You are on page 1of 6

Is It Outlier Deletion or Is It Sample Truncation?

Notes on Science and Sexuality


Author(s): Guillermina Jasso
Source: American Sociological Review, Vol. 51, No. 5, (Oct., 1986), pp. 738-742
Published by: American Sociological Association
Stable URL: http://www.jstor.org/stable/2095497
Accessed: 07/06/2008 11:55
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=asa.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We enable the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.

http://www.jstor.org

AMERICAN SOCIOLOGICALREVIEW

738
IS IT OUTLIERDELETIONOR IS IT SAMPLE
TRUNCATION?NOTES ON SCIENCEAND
SEXUALITY*
(Reply to Kahn and Udry, ASR, October, 1986)
GUILLERMINAJASSO

Universityof Minnesota

I argue that the estimates of the effects of spouses'


ages, marital duration, and contemporaneous
period influences on martial coital frequency
reported in Jasso (1985) are superior to the
outlier-deleting estimates reported in Kahn and
Udry (1986)-for chiefly one reason: If the
outlying observations in question are indeed
erroneous, as claimed by Kahn and Udry, then
(assuming correct specification) both the (fullsample) Kahn and Udry (1986) and the Jasso
(1985) estimates are unbiased estimates of the
underlyingparameters;in contrast, if the outlying
observations are correct, then only the Jasso
estimates are unbiased, the Kahn and Udry
estimates now reflecting biases associated with
truncation of the dependent variable. The Kahn
and Udry estimates on partitions of the sample,
even ignoring the truncation bias, cannot be
evaluated since the equations estimated do not
represent the process proposed in their verbal
argument (interactionswith marital duration) but
instead somethingquite different(interactionswith
date of marriage), for which no rationale is
offered.

numerical results which appear, variously, "intuitively appealing" or "appealingly counterintuitive" or, conversely, may, like Kahn and Udry, be
"troubled by . . . substantive findings." Relatively objective criteriasuch as those embodied in
the classical properties-of-estimatorsliteratureoperate to guard against too heavy a reliance on
subjectivejudgments.
Questions and Definitions

Since the phenomena under consideration are


exactly linearly related and hence ordinary language may lead to unfortunateambiguities, it is
useful to review, as precisely as possible, the main
questions of interest.
1. Among intact U.S. married couples observed
in 1970 and 1975, what is the relation between
coital frequency (CF) and the passage of time?
Letting "passage of time" denote the combined
effects of all exactly linearly related relevant
factors measured on the time dimension, Jasso
(1985) provides three answers: First, the mean
coital frequency per four-week period declined
from 8.94 to 8.23 (Table 2), indicating a mean of
the couples' change in CF of -0.71 (Table 3).
Second, not all couples experienced a decline; 41
percent experienced an increase and 13 percent no
change (Table 3). Third, holding constant' the
measured time-varying factors which both Jasso
and Kahn and Udry agree constitutethe core of the
relevant set and as well all, measured and
unmeasured,time-invariant(across the observation
PRELIMINARIES
periods) factors, the linear approximationindicates
thatthe ceteris paribuseffect of the passage of time
To properly evaluate the Kahn and Udry claims,
is a decline per year of about .22 to .24 times per
and
earlier
on
which focus on statisticalprocedure
four-week period (Table 4).
findings, it is necessary (i) to establish criteriafor
2. Among intact U.S. married couples observed
judging estimatorperformance,and (ii) to establish
in 1970 and 1975, what is the relation between
basic definitions of the terms in the substantive
coital frequency and each of the componentsof the
questions, so that previous research is not
passage of time? This question is as difficult to
misinterpreted.
discuss as it is to answer, given that-notwithstanding literary attention to sleeper legends and time
Criteriafor Judging EstimatorPerformance
machines as well as physicists' theoretical and
empirical investigation of time dilation proGiven that changing the specification, the data,
cesses-it is as yet outside the realm of direct
and/or the estimation rule can lead to sharply
human experience to observe, say, a given couple
disparate results, the question arises: How does
in which the aging of both spouses is not exactly
one judge empirical procedures?It would appear
linearly related. Ordinarylanguage is ill equipped
that most social scientists subscribeto the classical
to convey the change due to wife's advancing age
statistical criteria, such as the sampling-theory
had the husband's age somehow remained concriteria of unbiasedness, consistency (if unbiasedstant. Terms such as "marital duration" and
ness fails), and efficiency. It would appear also
"aging" appearat times to be used synonymously
that the final test in scientific work is something
with what we have termedthe "passage of time";
like an intersubjectively ascertainable external
a verb's tense (e.g., Kahn and Udry's use of the
reality-notwithstanding the fact that, as individupast tense in their opening description of Jasso's
als, investigators may be drawn to certain
results) can lead to imprecise and misleading
characterizations.Moreover, as sociologists have
* Please address all correspondence to: Guillermina known since the pioneering work of Blalock
Jasso, Departmentof Sociology, University of Minne- (1966) and Duncan (1966), ordinaryleast squares
sota, Minneapolis, Minnesota 55455.
(OLS) estimates of an equationcontaininga trio of

