Respondent Generated Intervals (RGI) S. James Press Department of Statistics University of California at Riverside Riverside, CA 92521-0138 jpress@ucr.edu Judith M. Tanur Department of Sociology State University of New York at Stony Brook Stony Brook, NY 11794-4356 jtanur@notes.cc.sunysb.edu 1 Respondent-Generated Intervals (RGI) Definition The Respondent-Generated Intervals, or RGI, protocol for asking questions in sample surveys involves asking respondents not only for a basic answer to a recall-type question, but also, the respondent is asked for a smallest value his/her true answer could be, and a largest value his/her true answer could be. We’ll refer to these values as the lower and upper bounds provided (sometimes called brackets). The result of the RGI protocol is that the respondents themselves generate the intervals in which their true beliefs lie, instead of having their quantitative beliefs forced into intervals pre-assigned by the survey designer. Other survey protocols sometimes provide pre-assigned intervals for the response, and the respondent must choose one of the intervals. Interval-Response Surveys Survey protocols that permit the respondent to give answers in intervals, selfdetermined, or pre-assigned by the survey designer, are often preferred by respondents for sensitive questions because the respondent need not be specific about the exact value being requested. Interval response protocols are also often preferred by respondents for questions for which the answers are not very well known. By responding in intervals for such questions, respondents need not be precise about the exact answer. Respondents prefer the RGI technique because it 2 allows them to have control over their disclosures, and RGI allows respondents to feel confident about the accuracy of the information they provide. The intervals RGI respondents provide tend to be narrower than pre-defined intervals. Genesis of RGI The RGI protocol for questionnaire design has its origins in Bayesian assessment procedures. In that context, for a specific individual, we might assess an entire prior distribution about an unknown parameter. That prior distribution represents the individual’s degrees of uncertainty about that unknown parameter. In certain contexts, we might assess many points on the individual’s subjective probability distribution for that parameter by means of a sequence of elicitation questions, and then connect those points by a smooth curve that represents the underlying distribution. For example, using some purely hypothetical numbers, suppose an individual has a normal subjective probability distribution representing “ 0 ”, the true (but unknown) change in the number of doctor visits he/she believes he/she made last year, compared with the previous year, so that 0 ~ N (4,1). In such a case, the individual believes that it is most likely that he/she visited a doctor 4 more times last year than the previous year, with a standard deviation of 1. So this individual equivalently believes that there is a 99.7% chance that he/she visited a doctor between 1 and 7 more times last year, or that there is really almost no chance that the true number of excess times was less than 1 or greater than 7. This probability distribution is subjective, in that it represents a specific individual’s degrees of belief about his/her uncertainty about the underlying quantity, in the case of this example, the individual’s uncertainty about how many more visits 3 he/she believes he/she truly made to the doctor last year compared with the previous year. We postulate that: in a factual survey each respondent has a distinctive recall distribution, and in an attitude or opinion survey he/she has an underlying probability distribution for his/her opinion or attitude about some issue. In the case of a recall-type question, we assume that the respondent knew the true value at some point (or had enough information to construct it) but because of imperfect recall, he/she is not certain of the true value. He/she may feel confident that he/she knows the true value (but may be wrong in spite of high confidence), or he/she may be quite uncertain of the true value (and conceivably could be correct about the true value, but not realize it). We furthermore assume that the respondent is not purposely trying to deceive. In the case of opinion or attitude questions, the respondent may have a very fuzzy idea of his/her attitude about an issue, or he/she may feel quite strongly and specifically about it. Benefits of RGI to the Survey Administrator The benefits of the RGI survey protocol to the respondents were discussed above. Now we discuss its benefits for the survey administrator. There are issues of (1) response rate, (2) measurement of range-of-belief, and (3) accuracy of estimating the population mean (in the case of factual questions). We treat these issues in turn. (1) Response Rate-- Interval protocol methods tend to reduce item nonresponse in that such techniques present a viable method for obtaining 4 some information from respondents who might otherwise provide none. (See Item Non-Response, Item Non-Response Remedies.) (2) Measurement of Range-of-Belief—Suppose ai , bi denote the lower and upper