Respondent Anxiety Reduction With the Randomized Response Technique

Donald E. Stem, Jr., Washington State University
Carl S. Bozman, Washington State University
ABSTRACT - This paper reports the results of an experiment evaluating the relative effectiveness of three randomized response model techniques in reducing respondent anxiety for questions on sensitive issues. Skin conductance was used as a measure of state anxiety. A conventional direct questioning technique was used as a control. The results clearly indicate that the Hopkins flask method and the spinner method were more effective than the telephone randomization device or direct questions in reducing respondent state anxiety due to question sensitivity.
[ to cite ]:
Donald E. Stem, Jr. and Carl S. Bozman (1988) ,"Respondent Anxiety Reduction With the Randomized Response Technique", in NA - Advances in Consumer Research Volume 15, eds. Micheal J. Houston, Provo, UT : Association for Consumer Research, Pages: 595-599.

Advances in Consumer Research Volume 15, 1988      Pages 595-599


Donald E. Stem, Jr., Washington State University

Carl S. Bozman, Washington State University


This paper reports the results of an experiment evaluating the relative effectiveness of three randomized response model techniques in reducing respondent anxiety for questions on sensitive issues. Skin conductance was used as a measure of state anxiety. A conventional direct questioning technique was used as a control. The results clearly indicate that the Hopkins flask method and the spinner method were more effective than the telephone randomization device or direct questions in reducing respondent state anxiety due to question sensitivity.


The purpose of this study is to evaluate the ability of three randomized response methods and conventional direct questioning to provide reliable data for sensitive consumer behaviors and attitudes.

Two specific error types have a unique impact on total error in consumer research on sensitive issues. Response error occurs in circumstances where data provided by respondents does not correspond to their true states of nature, behavior or attitudes. This problem has a number of causes. Subjects may answer a question without accurate personal knowledge, or they may deliberately provide misleading details in their response.

Nonresponse error, on the other hand, results when a member of the sample cannot be reached for comment, or when a respondent refuses to answer at least some part of the measurement instrument. There are three major reasons for these behaviors. Either they regard the information requested as extremely personal, or they feel revealing certain items would be embarrassing, or they attempt to guess and satisfy the perceived expectations of the experimenter (Tull and Hawkins 1984).

Both forms of error are aggravated by situations where information on sensitive issues is requested from individuals. Respondents have a tendency to provide normative responses, or refuse to respond at all, when there is a perceived risk in giving a truthful answer.

This tendency is a function of deviation of their true response from the relevant norms, the perceived probability of disclosure and the perceived consequences of disclosure. Any method which can reduce this perceived risk should logically lead to a reduction in total error for survey research on sensitive issues. One class of sensitive consumer behavior involves illegal behaviors such as shoplifting or drug usage. These methods can also be applied to consumption of products or services which are socially stigmatized. Abortion and contraceptives are good examples of this class of consumer behavior.

Randomized Response Method

In 1965 Warner proposed the randomized response technique as a me hod for reducing respondent concerns about answering sensitive questions. This device allows subjects to report their true behavior and attitudes toward sensitive issues while at the same time eliminating their risk of disclosure.

In the simplest case, an individual first observes the result of a randomization procedure with dichotomous outcomes. This process then determines whether the respondent answers the actual question or provides a fake response. The interviewer remains unaware of which reply was selected. Only the knowledge of the sample size, the number of yes or no responses, and the probabilities assigned to each condition are necessary to estimate the population parameters. The security offered to subjects is expected to enlist greater cooperation and result in estimates with reduced error variance. Lamb and Stem (1978) have demonstrated the validity of this conclusion empirically. Their study indicates that the model provides reductions in error variance beyond conventional methods in sensitive situations.

Randomized response techniques have been widely used to study certain types of consumer behavior. Consumption of illegal drugs has been the subject of a number of studies using randomized response methodology (Brown, 1975; Brown and Harding, 1973 and Dawes, 1974). Shoplifting behavior in the retail environment has been investigated (Geurts, Andrus and Reinmuth, 1975 and Reinmuth and Geurts 1975), along with the use of products and services which are stigmatized. Numerous studies have investigated the utilization of abortion services and the purchase behavior of contraceptives (Abernathy, Greenberg and Horvitz, 1970; Greenberg, Abernathy and Horvitz, 1970; Shimizu and Bonham, 1978 and Liu and Chow, 1976).

