PREFACE 2002 NSAF Nonresponse Analysis is the seventh report in a series describing the methodology of the 2002 National Survey of America’s Families (NSAF). The NSAF is part of the Assessing the New Federalism project at the Urban Institute, conducted in partnership with Child Trends. Data collection for the NSAF was conducted by Westat. The NSAF is a major household survey focusing on the economic, health, and social characteristics of children, adults under the age of 65, and their families. During the third round of the survey in 2002, interviews were conducted with over 40,000 families, yielding information on over 100,000 people. The NSAF sample is representative of the nation as a whole and of 13 states, and therefore has an unprecedented ability to measure differences between states. About the Methodology Series This series of reports has been developed to provide readers with a detailed description of the methods employed to conduct the 2002 NSAF. The 2002 series of reports includes the following: No. 1: No. 2: No. 3: No. 4: No. 5: No. 6: No. 7: No. 8: No. 9: No. 10: No. 11: No. 12: An overview of the NSAF sample design, data collection techniques, and estimation methods A detailed description of the NSAF sample design for both telephone and inperson interviews Methods employed to produce estimation weights and the procedures used to make state and national estimates for Snapshots of America’s Families Methods used to compute and results of computing sampling errors Processes used to complete the in-person component of the NSAF Collection of NSAF papers Studies conducted to understand the reasons for nonresponse and the impact of missing data Response rates obtained (taking the estimation weights into account) and methods used to compute these rates Methods employed to complete the telephone component of the NSAF Data editing procedures and imputation techniques for missing variables User’s guide for public use micro data 2002 NSAF questionnaire i About This Report Report No. 7 describes analysis conducted to gain some insight into the characteristics of nonrespondents to the 2002 NSAF, and to assess the impact of nonresponse on the NSAF statistics. This analysis includes a detailed breakdown of the call attempts to determine how effective the sample was worked and an investigation into the level of effort both in terms of refusal conversion and call attempts. In addition, a sociodemographic comparison of nonrespondents was done using census block information obtained for nonrespondents and respondents. For More Information For more information about the National Survey of America’s Families, contact: Assessing the New Federalism Urban Institute 2100 M Street, NW Washington, DC 20037 E-mail: nsaf@ui.urban.org Web site: http://anf.urban.org/nsaf Tim Triplett ii CONTENTS Chapter 1 Page INTRODUCTION ............................................................................................... 1-1 Demographic Comparisons of Hard-to-Interview Respondents.......................... 1-1 Recent Trends and Developments ....................................................................... 1-3 2 NSAF RESPONSE RATE OVERVIEW............................................................. 2-1 3 EVALUATION OF THE CALL ATTEMPTS EFFORT.................................... 3-1 4 EVALUATION OF THE EFFORTS TO CONVERT REFUSALS ................... 4-1 5 SOCIODEMOGRAPHIC STUDY OF SURVEY NONRESPONDENTS......... 5-1 Comparison of Screener Respondents versus Nonrespondents ........................... 5-2 Comparison of Extended Interview Respondents versus Nonrespondents.......... 5-5 Conclusions.......................................................................................................... 5-8 6 SUMMARY......................................................................................................... 6-1 REFERENCES ................................................................................................................R-1 Tables Table 2-1 3-1 3-2 4-1 5-1 5-2 5-3 Page National Response Rates, by Round.................................................................... 2-1 Call Attempt Regression Output.......................................................................... 3-6 Across-Round Call Attempt Comparisons........................................................... 3-7 Initial Refusal Regression Output........................................................................ 4-4 Household Screener Response, Refusal, and Other Nonresponse Rates by Neighborhood Characteristics......................................... 5-3 Household Extended Response, Refusal, and Other Nonresponse Rates by Neighborhood Characteristics......................................... 5-6 Linear Regression ................................................................................................ 5-8 Charts Page Chart 3A 3B 3C 4A Call Attempt Efforts and Adult Response Rates.................................................. 3-1 Screener: Average Number of Call to Complete ................................................. 3-3 Extended: Average Number of Call to Complete ................................................ 3-3 Adult Response Rates Before and After Refusal Conversion Attempts.............. 