COMMENTS
exactly linearly dependent regressors produces at
most two estimatesjointly containingthe effects of
the three elements of the trio; hence, such
estimates may be interpreted as indicating the
corresponding effects only if zero-restrictionassumptions are imposed, that is, only if it is
reasonableto argue that one of the three elements
in the trio has no effect on the dependentvariable.
Thus, due to the constraintscurrentlyimposed by
the phenomena, combined with the associated
constraintsof ordinarylanguage and of statistics,
interpretationof the results of demographicstudies
before Jasso (1985) is not always unambiguous.
Jasso's procedureexplicitly eschews zero restrictions; the estimates are obtainedby (i) specifying a
fixed-effects model so as to separate the timevarying from the time-invariant(including cohort)
factors, and (ii) using nonlineartransformationsin
orderto breakthe exact linear relationsinduced by
the fixed-effects model in the time-varyingfactors
measured along the time dimension. Therefore,
Kahn and Udry's claim that Jasso's results
"contradictall previousresearch" would appearto
be somewhat exaggerated.
Further, it is noteworthy that the prevailing
conjecture in the fields of medicine and allied
clinical specialties, as among poets, novelists, and
confessors-and observe that clinicians, poets,
novelists, and confessors attend to the behavior of
particularindividualsover long periods of time-is
exactly consistent with Jasso's estimates.' In brief,
the effects of age (pure age, not the passage of
time) on sexual responsiveness are believed to be
nonmonotonic (initially positive, subsequently
negative) for both males and females; the curves
differ, however, in both the shape and the timing
of the maximum, the latterbelieved to occur rather
early in the male (before age twenty) and rather
late in the female (possibly as late as age forty).
Discussion and some graphic quantification are
provided in Kinsey et al. (1948, 1953), Boalt
(1978), Gebhard(1978), and Kaplan (1986). Note
that, even if this view is correct, it is not
inconsistent with a process in which female
physical decline, though similar to the male's, is
masked by diminution of sexual inhibitions
(Gebhard 1978) and/or increased self-awareness
(Kaplan 1986) and contraceptivecompetence; nor
is it inconsistent with different processes in
different societies and/orhistorical periods.
With respect to the effects of marital duration,
Jasso's results are indeed negative and strongly
statistically significant and thus, contraryto Kahn
1 Serious issues would be raised for the sociology of
science were it the case that individuals who choose
medicine as a vocation and individuals who choose
demographyas a vocation differ systematically in their
basic views of human nature.