A comprehensive review of the RRM and further examples of its applications can be found in a monograph by Fox and Tracy (1986), or in a review article by Horvitz, Greenberg and Abernathy (1976).


This paper relies on a theoretical framework (Spielberger, 1972), which explains observed behavior resulting from the use of the randomized response technique. Since anxiety reduction is an implicit assumption of the randomized response model, it will be the focus of the analysis which follows.

An anxiety theory composed of two distinct constructs is frequently encountered in the psychology literature. First proposed by Spielberger (1972), this model has been incorporated in numerous studies. State anxiety is defined as an individual manifestation of temporary emotional reactions during threatening circumstances. This form of anxiety depends on personal perceptions of danger and elicits autonomic nervous system responses. Trait anxiety, on the other hand, refers to a subject's general tendency to have enduring behavioral dispositions over time. It should be possible to distinguish between underlying tendencies and arousal due to an external stimulus using this model.

A state anxiety reaction would be expected to occur when a socially, psychologically or physically dangerous stimulus is presented to an individual. The degree of physiological reaction is hypothesized to correspond directly with the perceived level of threat the situation generates. Cognitive appraisal and learned defense mechanisms will serve to mediate this process. Threatening situations may be reevaluated and/or avoided entirely. For example, a respondent can either refuse to answer a question or can provide an untrue response. In either case, a proportional change in involuntary nervous system activity would be anticipated. Individual differences in anxiety proneness may also influence the perception of a given circumstance.


State anxiety in a sensitive question situation is a function of the respondents perceptions of personal risk from a truthful response to the question. As with any perceived risk, the state anxiety is the product of the consequences of disclosure, times the probability of disclosure (Cunningham 1967).

In a given question situation total anxiety can be described as follows:

          At = At + As       (1)

and    As = S + (C x P)      (2)

or     AT = At + S + (C x P)       (3)


AT = Total anxiety for a given individual/situation.

At = Trait anxiety for a given individual respondent,

As = State anxiety for a given individual/situation,

S = The anxiety introduced by the interviewer, the instruments and other situational factors,

C = The perceived consequences of disclosure of the respondents answer,

P = The perceived probability of disclosure of the respondents answer for a given question method.

In order to evaluate the effectiveness of a randomized response method all effects except the portion of state anxiety due to the method must be controlled. Individual differences in trait anxiety and that portion of state anxiety attributable to interviewer effects and experimental situation, question sensitivity, consequences of disclosure can all be controlled by randomization or by using identical stimuli. With all other factors controlled any variation in total anxiety can be attributed to differences in the method or perceived probability of disclosure.

When confronted with a non-sensitive question, since consequences of disclosure are zero, differences in anxiety are due wholly to the questioning technique. On the other hand, when confronted with a sensitive question, anxiety is composed of the combination of anxiety due to the technique plus that due to the perceived probability of disclosure. The increase in anxiety between non-sensitive questions and sensitive questions is a measure of a questioning techniques ability to minimize perceived probability of disclosure, and therefore the effects of question sensitivity.

Skin Conductance

Electrodermal response was chosen as the physiological measure of state anxiety based on its reliable use by previous researchers (Born, Mainka and Stolting 1980; Hentschel and Ternes 1984; Knight and Borden 1979; McLeod, Hoehn-Saric and Stefan 1986).

Although there are a number of electrodermal measurements available, skin conductance has become the preferred measure of arousal (Edelberg 1972; Fowles et al. 1981; and Venables and Christie 1973) In addition, skin conductance has been found to be normally distributed (Edelberg 1972; and Schlosberg and Stanley 1953).

An E & M Instrument Company physiograph, Model DMP 44, was used to measure electrodermal activity in a manner consistent with generally accepted and recommended procedures (Fowles et al 1981). SCR is reported in Micro-mhos, a measure of electrical conductance. Two silver/silver chloride electrodes were attached to the non-dominant hand of all subjects during the questioning period. The first five questions were used to calibrate the equipment, stabilize the readings and set a base level for each respondent.