bounds, respectively, for respondent i, i = 1, …, n, where n denotes the number of respondents in the sample. Then, if a0 min ai , and b0 max bi , a measure of range of belief for the i i sampled respondents is given by: (b0 a0 ). (3) Accuracy—Data collected from a factual RGI survey can be used to generate an estimate of the population mean that has smaller non-sampling error (bias) than that associated with estimators obtained from more conventional surveys. The proposed estimator is model-based and is Bayesian. To obtain an estimate of the mean in a population we typically just form the simple average of the opinions or attitudes or recalled values from individuals who may have very different abilities to carry out the task required for a response. But such a simple average may not necessarily account well for typical unevenness in assessment of opinion or attitude, or in recall ability. The RGI protocol can be used to improve upon more traditional population estimates by its ability to learn more about the different abilities respondents have to quantify their fuzziness about their beliefs, and then take that fuzziness into account in the estimation process. Ideally, we would ask the respondents many additional questions about their beliefs. That would permit us to assess many fractile points on each of their underlying opinion or attitude or recall distributions. Owing to the respondent burden of a long questionnaire, the sometimes heavy cost limitations of adding questions to 5 a survey, the cost of added interviewer time, etc., there may sometimes be a heavy penalty imposed for each additional question posed in the survey questionnaire. (See Respondent Burden.) In such cases, RGI might be used for just a few questions. The RGI protocol obtains three points on each respondent’s underlying distribution. A hierarchical Bayesian procedure is suggested below for using the collection of triples of points from the respondents for estimating the population mean. The resulting RGI estimator can often improve upon the accuracy of the usual sample mean. Let the triple: ( yi , ai , bi ) denote the basic response, and, as above, the lower bound response, and the upper bound response, respectively, for respondent i, i = 1,…, n. Suppose that the yi ’s are all normally distributed N ( i , i2 ) , that the i ' s are exchangeable, and i ~ N ( 0 , 2 ) . i denotes the true (factual, or opinion, or attitude) value for respondent i. It has been shown (see Press, 2003, Appendix) using a hierarchical Bayesian model, that in such a situation, the conditional posterior distribution of the population mean, 0 , is normal, and is given by: ( 0 data, i2 , 2 ) ~ N ( , 2 ), (1) where the posterior mean, , conditional on the data and ( i2 , 2 ) , is expressible as a weighted average of the usage quantities, the yi ’s, and the weights are expressible approximately as simple algebraic functions of the 6 bounds. The conditional posterior variance, 2 , drives the credibility (Bayesian confidence) interval, and is discussed below. We assess values for the i2 parameters from: k1i bi ai ri , the respondent interval lengths. Analogously, we assess a value for 2 from: k2 b a , the average respondent interval length. For normally distributed data it is commonly assumed that lower and upper bounds that represent extreme possible values for the respondents can be associated with 2 standard deviations below, and above, the mean, respectively. We’ll generally assume therefore that k1 k2 k 4 (corresponding to standard deviations above and below the mean). The conditional posterior mean was shown (Press, 2004a) to be given by: n i yi , (2) 1 where the i ’s are weights depending upon the interval lengths that are given approximately by: 1 2 2 r (b a ) i n i . 1 1 r 2 (b a )2 i (3) 7 We note the following characteristics of this RGI (weighted average) estimator: 1) The weighted average is simple and quick to calculate, without requiring any computer-intensive sampling techniques. 2) It will be seen in the examples below that if the respondents who give short intervals are also the more accurate ones, the RGI estimator will tend to give an estimate of the population mean that has smaller bias than does the sample mean. In the special case in which the interval lengths are all the same, the weighted average reduces to the sample mean, y , where the weights all equal (1/n). In any case, the lambda weights are all positive (although some may be extremely small), and must sum to one. 3) The longer the interval a respondent gives, the less weight is applied to that respondent’s usage quantity in the weighted average. The length of respondent i’s interval is a measure of his/her degree of confidence in the basic response he/she gives, so that the shorter the interval, the greater degree of confidence that respondent has in the basic response he/she reports. Of course a high degree of confidence does not necessarily imply an answer close to the true value. 