the score predicted from ysis indicates that the patient’s performance on semantic fluency
his initial letter fluency score ( p ⫽ .076, one-tailed). The point has declined over the interval between the two testing occasions.
estimate of the abnormality of this discrepancy (i.e., the percentage That is, it is unlikely that a member of the cognitively intact elderly
of the population that would be expected to exhibit a discrepancy population would exhibit this large a decline in performance.
larger than that observed) is thus 7.60%, and the 95% confidence Finally, Study 4 provides an example of the use of an equation
interval is from 3.94% to 12.63%. In summary, the patient’s built in a clinical rather than healthy sample. Having found evi-
semantic fluency is considerably lower than expected given his dence of a decline with the equation built by using data from
initial letter fluency performance; although not statistically signif- Study 3, the data from Study 4 are used to examine whether the
 CRAWFORD AND GARTHWAITE 617

change from test to retest is unusual for patients with AD. Note Finally, the examples featured have primarily had a clinical
also that, in this last example, the study did not report the corre- focus. However, the use of regression equations to draw inferences
lation between test and retest scores for semantic fluency but did concerning the individual case (whether these equations are pro-
report the results form a paired-samples t test used to test for a vided by a third party or are built from summary data by using the
significant difference between mean scores on the two occasions methods set out here) has numerous applications in neuropsycho-
(see Table 1). logical research. Most obviously, there has been a massive resur-
There are much published data of this form in neuropsychology gence of interest in single-case studies in neuroscience as a means
(McCaffrey et al., 2000), and it is therefore fortunate that we can of specifying the functional architecture of cognition (Coltheart,
still build the equation in the absence of a reported correlation 2001). Although the focus of these studies is on a single case, the
between the predictor and criterion variable. By using Formula 8, patient’s pattern of performance is usually interpreted by referring
or by using the program that implements it (regbuild_t.exe), we to control values. The use of regression equations can play a useful
can recover the correlation between the scores at test and retest by role in providing such comparisons (i.e., through analysis of the
using the data provided. In this example, the correlation is .72, and discrepancy between a patient’s obtained scores and those pre-
we now have the information required to build the equation to dicted by equations built in the healthy control sample). Note that,
predict the patient’s score at retest from his initial score (see as the control samples in single-case studies are normally modest
Table 1 for the regression equation statistics). Had the study used in size (Crawford & Garthwaite, 2005), this is an example of
a two-way analysis of variance to compare means at test and retest where the exact methods set out here are clearly to be preferred
(with each participant as a block) then, as noted earlier, we can still over the approximate method.
recover the correlation. For this example, the F value from the As pointed out by a reviewer, the methods can also usefully
analysis of variance for the difference between test and retest supplement results obtained in group studies. For example, in a
would be 22.94, and taking its square root gives us the necessary clinical trial of a new drug or other form of intervention, research-
t value for entry into Formula 8 or into the program. ers are interested not only in the magnitude of the group differ-
By using the equation built with the data from Study 4, the ences between the active and placebo groups but also in the
number of individuals that show significant improvement from
patient’s predicted semantic fluency score at retest is 15.44 (based
baseline (or attenuation of decline). The methods set out here could
on his initial score of 21) compared with his obtained retest score
be used to set the criterion for improvement or attenuation in the
of 11. Dividing the raw discrepancy between the obtained score
individual case, which in turn could be used as a basis for calcu-
and predicted score (– 4.44) by sN⫹1 yields a standardized differ-
lating relative risk (via odds ratios) or “numbers needed to treat.”
ence of – 0.532. Evaluating this t value reveals that the patient’s
retest score is not significantly below the score predicted from his
score on first testing ( p ⫽ .597, two-tailed). The point estimate of Advantages of the Proposed Method of Drawing
the abnormality of this discrepancy is 29.9%, and the 95% confi- Inferences Concerning an Individual’s Discrepancy Over
dence interval is from 20.23% to 40.68%. Note that the population
Existing Methods
referred to on this occasion is the population of patients with early
AD, not the general (healthy) elderly population as in the previous As noted, regression equations are already widely used to draw
examples. Note also that in the present example a two-tailed test is inferences concerning an individual’s neuropsychological status.
used: Even if the neuropsychologist had independent grounds to Currently, the “standard” approach in neuropsychology to analyz-
believe that the patient’s cognitive decline would be atypically ing the discrepancy between an individual’s predicted and ob-
rapid for AD, or atypically slow, it is unlikely that they would have tained score is simply to divide the discrepancy by the standard
sufficient confidence in this to rule out the alternative possibility. error of estimate for the equation, treat the resulting quantity as a
In conclusion, the present analysis indicates that, although the standard normal deviate (i.e., z), and convert this quantile to a
patient has shown a decline that is somewhat greater than the probability by using a table of areas under the normal curve
average decline found in the AD sample, the indications are that (Crawford & Garthwaite, 2006a). Just as is the case with the
the difference in the patient’s performance on the two occasions approach presented here, it is possible to apply this method when
will nevertheless not be unusual in patients with early AD. only summary data are available; that is, a pre-existing equation is
The foregoing example of the use of equations built by using not required. It is therefore appropriate to briefly contrast the two
data from clinical samples is only one of many potential uses. approaches.
Indeed, given the vast number of clinical studies in the literature, In essence, the alternative method treats the sample used to build
this process is limited only by the ingenuity of the neuropsychol- the equation as if it were the entire population; that is, the regres-
ogist and by the time involved in conducting a search for published sion statistics are treated as population parameters rather than
studies relevant to the question in hand. For example, data such as sample statistics. Monte Carlo simulations have demonstrated that,
that in Study 4 could also be used to study the potential effective- in contrast to the method used in the present article, the Type I
ness of a pharmacological (or other form of) intervention in the error rate is inflated above the nominal rate and, relatedly, the
individual case. That is, in the example, the data were obtained abnormality of discrepancies between individuals’ obtained and
from untreated early AD cases and thus, if a treated early AD predicted scores is exaggerated (Crawford & Garthwaite, 2006a).
patient’s score at retest substantially exceeded that predicted by the These effects can be marked when the sample used to build the
regression equation (i.e., if the discrepancy was estimated to be equation is modest in size and when an individual’s score on the
unusual among untreated AD cases), this would be consistent with predictor variable is extreme. This latter phenomenon is a direct
a beneficial effect of the intervention. consequence of failing to allow for the uncertainty in estimating
618 REGRESSION EQUATIONS FROM SUMMARY DATA

