ALFRED P. SLOAN SCHOOL OF MANAGEMENT
William M, Evan and James R. Miller III
Sloan School of Management
Massachusetts Institute of Technology
2oC~^^
MASSACHUSETTS
INSTITUTE OF TECHNOLOGY
50 MEMORIAL DRIVE
CAMBRIDGE, MASSACHUSETTS 02139
Preliminary Draft
October, 1965
DIFFERENTIAL EFFECTS ON RESPONSE BIAS OF COMPUTER VS. CONVENTIONAL
ADMINISTRATION OF A SOCIAL SCIENCE QUESTIONNAIRE:
•\ A LABORATORY EXPERIMENT
William M, Evan and James R. Miller III
Sloan School of Management
Massachusetts Institute of Technology
JloC-^^
APR 18 1^67
(\/l. I. T. LlDi\MUIt-S
DIFFERENTIAL EFFECTS ON RESPONSE BIAS OP COMPUTER VS. CONVENTIONAL
ADMINISTRATION OF A SOCIAL SCIENCE QUESTIONNAIRE:
A LABORATORY EXPERIMENT
Response bias Is a subtle phenomenon which is difficult to de-
tect and even more difficult to control. It consists of a tendency
on the part of respondents, whether in the role of an interviewee, a
testee, or an experimental subject in social science research situations,
to give responses which involve either conscious or unconscious distor-
tion of data. The more common term, "response set," refers to specific
types of biasing behavior and, hence, is a special case of response
bias.
Social scientists have encountered and have sought to control this
phenomenon in applying diverse methodologies. Psychologists have strug-
gled with response bias in constructing and administering psychological
tests. Sociologists engaged in survey research have also attended to
this phenomenon, as have social psychologists who conduct laboratory
experiments. It has also received dramatic attention in an appendix,
"How to Cheat on Personality Tests," in William Whyte's Organization
Man . Whatever the form of response bias, we run the risk of gathering
dftta whose validity is impaired unless such bias can be detected and
controlled.
The purpose of this paper is to report the findings of a laboratory
experiment on the differential biasing effects of filling out a social
-2-
science questionnaire by conventional, pencil and paper procedures
versus a new method of typing answers directly onto a teletype con-
sole attached to a real-time computer,
DIMENSIONS OF RESPONSE BIAS
In two pioneering articles, Cronbach identified various "response
sets." By response sets, Cronbach means "any tendency causing a per-
son consistently to give different responses to test items than he
would when the same content is presented in a different form." Among
the response sets he identified are the tendency to gamble, i.e., to
respond when doubtful; the acquiescence set or the tendency to respond
"true" or "agree" to true-and-false questions and agree-or-disagree
questions; and the speed versus accuracy set, namely, an inclination
to respond rapidly versus carefully in order to achieve a high score.
He summarized extensive evidence demonstrating that such response sets
as bias In favor of a particular alternative, tendency to guess, and
working for speed rather than accuracy operate in many psychological
tests. Another response set, social desirability, or the tendency
to answer in a manner to conform to social norms or to give a favorable
Impression of oneself to others, has been studied at length by Edwards.
Social scientists concerned with survey research methodology have
observed the effect of the auspices under which a questionnaire is
administered. If the questionnaire is administered by a particular
member of an organization, e.g., by a supervisor or by some other re-
-3-
presentative of management, responses are likely to be more favorable
about the organization than if the same questionnaire is administered
by a non-partisan or by a non-member of the organization, such as a
professor. In other words, the social situation surrounding the ad-
ministration of a questionnaire may be conducive to bias on the part
of respondents. If the respondent is doubtful about the confidential-
ity of the data he is asked to provide, or if he is doubtful about the
use to which the information will be put, he may be inclined to bias
5
the results in order to protect his interests in the social situation.
Obviously, if the tendency to bias responses is conscious, as in the
case of Whyte's advice to cheat on personality tests, or even if it
is unconscious, in either case, the validity of the data is reduced.
Social psychologists conducting laboratory experiments have en-
countered the tendency of experimental subjects to cooperate or help
the experimenter confirm his hypotheses. This disposition to do the
experimenter's bidding, as Orne discovered, introduces a biasing ef-
fect which can distort the results of an experiment. Sociologists
and anthropologists engaged in field research have observed a similar
tendency on the part of respondents. As Vidich and Bensman put it:
"The respondent attempts to form an image of the
interviewer and the organization he represents and
to form a basis of response with respect to the inter-
viewer,.,. As a result of this image, respondents
provide information which will enable the researcher
to solve the particular problem.""^
Analogous to the findings of Orne and Vidich and Bensman is a
study by Jones of the "courtesy bias" encountered in surveys in
-4-
Southeast Asia. For a variety of cultural reasons, social scientists
have encountered the courtesy bias, or the wish to please. In conduc-
ting Interviews and in administering questionnaires. Jones considers
various questions concerning the design of research instruments that
counteract the courtesy bias. For example, projective drawings have
been used to overcome the disinclination to express negative evalua-
tions. Also, phrases designed to establish the acceptability of im-
polite responses have been used as introductory remarks. An example
would be, "Many people around here tell me that..,," In addition to
concern with questionnaire design, Jones indicates the Importance of
reassuring anonymity.
"Preventing the undesirable Influences of courtesy upon
survey responses Involves strict attention to other as-
pects of survey work beyond questionnaire design. The
standard reassurances of anonymity must be enforced
throughout the Interview, since it can be assumed that
the respondent's willingness to be 'impolite' will be
in direct proportion to the belief that this impolite-
ness will never be reported in relation to him personal-
ly. Interviewers are instructed to Interrupt respondents
who may wish to give their names; and in some Instances
of the use of administered questionnaires to special
audience groups, the phrase, 'You are not allowed to
sign your name' has been inscribed at the top of the
questionnaire,""
Cronbach and others have recommended a variety of procedures to
combat response sets. For example, the use of a multiple-choice test
instrument helps to overcome the acquiescence set observed in true-
and-false or agree-and-dlsagree questions. As Cronbach puts it,
"response sets are reduced by any procedure that Increases the struc-
turation of the test situation."
