1 A Study on the Characteristics of the Asset Item Nonrespondents in a Panel Survey Youngshil Park (Researcher, Statistics Korea) - Statistical Research Institute, Statistical Center, 713 Hanbatdaero, Seo-gu, Daejeon, 302-847, Korea - 82-42-366-7210 , youngshil@korea.kr Sunghee Lee (Assistant Research Scientist, University of Michiga) - Survey Research Center, 426 Thomson St., Ann Arbor, MI 48104, the United States - 1-734-615-5264, sungheel@isr.umich.edu Abstract Assets and income questions that ask respondents to report exact value with an open-ended question are typically associated with high nonresponse in surveys in western countries. However, there is little research on assets and income nonresponse in Korea because nonresponse rate is not as high comparing with western countries. Interestingly, there are high nonresponse rates in specific asset items of the KLIPS, and the KLIPS has tried to reduce nonresponse rate by applying with a follow-up range question, asking categorical information, among those who did not provide an exact amount. In this study, we show the asset nonresponse trends in Korea over a decade, and evaluate the effectiveness of the range question toward nonresponse. Also, we examine whether or not there is a persistent tendency in nonrespondents Key Words: nonresponse rate, panel survey, range questions, nonresponse tendency 2 A Study on the Characteristics of the Asset Item Nonrespondents in a Panel Survey RESEARCH BACKGROUND Assets and income questions that ask respondents to report an exact value with an open-ended question are typically associated with high nonresponse rate in surveys (Moore, Stinson, and Welniak, 2000; Riphahn and Serfling, 2005; Schrapler, 2006), and this high nonresponse rate can cause bias when estimating of assets and income values as well as other statistics in models utilizing those variables (Yan, Curtin, and Jans, 2010). There have been efforts to reduce the bias through the improving of imputation. Also, some survey designers have tried to diminish the bias by employing various question types such as “range cards” or “unfolding brackets”. Both are follow- up questions that ask respondents who did not answer an exact value to place assets and income amounts in a specific range (Juster and Smith, 1997). Unfolding brackets ask nonrespondents and to place an amount above or below of a specific range. Surveys including the Health and Retirement Survey (HRS) and the study of Asset and Health Dynamics Among the Oldest Old(AHEAD) have adopted this technique since it originated from the wealth module of the Panel Study of Income Dynamics (PSID) in 1984, and it has been proven to substantially decrease nonresponse rates (Heeringa, Hill, and Howell, 1995; Juster, Cao, Perry, and Couper, 2006). On the other hand, the range card technique is a range of letters that correspond to quantitative intervals to show for nonrespondents. This is used effectively in face-to-face interviewing rather than telephone 3 interviewing (Juster and Smith, 1997). Unlike in western countries, there are few researches on assets and income nonresponse rates in Korea. While the income nonresponse rate is between 20~40% in western countries according to survey literature (Moore, Stinson, and Welniak, 1999), the data analysis from the Korean General Social Survey (KGSS) and the Korean Labor and Income Panel Survey (KLIPS) shows that the income nonresponse rates in both surveys are less than 5%. We guess the reason why the nonresponse rate is low is that income is not as sensitive to Korean comparing with western people. Another possibility is that interviewers are trained to obtain at least categorical information when respondents refuse to answer. However, interestingly, the nornesponse rates are high in specific asset questions of the KLIPS, and the KLIPS has tried to mitigate the nonresponse rate by adding follow-up range questions. These questions ask about categorical information for those who did not provide an exact amount. The range question strategy in the KLIPS is different from the range cards. Whereas range cards are separated from the main questionnaire and shown to respondents to answer after interviewers ask questions in other surveys like the HRS, range cards in the KLIPS are included in the questionnaire. The category may be sometimes shown, but interviewers usually read a question and the response categories. This technique has been used for over a decade. In this study, by using the KLIPS data, we