the slope of the population regression line from a sample: As noted methods for calculating and analyzing the results of multiple
earlier, error in estimating the slope will have a greater effect on regression equations by using summary data, nor have we imple-
extreme values. mented these methods in a program.
It should be acknowledged, however, that the two methods will Finally, note that, when a multiple regression is available (i.e.,
yield very similar results when the sample used to build the from a published study or from a psychologist’s analysis of her or
regression equation has a moderately large n, particularly when the his own raw data set), then it is possible to apply methods for
individual’s score on the predictor variable is not extreme (see the analyzing the discrepancy between an individual’s obtained and
worked example for Study 1 in which sN⫹1 and sY䡠X are very predicted score that are directly analogous to those covered here
similar in magnitude). As noted by a reviewer, in such circum- for the bivariate case. These methods have been set out by Craw-
stances the sY䡠X method could be used as in place of the technically ford and Garthwaite (2006a) and have been implemented in an
correct method presented here. On the other hand, if the computer accompanying computer program (this program requires direct
programs that accompany this article are used, there is little to be entry of the intercept and the regression coefficients for each
gained by such a substitution as the correct calculations can be predictor variable).
applied in much less time than would be required for hand calcu-
lation of the approximate method.
Caveats on the Use of These Methods and Some
A further difference between the two methods is that, unlike the
method advocated here, the alternative method cannot provide
Pragmatic Considerations
confidence limits on the abnormality of an individual’s discrep- The validity of inferences made with the methods set out here is
ancy: It does not acknowledge any uncertainty in the sample’s dependent on the quality of the data used to build the equation; that
regression statistics and therefore cannot quantify the effects of is, the methods will not provide accurate results if the assumptions
this uncertainty on the estimate of the individual’s level of abnor- underlying regression analysis have been violated (see Tabachnick
mality. This can be seen as a significant disadvantage given that (a) & Fidell, 1996, for a succinct treatment of this topic). For example,
confidence limits remind us that our data are fallible and quantify one assumption underlying the use of linear regression is that of
the effects of this fallibility, and (b) many bodies, such as the APA homoscedasticity of the residuals. If the size of the residuals
(Wilkinson & APA Task Force on Statistical Inference, 1999), increases as scores on the predictor variable increase (as indicated
stress the importance of incorporating such limits into research and by a fan-like appearance on a scatterplot), then this assumption
practice. would be violated. Another assumption is that the relationship
between the predictor and criterion variable is linear.
In the case of regression equations published in peer reviewed
Multiple Regression Equations From Summary Data?
journals or in test manuals, it is probable (but not guaranteed) that
Compared with equations with a single predictor, multiple re- these threats to validity will have been identified (by examination
gression equations provide a more flexible and potentially more of residual plots and so forth) and rectified or ameliorated (e.g., by
sensitive means of testing hypotheses concerning an individual. transforming the Y variable in the case of heteroscedasticity). In
For example, when testing for change in an individual’s neuropsy- the absence of the raw data, such strategies are not possible.
chological functioning, if age is related to the magnitude of prac- These concerns need to be balanced by two pragmatic consid-
tice effects, then age can be incorporated into an equation along erations. First, with many of the combinations of predictor and
with the initial test score to obtain a more precise estimate of an criterion variables likely to be used in practice there is little
individual’s expected change. evidence that heteroscedasticity and/or nonlinearity is a pervasive
Although the computations involved are considerably more problem. For example, if the predictor and criterion variables are
complex than those involved in the bivariate case, it is possible to both standardized psychological tests (as is the case when attempt-
build multiple regression equations from summary data alone. ing to infer change from test to retest or when comparing an
However, there is a major impediment to actually carrying this out estimate of an individual’s premorbid functioning with their cur-
in practice: Published studies rarely report the full matrix of rent functioning) such problems do not appear to be very common.
correlations between the criterion and predictor variables, but this Second, in an ideal world, neuropsychologists would have ac-
matrix is required to build the equation (the means and standard cess to regression equations that had been built by using large
deviations for all variables are also required, but these are usually samples and had been carefully evaluated. However, it is clear that
available). the number of such equations is very limited in comparison with
The studies most likely to include a full correlation matrix are the wide variety of hypotheses that neuropsychologists may wish
those reporting the results of a multiple regression analysis. For to test. Therefore, in the absence of an existing equation, and when
example, the American Psychological Association recommends relevant summary data are available, the approach suggested here
that such studies report the full matrix (e.g., see APA, 2001), but needs to be contrasted with the alternatives open to the neuropsy-
of course there is no need to calculate the coefficients of the chologist. These are that the neuropsychologist will either simply
regression equation from summary data in these circumstances as ignore the existence of such data despite its relevance to the
they will already be available. assessment question or will attempt to use the data informally to
A further practical difficulty is that, even if the correlation generate a “guesstimate.” For example, in the latter case the
matrix is available, the precision with which the correlations are reasoning might proceed along the following lines: “Given that
reported would be much more of a problem in using multiple this test appears to have fairly high test–retest reliability and is
regression than it is in the bivariate case (Sokal & Rohlf, 1995). In subject to a moderate practice effect, the difference between this
view of these practical considerations, we have not set out the individual’s scores looks fairly unusual.” It is well known that our
 CRAWFORD AND GARTHWAITE 619