-5-
Another method of structuring the test situation which thus far
has not been Investigated by social scientists as a means of reducing
response bias Is the use of a computer In administering a psychological
test or a survey questionnaire. If such biases as carelessness, lack
of candor, and disposition to respond In a socially desirable way for
fear of negative evaluation and/or other forms of reprisal operate In
psychological tests and survey questionnaires. In part because the
interview or the research situation Is not confidential or is not per-
ceived as confidential and anonymous, then any procedure that conveys
a feeling of privacy could very well reduce response bias. It was this
consideration that led us to design our experiment. Our over-all pur-
pose in conducting this experiment was to determine whether or not the
administration of questionnaires by means of a computer would elicit
more honest and more candid responses than when the same questions were
answered in the conventional, pencil and paper manner. Our major hypo-
theses were as follows:
1) Whenever the content of a question Is regarded by the subject
as highly personal and possibly disturbing, he will respond with
greater honesty and candor under computer administration as op-
posed to conventional administration of the questionnaire. This
hypothesis is based on the assumption that subjects will regard
typing answers directly into a computer as a situation guaran-
teeing them greater privacy, greater anonymity, greater confi-
dentiality, and greater freedom from negative evaluations than
the conventional situation wherein test papers must be scrutinized.
-b-
scored, and Interpreted directly by other human beinps.
2) When an emotionally neutral and impersonal question is asked»
no significant differences in response tendencies attributable
to mode of questionnaire administration will appear.
PROCEDURE
Questionnaire Construction
Our first task was to select an appropriate set of questions to
ask. We purposely set out to construct a questionnaire whose contents
would be regarded by subjects as highly personal and which would suc-
ceed in arousing some anxiety and reluctance on their part to respond
openly and honestly. Only in this manner could we expect to detect
whatever differences in response tendencies might exist between the
two modes of administration. In addition, we wished to draw as much
as possible from the body of existing psychological test instruments
so as to benefit from the experience of previous researchers in ques-
tionnaire construction. This led us to select the following items for
inclusion in our questionnaire:
1) The entire Allport-Vernon-Lindzey (AVL) test of individual
values (45 multiple-rank items);
2) Questions from the Minnesota Multiphasic Personality Inven-
tory (MMPI) relating to manifest anxiety (20 true-false items);
3) Questions from the MMPI Lie Scale designed to detect dishonest
responses (15 true-false items);
-7-
h) Questions constituting the Srole Scale of perceived socio-
cultural anomie (5 multiple-choice items).
Fifteen true-false questions were then formulated by the authors
to determine an individual's felt need for privacy (see Appendix I),
The purpose of Including these questions was to find out whether dif-
ferences in response bias tendencies might be correlated with dif-
ferent levels of the need for privacy.
In addition, it was decided to introduce a degree of experimental
control into the questionnaire itself, i.e., certain neutral questions
were formulated which were relatively impersonal in their content and
V'/hich were not expected to arouse much anxiety or reluctance to answer
on the part of subjects. It was presumed that non systematic differ-
ences between modes of administering the questionnaire would be de-
tected on these neutral question items. If differences appeared in
the previously listed items where we expected such differences, but
failed to appear on these neutral items where we expected no differ-
ences, this outcome would lend substantial credibility to our original
hypotheses. Neutral items included in our questionnaire were:
1) Factual questions constructed by the authors which were
considered relatively easy to answer (10 true-false items, see
Appendix II).
2) Questions constructed by the authors to elicit relatively
impersonal opinions (10 true-false items, see Appendix III).
-8-
The entire questionnaire contained 120 items — 45 items from the
AVL test, 35 Items from the MMPI, 35 items constructed by the authors,
aiid 5 items constituting the Srole Scale. The ^5 items from the AVL
test were presented first in their standard order, then the 35 MMPI
Items and the 35 items constructed by the authors were presented in
12
a scrambled order, and the questionnaire ended with the 5 Srole items.
Sample Selection and Experimental Design
Sample selection was guided by two over-all considerations. First,
v;e wished to collect as homogeneous a sample as possible in order to
minimize extraneous variation in differences between modes of admini-
stering the questionnaire. Second, for ease of administration, we
v.ished to draw our sample from an available population. We, therefore,
decided to recruit M.I.T. undergraduates and to limit the experiment
to seniors whose fathers were business or professional men and whose
own career objectives were similarly directed.
Notices were sent to each of the undergraduate fraternities re-
questing volunteers for the experiment. Volunteer forms were distri-
buted to the various fraternities to gather background information on
the subjects and to aid in scheduling the experiment. On the basis of
volunteer forms returned, a total sample of 60 subjects was chosen, all
of whom satisfied the above requirements. The next task was to split
this total sample into two sub-samples — one to be given the experimental
treatment or computer administration of our questionnaire, the other
to be given the control treatment or standard, pencil and paper ad-
-9-
ministratlon. However, this raised a methodological problem.
It was feared at the outset of the experiment that, even if our
hypothesis about the bias-reducing effect of computer administration
were confirmed, the real significance of such a finding would be miti-
gated as the general public becomes increasingly more knowledgeable
of and experienced with computers. Increased knowledge and experience
would probably reduce credibility in the computer's claim to preserve
anonymity and confidentiality. Consequently, we considered a two-
factor experiment with computer knowledge and experience as the second
factor. However, in order to assign subjects in advance to the experi-
mental and control treatments, respectively, and yet still insure a
reasonable balance between these two groups in terms of computer know-
ledge and experience, it would have been necessary to ascertain the
degree of each subject's familiarity with computers prior to the ex-
periment. This, we feared, might bias subjects' responses by indicating
to them what the experiment was really about. Therefore, we abandoned
the idea of a two-factor experiment and relied upon a modified random
assignment procedure to balance the two treatment samples with respect
to computer knowledge and experience.
The modified random assignment procedure was carried out as follows
First, each subject's major field of concentration was ascertained from
his volunteer form. Then, subjects were divided into two groups accor-
ding to whether thie major was electrical engineering or something else.
This separation was made on the assumption that electrical engineering
majorswould probably possess a relatively high degree of computer know-
-10-
ledge and experience as compared to others In the sample. Electrical
engineering; majors were then divided randomly, but evenly, between the
two treatments, and the rest of the subjects were assigned at random.
This procedure generated. an experimental and a control group, each con-
taining 30 subjects.