739
and Udry's reading, do not contradictthe previous
work.
ESTIMATIONPROCEDUREI: OUTLIERS
Outliers and Influentials:Diagnostics and
Treatment
What is an outlier? How can one tell whether an
outlier (or, for that matter, any other observation)
is influential?and what does one do about it? The
problem arises because ordinary least squares,
being a solution which minimizes the sum of
squared residuals, thereby allows observations
with large residuals to be relatively more influential in the estimate of the slope. It is importantto
note that an outlier-which Beckman and Cook
(1983, p. 121) and Barnett(1983, p. 150) describe
as an observation "that appears surprising or
discrepantto the investigator"2-is not necessarily
influential and as well that inliers may be
influential. Contraryto Kahn and Udry, the degree
of influence depends on the residualand not on the
value of a single measuredvariable. Detecting the
presence of influentials is not as easy as noticing
outliers, in part because "influence" has.meaning
only relative to a model, in part because detection
is an iterative process, each perturbationpotentially generating a new set of influentials, and in
part because the available formal diagnostics,
numerical(of which Hocking (1983) lists nine) as
well as graphical, are sensitive to a variety of
assumptions (e.g., about the number of outliers
present in the data), appear to target different
aspects of the data, and are subject to "masking"
and "swamping" processes (whereby some
noninfluentialsare falsely identified as influentials
and some influentials are falsely absolved). (See
Belsley et al. 1980; Cook and Weisberg 1982;
Beckman and Cook 1983; Barnett 1983; Hocking
1983; Weisberg 1983.)
Even if diagnostics unambiguosly indicate the
presence of influential, the appropriatecourse of
action is far from clear; alternatives include
retention, downweighting, and, as Barnett (1983,.
p. 150) puts it, "the extreme resort of rejection."
Given that, as Beckman and Cook (1983, p. 145)
point out, "Little informationis available on the
performanceof the usual least squares estimators
in combination with rejection via formal tests,"
statisticians' misgivings about deletion (e.g.,
Freedman et al. (1978); Beckman and Cook
(1983); Prescott (1983); Hocking (1983); Zellner
(1984)) would appearreasonable.3Indeed, Belsley
2
The subjectivityof outlier identificationand rejection
is highlighted in Collett and Lewis (1976), who
document both between- and within-individualvariation
in judgments concerning outliers and investigate the
determinantsof the decision that a particularobservation
is an outlier.
3 See also Bollen and Jackman(1985, p. 538).

740
et al. (1980, 16) caution: "It should be obvious
that an influential point is legitimately deleted
altogetheronly if, once identified, it can be shown
to be uncorrectablyin error. Often no action is
warranted,and when it is, the appropriateaction is
usually more subtle than simple deletion."
OutlierDeletion: Induced Sample Truncation?
Considernow the Kahn and Udry procedure.They
begin by assertingcertitudethat the four rightmost
values of CF are the result of keypunchingerrors
and, based on the results from two diagnostics,
proceed to obtain two sets of estimates, labeled (3)
and (4) in their table, in which they eliminate from
the sample, first, those four cases, and, second, an
additionalfour observations.4To learn what effect,
if any, these two procedureshave on the properties
of the obtained estimates, we write the estimating
equation in two equivalent ways:

AMERICAN SOCIOLOGICALREVIEW
(of which valuable exposition is found in Maddala
(1983)), arises as follows: When the sample
contains only cases in which the dependent
variable is confined to a specified segment of its
range, there arises a correlationbetween the error
term and the regressors. In the presence of such
correlation, OLS estimates are biased. Therefore,
if the outlying observationsare correct, then all of
Kahn and Udry's specifications yield biased
estimates of the underlyingparameters.
OutlierRetention:Errorsin Variables?