subjective estimates of such probabilities are not very accurate and Crawford, J. R., Moore, J. W., & Cameron, I. M. (1992). Verbal fluency:
are prone to systematic biases (Beach & Braun, 1994; Tversky & A NART-based equation for the estimation of premorbid performance.
Kahneman, 1971); for example, we typically underestimate the British Journal of Clinical Psychology, 31, 327–329.
Franzen, M. D., Burgess, E. J., & Smith-Seemiller, L. (1997). Methods of
magnitude of the differences expected by chance.
estimating premorbid functioning. Archives of Clinical Neuropsychol-
Finally, although it has been demonstrated that it is relatively ogy, 12, 711–738.
straightforward to build an equation from summary data (particu- Harris, R. J. (1994). ANOVA. An analysis of variance primer. Itasca, IL:
larly if the computer programs accompanying this article are used), Peacock.
we also hope this article will encourage researchers to consider Heaton, R. K., & Marcotte, T. D. (2000). Clinical neuropsychological tests
including regression equations in their articles. By doing so, they and assessment techniques. In F. Boller & J. Grafman (Eds.), Handbook
of neuropsychology (2nd ed., Vol. 1, pp. 27–52). Amsterdam: Elsevier.
would provide a means whereby their results can be directly
Henry, J. D., & Crawford, J. R. (2004). A meta-analytic review of verbal
applied by other neuropsychologists to draw inferences in the fluency performance following focal cortical lesions. Neuropsychol-
individual case. ogy, 18, 284 –295.
Henry, J. D., Crawford, J. R., & Phillips, L. H. (2004). Verbal fluency
performance in dementia of the Alzheimer’s type: A meta-analysis.
References Neuropsychologia, 42, 1212–1222.
Howell, D. C. (2002). Statistical methods for psychology (5th ed.). Bel-
American Psychological Association. (2001). Publication manual of the
mont, CA: Duxbury Press.
American Psychological Association (5th ed.). Washington, DC: Author.
Lezak, M. D., Howieson, D. B., Loring, D. W., Hannay, H. J., & Fischer,
Beach, L. R., & Braun, G. P. (1994). Laboratory studies of subjective
J. S. (2004). Neuropsychological assessment (4th ed.). New York: Ox-
probability: A status report. In G. Wright & P. Ayton (Eds.), Subjective ford University Press.
probability (pp. 107–128). Chichester, United Kingdom: Wiley. McCaffrey, R. J., Duff, K., & Westervelt, H. J. (2000). Practitioner’s guide
Chelune, G. J. (2003). Assessing reliable neuropsychological change. In to evaluating change with neuropsychological assessment instruments.
R. D. Franklin (Ed.), Prediction in forensic and neuropsychology: Sound New York: Kluwer.
statistical practices (pp. 123–147). Mahwah, NJ: Erlbaum. Morris, R. G. (2004). Neuropsychology of older adults. In L. H. Goldstein
Chelune, G. J., Naugle, R. I., Lüders, H., Sedlak, J., & Awad, I. A. (1993). & J. E. McNeil (Eds.), Clinical neuropsychology: A practical guide to
Individual change after epilepsy surgery: Practice effects and base rate assessment and management for clinicians (pp. 345–358). Chichester,