Questionnaire Administration
Administration of the questionnaire was carried out in three steps.
First, both experimental and control subjects were given pre-experimental
instructions, For the control subjects, this was a simple statement
that they would be asked to fill out a questionnaire with pencil and
paper and that their answers would be treated anonymously and confiden-
tially. Without further ado, control subjects were then directed to a
private office to complete their task.
For the experimental subjects, a somewhat more elaborate procedure
was required. They were told that they would be asked to answer some
questions, and they were assured of the same anonymity and confidential-
ity as were the control subjects. However, experimental subjects were
told that their answers were not to be written down on the questionnaire
Itself. Instead, their answers were to be typed directly into a com-
puter by means of an on-line teletype console. It was further stressed
that the printed record of each experimental subject's responses to the
various questions could be removed from the teletype console at the end
of the experiment and taken home by the subject. This statement was
-11-
designed to give experimental subjects additional assurance of anony-
mity and confidentiality. Next, instructions were given to each sub-
ject concerning the proper use of the teletype console. A short prac- .
tice program was written in FORTRAN to give experimental subjects
experience with the console, to test their understanding of the instruc-
tions, and to point out any errors in usage.
The second step was carried out by the subjects themselves. Con-
trol subjects were left alone in a private office to complete their
questionnaires at their own speed. They were requested to inform the
experimenter when they were finished and to give him their written re-
sults at that time. Experimental subjects were also left alone in a
private office with a teletype console. They were requested to type
answers directly into the console as instructed by the computer program
and to inform the experimenter when they were finished,
A third step consisted of a post-experimental questionnaire ad-
ministered uniformly to both experimental and control subjects. The
post-experimental questionnaire (see Appendix IV) requested every
subject to review each of the preceding questions and to Indicate:
1) Whether or not he had found that question disturbing in any
way;
2) Whether or not he had been reluctant to give a candid answer.
In addition, questions were asked to ascertain:
1) How persuaded each subject was that his answers would in fact
be anonymous and confidential;
2) How serious each subject had been in expressing his opinions;
3) How many semester-length courses each subject had taken which
-12-
were directly concerned with digital computer technology;
4) How extensive each subject would rate his knowledge of
computers;
5) How extensive each subject would rate his first-hand experi-
ence with computers.
After completing the post-experimental questionnaire, subjects were
given an explanation of the rationale of the experiment, thanked for
their cooperation, and paid for their efforts.
RESULTS
Specific Predictions
Experimental measures relevant to testing our primary hypotheses
were of two types. First, there were personal questions included within
the questionnaire for the express purpose of inducing anxiety and re-
luctance on the part of all subjects to answer honestly. We anticipated
that these questions would in fact be perceived by everyone as personal
and at least somewhat disturbing, but we anticipated that experimental
subjects, feeling assured of greater privacy, anonymity, and confiden-
tiality, would answer these questions with greater honesty and candor
than would control subjects.
Neutral questions included within the questionnaire constituted a
second set of relevant measures. Few subjects were expected to per-
ceive these items as either personal or disturbing, nor was much re-
luctance to respond with candor anticipated. Consequently, very little
-13-
response bias was anticipated, and, therefore, no differences In re-
sponse bias between experimental and control subjects were predicted
on these neutral items »
Listed below for easy comparison with actual results are the
specific predictions made concerning these two types of measures prior
to our experiment
1) Experimental subjects would have a higher average score than
control subjects on the religious value orientation or the AVL
testo In the social environment of a college undergraduate (par-
ticularly at an engineering school), religious values and senti-
ments are probably attenuated » Consequently, the more honest and
candid response would show up as a greater willingness to admit
a positive orientation toward religious values o Uo predictions
were made concerning AVL results on other than the religious value
orientation,
2) Similarly, experimental subjects would admit symptoms of
manifest anxiety more freely than would control subJectSo This
would show up as a higher proportion of favorable responses for
experimental subjects on the Manifest Anxiety Scale of the MMPI.
3) Experimental subjects would register a smaller proportion of
dishonest responses than control subjects on the Lie Scale of the
MMPI, This constitutes our most direct test of the primary hypo-
thesis regarding differential response bias.
-14-
k) Experimental subjects would agree more readily than control
subjects to socially undesirable statements Indicating socio-
cultural anomie. This would manifest itself as a higher agreement
score on the Srole Scale „
5) There would be no significant differences between experimental
and control subjects on either the factual questions or the im-
personal opinion questions purposefully introduced into the ques-
tionnaire as a neutral comparison base.
Test of Predictions
Our findings on the religious value score of the AVL test agreed with
our first prediction^ Experimental subjects achieved a mean score of
3^»87, while control subjects registered a mean score of 27<.50, The
difference of 7,31 between the two group means resulted In a t-value
of 3.51, which is statistically significant at the ,0005 level (one-
tail) o
Results relating to our other four predictions are summarized in
Table lo The statistics reported are the proportion of responses
Insert Table 1 about here
favorable to the indicated scale given by the experimental group (first
column), the proportion of responses favorable to the same scale given
by the control group (second column), the difference in proportions
(third column), and the statistical significance of the indicated dif-
ference (last column)
-15-
Inspection of Table 1 shows that all but one of the
predictions were confirmed o The experimental subjects admitted more
symptoms of manifest anxiety, told fewer lies, and agreed more readily
with the Srole statements of anomie^ — all as predicted „ Non-systematic
differences appeared between experimental and control subjects on the
Factual and Impersonal Opinion Scales — also as predicted. However,
a relatively large difference did appear on the Impersonal Opinion
Scale which turned out to be significant at the ^011 level. This was
definitely not predicted, nor could it be interpreted in the framework
of our original hypotheses, We shall revert to this result later when
we discuss the impact of the acquiescence set,
A more sensitive test was designed to distinguish the cases where
differences were predicted from the cases where no differences were pre-
dictedo The Lie Scale was first reversed (i,e,, to indicate number of
truths told rather than number of lies told), and then pooled with the
Manifest Anxiety and Srole Scales, A weighted average proportion of
favorable responses on this pooled sample was computed both for ex-
perimental and for control subjects o Similarly, the Factual and Im-
personal Opinion Scales were pooled, and a weighted average proportion
of favorable responses was computed both for experimental and for
control subjects. These results appear in Table 2,
Insert Table 2 about here
-16-
Inspection of Table 2 affords clear confirmation of our hypo-
thesized resultSo Where directional predictions were made (ioe,, on
the Manifest Anxiety, Lie, and Srole Scales), a small but highly
significant weighted average difference emerged between the two pro-
portions, and this difference was in the predicted direction. In
contrast, where no differences were predicted (i.e., on the Factual
and Impersonal Opinion Scales), a much smaller and insignificant dif-
ference emerged.