The foregoing makes clear that a high price-loss


of unbiasedness-is paid for deleting cases which
are in fact correctly measured if such deletion
results in truncationof the dependentvariable. We
must now ask what price is paid by retaining
erroneousmeasures. If the four cases reportingan
"88" on CF in fact representkeypunchingerrors,
as Kahn and Udry believe, how are Jasso's
(L.a) CFit = 1oi + Y-PkXkit+ Eit (i= 1,,,N;
estimates affected? The literature on errors in
t = 1,T)
variablesdemonstratesthat the presence of random
(L.b) (CFi2-CFi1) = l43k(Xki2-Xkil) + (Ei2-Eil)
measurement
error6in the dependentvariable does
(i= 1,,,N).
not affect the unbiasedness of the obtained
Equation (L.a) representsthe differential-intercept estimates (Kmenta 1971). Ordinaryleast squaresis
model, each couple possessing its own unique designed, as it were, to withstand measurement
intercept Poi Equation (L.b) is written in the errorin the dependentvariable. Thus, inclusion of
first-differences version applicable in the case erroneously-measured values of the dependent
where the number of time periods T is equal to variable, even if it induces influential, does not
two.
bias the parameterestimates.
Now notice two things. First, Kahn and Udry's
Therefore, if the outlying observations are
specifications (3) and (4), by discarding the four erroneous, then both the Jasso and the Kahn and
largest observations of CF, have the effect of Udry estimates are unbiased, while if the outlying
truncating the sample at a right limit of the observations are correct, then only the Jasso
dependent variable in equation (L.a). Second, estimates are unbiased.
Kahn and Udry's specifications also truncate the
The argumentfor outlier inclusion based on the
dependent variable in (L.b), since the four cases propertiesof the obtained estimates would appear
with CF= 88 correspond to the four leftmost to be decisive. Yet, to borrow Mr. Jefferson's
values of first-differencedCF (less than or equal to words, a decent respect for the opinions of fellow
-78).5
scholars dictates some elaborationof why, unlike
Thus, it would appearthat Kahn and Udry, by Kahn and Udry, I have been willing to believe that
eliminating all cases where the dependentvariable the "88" measures may be correct. First, it is not
exhibits a value greater(or lesser) than a specified true that, as Kahn and Udry state, "88" was the
constant, have artificially, as it were, generateda highest value observed in 1970; in the complete
new problem-sample truncationbias. This situa- NFS sample of which the panel is a part, of the
tion, a special case of limited dependentvariables 5981 women who were asked the CF item, two are
coded "98," which according to the codebook
4 Since the estimation procedureaffects the precision denotes "98 or more times" (and twelve cases are
coded "88"). Second, the data indicate a tendency
of the parameterestimates and since readersmay wish to
for responses to the CF question ("In the past four
know the quantitativeestimate of the precision in orderto
guage statistical discernibilityor to choose one-tailed or
weeks, how many times have you had intertwo-tailed tests, it would have been helpful for Kahn and course?") to be multiples of four (Trussell and
Udry to report either the standarderrors of the t-ratios Westoff 1980). Third, it is widely believed that
(instead of the four-category significance variable, there are quite large individualdifferences in coital
"asterisk"). Absent this information, it is difficult to
frequency (see, especially, Kinsey et al. 1948,
evaluate Kahn and Udry's claims that their procedures
"wipe out . . . significance."
5
Additionally, if, as their discussion implies, specification (4) deletes the four observations with the next
largest (in absolute value) first differences, then the other
tail of first-differencedCF is truncatedas well.

6 Keypunching errors are approximatelyregarded as


random.The social science enterprisewould be in serious
difficulty if one did otherwise.