information. Neuropsychology, 7, 41–52. United Kingdom: Wiley.
Coltheart, M. (2001). Assumptions and methods in cognitive neuropsy- O’Carroll, R. (1995). The assessment of premorbid ability: A critical
chology. In B. Rapp (Ed.), The handbook of cognitive neuropsychology review. Neurocase, 1, 83– 89.
(pp. 3–21). Philadelphia: Psychology Press. Sherman, E. M. S., Slick, D. J., Connolly, M. B., Steinbok, P., Martin, R.,
Crawford, J. R. (2004). Psychometric foundations of neuropsychological Strauss, E., et al. (2003). Reexamining the effects of epilepsy surgery on
assessment. In L. H. Goldstein & J. E. McNeil (Eds.), Clinical neuro- IQ in children: Use of regression-based change scores. Journal of the
psychology: A practical guide to assessment and management for cli- International Neuropsychological Society, 9, 879 – 886.
nicians (pp. 121–140). Chichester, United Kingdom: Wiley. Sokal, R. R., & Rohlf, F. J. (1995). Biometry (3rd ed.). San Francisco, CA:
Crawford, J. R., & Garthwaite, P. H. (2002). Investigation of the single W. H. Freeman.
case in neuropsychology: Confidence limits on the abnormality of test Strauss, E., Sherman, E. M. S., & Spreen, O. (2006). A compendium of
scores and test score differences. Neuropsychologia, 40, 1196 –1208. neuropsychological tests: Administration, norms and commentary (3rd
Crawford, J. R., & Garthwaite, P. H. (2005). Evaluation of criteria for ed.). New York: Oxford University Press.
classical dissociations in single-case studies by Monte Carlo simulation. Tabachnick, B. G., & Fidell, L. S. (1996). Using multivariate statistics (3rd
Neuropsychology, 19, 664 – 678. ed.). New York: Harper Collins.
Crawford, J. R., & Garthwaite, P. H. (2006a). Comparing an individual’s Temkin, N. R., Heaton, R. K., Grant, I., & Dikmen, S. S. (1999). Detecting
predicted test score from a regression equation with an obtained score: significant change in neuropsychological test performance: A compari-
A significance test and point estimate of abnormality with accompanying son of four models. Journal of the International Neuropsychological
confidence limits. Neuropsychology, 20, 259 –271. Society, 5, 357–369.
Crawford, J. R., & Garthwaite, P. H. (2006b). Methods of testing for a Tversky, A., & Kahneman, D. (1971). The belief in the law of small
deficit in single case studies: Evaluation of statistical power by Monte numbers. Psychological Bulletin, 76, 105–110.
Carlo simulation. Cognitive Neuropsychology, 23, 877–904. Wilkinson, L., & American Psychological Association Task Force on
Crawford, J. R., & Howell, D. C. (1998). Regression equations in clinical Statistical Inference. (1999). Statistical methods in psychology journals:
neuropsychology: An evaluation of statistical methods for comparing Guidelines and explanations. American Psychologist, 54, 594 – 604.
predicted and obtained scores. Journal of Clinical and Experimental Zachary, R. A., & Gorsuch, R. L. (1985). Continuous norming: Implica-
Neuropsychology, 20, 755–762. tions for the WAIS-R. Journal of Clinical Psychology, 41, 86 –94.