show the asset nonresponse trends in Korea for over a decade. Also, we evaluate the effectiveness of the range questions toward nonresponses. In addition, we examine whether or not there is a persistent tendency in nonrespondents. RESEARCH QUESTIONS 4 This study investigates the following specific questions. First, what is the nonresponse pattern of the asset items applied to the range questions in a panel survey? We focus on examining how the use of the range question reduces the nonresponse rate and how the amount of the reduction has changed over a decade in the panel survey. The effectiveness of range questions has been proven to work only for short periods like one or two waves (Juster and Smith, 1997; Juster, Cao, Perry, and Couper, 2006; Heeringa, Hill, and Howell, 1995), so we don’t know the trend of its effectiveness over a longer period of time. Although some studies (Kennickell, 1997; Yan, Curtin, and Jans, 2010) examined the trends for the longer period, the data used in the studies were not a panel survey. A second question has to do with the characteristics of respondents and nonrespondents, especially focusing on the range data respondents among the nonrespondents. If the characteristics of respondents differed from the characteristics of nonrespondents, we can expect that the asset estimates between the two groups are statistically different. Moreover, if it proved, developing a proper imputation program would be quite important (Juster and Smith, 1997; Juster, Cao, Perry, and Couper, 2006). Some surveys have utilized range data gathered from the bracket questions to impute assets or income in to the nonrespondents of the surveys. Even though the range data are not be used for imputation purposes in the KLIPS, we need to find out whether or not there is a systematic difference between the data of two groups for further developmental studies. Third, do the range data respondents tend to not respond to an open-ended question in next waves? Kennickell (1997) showed that there was a substantial persistence in the use of ranges by individual respondents, but that result was only for questions from the same wave. We are not sure whether or not this persistency is valid over waves. We attempt to investigate this nonresponse tendency by showing the effect of range data experience over a 10 year period. If the propensity is consistent, interviewers can anticipate the probability of 5 that the range data respondents don’t answer to the open-ended question in future waves. DATA This study used data from the KLIPS, a panel survey of the labor market, income activities of households, and individuals residing in urban areas. It has been conducted every year with face-to-face interviewing by the Korean Labor Institute since 1998. The KLIPS sampling method is an equal probability of households from urban areas. It started with 5,000 household samples which contain 13,000 people (aged 15 and older). Individuals that have blood or other legal ties to the original panel members are also recruited to the original sample from wave 2. For example, if a panel member gets married and forms an independent household with his/her spouse, then, his/her spouse becomes a ‘new respondent’ for the original panel (www.kli.re.kr). The household response rate of the KLIPS from wave 2 to wave 11 are as follows: 87.9%(wave 2), 81.2%(wave 3), 78.1%(wave 4), 77.0%(wave 5), 78.7%(wave 6),78.6%(wave 7), 77.6%(wave 8), 77.5%(wave 9), 76.5%(wave 10), 75.2%(wave 11). It is also important how many original samples remain in the current wave because the same households and individuals are interviewed repeatedly. Household retain rates are as follows: 87.6% (wave 2), 80.9%(wave 3), 77.3%(wave 4), 76.0%(wave 5), 77.2%(wave 6),77.3%(wave 7), 76.5%(wave 8), 76.5%(wave 9), 75.5%(wave 10), 74.2%(wave 11). Overall, the response rates and retain rates are stable after wave 4. The questionnaire is divided into two parts: household and individual characteristics. The asset module is included in the household section. For the household questions, interviewers were trained to collect data from the heads of household or their spouses to gain 6 more exact information (KLI, 2010). The asset module is composed of three real estate items, six financial items, and several items related to cars and other possessions. Real estate and financial asset items have been surveyed from wave 2, but the items related to cars and other asset information was started since wave 10 and wave 11 