Although highly encouraging, the above results constitute only a
partial confirmation of our hypotheses. Merely to ^how that
predicted results actually occurred is not enough. It must also be
shown that the predicted results occurred for the reasons stated in
the original hypotheseSr^ot just by accident or for some other reasons.
In the case of our hypotheses, this means demonstrating that:
1) Questions contained in the religious portion of the AVL test,
the Manifest Anxiety Scale, the Lie Scale, and the Srole Scale
really were regarded by the entire group of subjects as more
personal and disturbing than questions contained in the rest of
the AVL test, the Factual Scale, and the Impersonal Opinion Scale.
2) Experimental subjects really were more thoroughly persuaded
that their responses would be treated as anonymous and confiden-
tial than were control subjects.
Responses to some of the post-experimental questions were used to test
whether or not each of the above two conditions did in fact hold during
-17-
the experimental situation.
The reader will recall that all subjects were asked to review the
entire questionnaire after completion of the experiment and to indicate
which of the component questions had disturbed them and which ones they
were reluctant to answer honestly. As anticipated, the religious por-
tion of the AVL test, the Manifest Anxiety Scale, the Lie Scale, and
the Srole Scale were regarded as far more disturbing than the rest of
the AVL test, the Factual Scale, and the Impersonal Opinion Scale,
Whereas the former set of personal questions succeeded in disturbing
subjects 584 out of 2200 instances, or 26.6 percent of the time, the
latter set of neutral questions were regarded as disturbing only 3^3
out of 2400 instances, or 14,3 percent of the time. This difference
was significant at well beyond the .001 level and served to demonstrate
the first condition listed above.
Regarding the second condition, experimental subjects appeared
to have been slightly more persuaded than controls that their answers
would be treated anonymously and confidentially. On the persuasion
question in the post-experimental questionnaire, where individual
scores could range from one (meaning very much persuaded) to seven
(meaning not at all persuaded), the mean score registered by experi-
mental subjects was 2.655, while the mean score registered by control
subjects was 2.933. The difference between these two means was in
the predicted direction. However, it was neither large nor very
significant. A t-test indicated a t value of .69> significant at the
-18-
.247 level (one-tall).
One further measure of the validity of subjects' responses was
provided by the seriousness question in the post-experimental ques-
tionnaire. On a scale similar to the one associated with the per-
suasion question, the entire group of subjects registered an average
score of 1,609, suggesting that our subjects were quite serious in
expressing their opinions. This result appeared to hold equally well
for both experimental and control subjects. Their group means dif-
fered by only ,225 points.
Having achieved reasonable confirmation of our original hypotheses,
we then proceeded to test for a second-order effect. This involved
the privacy questions included in the experimental questionnaire. We
reasoned that, if a feeling of greater privacy, anonymity, and con-
fidentiality induces greater honesty and candor, then perhaps the
extent to which this effect occurs might be directly related to an
individual's felt need for privacy. That is, the greater an indivi-
dual's felt need for privacy, the greater might be the effect of
giving him additional privacy upon inducing more honest and candid
responses .
To test this idea, a rank-order correlation was performed between
each individual's total score on the Privacy Scale and his total
scores on the Manifest Anxiety Scale, the Lie Scale, and the Srole
Scale, respectively. The results were not confirmatory. Kendall
tau correlation coefficients ranged in magnitude from -.169 to +,130,
-19-
and associated significance levels ranged from .115 to .58O. In other
words, an essentially random pattern of correlations emerged. Hence,
we set aside the notion of this second-order effect.
Impact of the Acquiescence Set
Let us now return to the anomalous significant difference on
the Impersonal Opinion Scale reported in Table 1, As stated pre-
viously, no significant difference was predicted on this scale; yet
one occurred. This raised the possibility of some Joint effect brought
about by the conditions of the experimental treatment in conjunction
with the acquiescence set. Specifically, we reasoned as follows. A
response was scored as favorable to the Impersonal Opinion Scale if
the subject agreed with the stated opinion and indicated this by
checking "true." Otherwise, if he checked "false," that response
was scored as unfavorable to the scale. Consequently, it would be
impossible to separate an acquiescence response (i.e., blind accep-
tance without regard to the content of the opinion) from a genuine
agreement response on any of the impersonal opinion questions.
If computer administration of the questionnaire should happen to
induce or intensify the acquiescence set, then a significantly larger
number of agreements by experimental subjects than by control subjects
on the Impersonal Opinion Scale might very well indicate a differential
tendency to acquiesce rather than genuine differences in opinion be-
tween the two groups. If true, this would serve to confuse our other
findings, since the Lie Scale, the Srole Scale, and all but five items
-20-
on the Manifest Anxiety Scale were similarly scored. I.e., checking
"true" was uniformly scored as favorable to the scale. For this
reason. It became quite important to investigate the possible con-
nection between acquiescence and the computer administration of the
questionnaire .
The first step in isolating the possible effects of the acquiescence
set was to find a way of distinguishing a blind acceptance response from
a genuine agreement. This could easily be done on the Factual Scale,
since, in constructing that scale, factually correct answers had been
randomized with respect to the true-false categories. Therefore, a
direct test of the acquiescence phenomenon and, more particularly, of
the connection between acquiescence and the computer administration of
the questionnaire could be performed as follows.
1) Focus attention upon responses to questions contained in the
Factual Scale.
2) Identify those items in the Factual Scale where the factually
correct response is to disagree with the statement or to check it
"false."
3) Cross-classify responses to th6se items first according to
whether they were given by an experimental or by a control sub-
ject, and then according to whether they were checked "true" or
"false."
^) If there were any differential tendency for computer admini-
stration to induce acquiescence, this should manifest itself as
a greater frequency on the part of experimental subjects to check
-21-
"true," even though this is a factually Incorrect response.