(Appendix follows)
620 REGRESSION EQUATIONS FROM SUMMARY DATA

Appendix

A Proof That, for Crawford and Garthwaite’s (2006a) has a t distribution on n – 2 df, so that
Method, the Estimated Proportion of the PopulationA1
That Exhibits a Larger Discrepancy in the Same Direction
as a Case Equals the One-Tailed p Value of the
冉
Pr共d ⬎ d * 兲 ⫽ Pr tn⫺2 ⬎
d*
sN⫹1 冊 . (11)

Significance Test
Also, the test statistic for testing if d * is from the same normal
*
Let d be the discrepancy between the obtained score (Y0) and distribution as the control ds is
predicted score (Ŷ) for the case. Then the proportion of control
participants who have a discrepancy that is both in the same
d*
direction as the case and larger than that of the case is , (12)
s N⫹1

Pr共d ⬎ d * 兲, (9) and this is compared with a t distribution on n – 2 df. Comparison

of Equations 11 and 12 shows that Pr(d ⬎ d *) is equal to the p
value for the one-tailed test.
where d is the discrepancy of a control participant. Because the present article is concerned with the use of regres-
Dividing both quantities by sN⫹1 (defined in Equation 5), sion equations built from summary data and, as noted, this is not
practical in the vector case (i.e., when there is more than one
predictor variable), the proof is couched in terms of the bivariate

冉 冊
case. However, it is easily extended to the vector case by simply
d d* substituting n – k – 1 df for n – 2 df in the relevant equations, where
Pr共d ⬎ d * 兲 ⫽ Pr ⬎ . (10)
sN⫹1 sN⫹1 k is the number of predictor variables.

A1
Now Note that the estimated proportion of the population is the proportion
of the population with the same score on the predictor variable as the case.

Received June 26, 2006

d Revision received March 19, 2007
s N⫹1 Accepted March 27, 2007 䡲