4) If we define the precision of a distribution as its reciprocal variance, the quantity { [ri 2 (b a ) 2 ] /(4) 2 } may be seen to be the conditional variance in the posterior distribution corresponding to respondent i, and 8 therefore, its reciprocal represents the conditional precision corresponding to respondent i. Summing over all respondents’ precisions gives: total conditional posterior variance = 2 1 . n 16 1 r 2 (b a )2 i (4) The number “16” is the square of the 4 that arises from 4 i bi ai ri , and 4 b a . So another interpretation of i in equations (2) and (3) is that it is the proportion of the total conditional posterior precision in the data attributable to respondent i. The variance of the conditional posterior distribution is given in eqn. (4). The posterior variance is the reciprocal of the posterior total precision. Because the posterior distribution of the population mean, 0 , is normal, it is straightforward to find credibility intervals for 0 . For example, a 95% credibility interval for 0 is given by: ( 1.96, 1.96 ) . (5) That is, 9 P{( 1.96 0 1.96 ) data} 95%. (6) An alternative interval estimate of the population mean that is more likely than the credibility interval to cover the true population value is the Average Respondent-Generated Interval, ARGI b a . The range-ofbelief for the sample of respondents is b0 a0 . Examples We next present some numerical “theoretical examples” of the use of RGI estimates. 1) Example For Recall Questions We examine an artificial extreme case. Suppose we have a sample survey of size n = 100 in which the RGI protocol has been used. Suppose also that the true population mean of interest that we are trying to estimate is given by 0 1000. In this example we fix both the basic response quantities and the respondents’ bounds, (ai , bi ) , ri bi ai , i 1,..., n, artificially. Assume the respondents have been ordered so that the first 30 respondents all have excellent memories, are quite accurate, and have individual truths of 900.. Suppose the intervals these accurate respondents give are: (a1, b1 ),...,(a30 , b30 ) (975,975),...,(975,975) . 10 That is, they are all not only pretty accurate, but they all believe that they are accurate, so they respond to the bounds questions with degenerate intervals whose lower and upper bounds are the same. Accordingly, these accurate respondents all report intervals of length ri 0 , and usages of equal amounts, yi 975 (compared with the true population value of 1000., In addition, suppose the next 20 respondents also are quite accurate, but their individual truths are above 1000, say 1100. Suppose the intervals these respondents give are: (a31, b31 ),...,(a50 , b50 ) (1125,1125),...,(1125,1125) . Next suppose that the last 50 respondents all have poor memories and are inaccurate. They report the intervals: (a51 , b51 ),...,(a100 , b100 ) (500,1500),...,(500,1500) , that have lengths of ri 1000 , and they report unequal basic response quantities of 550, for the first 30, and 1450 for the last 20. We now find: a 767.5, and b 1267.5, so b a 500. We may now calculate that the weights are given by: .0167, i 1,...50 .0033, i 51,...,100 i 11 100 It is easy to check that: i 1. We may now readily find that the conditional 1 posterior mean RGI estimator of the population mean, 0 , is given by: 100 i yi 1014.17. i 1 The corresponding sample mean is given by: y 972.5. The numerical error (bias) of the posterior mean is given by: 1000 1014.17 1000 14.17. The numerical error (bias) of the sample mean is given by y 1000 972.5 1000 27.5. The RGI estimator has reduced the bias error by 27.5 - 14.17 = 13.33, or about 48%, compared with the error of the sample mean. It is also interesting to compare interval estimates of the population mean by comparing the standard error of y , with , the standard deviation of the 12 conditional posterior distribution of . These estimates give rise to the corresponding confidence and credibility intervals for 0 , respectively. We may readily find that for the data in our theoretical example, 16.14. It is also easy to check that for our data, the standard deviation of the data is 323.80. So the standard error for a sample of size 100 is 323.80/10 = 32.38. Thus, the RGI estimate of standard deviation is about half the standard error of the sample mean. Correspondingly, the length of the 95% credibility interval 2(1.96) ( ) = 63.27, while the length of the 95% confidence interval is 2(1.96)(32.38) = 126.93. The 95% confidence interval is about twice as long as the 95% credibility interval. The 95% credibility interval is given by: (982.53, 1045.80). The 95% confidence interval is given by: (909.04, 1035.96). We note in this theoretical example that: 1) Both the RGI credibility interval and the confidence interval cover the true value of 1000, but the credibility interval is smaller. 2) we expect to find many situations for which the bias error of the RGI estimator is smaller than that of the sample mean; however, the differences may be more, or less, dramatic compared with their values in this example; 3) in large samples, we expect that both the sample mean and the RGI estimator will converge to one another, and to the true population mean. 4) We also note that the success of the theoretical example depends on the fact that the accurate respondents give short intervals and so have heavier 13 weights than the inaccurate respondents who give longer intervals. In order to help assure that such conditions hold in operational studies, we can ask respondents a preliminary question concerning their confidence in their accuracy and use their confidence weightings to moderate their responses. 