4-2 iii 1. INTRODUCTION This report analyses the characteristics of nonrespondents to the National Survey of America’s Families (NSAF). Conducted in 1997, 1999 and 2002, the NSAF provides a comprehensive look at the well-being of children and nonelderly adults, and it reveals sometimes striking differences among the 13 states studied in depth. The survey provides quantitative measures of child, adult, and family well-being in America, with an emphasis on persons in low-income families. To garner respondents living with and without household landline telephones, the NSAF used a random-digit dial (RDD) sample of telephone households and an area sample of nontelephone households. While numerous survey strategies were employed to reduce the number of nonrespondents, there was a steady decline in response rates throughout the three years of the NSAF data collection. As a result, this paper assesses the impact of nonresponse on the NSAF statistics, particularly the 2002 estimates. The next section of this report provides more information on the NSAF response rates, while this introduction reviews the nonresponse literature. Response rates in many respects measure people’s willingness to participate in a study. The most important factor in getting a good response rate is making additional contact attempts (Newcomer and Triplett 2004). Many organizations use a variety of contact attempt and refusal conversion strategies for each of their studies. There is no consensus, however, among survey organizations on the maximum number of contact attempts or the amount of refusal conversion needed. Efforts to reduce the number of nonrespondents in a study are usually related to two factors: budget and time. Given a limited budget, making a large number of contact attempts or attempting to convert refusals can prove economically unfeasible. Similarly, a study that must be completed in a short period may not allow for enough time to make a large number of attempts. Numerous studies have investigated the issue of how much additional effort an organization should make in attempting to reduce nonresponse. They support Kish’s (1965) dicta that new responses must be numerous enough to justify the effort and that decreasing the proportion of nonresponse is important only if it also reduces its effect. The first assumption is an issue of cost and is addressed in this paper by looking at the costs of calling back telephone numbers versus using additional RDD telephone numbers to obtain the required number of completed interviews. The second assumption is an issue of whether the respondents reached after multiple call attempts or refusal conversion differs from respondents never refusing to cooperate. In sections 3 and 4, it will be become evident that those reached after refusal conversion and multiple call attempts do indeed differ. Demographic Characteristics of Hard-to-Interview Respondents Research by James Massey and colleagues (1981) found that respondents who initially refuse, but later in the study complete an interview are disproportionately persons 65 years or older. This study also found that male respondents were more difficult to reach than female respondents. William C. Dunkelberg and George S. Day (1973) found that the first attempt to contact respondents yields only 25 to 30 percent of the final sample. They also observed a rapid decline in completing interviews after three attempts. Dunkelberg and Day (1973), however, found that over 20 percent of interviews required more than four attempts. They also found that the 1- 1 demographic characteristics of respondents reached on the first few contacts differed from those found at home later in the interviewing process. Specifically, they note that younger adults, higher-income adults, and respondents from larger cities require more contact attempts. The findings from their research are consistent with the findings in this report. What makes that somewhat surprising is that Dunkelberg and Day’s (1973) research was based on “personal” interview data from the 1967 and 1968 “Survey of Consumer Finances” conducted by the University of Michigan—more than 35 years ago. Research on telephone studies has consistently shown that with more call attempts the final sample becomes younger and includes higher proportions of male and black respondents (Blair and O’Rourke 1986; Traugott 1987; Merkle 1993; Shaiko 1991; Triplett 2002). The data analyzed in this report support these findings. Some of the call attempt research has found that highly educated respondents require more call attempts, but this finding is not consistent with the results from the analysis done in this report. A report from the 1986 U.S. National Crime Survey did not find differences in gender and race between the initial households and the follow-up households in which there were unanswered telephone numbers (Sebold 1988). The unanswered telephone group could explain the reason for the non-finding of differences by gender and race. The U.S. National Crime Survey, however, did find that the initial interviews contained proportionally more respondents 65 years or older. The NSAF was conducted in 1997, 1999, and 2002, while many of the studies reviewed for this research were from telephone studies conducted in the 1970s and 1980s. However, results regarding the number of calls needed to complete an interview, difficulties reaching certain demographic groups, and percentage of the sample completed in the first three calls have maintained similar patterns. These results are quite similar to Groves and Kahn’s (1979) findings in their comparison of telephone versus personal interviews. Over time, however, the number of call attempts needed to reach 18- to 24-year-old respondents has increased significantly; nonwhite respondents are also becoming harder to contact. In addition, fewer