COMMENTS
1953). Fourth, it is widely believed that there are
also quite large culturallregionaldifferences (Ford
and Beach 1951);7 a probability sample from a
nation of immigrants will thus exhibit this
additional source of variation. Fifth, a figure of
"88" represents a mean weekly figure of 22,
which, in turn, representsthreedaily plus a weekly
lagniappe-well within the bounds of the Aranda
regimen and perhaps not warranting explosion
from the overworld.8
ESTIMATIONPROCEDUREII: SAMPLE
PARTITION
Kahn and Udry propose that the determinationof
coital frequencydiffers systematicallybetween two
subsets of the population, those couples married
less thantwo years and those couples marriedmore
than two years. Space does not permit analysis of
their rationale-of why marital duration (rather
than, say, an exogeneous variable or several
variables)was selected as the criterionvariablefor
the existence of different parameterregimes or of
why they assumed discontinuous regimes or of
why they fixed the numberof regimes at two or of
why they pre-selected the cutoff point or of why
thatpre-selectedpoint was set at two (the statistical
issues and formal tests associated with such
decisions are reviewed in Judge et al.'s (1985)
discussion of varying parametermodels). InsteadI
show that their rationale would lead to equations
differentfrom the ones they estimate, that, in fact,
their specifications (5) and (6) embody a rather
different view one for which they provide no
rationale.
To see this, note that Kahn and Udry's view that
the process of CF determinationdiffers between
couples married less than two years and couples
married more than two years would lead to
partitioning the sample by marital duration; one
subset would contain only observations in which
maritaldurationis less than two, the other subset
only observations in which marital duration
exceeds two. But what they do is very different.
Presumablydue to unwillingness to relinquishthe
differential-interceptmodel (for good reason, in
my view), they partitionthe sample not by marital
duration but rather by date of marriage. Both
subsets contain observations of marital duration
larger than two; they differ only with respect to
date of marriage. Thus, the specifications they
estimate are based on the implicit argumentthatCF
determinationdiffers systematically between two
groups of couples, those who marriedbefore 1968
7 For example, the normal regimen of the Keraki is
once a week, while that of the younger Arandais three to
five times nightly (Ford and Beach 1951).
8 This regimen sets the stage for Graham Green's
classic seduction in his short story, "Chagrin in Three
Parts."

741
and those who married after 1968. I will not
presume to make such an argument. However,
observe that (ignoring truncationbias) interpretation of the estimates obtained from Kahn and
Udry's specifications (5) and (6), given that no
attempt was made to learn the regime-defining
variables and cutoff points empirically, hinges on
the reasonablenessof the assumptionthat couples
who chose to marry before 1968 differ in
fundamental ways from couples who chose to
marryafter 1968.
ESTIMATIONPROCEDUREIII: THE PERIOD
EFFECT
There is a straightforwardway to check whether
the period effects reportedin Jasso (1985) embody
a month effect: in equation (L.a) replace the
decimal-year date-of-observationvariable with an
integer-yearvariable; or in equation (L.b) replace
the (variable) time between interviews by the
(constant)numberof years between waves.9 Early
versions of Jasso (1985) had integer-year in the
specifications;its effect in the equationscorresponding to the six published specifications ranges from
-0.23 to -0.22 in the equations with the single
passage-of-time factor and from -0.72 to -0.70
in the equations with the decomposition-compared to estimates of -0.24 to -0.22 and -0.75
to -0.72, respectively, obtained with the more
refined coding. This almost perfect stability
indicates that Jasso's estimated period effects are
not capturing a month effect but rather the
contemporaneousinfluences associated with each
of the two survey periods approximatelyfive years
apart.
REFERENCES
Barnett, Vic. 1983. "Discussion of Beckman and
Cook." Technometrics25:150-52.
Beckman, R.J., and R.D. Cook. 1983. "Outlier......
.
s." Technometrics25:119-49.
Belsley, David A., Edwin Kuh, and Roy E. Welsch.
1980. Regression Diagnostics: Identifying Influential
Data and Sources of Collinearity. New York: Wiley.
Boalt, GunnarH.R. "Family and Marriage." 1978. Pp.
155-61 in The New Encyclopedia Britannica,
Macropaedia 7. Fifteenth edition. Chicago: Britannica.
Blalock, Hubert M., Jr. 1966. "The Identification
Problem and Theory Building: the Case of Status
Inconsistency."AmericanSociologicalReview31:52-61.
Bollen, Kenneth A., and Robert W. Jackman. 1985.
"Regression Diagnostics: An ExpositoryTreatmentof

9 An equivalent procedure is to estimate the firstdifferencesequationomitting the period term and without
suppressing the constant. The estimated intercept then
representsthe time trend;it is equal to the productof the
coefficient of the integer-yearvariableand a factor equal
to the numberof years between waves.