Establishing a base level standardized the dependent variable so that any individual differences in skin resistance due to trait anxiety are minimized. This acts, along with random assignment of subjects, to preclude the possibility that variation in respondent reactions to stimuli is an artifact of trait anxiety.


A convenience sample of 137 undergraduate students was used to examine the relationship between questioning technique and state anxiety. Students were selected in an effort to obtain homogeneous treatment groups and thereby provide as rigorous a theory test as possible. This type of sample was preferred because heterogeneous respondents would have exhibited greater variation in behavior. Furthermore, the chance of making an incorrect statistical inference increases when the error variance is larger. The probability of rejecting the null hypothesis due to uncontrolled factors, when it is true, was reduced by employing students as subjects (Calder, Phillips, and Tybout 1981).


Experiments were conducted on a one to one basis in a research laboratory by four trained assistants. Upon entering the lab, a subject was randomly assigned to one of four treatment conditions. Each person was then informed that the research was concerned with evaluating the effectiveness of alternative survey methods, that they could terminate the interview at any time, and that all their answers would remain confidential. Every respondent, except one person, agreed at this point to continue the interview. This particular individual indicated that he did not have the time required to participate in the study. The remaining one hundred and thirty six subjects responded to a series of fifteen questions.

The first five questions were used for the warm up. The remaining ten questions were equally divided into queries that had been classified as either sensitive or non-sensitive in nature. The process consisted of a number of surveys asking for ratings of question sensitivity. The selection of sensitive and nonsensitive questions was judgmental, and was based on high or low mean ratings along with low variance.

At the end of the session, participants completed a short self administered questionnaire. Measures of question sensitivity were obtained using a seven point bipolar adjective scale. A rating of one indicated that the question was not sensitive while a rating of seven indicated the question was regarded as very sensitive. Demographic questions were also included.

The participants in the first condition acted as the control group and did not use a randomized response device to determine whether they gave actual or fake answers. They simply answered yes or no to a series of fifteen direct questions. Those respondents placed in the remaining three experimental groups each used a different type of technique to dictate the form of their response. However, all three groups had the same probabilities associated with the stipulated procedure. The likelihood of giving an actual answer for any question was 70% while the chance of giving a fake yes or no reply was 10% and 20% respectively.

The second group used a spinner randomization device developed by Stem and Steinhorst (1984) which provided them with specific directions. A window on the face of the spinner opened on a fake answer or a gray area which indicated the respondent should provide a rue response.

The third group utilized the Hopkins flask model randomization device (Liu and Chow, 1976). This device consists of a flask containing different colored marbles. When a white marble was shaken into the neck of the bottle people gave the fake answer written on its surface. If a green marble appeared, the true answer was given.

For the final group, a randomization device developed for use over the telephone was used to determine responses. Specification of a phone book page was the first step in this randomization device. Next, a subject selected a phone number as a starting point and did not reveal it to the interviewer. Successive numbers were then used to determine the form of reply (Stem and Steinhorst, 1984). In this study, strings of digits ending in three through nine resulted in an actual answer being given while an automatic response of yes was given when the number equaled two. The fake answer no was the proper response in all other circumstances.


A repeated measure ANOVA was used to test for overall differences between the methods; for differences between methods for sensitive questions; for differences between methods for non-sensitive questions; for differences between methods for the increase due to question sensitivity. The mean of each subjects responses to the five sensitive and five non-sensitive questions were entered as repeated dependent measures. Treatment level was the sole independent variable.


Differences in subjects age and sex have been found to significantly effect the amplitude of physiological measures of anxiety (Venables and Christie 1973). For this reason, a Chi-square analysis was conducted on these demographic characteristics before any investigation of state anxiety took place. As Table 1 indicates, no significant differences were identified between participants assigned to alternative treatment conditions with respect to either variable. Any bias that may have occurred from these potential confounding influences appears to have been effectively controlled by the random assignment of subjects.

The exit interview sensitivity ratings provided by each respondent for each of the ten questions of interest were correlated with the mean SC score for that question. This process provides an estimate of the criterion validity of using skin conductance as a measure of state anxiety resulting from question sensitivity.