5) If no such connection exists between acquiescence and com-
puter administration, then experimental and control subjects
should check "true" with approximately the same frequency.
Table 3 contains a cross-tabulation of the above information.
Insert Table 3 about here.
Inspection of Table 3 indicates that experimental subjects checked
"true" slightly more frequently than did control subjects, even though
this was the wrong answer. However, a Chi Square analysis performed
on Table 3 resulted in a Chi Square value of .379, which was signifi-
cant at only the .269 level (one-tall). A binomial test performed on
the top-row frequencies of Table 3 yielded an almost identical one-
tail significance level of .279. These results did not give adequate
support to the existence of a tendency for computer administration to
Induce acquiescence.
One additional test was performed using five items contained in
the Manifest Anxiety Scale. Whereas all of the Lie Scale and Srole
Scale Items, as well as fifteen out of the twenty Manifest Anxiety
items, were constructed so that a "true" response was scored as
favorable to the scale, five of the Manifest Anxiety items were scored
in reverse, viz., a response checked as "false" was scored as favorable
to the scale. This provided the basis for an additional check on the
possible connection between computer administration of the question-
naire and acquiescence.
-22-
If the results reported in Table 2 were really the result of ac-
quiescence. Induced by computer administration, then the weighted
average difference of .0^1 between experimental and control subjects
should at least disappear and perhaps even reverse sign when attention
is focused exclusively upon the five reverse-order Manifest Anxiety
items. But actual analysis of these five items showed a weighted
average difference of .033. The difference and its sign were pre-
served, although at a slightly reduced magnitude. The significance
of this difference was ,246 (one-tail).
In light of the above three results, it did not seem reasonable to
attribute the anomalous difference on the Impersonai -Opinion Scale to
an acquiescence set. More than likely, it arose from some unknown and
uncontrolled differences between the experimental and control groups
which were unrelated to the experiment proper. Alternatively, it may
have been due to pure chance.
Impact of Computer Knowledge and Experience
One final issue relates to computer knowledge and experience.
Analysis of the post-experimental questionnaire indicated that our
attempt to achieve an approximate balance between treatment groups
on this factor had been unsuccessful. We reviewed answers given to
the three post-experimental questions concerning a) self-rated com-
puter knowledge; b) self-rated computer experience; and c) number of
semester-length courses related to computer technology. A Joint index
-23-
was constructed by computing a weighted average of each subject's
three numerical answers. Weights were chosen for each question in
proportion to the mean score (averaged over all subjects) on that
question so as to make all three questions contribute about equally
to the Joint index. Fifty-nine of the sixty subjects were ranked on
this index (one experimental subject neglected to answer any of the
questions), and subjects were classified as "high" if they fell above
the median and "low" if they fell at or below the median. This classi-
fication was then cross-tabulated with the two treatment groups.
In Table ^ , vie observe a highly uneven distribution of knowledge
and experience among the two treatment groups. A relatively large
Insert Table 4 about here.
number of subjects with high knowledge and experience fell into the
experimental group, while many of those possessing less knowledge and
experience were assigned to the control group, A Chi Square test was
performed on these frequencies which resulted in a Chi Square value of
2.0'4, significant at the ,15'! level (two-tail). A more sensitive
statistical test, the Mann-Whitney test, was performed directly upon
the two samples of index values. This test indicated that such an
uneven distribution among the experimental and control gorups could
only have occurred by chance alone with probability .023 (two-tail).
On the basis of these somewhat surprising results, we suspected
that the index constructed out of the three post-experimental questions
-2i<-
might have been influenced excessively by the recent experimental ex-
perience itself. In other words, we suspected that experimental sub-
jects, fresh from working first-hand with a computer, might rate them-
selves as excessively knowledgeable and experienced. If true, this
would have accounted, at least in part, for the unbalanced frequencies
shown in Table ^. However, this did not seem to be the case Judging
from a cross-tabulation of the number of semester-length computer
courses by treatment groups (see Table 5).
Insert Table 5 about here.
An analysis of the frequencies in Table 5 yielded a Chi Square
value of 4.32, which is significant at the .038 level (two-tail). A
Mann-Whitney test performed directly upon the numerical responses given
by the two groups yielded an analogous difference significant at the
.047 level (two-tail). The biasing effect of recent experience with
the computer would seem far less likely to generate these types of
results, since the question asked here was a factual one rather than
a request for the subject to rate himself. In light of this, it was
concluded that our attempt to create a balanced experimental design
had been genuinely unsuccessful.
Fortunately, however, the kind of imbalance created, in conjunc-
tion with the previously reported results of the experiment, turned
out to be a blessing in disguise. Recall that our original purpose
in trying to balance the samples with respect to computer knowledge
and experience was to control for the mitigating influence of this
-25-
factor upon the predicted effects. Yet the predicted effects were
observed, even though the experimental group demonstrated a signifi-
cantly greater degree of knowledge and experience. This afforded
greater rather than less confirmation of our original hypotheses.
SUMMARY AND CONCLUSIONS
An experiment was performed to determine the bias-reducing effect
of administering a social science questionnaire by means of a computer.
It was assumed that subjects would regard typing answers directly
llv^O a computer as a situation guaranteeing them greater privacy
and a greater sense of anonymity and confidentiality than the conven-
tional situation wherein questionnaires or psychological tests are
filled out with pencil and paper and must then be scrutinized, scored,
and interpreted directly by other human beings. Based upon this as-
sumption, two specific hypotheses were formulated. First, whenver the
content of a question is regarded by a subject as highly personal and
possibly disturbing, he will respond with greater honesty and candor
under computer administration as opposed to conventional administration
of the questionnaire. Second, when an impersonal and emotionally neu-
tral question Is asked, no such difference in response tendency will
occur.
To test the above pair of hypotheses, a questionnaire was formu-
lated containing two types of questions. Some of the questions were
specifically designed to be highly personal and disturbing. Others
-26-
were designed to be Impersonal and emotionally neutral. Measures of
honesty and candor were constructed out of potential responses to these
two types of questions. Specific predictions were then made In terms
of these measures concerning differences In the honesty and candor with
which subjects would respond depending upon whether the questionnaire
was administered by a computer or In the conventional, pencil and paper
manner «
A sample of sixty undergraduate engineers was drawn from the
senior class of the Massachusetts Institute of Technology. This sample
was divided randomly, but evenly, into an experimental group, to re-
ceive computer administration of the above questionnaire, and a control
group, to receive conventional administration of the same questionnaire.