respectively. The range question for reducing nonresponse rates is adopted by only real estate assets which consist of three items: total real estate other than housing, real estate leasing including housing, and real estate renting other than housing. <Figure 1> shows a flow chart of these questions. First, ownership status is obtained with a yes or no question. If an asset is owned, its exact amount is asked with an open-ended question. Then, if the exact amount is not answered, a range question is asked with a closed-ended question. There are eleven categories in the range question, and these categories are all the same for each nonrespondent and the three real estate items. Meanwhile, the KLIPS has requested whether or not an owner knows the exact information if an asset is owned since wave 6. After that, the question asks exact information only for the respondents who know that information. Figure 1 here RESULTS 1. Nonresponse Trends of Real Estate Asset Items 1) Nonresponse Rates by Item The final response disposition codes of a real estate asset item are followed by: the continuous data response to the open-ended question which requests an exact value (the exact data response), the categorical data response to the range question which asks rough information among nonrespondents (the range data response), and the final nonresponse. The 7 range data response plus the final nonresponse is the initial nonresponse to the open-ended question. Table 1 here <Table 1> presents response patterns for the three kinds of real estate asset items from wave 2 to wave 11. The initial nonresponse rate are different according to asset items. For the total real estate value other than housing item, 70.5% of the owners didn’t give exact values in wave 2. Comparable figures in the other two items were 11.2% and 36.5%. The difference of the initial nonresponse rates by item has been consistent through wave 11 although the gap between the other two items about leasing and renting real estate disappeared. The reason why nonresponse rates of the total real estate other than housing item is higher than the other two asset items is that the market value of total real estate other than housing is subject to more uncertainty than deposits for leasing and renting. This interpretation is supported by the result of <table 2>. We also see that item nonresponse rates are different according to housing asset items. While the KILIPS asks housing market value to people who own with using the first item, it asks about deposit amounts to people who rent house monthly and yearly through the second and the third item. The latter information is more easily accessible than housing market value of owners. Consequently, the nonresponse rates of the latter items are lower than the first housing asset item. 2) The Response Trends of the Range Question <Table 1> shows the exact data response rate and the initial nonresponse rate (range data response rate + final nonresponse rate). Overall, the exact data response rate have increased, but the initial nonresponse rates have declined. Specifically, it reveals that large 8 amounts of the initial nonresponse rate are substantially reduced if the questions ask categorical information. In wave 2, the reduced nonresponse rate of the range question, that is range data response rate, is 65.6% for the total real estate other than housing item, 8.6% for the leasing real estate including housing item, and 32.2% for the renting real estate other than housing item. Over a decade, the range data response rates from wave 2 to wave 11 report range from 65.6% to 46.3% for the first item, 8.6% to 0.7% for the second item, and 32.2% to 1.1% for the third item. We find that the reduction rates have decreased for 10 years, and the rate for the latter two items is close to 0% in the most recent wave.1 Also, the final nonresponse rate has decreased gradually. It has started with low nonresponse rate from wave 2, and it has decreased much slower than the initial nonresponse rate. Thus it looks stable compared with the initial nonresponse rate, but the final nonresponse rate became near 0% in wave 11. For the first asset item, the final nonrespons rates from wave 2 to wave 11 range from 4.9% to 0.2%. The reduced nonresponse rate in the range question has decreased over the last decade, yet a dramatic change happened in wave 7. We plotted the initial nonresponse rate and the final nonresponse rate of the total real estate other than housing item dividing