Table 2 illustrates a strong correlation between these measures and implies there is sufficient justification to warrant using this procedure to identify questions that are perceived as threatening. While recognizing levels of state anxiety are not directly linked to survey error, it does suggest the hypothesized relationship between item sensitivity and autonomic nervous system activity is valid.




An inspection of composite SC indices, within treatments, indicated that their respective distributions were normal and exhibited constant error variance. A repeated measures ANOVA was used for the analysis (Keppel 1982). As Table 3 indicates, significant variation was found between treatments and between sensitive and non-sensitive questions. More importantly a highly significant interaction between treatment and sensitivity was established. These results indicate that anxiety varied between different questioning techniques, and as expected there was also a variation in anxiety between sensitive and non-sensitive questions. The indirection effect establishes that different questioning techniques have different effects on the anxiety level due to question sensitivity.



As shown in Table 4, the basal level of anxiety (situation) for non- sensitive questions varied significantly. This measure indicated the state anxiety due to the method plus the experimental situation. Differences between treatments are therefore due wholly to the method. The telephone method introduced the highest level of anxiety. Direct questioning and the flask method were significantly lower, with the spinner falling between the two extremes.

The mean SCR scores for the sensitive question measures total state anxiety due to both method, experimental situation and question sensitivity. Again the telephone method exhibited the highest anxiety level. Direct questioning was next, as expected, since it provides no protection against disclosure. Both flask and spinner provided relatively low levels of total anxiety, with the spinner performing slightly better than the flask.



The difference between the two measures is a measure of the methods ability to minimize any increase in anxiety due to question sensitivity. Both the flask RRM and the spinner RRM were very successful at this important objective. The telephone method provided some slight level of protection over direct questioning, however the difference was not significant.


While there is no direct link between anxiety level and response error or non-response error, the results of this study provide strong evidence that the randomized response method can significantly reduce the level of anxiety due to question sensitivity. The high correlation between anxiety as measured by SC and rated sensitivity indicates a strong relationship.

A significant difference in state anxiety does seem to exist between question methods. The results point to a clear distinction between two of the randomized response methods and direct questioning in terms of anxiety reduction. While the flask and spinner methods were, in part, validated by the results, the telephone method performed quite poorly. Total anxiety was higher for the telephone than for even direct questioning, and only slightly better for minimizing the increase in anxiety due to question sensitivity. While the external validity of using a telephone technique in the laboratory is suspect, it is recommended that the telephone method be used with caution until further study is accomplished.

This area of investigation offers many opportunities for further research. There are other randomized response methods which were not included in this experiment. Studies including these methods on similar as well as other populations should prove interesting. In addition, the perceived risk model presents an opportunity for empirical validation. A final suggestion for further research would be to replicate the study with a sensitive issue for which population parameters are available for comparison with the statistics produced by the randomized response methods. The relationship between anxiety and response error could then be evaluated.


Abernathy, James R., Gernard G. Greenberg and Daniel Horvitz, (1970) "Estimates of Induced Abortion in Urban North Carolina," Demography, 7, 1, 19-29.

Born, Claudia, Gerda L. Mainka, and Elke Stolting (1980), "The Paradox Reactions of Low-Anxiety Persons," Zeitschrift fuer Experimentelle und Angewandtc Psychologie, 27(3), 359-383.

Brown, G. H. (1975), "Randomized Inquiry vs. Conventional Questionnaire Methods in Estimating Drug Usage," HUMRO Technical Report 75-14, Alexandria, VA: Human Resources Research Organization.

Brown, G. H., and F. D. Harding (1973), "A Comparison of Methods of Studying Illicit Drug Usage," HUMRO Technical Report 73-9, Alexandria, VA: Human Resources Research Organization.

Calder, Bobby J., Lynn W. Phillips, and Alice M. Tybout (1981), "Designing Research for Application," Journal of Consumer Research, 8 (September), 197-207.

Cunningham, Scott M. (1967), "The Major Dimensions of Perceived Risk," in D. F. Cox (ed), Risk Taking and Information Handling in Consumer Behavior, Boston: Harvard University Press, 82-108.

Dawes, R. M. (1974), "Guttman Scaling Randomized Responses: A Technique for Evaluating the Underlying Structures of Behaviors to Which People May not Wish to Admit," University of Oregon, Oregon Research Institute.