742
Outliers and InfluentialCases." Sociological Methods
and Research 13:510-42.
Collett, D., and T. Lewis. 1976. "The SubjectiveNature
of Outlier Rejection Procedures." Applied Statistics
25:228-237.
Cook, R. Dennis, and Sanford Weisberg. 1982.
Residuals and Influence in Regression. New York:
Chapmanand Hall.
Duncan, Otis Dudley. 1966. "Methodological Issues in
the Analysis of Social Mobility." Pp. 90-95 in Social
Structure and Mobility in Economic Development,
edited by Neil J. Smelser and Seymour M. Lipset.
Chicago: Aldine.
Ford, Clellan S., and FrankA. Beach. 1951. Patterns of
Sexual Behavior. New York: Harper.
Freedman, David, Robert Pisani, and Roger Purves.
1978. Statistics. New York: Norton.
Gebhard, Paul H. 1978. "Sexual Behavior, Human."
Pp. 593-601 in The New Encyclopedia Britannica,
Macropaedia 16. Fifteenth edition. Chicago: Britannica.
Greene, Graham. 1978. "Chagrin in Three Parts." Pp.
45-52 in May We Borrow YourHusband? and Other
Comediesof the Sexual Life. New York: Penguin. First
published in 1967.
Hocking, Ronald R. 1983. "Developments in Linear
RegressionMethodology: 1959-1982." Technometrics
25:219-30.
Jasso, Guillermina. 1985. "MaritalCoital Frequencyand
the Passage of Time: Estimating the SeparateEffects
of Spouses' Ages and Martial Duration, Birth and
MarriageCohorts, and Period Influences." American
Sociological Review 50:224-41.
Judge, George G., W.E. Griffiths, R. Carter Hill,

AMERICAN SOCIOLOGICALREVIEW
Helmut Lutkepohl, and Tsoung-ChaoLee. 1985. The
Theoryand Practice of Econometrics. Second Edition.
New York: Wiley.
Kahn, Joan R., and J. Richard Udry. 1986. "Marital
Coital Frequency:Unnoticed Outliers and Unspecified
InteractionsLead to ErroneousConclusions (Comment
on Jasso, ASR, April 1985)." American Sociological
Review 51:000-000.
Kaplan, Helen S. 1986. "Sexual Relationshipsin Middle
Age: ComparativePhysiologic Changesin Women and
Men." The Journal of Clinical Practice in Sexuality
2:21-28.
Kinsey, Alfred C., Wardell B. Pomeroy, and Clyde E.
Martin. 1948. Sexual Behavior in the Human Male.
Philadelphia:Saunders.
Kinsey, Alfred C., Wardell B. Pomeroy, Clyde E.
Martin, and Paul H. Gebhard. 1953. Sexual Behavior
in the Human Female. Philadelphia:Saunders.
Kmenta, Jan. 1971. Elements of Econometrics. New
York: Macmillan.
Maddala, G. S. 1983. Limited-Dependentand Qualitative Variablesin Econometrics.Cambridge:Cambridge
University Press.
Prescott, Phillip. 1983. "Discussion." Technometrics
25:156-57.
Trussell, James, and CharlesF. Westoff. 1980. "Contraceptive Practice and Trends in Coital Frequency."
Family Planning Perspectives 12:246-49.
Weisberg, Sanford. 1983. "Some Principles for Regression Diagnostics and Influence Analysis (Discussion of
Hocking)." Technometrics25:240-44.
Zellner, Arnold. 1984. Basic Issues in Econometrics.
Chicago: University of Chicago Press.

You might also like