into two parts from wave 2 to wave 6 and from wave 7 to wave 11 in <figure 2>. It shows the trends of the initial nonresponse rate and the reduced nonresponse rate in each part are steady. Interestingly, the final nonresponse rate in wave 7 is higher than the waves before and after it (wave 6 and wave 8) in <table 1>. We found that the discontinuity between wave 6 and wave 7 is only for real estate asset items by comparing with other housing asset items in <Table 2>. Yan, Curtin, and Jans (2010) examined the trends in income nonresponse for about 20 years from the Survey of Consumers from 1986 to 2005. This survey adopted for unfolding brackets technique like the range questions. They found that the nonresponse reduction rate by the unfolding brackets decrease after a peak in 1990. 1 9 There is no fluctuation in the nonresponse rate between wave 6 and wave 7 in the other items. Figure 2 here We have talked several times with the KLIPS manager, and there was no change between wave 6 and wave 7 in the survey protocol, wording in the asset items, the training manual and so on. We can only assume that interviewers may have persuaded respondents to answer continuous data instead of allowing them to accept ‘Don’t Know(DK)’ answer. Usually, interviewers were trained to obtain at least categorical information rather than DK. 2. The Characteristics of Respondents to the Range Question Here, we compare how different the range data respondents among nonrespondents are from the exact data respondents for total real estate other than housing item. Generally, the variables including respondent’s sex, age, race, education, and marital status are considered to be correlated with the income nonresponse (Yan, Curtin, and Jans, 2010). Table 3 here <Table 3> shows the result of the mean difference between the exact data respondents and the range data respondents by the heads of household’s sex, age, level of education, marital status, and employment status, the number of household member, and the total household income. It displays only three waves selectively (wave 2, 7, 11), but we could read the consistent pattern for 10 waves. The characteristics of the range data respondents are placed in the middle between the exact data respondents and the final nonrespondents, although we can not be sure of sure of this pattern because the number of the final nonrespondents is low. However, there is a distinct pattern between the characteristics of the exact data respondents and the range data respondents. On the whole, the mean difference of these two groups is statistically significant on the heads of household’s age and the level of 10 education and household income variables. Those who answer to the range question are likely to be older, and less educated. Also, they are likely to have a lower income level than those who answer to the open-ended question. The result suggests that this fairly distinct pattern will be potentially useful in a model-based imputation procedure (Hoynes, Hurd, and Chand , 1998). Also, it could be a good reason to redesign the question for people whose cognitive skill is lower than average.2 This selectivity in group characteristics may cause a systematic difference in asset distribution. <Table 4> presents that there is a clearly different asset distribution pattern between the exact data respondents and the range data respondents. Whereas 2.1% of those who answered exact data placed in the category of less than 10 million Korean won (krw), 8.0% of those who answered range data placed in the same category in wave 3. Until about less than 100 million krw, the proportion of range data respondents is higher than the proportion of exact data respondents. This tendency is consistent over 10 waves. It means that the assets of the range question respondent is relatively distributed lower level compared to the asset of the exact data respondents. Furthermore, it suggests that the asset is likely to be overestimated without including range data response. It coincides with the range data characteristics. The range data respondents are less likely to be higher status in aspects of 2 Several studies shows the similar results (Juster and Smith, 1997; Hoynes, Hurd, and Chand ,1998; Yan, Curtin, and Jans, 2010). Even though these studies focused on all kinds of assets, bracketed respondents who self-select in HRS or AHEAD have different characteristics from the respondents who gave actual dollar values. For example, in AHEAD wave 1 those who gave a DK response when asked about asset values were less likely to be married, and more likely to be old than those who gave actual dollar values (Hoynes, Hurd, and Chand 1998). The variables having an effect on the nonresponses are related to the cognitive ability to give correct income or asset information. Among respondents who have lived in their current home for many years, unwillingness to report housing values may reflect on uncertainty rather than sensitivity about values, especially among less-educated and lower-income respondents (Juster and Smith, 1997). 