Results of the experiment confirmed both of the original hypo-
theses. A small, but statistically significant difference between the
manner in which experimental and control groups answered the personal
questions did emerge. Furthermore, this difference was in the predic-
ted direction (i.e., the experimental group displayed greater honesty
and candor in their responses). In contrast, no systematic differ-
ence emerged between the experimental and control groups on the im-
personal and emotionally neutral questions.
Additional analyses were performed on three distinct fronts.
First, an attempt was made to verify that the experimental situation
had in fact succeeded in creating the differential preconditions to
which observed differences in response bias had been attributed.
-27-
Second, an attempt was made to rule out the possibility that these same
results might have been generated by an altogether different phenomenon-
the operation of an acquiescence set. Third, elaborations and Implica-
tions of the original hypotheses were sought. On the basis of these
additional analyses, it was concluded that:
1) The personal questions included in the questionnaire for
the purpose of disturbing subjects and Inducing reluctance on
their part to answer honestly had in fact been regarded by all
subjects as distinctly more disturbing than the impersonal ques-
tions designed to be emotionally neutral.
2) Experimental subjects indicated that they were slightly more
persuaded than control subjects that their answers would be
treated anonymously and confidentially.
3) There was little evidence to support attribution of observed
differences in response bias to the operation of an acquiescence
set .
^) There was little evidence to support a possible connection
between an individual's felt need for privacy and his tendency
to bias responses to personal questions.
5) Predicted differences in response bias were observed even
though the experimental subjects appeared distinctly more know-
ledgeable of and experienced with computers.
-28-
Although the experiment reported herein succeeded In verifying
our two hypotheses concerning response bias, the results obtained can-
not be regarded as highly definitive. Recall how small most of the
reported differences turned out to be. The fact that they achieved
a fair degree of statistical significance merely indicates that a
phenomenon of the type we hypothesized does exist. However, mere
statistical significance does not necessarily indicate that this
phenomenon is either large in magnitude or important in its implica-
tions. Further research is required to answer both of these more
Interesting questions.
Concerning the possibilities for further research, at least
three are noteworthy. First, even though there was little evidence
to support the operation of an acquiescence set in our experiment, a
future questionnaire might well be designed in such a way as to pro-
vide specific control of this factor in advance. Substitution of
multiple-choice questions for true-false questions would be one way
to accomplish this end. Randomization of the scoring of any remaining
true-false questions would help achieve the same results.
Second, subjects should be used as their own controls
in a test-retest experimental design. That is, instead of assigning
two separate groups of subjects to the experimental and control treat-
ments, respectively, and then comparing their responses, we would
suggest the following procedure? ^
1) Select a sample of subjects.
2) Divide the sample randomly, but evenly, into an experimental
and control group.
-29-
3) After performing the experiment, allow sufficient time to
elapse so that subjects will forget the answers they gave on
the questionnaire.
4) Repeat the experiment, but with the former experimental sub-
jects now playing the role of controls and vice versa.
Thi? procedure would provide greater control over extraneous between-
treatment variations in observed differences and, therefore, permit
more sensitive tests of the original hypotheses.
Third, and most important of all, we would suggest transforming
the experiment from a single factor to a two-factor design. The first
factor would be "the extent to which questions are regarded as per-
sonal, disturbing, and emotionally charged." The second factor would
be "the extent to which respondents perceive themselves as protected
from any ill effects which might result from giving honest and candid
answers." The phenomenon to be studied would once again by a form of
response bias such as lack of honesty and candor and the role of com-
puters in reducing bias. By restructuring the problem in this manner,
it would be possible to observe the individual effects of both factors
at various levels, as well as any Interaction effects which might exist
between the two. In particular, effects of the first factor (degree
of personalness ) might be investigated at much higher levels of in-
tensity. Subordinates might be asked to evaluate the competence of
their supervisors with and without their supervisors' knowledge. This
might be particularly effective in generating large differences if
supervisor-subordinate pairs were drawn from large, formal organizations
-30-
such as the military.
Finally, if subsequent research confirms the findings of this
exploratory experiment which, to our knowledge, is the first of its
kind, it could have a general implication for social science research.
Whenever questions of a personal and salient nature are asked in a
sample survey, in a psychological test, or in a laboratory experiment —
and this is almost invariably the case in social science research — it
would be important to consider the use of computers as a data-gathering
Instrument, Thus, for example, social scientists engaged in research
on formal organizations as well as members of personnel departments
conducting employee surveys might profit from using computers for this
purpose.
Another area of potential application, if the findings of this
experiment are substantiated, is in public opinion surveys. Unlike
market research surveys, which are concerned with attitudes and be-
havior patterns that are not very personal in nature, public opinion
surveys often involve delicate and controversial matters. In line
with the hypotheses and findings of this study, we would expect more
response bias in the case of public opinion surveys than in market
research surveys. Hence, the use of computers in administering public
opinion surveys — possibly with the help of a special mobile data-phone set-
might counteract response bias.
The use of computers in social science research has widely dif-
fused in recent years. Thus far, however, it has been employed almost
-31-
exclusively as a data-processing instrument. A new use of computers
has been recently explored to provide immediate feedback to respon-
17
dents of the results of survey research. However, if our findings
are confirmed in further experiments, we will have uncovered, in
effect, a new function for computers, namely, as a data-gathering
instrument for social science research.