interviews are now completed in the first three call attempts. Another clear difference with telephone studies prior to the 1990s is that large number of telephone numbers that were never answered (categorized at ring no-answer) have subsequently been replaced by an increase in telephone numbers that are coded as having reached an answering machine. This paper does not examine this issue in detail, but the results tend to support the hypothesis that answering machines have not had much effect on telephone surveys. Basically, home recorders seem to have substituted for the ring no-answer group. Since we can often determine the residential status of most answering machines, the answering machine has improved our ability to estimate response rates. Several papers look more closely at the effects of home recorders; raising some concerns over the effect home recorders may have on the representativeness of general population samples (Oldendick 1993; Piazza 1993; Tuckel and O’Neil 1996; Link and Oldendick 1999). It took approximately nine months to complete data collection for each round of the NSAF, always starting in February. Therefore, there was little or no data collected in the months of November, December, and January. Prior research has shown that response rate varies depending on which month you are calling (Steeh et al. 1983). A more recent study, however, has found that the time of the year the data collection takes place has little effect on the sample efficiency and response rates (Losch et al. 2002). Results from previous call attempt research (Triplett 2002) have shown that while studies conducted in the winter and summer do not differ 1- 2 significantly in terms of response rates, they do provide a few unique call attempt results (such as reaching young people is easier in winter and reaching black respondents is easier in summer). Research on call attempts is a necessary procedure for reducing nonresponse. In determining the optimal number of call attempts, one should be aware of the work that Michael Weeks and James T. Massey have done in determining the optimal times to contact sample households (Weeks et al. 1980; Massey et al. 1996). This work has led to the development of optimal time scheduling for telephone surveys (Weeks et al. 1987; Greenberg and Stokes 1990). To determine the optimal number of call attempts for a general population random digit dialing study, a survey research operation should be aware of the optimal times to reach the majority of its sample. It is from this work that we get clear data showing the advantages of contacting respondents during the evening on weekdays or on weekends rather than calling during the day on weekdays. By following the recommendations of this research, an organization can begin to minimize the overall call attempts needed on a study. Refusal conversion increases response rates and usually changes the final demographic sample distribution. Several studies, however, show issues related to data quality when comparing the information provided by reluctant respondents (respondents that initially refused to participate in a study) to respondents who never refused to cooperate (Cannell and Fowler 1963; Blair and Chun 1992; Lavrakas et al. 1992; Triplett et al. 1996; Teitler et al. 2003). Therefore, in determining the appropriate refusal conversion effort, an organization needs to consider more than just the response rate. While these studies do not argue for the elimination of refusal conversion, they do raise a concern over the accuracy of reporting on difficult survey questions. Refusal conversion involves waiting for a better time to call and try again (Groves and Couper 1998). One study argues that the optimal number of days to wait is approximately one week for most refusals, and two weeks for refusals where the actual respondent refused the interview (Triplett et al. 2001). The need to wait before attempting refusal conversion makes it more difficult to convert refusals with people who refuse late in the study or are hard to contact by phone. Recent Trends and Developments In the most recent decade, the telephone system in the United States has undergone rapid change, making it more difficult for survey research firms using RDD sampling methods to identify residential households (Roth et al. 2001; Triplett and Abi-Habib 2005). As the telecommunication world changes it is important to readdress the efficiency of RDD sampling (Tucker et al. 2002). Since 2002, new laws allow cellular phone subscribers to transfer their landline phone numbers over to their cell phones. There is also a general increase in the number of households that use only cellular phones. As a result, future telephone nonresponse analysis will need to consider telecommunication coverage issues. Fortunately for the 2002 NSAF, the estimated coverage rate (percentage of households that had a non-cellular phone number or landline phone number) for the RDD sample was higher than 95 percent.1 Thus, this report does not address these coverage issues. 