11 education and income than exact data respondents.3 Table 4 here 3. Nonresponse Tendency of the Range Data Respondents We found that the exact data response rate and the range data response rate have been changed over waves. Does it mean that individual respondents changed their attitude to respond exact data in later wave? However, distinct characteristics of the two groups have been constant. How can we explain this pattern? Table 5 here <Table 5> presents that the response rate in the next wave depend on response or nonresponse type in current wave. The range data respondents in wave 2 are likely to not respond to the open-ended question in next wave than the exact data respondents do. While 59.1% of owners didn’t answer to the exact data question in wave 3 among exact data respondents in wave 2, 77.0% of owners didn’t answer to exact data question in wave 3 among the range data respondents in wave 2. Moreover, 71.7% of owners didn’t answer to exact data question in wave 3 among the nonowner of asset in wave 2. Nonresponse rates to the open-ended question in next wave are larger in order of range data respondents, nonowner, and exact data respondents. This tendency has been uniform for 10 waves, although the nonresponse rates to the open-ended question rapidly changed between wave 6 and wave 7. We plotted the nonresponse rates to the open-ended question splitting into wave 6 and wave 7 According to the first wave of AHEAD data, for real estate other than housing, respondents who provide brackets have lower real estate assets than respondents who give continuous amounts (Hoynes, Hurd, and Chand, 1998). 3 12 in <figure 3>. As a result, we see the nonresponse rate of the range data group has been more stable than the exact data and nonowner group since wave 7. It may be interpreted people have persistent nonrespons tendency of range data group rather than exact data respondents and nonowner group. Figure 3 here Unlike range data group, the response rate to the open-ended question of exact data respondents increases gradually. In addition, response rates to the open-ended question of nonowner excess nonresponse rate to the open-ended question since wave 7. We propose that encouraging the exact data and nonowner group is more effective way to increase item response rate. Table 6 here <Table 6> shows the result of logistic analysis whether or not the experience of the range data response affects to later nonresponse to the open-ended question asking exact amount. There are two models according to the kind of independent variables. In the first model in left side of the table 6, an independent variable is whether or not respondents answer to the range question one wave ago. In the other model, an independent variable is cumulative frequency of range data response for all previous waves. We examine the effect of independent variables while controlling socio-demographic variables of same wave of an independent variable. The logistic regressions suggest that the experience of the range data response is related to the nonresponse in next wave, and this result is consistent for all waves regardless of the kind of independent variable. However, odds ratio in the first model is larger than it in the second model. It implies that the effect of range data experience one wave ago is larger than the number of range data response experience. 