,-^
,-v
^^
^-^
,-v
0) *
r-^
r-\
rH
,-i
rH
U in
•H
•H
•H
•H
•H
C 0)
K)
cd
rt
cd
a
bO
nJ o
-P
-p
■P
+i
•p
c
o C
1
1
1
1
1
••H
•H 4-1 0)
rH
r-i
r-H
CM
r\j
0) 4J
^ O ?H
— '
■ — '
■*-^
—
V '
^ (d
•H Q)
3 o
C <«
a\
in
on
^
r-i
w ^
U) ««
rH
t—
cr>
rH
r-i
(0 73
■H -H
o
o
CM
r-i
O
(1) C
CO Q
•
•
•
•
•
to m
one i:
(0
to
0) c
-p d)
o o
U J=i
C -H
oo
rH
o
U3
r—
o
Q) +J
Ln
OO
(M
OO
t>-
0.73
fn C (<
o
o
O
O
o
U rH
0) -H O
•
•
•
*
•
3 3
<;h a
1
1
C O
«M o
S
■H t^
>
0)
nc C
t>-
o
o
OO
OO
4J -P
E
rH -H o a
■^
rH
>-i
cr\
in
x: "-i
a
Xi 4-> tn 3
•
•
•
•
•
o «
0)
CO tn -P O
•H C
u
t. O C fn
II
II
II
II
II
■^<
H
o a o o
s
> O O
CO o
rHlO
colo
^ o
LO O
4->
>s
a tn
vr> o
OO s\
i-i \S\
f-O
t^ O
o CO
£1
Et. Cu
cmVo
\=r
r-i
f\J c^
rH OO
•H 0)
to -H
^H cd
3
• rH
(US
to
w rt
+i
(U
« C -P
o
cr\
O
t—
O
O 0)
cc
(1) o c a
o
.=r
^
t~-
\D
cd x:
rH -H 3
LP*
rH
rH
CO
\D
U 4J
(U
£1 4-> E O
•
*
*
•
•
cd
rH
rt tn -H t,
x: o
05
t. o tn a
II
II
II
II
II
O -P
O
o a 0)
CO
> o a
olo
t— |o
rH lO
ooo
CO o
(U =
aj t. X
olo
KD LTV
C\J LA
VD O
CJs O
x: (u
P- (X, U
ro|^
l=r
rH
c\j oo
r-i OO
-P rH
•
£1
£1
> cd
rt
•H Cm
Eh
M
cm
s
s
73
M
M
K
0)
rH
Indicat
ponse "
a>
•H
•H
,
•
o
-p
0)
cd P
-*
CO
O
a
a
c
o
i
^
(J
<
•H
• X
0)
4->
M
•H
N^
C'--
= (!)£:
-p
(U
Di
73
•H M
(U rH cd
o
•H
S
0)
O.M
rH a
c
0)
X
s
«
l-l
O M
£1 E -P
4->
r-i
c
0)
cd
Cd cd to
o
a
<
rH
o
rH X
tn X 0)
o
o
rH
o
C 73
> -H
(0
rt
CO
rH
O G
Cd fn C
0)
!
P-i
£)
Cm
w
O
c
o
W
•H
+J
-P
rH
W
3
0)
CO
Zl
0)
CJf
en
X3
OJ
rH
O
O
cu
•
r\j
a>
rH
£1
rt
&^
^^
^-^
V '»
rH
rH
o *
■H
•H
C
o C
1
1
■H ti 3
rt t. +j o
H
II
u o c u
o a o o
in
o
C7\|0
> o o
in
o
^ O
rt t.
VD
C\J
^Vo
fc c^
^
iH
• E 3
rt tn -H O
II
II
t. o t. t^
O D. o a
o
o
vO O
a s- X
t^
r\j
^ko
fc 0-. u
r-\
"O
0)
1
-p
OS
o
0)
•H
3
13
cr
a>
t.
rH
a
Cd
C
a>
o
o
to
c
^
0)
0)
u
a
d)
Cm
(;-(
Im
o
•H
X)
CO
4J
(d
rH
3
x:
CO
-p
(U
•H
U
S
-o
CO
0)
C
rH
o
o
•H
o
+J
Oa
ta
=
4J -P
(U
X. -H
O X
■H C
x: <
^
-p
o CO
•H
4-> CUi
CO vH
•H C
U (d
+j <-<
XI
Cm Cd
o u
o
0) >
> cd
•H Cm
■P ~
Cd
O (1)
•H CO
-o c
c o
•H Q.
CO
CO .
•
Cd -p
-^
Q) cd (U
r^
E -H
• X
0}
= (U c
•p
0) rH td
o
rH Q.
c
X) E -P
•p
(d Cd to
o
^ X (1)
o
O (U Cm
Cm
> -H
Cd fn C
(U
fc O Cd
(U
s fc E
CO
Ml
41
*
Table 3 » Responses on the Factual Scale Where the Factually
Correct Answer Was to Check "False"
Number of Responses
Checked "True"
Number of Responses
Checked "False"
Experimental
Group
Control
Group
15
11
135
139
Table ^ . Computer Knowledge and Experience by Treatment
Degree of Knowledge and Experience
Experimental Group
Control Group
High
Low
16
13
10
20
Table 3 . Number of Semester-Length Computer
Courses by Treatment
None
At Least One
Experimental Group
Control Group
6
23
15
15
APPENDIX I
Privacy Scale
1. I would not like to take a Job which required sharing an office
with somebody else,
2. I often find it helpful to take a walk by myself so I can think
without interruptions.
3. I don't mind telling other people about my salary.
4. I think everybody in the United States should be fingerprinted.
5. When I am in a crowd, I like to lose myself,
6. I enjoy getting telephone calls from survey organizations asking
me whether I am watching a particular TV show or listening to a
particular radio station,
7. I worry about all the information employers and the government
have about people in this country,
8. I don't mind getting unsolicited mail inviting me to Join book
clubs or buy insurance or things like that,
I keep all my correspondence under lock and key.
I feel more confident knowing that I am talking on a telephone
that has no extensions.
I get very upset when someone opens my mail.
I don't tell my friends anything about my family.
In selecting and promoting executives, I think an employer has
the right to consider the personality of each perspective can-
didate's wife.
The Internal Revenue Service should have unrestricted access to
any and all information pertaining to business transactions
among citizens.
Administrative officers in universities should have the right
to ask professors for personal information about their students
when handling disciplinary matters.
APPENDIX II
Factual Scale
1. There are seven continents on earth.
2. The earth is closer to Mars than to the moon.
3. The number of senators from each state is uniform, but the
number of members of the House of Representatives depends
upon the population of a state.
^. Doctors are not in short supply in rural areas.
5. The American buffalo is becoming extinct in the Western states,
6. The printing press is over a thousand years old.
7. There were originally 13 states in the United States.
8. Michelangelo was a contemporary of Julius Caesar.
9. The French Revolution preceded the American Revolution.
10, Head hunting is still practiced in some parts of the world.