Edelberg, Robert (1972), "Electrical Activity of the Skin: Its Measurement and Uses in Psychophysiology," in Handbook of Psychophysiology, ed. N. S. Greenfield and R. A. Stenback, New York: Holt, Rinehart, and Winston, 367-418.

Fox, James A. and Paul E. Tracy (1986), Randomized Response: A Method for Sensitive Surveys, Beverly Hills: Sage Publications.

Fowles, Don C., Margaret J. Christie, Robert Edelberg, William W. Grings, David T. Lykken, and Peter H. Venable (1981), "Publication Recommendations for Electrodermal Measurements," Psychophysiology, 232- 239.

Geurts, M. D., R. R. Andrus and J. Reinmuth (1975), "Researching Shoplifting and Other Deviant Customer Behavior Using the Randomized Research Design," Journal of Retailing, 51, 43-48.

Greenberg, Bernard G., Roy R. Kuebler, James R. Abernathy, and Daniel G. Horvitz (1971), "Application of the Randomized Response Techniques in Obtaining Quantitative Data," Journal of the American Statistical Association, 66 (June), 243-250.

Greenberg, Bernard G., J. R. Abernathy, and D. G. Horvitz (1970), "A New Survey Technique and its Application in the Field of Public Health," Milbank Memorial Fund Quarterly, 68, 39-55.

Horvitz, D. G., B. G Greenberg and J. R. Abernathy (1976), "Randomized Response: A Data-Gathering Device for Sensitive Questions," International Statistical Review, 44, 181-196.

Hentschel, Uwe and Gaby Ternes (1984), "Moderating Effects of Trait Anxiety on Electrodermal Reactions in Different Learning Conditions," Psychological Reports, 54, 803-809.

Keppel, Geoffrey (1982), Design & Analysis A Researcher's Handbook, Engelwood Cliffs: Prentice Hall, Inc.

Knight, Martha L. and Richard J. Borten (1979), "Atonomic and Affective Reactions of High and Low Socially-Anxious Individuals Awaiting Public Performance", Psychophysiology, 16, 3, 209-213.

Lamb, Charles W. and Donald E. Stem, Jr. (1978), "An Empirical Validation of the Randomized Response Technique," Journal of Marketing Research, (November), 616-621.

Liu, P. T. and L. P. Chow (1976), "A New Discrete Quantitative Randomized Response Model," Journal of the American Statistical Association, 71 (March),72-73.

McLeod, Daniel R., Rudolf Hoehn-Saric, and Robin L Stefan (1986), "Somatic Symptoms of Anxiety: Comparison of Self-Report and Physiological Measures," Biological Psychiatry, 21, 301-310.

Reinsuth, James E. and Michael D. Geurts (1975), 'The Collection of Sensitive Information Using a Two-Stage Randomized Response Model," Journal of Marketing Research, 12, (November), 402-407.

Schlosberg, Harold and Walter C. Stanley (1953), "A Simple Test of the Normality of Twenty-Four Distributions of Electrical Skin Conductance," Science, 117, (January), 35-3.

Shimizu, Iris M., and Gordon S. Bonham (1978), "Randomized Response Technique in a National Survey" Journal of the American Statistical Association, 73 (March) 35-39.

Spielberger, Charles D. (1972), "Anxiety as an Emotional State," in Anxiety Current Trends in Theory and Research, ed. Charles D. Spielberger, New York: Academic Press.

Stem, Donald E. Jr. and R. Kirk Steinhorst (1984), "Telephone Interview and Mail Questionnaire Applications of the Randomized Response Model," Journal of the American Statistical Association, 79 (September) 555-563.

Tull, Donald S. and Del I. Hawkins (1984), Marketing Research Measurement and Method, New York: MacMillan Publishing Co.

Venables, P. H. and M. J. Christie (1973), "Mechanisms, Instrumentation, Recording Techniques, and Quantification of Responses," in Electrodermal Activity in Psychological Research, ed. William F. Prokasy and David C. Raskin, New York: Academic Press.

Warner, Stanley L. (1965), "Randomized Response: A Survey Technique for Eliminating Evasive Answer Bias," American Statistical Association Journal, (March), 63-69.