13 CONCLUSION This study shows the nonresponse trends of the real estate asset items and the characterisicts of the nonrespondents, in particular, respondents to the range question among nonrespondents in the panel survey of Korea. First, it appeared that nonresponse rates has decreased over waves and the range questions have contributed to reduce nonresponse rate dramatically. However, the reduced nonresponse rate by the range question has declined over a decade and for even some items (real estate leasing including housing and renting other than housing items), the rate disappeared in recent waves. Second, we found that the characteristics of the exact data respondents and the range data respondents are different and it leads to systematic differences in the asset distribution between two groups. Range data respondents relative to exact data respondents are placed lower asset level than the exact data respondents. Third, it suggested that the range data respondents have a persistent tendency toward nonresponse in next wave comparing with other groups (the exact data respondents and nonowner). These findings have important implications for survey quality and practical aspects. First, the fact that the range data response rate has decreased means the range question may not be needed in specific time in a panel survey. The KLIPS should test this effect to save questionnaire space and cost. Second, it should be considered imputation of the real estate asset data value with range data information. If we use only exact data in analyzing, it causes bias in asset estimates. The KLIPS should consider Imputation based on covariates like the head of household’s age, education, and income may provide an important gain in assigning assets at the individual level even though the effect on the population may not be large (Hoynes, Hurd, and Chand, 1998). Third, by showing range data respondent’s persistent tendency toward the nonresponse, we can predict probability of respondent’s behavior in the next wave and prepare the effective way to reduce item nonresponse. For example, it will be 14 efficient to persuade the people who don’t experience nonresponse in previous wave to increase response rate. At the same time, we also pay attention to a nonrespondents group, because this group has differential characteristics from the respondents group. 15 REFERENCE Heeringa, S. G., D. H. Hill, D. A. Howell, 1995, “Unfolding Brackets for Reducing Item Nonresponse in Economic Surveys,” Panel Study of Income Dynamics Technical Paper Series #95-01, University of Michigan. Hoynes, H., M. Hurd, and H. Chand, 1998. “Household Wealth of the Elderly Under Alternative Imputation Procedures.” In David Wise (ed.), Inquiries in the Economics of Aging, PP. 229-254. Chicago: University of Chicago Press. Juster, F. T. and J. P. Smith, 1997, “Improving the Quality of Economic Data: Lessons from the HRS and AHEAD,” Journal of the American Statistical Association 92(440): 1268-1278 Juster, T. F., H. Cao, M. Perry, and M. Couper, 2006, “The Effect of Unfolding Brackets on the Quality of Wealth Data in HRS,” Working Paper 2006-113. University of Michigan. Kennickell, A. B. 1997. Using Range Techniques with CAPI in the 1995 Survey of Consumer Finance. KLI, 2010, User’s Guide, Korean Labor Institute. (www. Kil.re.kr) Moore, J.C., L.L. Stinson, and E. J. Welniak, 1999, “Income Reporting in Surveys: Cognitive Issues and Measurement Error,” In Cognition and Survey Research, M. Sirken, D. J. Heermann, S. Schechter, N. Schwartz, J. M. G. Tanur, and R. Tourangeau (eds). New York: Wiley. Moore, J.C., L.L. Stinson, and E. J. Welniak, 2000, “Income Measurement Error in Surveys: A Review,” Journal of Official Statistics 16(4): 331-361. Riphahan, R. T. and O. Serfling, 2005, “Item Nonresponse on Income and Wealth Questions,” Empirical Economics 30(2): 521-538. Schrapler, J. 2006, “Explaining Income Nonresponse - A Case Study by Means of the British Household Panel Survey,” Quality and Quantity 40 : 1013-1036. Yan, T., R. Curtin, and M. Jans, 2010, “Trends in Income Nonresponse Over Two Decades,” Journal of Official Statistics 26(1): 145-164. 16 <Table 1> Response Patterns of Real Estate Asset Items by Wave Real estate other than housing Owners Initial Nonresponse Wave Owners Exact Data Range Data Response Response Final Nonresponse Freq. % Freq. % Freq. % Freq. % 2 947 21.0 279 29.5 622 65.6 46 4.9 3 833 19.5 238 28.6 586 70.3 9 1.1 4 931 21.9 316 33.9 597 64.2 18 1.9 5 942 21.9 283 30.0 649 68.9 10 1.1 6 941 20.5 238 25.3 693 73.6 10 1.1 7 988 20.7 439 44.4 515 52.1 34 3.4 8 1096 22.6 600 54.7 481 43.9 15 1.4 9 1143 22.9 654 57.2 475 41.6 14 1.2 10 1193 23.5 700 58.7 484 40.5 9 0.8 11 1202 23.5 643 53.5 557 46.3 2 0.2 Real estate leasing including housing 2 430 9.5 382 88.8 37 8.6 11 2.6 3 550 12.9 458 83.3 82 14.9 10 1.8 4 597 14.1 503 84.3 84 14.0 10 1.7 5 618 14.4 518 83.8 97 15.7 3 0.5 6 604 13.2 535 88.6 65 10.7 4 0.7 7 623 13.1 598 96.0 15 2.4 10 1.6 8 738 15.2 719 97.4 11 1.5 8 1.1 9 801 16.0 784 97.9 13 1.6 4 0.5 10 812 16.0 