1 Federal Communications Commission, Telephone Subscribership in the United States: Data through Nov 2002. Released April 2003.http://www.fcc.gov/Bureaus/Common_Carrier/Reports/FCC-State_Link/IAD/subs1102.pdf. 1- 3 Another recent trend that also held true for the NSAF (Brick et al. 2003) is a decline in survey response and, in particular, telephone survey response rates (Steeh et al. 2001; Newcomer and Triplett 2004; Curtin et al. 2005). The current trend is much worse than any other historic period of falling response rates (Steeh 1981; Curtin et al. 2005). What is not clear is the effect of this decline in response rates on our survey estimates. The main purpose of this report is to begin to answer that question. Understanding the effect of nonresponse on the NSAF will be a good indicator of nonresponse issues in general. At the same time, the NSAF has a number of unique sample design features that makes it hard to generalize some of this study’s findings. The next section provides an overview of the NSAF. More information about the survey is available in other reports in the NSAF methodology series, particularly Report No. 1, 2002 NSAF Survey Methods and Data Reliability. 1- 4 2. NSAF RESPONSE RATE OVERVIEW The NSAF sample had two parts. The main sample consisted of a random-digit dial survey of households with telephones and a supplementary area probability sample of households without telephones. In both the RDD and area samples, interviews were conducted in two stages. First, a short, five-minute screening interview was conducted to determine household eligibility. Once the household was deemed eligible, a second, more detailed, 27- to 50-minute extended interview was conducted to ask about the main survey items of interest. Since the NSAF 2002 area sample was quite small (n = 649) and had a relatively high response rate (79.4 percent), this nonresponse analysis will focus on the RDD telephone component. The telephone component of the 2002 NSAF used a list-assisted method to select the RDD sample of telephone numbers, and it used computer-assisted interviewing (CATI) for screening and interviewing. Featuring large sample sizes in each of 13 states with additional sample drawn from the balance of the nation, the NSAF allowed researchers to produce national estimates. The sample of telephone households that completed the screener survey were subsampled for inclusion in the sample of households needed to complete the extended survey. The subsampling rates varied depending on the state, the presence of children in the household, and responses to income screening questions. All households with children that were classified as low-income households were sampled, while higher-income households with children and all households without children (but with someone under 65) were subsampled. The initial 2002 RDD sample consisted of 556,651 telephone numbers. Of the 556,651 telephone numbers called, 133,503 households were screened, and detailed extended telephone interviews were conducted in 39,220 households with 43,157 persons under the age of 65. Prior rounds of the NSAF yielded a few more extended interviews (46,798 in 1997 and 45,040 in 1999) that can be in part attributed to the lower third round response rate. The table below shows the NSAF response rates by round, with separate calculations for the screener interviews, extended adult interviews, and extended child interviews. In addition, this table shows how the response rate was much lower for the RDD telephone sample than the area probability sample. Table 2-1. National Response Rates, by Round Sample type All Screener Child extended Adult extended Round 1 Round 2 Round 3 77.8 84.1 76.9 76.7 81.4 77.5 66.0 83.5 78.7 RDD sample only Screener Child extended Adult extended 77.4 83.4 79.4 76.3 80.5 77.0 65.5 83.0 78.2 Area sample only Screener Child extended Adult extended 87.0 95.7 92.5 89.2 96.2 93.0 81.6 95.0 92.5 2- 1 The most notable thing from this table is the decline in the screener response rate for the third round of data collection. This occurred despite additional efforts made during the third round to improve response rates. The decline in the screener response rate in 2002 did mirror the issues and problems with declining responses other similar telephone surveys were experiencing. There were a number of small changes in the survey design between rounds that could have had some impact on response rates. Only a change in the incentive strategy, however, had as its primary goal an overall improvement in response rates. This change involved including a $2 bill with the advanced letter or offering the $2 to those telephone households in which an address was not obtained. Unfortunately, there was only a slight gain from the $2 advance and it was not a statistically significant gain. The strategy to sample refusals in the third round also had a positive impact on the responses rates by allowing only more experienced interviewers to work on refusal conversion efforts. However, this gain was also small and not statistically significant. Finally, only one reverse address matching service was used to obtain mailing addresses based on sampled telephone numbers in both rounds 1 and 2, whereas three services were used in round 3. This increased the percentage of listed addresses in which to send advanced letters. However, sending out a greater percentage of advance letters did not significantly affect the response rates, in part due to the lower reliability of the match that was found from alternative sources. 