2) Example for Opinions and Attitudes Suppose there is a population of opinions about some issue, say, “Issue A”. Perhaps the analyst would like to establish the mean of the opinions of all people living in the City of New York about Issue A. This is “truth”. But it isn't an absolute truth. It's a relative truth, a truth at this time, and under current circumstances. It's not an absolute truth that will have the same value next month, or next year, as would be the case for a factual or recall-type of question. The analyst would like to estimate that relative truth. Relative true values are generally different for each person in a sample, as is generally the case in a recall question. But accuracy in an opinion or attitude context has no meaning, other then with respect to sampling error, although respondents with fuzzy attitudes or opinions who give intervals as well as point values provide additional information with which their views may be sharpened. There is no “correct” answer for an opinion or for an attitude for a given respondent, as there would be for a person answering a recall-type of question. Similarly, response bias does not have the same meaning as in recall. (With a recall-type of question, one of the reasons for response bias arises out of faulty memory.) When using RGI for attitudes or opinions we can find both point and interval estimators, although we can't really say much about accuracy, per se. But we can 14 speak about the RGI point estimator coming closer to second-guessing the opinions of New Yorkers in the above example than a mere sample mean that includes some people with very fuzzy opinions, and some people who have very firm opinions. RGI can provide a measure of average strength-of-opinion, as measured by the average range supplied by all respondents , (b a ) . It can also supply a credibility interval measure of belief; a confidence interval can also supply an interval measure of belief. The confidence interval, however, only reflects sampling uncertainty, whereas the RGI credibility interval also reflects fuzziness of opinion. The range-of-belief also available with RGI, (b0 a0 ) , is somewhat different. For further discussion of the use of RGI surveys with opinion questions, see Press and Tanur, 2004a. Future Research Optimally efficient RGI used with factual questions requires that respondents who give narrow intervals also be the more accurate ones, and respondents who give wide intervals also be the more inaccurate ones. When RGI is used with opinion or attitude questions, firm and fuzzy opinions or attitudes should be analogously categorized. Future research with RGI should involve developing meta-cognitive procedures that will achieve these objectives. References Press, S. James (2004). “Respondent-Generated Intervals (RGI) for Recall in Sample Surveys”, Journal of Modern Applied Statistical Methods, May, 2004, Vol. 3, No. 1, pp.104-116. 15 Press, S. James and Tanur, Judith M. (2004a). “An Overview of the Respondent-Generated Intervals (RGI) Approach to Sample Surveys”, Journal of Modern Applied Statistical Methods, Nov. 2004, Vol. 3, No. 2, pp. 288-304. Additional Reading Chu, LiPing, Press, S. James, and Tanur, Judith M. (2004). “Confidence Eliciation and Anchoring in the Respondent-Generated Intervals Protocol”, Journal of Modern Applied Statistical Methods, Nov. 2004, Vol. 3, No. 2, pp. 417-431. Press, S. James and Judith M. Tanur (2000). "Experimenting with Respondent-generated Intervals in Sample Surveys" , with discussions, in Survey Research at the Intersection of Statistics and Cognitive Psychology, Working Paper Series #28, Monroe G. Sirken, Editor, National Center for Health Statistics, Center for Disease Control and Prevention, pp. 1-18. Available for downloading at: http://www.statistics.ucr.edu/press.htm Press, S. James and Kent H. Marquis(2002). "Bayesian Estimation in a U.S. Government Survey of Income Recall using RespondentGenerated Intervals", Proceedings of the Conference of the International Society of Bayesian Analysis, Crete, Eurostat. Available for downloading at:http://www.statistics.ucr.edu/press.htm 16 Press, S. James, and Tanur, Judith M. (2003). “The Respondent- Generated Intervals Approach to Sample Surveys: From Theory to Experiment”, in New Developments on Psychometrics, by Yanai, Haruo; Okada, Akinori; Shigemasu, Kazuo; Kano, Yutaka; Meulman, Jacqueline J. (Eds.), Tokyo: Springer-Verlag. Press, S. James and Tanur, Judith M. (2004b). “Relating Respondent-Generated Intervals Questionnaire Design to Survey Accuracy and Response Rate”, Journal of Official Statistics, Statistics Sweden, June, 2004, Vol. 20, No. 2, pp. 265-287. Special Issue on Questionnaire Development, Evaluation, and Testing Methods. S. James Press Judith M. Tanur Dr. S. James Press is Distinguished Professor in the Department of Statistics, University of California at Riverside, Riverside, California. Dr. Judith M. Tanur is Distinguished Teaching Professor in the Department of Sociology, State University of New York at Stony Brook, Stony Brook, New York. 17