796 98.0 15 1.9 1 0.1 11 804 15.7 798 99.3 6 0.7 0 0.0 Real estate renting other than housing 2 605 13.4 384 63.5 195 32.2 26 4.3 3 790 18.5 551 69.7 221 28.0 18 2.3 4 584 13.8 425 72.8 127 21.7 32 5.5 5 632 14.7 535 84.7 81 12.8 16 2.5 6 620 13.5 476 76.8 125 20.2 19 3.1 7 660 13.9 554 83.9 63 9.5 43 6.5 8 748 15.4 668 89.3 59 7.9 21 2.8 9 659 13.2 630 95.6 26 3.9 3 0.5 10 713 14.1 686 96.2 18 2.5 9 1.3 11 626 12.2 619 98.9 7 1.1 0 0.0 <Table 2> Nonresponse Rate of the Housing Asset Items in the KLIPS (percent) w2 w3 w4 w5 w6 w7 w8 w9 w 10 w 11 Item1 12.2 12.3 9.6 9.3 5.9 4.5 4.8 3.7 3.3 0.9 Item2 4.6 4.2 4.2 3.8 2.0 1.0 1.2 0.9 0.5 0.2 Item3 5.7 5.0 3.2 3.6 2.0 0.9 0.7 0.9 0.5 0.1 Note: The KLIPS asks Item 1 to housing owner, Item 2 to yearly renter, and Item 3 to monthly renter. 17 <Table 3> The Characteristics of the Respondents and Nonrespondents : Total Real Estate other than Housing Item Wave2 Exact Data Range Data Nonrespondents Exact Data - Range Data Respondents Respondents Respondents N Mean S.E N Mean S.E N Mean S.E t d.f. sig. Diff. S.E. Male 278 1.0 0.012 617 0.9 0.012 46 0.8 0.057 3.496 767.8 0.000 0.1 0.017 Age 278 49.6 0.705 617 51.6 0.498 46 57.4 1.960 -2.266 893.0 0.024 -2.0 0.880 Education 278 3.1 0.094 617 2.7 0.066 46 2.2 0.281 3.424 893.0 0.001 0.4 0.117 Married 278 0.9 0.016 617 0.9 0.013 46 0.8 0.061 2.450 660.8 0.015 0.1 0.021 Employed 279 0.8 0.025 622 0.8 0.016 46 0.7 0.071 -0.833 899.0 0.405 0.0 0.029 Number of HH 279 3.6 0.074 622 3.6 0.053 46 3.3 0.234 0.624 570.6 0.533 0.1 0.091 Total Income 278 2922.8 183.147 619 2458.8 96.806 45 2108.7 336.937 2.448 895.0 0.015 464.0 189.534 Wave7 Male 432 0.9 0.013 506 0.9 0.014 34 0.8 0.070 1.222 933.9 0.222 0.0 0.019 Age 432 52.0 0.581 506 54.6 0.557 34 56.6 2.031 -3.169 936.0 0.002 -2.6 0.807 Education 432 3.1 0.076 506 2.7 0.076 34 2.3 0.258 3.556 930.2 0.000 0.4 0.108 Married 432 0.9 0.015 506 0.8 0.016 34 0.8 0.070 2.004 935.7 0.045 0.0 0.022 Employed 439 0.8 0.020 515 0.7 0.019 34 0.7 0.079 1.157 940.0 0.247 0.0 0.027 Number of HH 439 3.3 0.056 515 3.2 0.059 34 3.0 0.223 1.171 951.2 0.242 0.1 0.082 Total Income 430 4398.5 222.202 506 3829.5 233.909 34 2836.0 438.417 1.745 934.0 0.081 568.9 326.104 Wave11 Male 639 0.9 0.013 548 0.9 0.015 2 0.5 0.500 1.349 1119.8 0.178 0.0 0.019 Age 639 53.4 0.505 548 56.8 0.571 2 42.5 5.500 -4.491 1185.0 0.000 -3.4 0.760 Education 639 3.3 0.062 548 2.9 0.073 2 4.5 1.500 3.556 1185.0 0.000 0.3 0.095 Married 639 0.9 0.014 548 0.8 0.016 2 0.5 0.500 1.921 1109.8 0.055 0.0 0.021 Employed 643 0.8 0.016 557 0.7 0.019 2 1.0 0.000 2.189 1141.7 0.029 0.1 0.025 Number of HH 643 3.1 0.047 557 3.0 0.054 2 3.0 1.000 2.085 1198.0 0.037 0.1 0.071 Total Income 643 5287.7 213.076 557 4534.3 185.584 2 2754.0 1014.000 2.627 1198.0 0.009 753.4 286.790 Note: Education: 0 no schooling, 1 elementary school, 2 middle school, 3 high school 4 college 5 university, 6 master degree, 7 doctoral degree 18 <Table 4> Asset Distribution of the Exact Data Respondents and the Range Data Respondents: Real Estate Asset other than Housing Item 1 2 3 4 5 6 7 8 9 10 11 total mean Exact 2.1 10.1 14.3 20.6 4.2 27.3 9.2 5.5 2.9 2.1 1.7 100.0 13,940.3 Range 8.0 14.8 18.3 10.8 13.0 20.5 7.5 2.2 1.7 2.4 .9 100.0 9,453.5 w3 Total 6.3 13.5 17.1 13.6 10.4 22.5 8.0 3.2 2.1 2.3 1.1 100.0 10,749.4 Exact 4.4 6.0 13.3 17.1 6.0 32.0 7.6 3.8 2.2 4.4 3.2 100.0 16,910.5 w4 Range 11.2 13.6 19.6 15.7 14.7 14.2 5.0 2.2 1.2 1.7 .8 100.0 10,077.9 Total 8.9 11.0 17.4 16.2 11.7 20.4 5.9 2.7 1.5 2.6 1.6 100.0 12,442.7 Exact 2.8 4.9 11.3 18.0 7.8 25.4 15.5 4.2 3.5 3.9 2.5 100.0 17,885.6 w5 Range 7.6 10.8 17.1 16.5 11.9 19.1 7.2 2.8 2.3 3.5 1.2 100.0 13,514.3 Total 6.1 9.0 15.3 17.0 10.6 21.0 9.8 3.2 2.7 3.6 1.6 100.0 14,841.6 Exact 1.3 8.0 11.8 13.4 7.1 25.2 14.3 7.6 3.4 6.3 1.7 100.0 18,459.7 w6 Range 9.4 12.1 19.3 10.7 9.4 17.2 7.8 3.0 3.2 4.9 3.0 100.0 16,195.2 Total 7.3 11.1 17.4 11.4 8.8 19.2 9.5 4.2 3.2 5.3 2.7 100.0 16,774.1 Exact 1.8 7.3 12.1 11.6 6.2 22.8 15.5 7.5 4.1 8.2 3.0 100.0 20,736.7 w7 Range 11.1 11.7 14.0 11.1 10.9 17.9 7.6 4.7 3.1 5.0 3.1 100.0 16,921.8 Total 6.8 9.6 13.1 11.3 8.7 20.1 11.2 6.0 3.6 6.5 3.0 100.0 18,677.3 Exact 1.3 5.7 13.0 14.2 4.3 25.2 13.7 6.7 4.7 8.0 3.3 100.0 21,843.7 w8 Range 9.1 12.9 14.3 10.0 9.1 16.2 9.4 6.4 1.7 5.8 5.0 100.0 19,364.3 Total 4.8 8.9 13.6 12.3 6.5 21.2 11.7 6.6 3.3 7.0 4.1 100.0 20,740.5 Exact 2.1 4.4 8.3 12.4 4.9 23.5 15.1 8.4 5.4 11.2 4.3 100.0 28,029.3 w9 Range 5.1 13.7 17.7 13.3 9.9 19.6 6.7 2.7 2.9 4.0 4.4 100.0 16,948.9 Total 3.4 8.3 12.2 12.8 7.0 21.9 11.6 6.0 4.3 8.1 4.3 100.0 23,367.5 Exact 2.4 4.9 7.9 11.7 5.0 24.6 12.3 7.1 5.1 11.4 7.6 100.0 30,885.5 w Range 5.8 7.0 17.1 12.8 14.5 15.7 8.9 4.8 3.3 4.8 5.4 100.0 19,524.3 10 Total 3.8 5.7 11.7 12.2 8.9 20.9 10.9 6.2 4.4 8.7 6.7 100.0 26,241.2 Exact .2 4.5 8.6 10.4 5.3 23.5 12.0 9.3 5.3 14.3 6.7 100.0 33,512.6 w Range 6.5 8.4 13.1 12.4 11.5 19.2 9.0 4.5 4.8 5.6 5.0 100.0 20,530.1 11 Total 3.1 6.3 10.7 11.3 8.2 21.5 10.6 7.1 5.1 10.3 5.9 100.0 27,486.5 Note: The response category of range question is followed. 