APPENDIX III
Impersonal Opinion Scale
1. I don't like shopping In supermarkets.
2. I don't like Sunday newspapers that take all day to read.
3. I sometimes find myself puzzled by some works of modern art
^. I think book reviewers are sometimes too brutal In their
criticisms ,
5. I think grade schools should teach children at least one
foreign language.
6. I think people should be patriotic.
7. I think too many people In America watch baseball games.
8. I think people should be honest in reporting their Income
to the government.
9. I think more effort should be made to cut down the number
of automobile accidents,
10. I think scientific progress will benefit all mankind.
APPENDIX IV
Post-Experimental Questionnaire
For each of the 120 questions you have Just answered, please check
your answers to each of the following pair of questions.
Did you find this Were you reluctant to
question disturbing give a candid answer
Question number In any way? to this question?
[120 spaces provided for Yes-No responses]
How persuaded are you that the answers you have given on the ques-
tionnaire will, in fact, by anonymous and confidential?
1234567
very much don't not at all
persuaded know persuaded
How serious do you think you were in expressing your opinions on the
items in the questionnaire?
1 2 3 '^ 5 6 7
very somewhat not at all
serious serious serious
How many semester-length courses have you taken which were directly
concerned with digital computer technology?
number of courses
How would you rate your degree of knowledge concerning computers?
■ 1 2 3 4 5 6 7
none moderate very
at all extensive
How much first-hand experience have you had using computers?
1234567
none moderate verv
at all extensive
FOOTNOTES
* Work reported herein was supported In part by Project MAC, an M.I.T,
research program sponsored by the Advanced Research Projects Agency,
Department of Defense, under Office of Naval Research Contract Number
Nonr-4102(01) . Reproduction In whole or In part Is permitted for any
purpose of the United States Government, The authors are Indebted to
the Sloan fund of the Sloan School of Management of M.I.T. for a grant
which made this study possible. They also wish to express their grati-
tude to John D, Roach for his valuable assistance in running the ex-
periment and to Jenne Brltell for Indispenslble research assistance.
1. William H, Whyte, Organization Man , New York: Doubleday-Anchor,
1956, pp. 449-i456. See also Marvin E. Shaw, "The Effectiveness
of Whyte's Rules: 'How to Cheat on Personality Tests,'" Journal
■ of Applied Psychology , ^6 (No. 1, 1962), pp. 21-25.
2. Lee J. Cronbach, "Response Sets and Test Validity," Educational
and Psychological Measurement , 6 (Winter, 1946), pp. 475-495.
See also, William E, Droen, Jr. and Robert D. Wint, "Varieties
of Response Sets," Journal of Consulting Psychology . 22 (June,
1958), pp. 237-240.
3. Lee J. Cronbach, "Further Evidence of Response Sets and Test
Design," Educational and Psychological Measurement , 10 (Spring,
1950), pp. 3-29.
-2-
k, Allen L. Edwards, "Social Desirability and the Description of
Others," Journal of Abnormal and Social Psychology , 59 (November,
1959), pp. 434-436; Allen L. Edwards, The Social Desirability
Variable in Personality Assessment and Research , New York: The
Dryden Press, 1957.
5. Cf. Herbert G. Heneman, Jr. and Dale Yoder, "Employee Opinion Sur-
vey by Remote Control," Personnel Journal , 32 (October, 1953),
pp. 169-173; Marvin D. Dunnette and Herbert G. Heneman, Jr.,
"Influence of Scale Administration on Employee Attitude Responses,"
Journal of Applied Psychology . 40 (April, 1956), pp. 73-77;
Leonard I. Pearlin, "The Appeals of Anonymity in Questionnaire
Response," Public Opinion Quarterly . 25 (Winter, 1961), pp. 640-
647.
6. Martin Orne, "On the Social Psychology of the Psychological Experi-
ment," American Psychologist , 17 (November, 1962), pp. 776-783.
7. A. Vidich and J. Bensman, "The Validity of Field Data," Human
Organization . 13 (Spring, 1954), pp. 22-23.
8. Emily L, M. Jones, "The Courtesy Bias in Southeast Asian Surveys,"
International Social Science Journal . 15 (No. 1, 1963), PP • 70-
7 6.
9. Ibid . , pp. 72-73.
10. Cronbach, "Response Sets and Test Validity, op. cit. . p. 488. See
also, Irwin A, Berg and Gerald M. Rapaport, "Response Bias in
an Unstructured Questionnaire," Journal of Psychology , 38 (October,
1954), pp. 476ff.
-3-
11. Leo Srole, "Social Integration and Certain Corollaries: An
Exploratory Study," American Sociological Review . 21 (December,
1956), pp. 709-716
12. A few additional items were included in the final questionnaire.
However, these were not really a part of this experiment, and
they were never analyzed.
13. This practice program carried experimental subjects through the
same sequence of operations as did the main program, also written
in FORTRAN, controlling the question and answer process. Subjects
were instructed by the program to type their answers directly
into the console, and answers were stored in tape files created
by the computer. These tape files were later retrieved for pur-
poses of analysis,
l^ . In computing statlBtlcal significance, the component questions in
any given scale were not pooled in order to get a single estimate
of the underlying response probability associated with that entire
•CftXe. Rather, each component question was treated separately,
experimental and control samples were pooled to arrive at a single
estimate of the underlying response probability associated with
that question, and then results were averaged across questions.
This made for a somewhat more sensitive test of proportion dif-
ferences, since it served to control some of the between-question
variability.
-14-
L5. For a related finding on the differential responses to threatening
and non-threatening questions, see, Dunnette and Heneman, op.
cit . . pp. 75-76.
16. We had originally Intended to perforin a similar analysis on re-
sponses to the reluctance question contained in the post-experi-
mental questionnaire. Unfortunately, however, there was sub-
stantial evidence that subjects refused to answer this question
carefully, if at all. Consequently, no analyses were performed,
not even on the very few responses which appeared valid,
Cf , William M. Evan, "Swift Feedback in Survey Research," The Technology
Review . 67 (June, I965), pp. 1-3.
a/
NOVl 1967)
APR I 1968
FEB 27 '69
SEP 25 *6i