2- 2 3. EVALUATION OF THE CALL ATTEMPTS EFFORT Making additional call attempts is an integral part of any effort to reduce nonresponse. For instance, if we used a maximum of 15-call attempts cutoff rule on the NSAF, response rates in all three rounds of data collection would have been less than 50 percent. Simply going from three to five call attempts improved the response rate 12.1, 18.1, and 14.2 percentage points in 1997, 1999, and 2002, respectively. Going from 10 to 15 call attempts increased the response rate 16.8, 13.1, and 12.5 percentage points in 1997, 1999, and 2002, respectively. Interestingly, there is not a perfect correlation between the response rate after five calls and the final response rate. For example, after five call attempts, the 1999 study had by far the highest response rate, with 2002 and 1997 roughly equal. Of the first and third round, however, the 1997 survey had the highest final response rate. This demonstrates that it is possible to accomplish a good final response rate on studies that do not get off to a good start. Chart 3A shows, for each round of data collection, how much response rates increase when additional call attempts are made. There was no firm maximum call attempt rule for the NSAF. Many households received more than 60 call attempts. In all three rounds, the data consistently shows that after 20 call attempts, there was a significant decline in additional call attempts’ success in improving the response rate. However, the final response rates for the NSAF increased on average 4 percentage points from making more than 20 call attempts. In contrast, after 15 call attempts, an additional five call attempts increased the NSAF response rate between 6 and 8 percentage points. From a purely cost-benefit perspective, it would seem that a cutoff point of 20 call attempts were reasonable. If 3- 1 a cutoff rule was imposed, however, additional sample would have had to be screened to reach the desired sample size. Since achieving high response rates on the screener interview was a much more difficult task, increasing call attempts beyond 20 calls to complete already screened households was the appropriate strategy. In addition, the study had a lengthy, nine-month data collection period, so it was not only possible to make more than 20 calls to hard-to-reach households but also to spread out these call attempts over different times and days. Finally, this research shows that the interviews completed after 20 call attempts differ from those that were completed with less than 20 calls, which again supports the decision to not impose a maximum call rule. What are the characteristics of people who are difficult to reach in telephone surveys? As with most RDD samples, male respondents required more call attempts to complete an interview than female respondents (Triplett 2002). Similar to the 10-year trend discovered in prior national call attempt research, the 2002 NSAF indicates that women are becoming just as difficult to reach as men (Triplett 2002). While respondents from the Northeast required more contacts on average to complete, it was the state of Florida whose respondents required the most call attempts to complete on the NSAF. Asian respondents required more call attempts, as did foreign-born residents. For both the 1997 and 1999 NSAF study, education of the respondent was not a factor in determining how easy or difficult a person was to reach. In the 2002 study, however, lesseducated respondents did require more call attempts to complete an interview. It also took more call attempts to complete interviews with respondents who have never been married or who are married but are currently separated or do not have a spouse in the household. As expected, reaching respondents who had full-time jobs required more call attempts than interviewing those not employed. Finally, interviewing respondents age 18 to 24 required significantly more call attempts. A somewhat surprising result is that other than finding it hard to contact respondents whose family income was below 50 percent of the federal poverty level, income was not an issue in terms of call attempts needed to complete the interview. Chart 3B shows the average number of call attempts it took to complete an interview with people who were more difficult to reach in order to complete the 2002 NSAF screener survey. Chart 3C shows groups that were harder to reach in order to complete the 2002 NSAF extended survey. For the most part, the same groups that required more call attempts to complete the screener also required more call attempts to complete the extended interview. How does the number of call attempts alter the final demographic distribution? Completing interviews in households that require more call attempts increased the number of interviews completed in urban areas and households with children under 18. Additional call attempts did not, however, have any effect on reaching households that were at or below the federal poverty level. Also, increasing the number of call attempts increased the percentage of respondents without a high school degree and increased the percentage of black respondents interviewed. Completing interviews in households that were previously ring no answer increased the number of interviews in rural areas, single-adult households, and households with no children under 18. In addition, reducing the number of ring no answers remaining in the non-completed sample increased the percentage of less-educated respondents and respondents age 55 and older interviewed. 