1: less than 10 million KRW, 2: 10 to less than 25, 3: 25 to less than 50, 4: 50 to less than 75, 5: 75 to less than 100, 6: 100 to less than 200, 7: 200 to less than 300 , 8: 300 to less than 400, 9: 400 to less than 500, 10: 500 million to less than 1 billion KRW, 11: 1 billion KRW or more 19 <Table 5> Response Tendency in the Next Wave: Total Real Estate other than Housing Item (n+1) wave Owner(100%) Respondent (100%) w2 w3 w4 w5 w6 w7 w8 w9 w 10 Exact Data Respondents Range Data Respondents Nonowner Exact Data Respondents Range Data Respondents Nonowner Exact Data Respondents Range Data Respondents Nonowner Exact Data Respondents Range Data Respondents Nonowner Exact Data Respondents Range Data Respondents Nonowner Exact Data Respondents Range Data Respondents Nonowner Exact Data Respondents Range Data Respondents Nonowner Exact Data Respondents Range Data Respondents Nonowner Exact Data Respondents Range Data Respondents Nonowner Size 279 622 3,561 238 586 3,433 316 597 3,316 283 649 3,356 238 693 3,651 439 515 3,774 600 481 3,753 654 475 3,858 700 484 3,876 Response 251 90.0 563 90.5 3,139 88.1 92.9 221 92.8 544 88.1 3,025 91.5 289 95.1 568 91.6 3,039 266 94.0 630 97.1 3,073 91.6 233 97.9 670 96.7 3,380 92.6 427 97.3 493 95.7 3,512 93.1 575 95.8 467 97.1 3,528 94.0 637 97.4 461 97.1 3,640 94.3 672 96.0 466 96.3 3,640 93.9 Nonown 87 34.7 206 36.6 2,867 91.3 26.7 59 30.0 163 89.7 2,712 36.3 105 26.6 151 91.0 2,764 108 40.6 185 29.4 2,822 91.8 81 34.8 174 26.0 3,095 91.6 108 25.3 110 22.3 3,176 90.4 136 23.7 123 26.3 3,234 91.7 144 22.6 123 26.7 3,332 91.5 139 20.7 121 26.0 3,349 92.0 Own 164 65.3 357 63.4 272 8.7 73.3 162 70.0 381 10.3 313 63.7 184 73.4 417 9.0 275 158 59.4 445 70.6 251 8.2 152 65.2 496 74.0 285 8.4 319 74.7 383 77.7 336 9.6 439 76.3 344 73.7 294 8.3 493 77.4 338 73.3 308 8.5 533 79.3 345 74.0 291 8.0 Exact Data Response 67 40.9 82 23.0 77 28.3 53.7 87 24.7 94 31.0 97 46.2 85 18.9 79 36.0 99 80 50.6 67 15.1 67 26.7 105 69.1 166 33.5 147 51.6 235 73.7 143 37.3 197 58.6 321 73.1 122 35.5 176 59.9 369 74.8 112 33.1 186 60.4 357 67.0 125 36.2 146 50.2 Initial Nonresponse Range Data Final Response Nonresponse 97 59.1 0 0.0 271 75.9 4 1.1 192 70.6 3 1.1 45.7 74 1 0.6 73.8 281 6 1.6 66.5 208 8 2.6 52.7 97 2 1.1 80.1 334 4 1.0 63.3 174 2 0.7 76 48.1 2 1.3 376 84.5 2 0.4 180 71.7 4 1.6 43 28.3 4 2.6 312 62.9 18 3.6 127 44.6 11 3.9 84 26.3 0 0.0 229 59.8 11 2.9 136 40.5 3 0.9 116 26.4 2 0.5 218 63.4 4 1.2 113 38.4 5 1.7 123 24.9 1 0.2 221 65.4 5 1.5 120 39.0 2 0.6 176 33.0 0 0.0 220 63.8 0 0.0 143 49.1 2 0.7 20 <Table 6> Logistic Regression: Total Real Estate other than Housing Item Independent var.: Experience of range data response ahead of Independent var.: Time of one wave Cumulative frequency of range data response Dependent Variable B S.E. Sig. Exp(B) B S.E. Sig. Exp(B) w3 0.696 0.204 0.001 2.006 0.696 0.204 0.001 2.006 w4 1.199 0.206 0.000 3.315 0.616 0.139 0.000 1.852 w5 1.308 0.198 0.000 3.698 0.576 0.106 0.000 1.779 w6 1.680 0.212 0.000 5.364 0.518 0.088 0.000 1.679 w7 1.282 0.200 0.000 3.605 0.118 0.060 0.048 1.125 w8 1.391 0.165 0.000 4.021 0.345 0.052 0.000 1.412 w9 1.542 0.160 0.000 4.673 0.252 0.043 0.000 1.286 w 10 1.746 0.161 0.000 5.733 0.281 0.040 0.000 1.324 w 11 1.224 0.149 0.000 3.402 0.207 0.034 0.000 1.230 Note: Socio-demographic variables of same wave with an independent variable were controlled. A dependent variable is nonresponse to open-ended questions asking exact value in a current wave. 21 <Figure 1> A Flow of the Asset Question 22 <Figure 2> Reduced Nonresponse Rate by the Range Question: Real Estate other than Housing Item 80.0 70.0 60.0 % 50.0 40.0 30.0 20.0 10.0 0.0 2 3 4 5 6 7 8 9 10 11 Wave Initial nonresponse Final nonresponse <Figure 3> Nonresponse Rate to the Open-ended question in the Next Wave by Response Group Type 90.0 Range Data Respondents 80.0 70.0 nonowner 60.0 % 50.0 40.0 Exact Data Respondents 30.0 20.0 10.0 0.0 w3 w4 w5 w6 w7 w8 w9 w10 w11