3- 2 3- 3 What about the cost of additional call attempts? By far the largest cost associated with telephone data collection is paying for the interviewers’ and supervisors’ time. Another significant cost is the phone charges. Both these cost factors are affected by how many telephone calls are made. To minimize costs, one could ignore response rate and try to minimize the total call attempts needed to complete the target number of interviews. Calling would stop when the probability of the next call attempt yielding an interview begins to fall. For the NSAF, this did not begin to occur until after the 20th call attempt. So from a purely cost-to-complete approach, one could argue to give up on completing an extended interview once you reach the 20th call attempt. The increase in nonresponse, had the NSAF design not permitted making more than 20 call attempts, would have been approximately 5 percentage points in 1997 and 4 percentage points in 1999. A better strategy for minimizing costs would be to allow call attempts beyond 20 for households in which a resident of that household requested that someone call back. This was, in fact, the strategy for the NSAF, where only a random sample of the numbers that were never answered was called more than 20 times for survival estimation.2 Telephone numbers that yielded a live contact, however, were never retired except in the event of multiple refusals or in situations in which the number stopped working. In round 1, 19,175 extended were completed (almost 20 percent of all completes) after 20 call attempts, while in round 3, 6,880 extended interviews were completed (16 percent of all completes) after 20 call attempts. The decline in the percentage needing more than 20 call attempts was a result of the implementation of a more efficient calling strategy that spread out the call attempts over a long period (see Cunningham et al. 2005). However, even 16 percent of the total completes occurring after 20 call attempts is quite large and certainly argues for continuing to try to reach identified residential telephone households well beyond the 20-call limit imposed on non-contacted sample. In fact, in 2002, 7.5 percent of all completes needed more than 30 call attempts, and 3.5 percent of completes needed more than 40 call attempts. This is why no upper call attempt limit was set in order to minimize nonresponse from family respondents who are difficult to reach. One problem with looking at call attempts needed to reach different demographic groups is that there is a known correlation between demographic groups. For instance, it took fewer call attempts to reach respondents with children and respondents who are married. Since there is a strong correlation between marital status and having children, it is difficult to tell which is more important in determining the average call attempts. One solution would be to create a multiple regression equation that would look at the joint effect of all the key demographic variables. In this paper, an ordinary least squares regression was run for the combined 1997 to 2002 NSAF data. The dependant variable was the total number of calls, and the independent variables were the demographic variables immigration status, marital status, census region, foreign-born status, age, gender, education, ethnicity, race, labor force status, poverty level, hours worked, and number in family of various age groups. Two other independent variables were the survey variables years of survey and year of survey. Means were substituted for missing data, and there were no missing data for total number of call attempts. The regression results are shown in table 3-1. The regression analysis shows that race and ethnicity were two of the most important factors in determining the number of call attempts with 2 Survival estimation was a method employed to approximate what percent of non-contacted telephone numbers were residential households. This estimate was important in calculating the final response rate. 3- 4 nonwhite and Hispanic respondents requiring more call attempts. In addition, the regression analysis showed that more call attempts were needed to reach unmarried respondents and employed respondents. Having family members over age 65 was also a factor in reducing the number of call attempts. Less important factors that still predicted more call attempts include being young, foreign born, working long hours, or having children age 6 to 17. Having young children 0 to 5 or being a male or female respondent was not a significant factor in estimating call attempts needed to complete the extended interview. 3- 5 Table 3-1. Call Attempt Regression Output Dependent variable: total number of calls needed to complete the interview Sample: all three rounds of the NSAF, 1997–2002 (R SQUARE=.161) Independent variables (Constant) B 335.56 Std. error Beta 26.72 T Sig T 12.56 0.00 UIMMSTAT Immigration status -0.606 0.180 -0.029 -3.366 0.001 UMARSTAT New marital status 0.224 0.017 0.044 12.860 0.000 -0.281 0.026 -0.029 -10.892 0.000 2.631 0.300 0.076 8.761 0.000 AGE Age -0.039 0.004 -0.042 -9.602 0.000 SEX Gender -0.083 0.065 -0.004 -1.284 0.199 0.277 0.050 0.016 5.574 0.000 UBETH Hispanic -2.041 0.102 -0.061 -19.949 0.000 UBRACE Race (2 category) -1.924 0.078 -0.068 -24.561 0.000 U_LFSR Labor force status recode 0.817 0.035 0.067 23.401 0.000 U_SOCPOV Social family income as % of poverty 0.144 0.028 0.017 5.039 0.000 U_USHRS Hours worked per week this year 0.027 0.002 0.030 11.027 0.000 UAGE1 No. family members age 0–5 -0.001 0.046 0.000 -0.018 0.985 UAGE2 No. family members age 6–17 0.144 0.030 0.015 4.856 0.000 UAGE3 No. family members age 18–24 -0.046 0.054 -0.003 -0.853 0.393 UAGE4 No. family members age 25–34 0.019 0.064 0.001 0.305 0.760 UAGE5 No. family members age 35–44 0.037 0.061 0.003 0.609 0.542 UAGE6 No. family members age 45–54 -0.092 0.062 -0.006 -1.475 0.140 UAGE7 No. family members age 55–64 -0.423 0.080 -0.020 -5.264 0.000 UAGE8 No. family members age 65+ -1.002 0.094 -0.029 -10.643 0.000 4.968 0.036 0.361 137.230 0.000 -0.162 0.013 -0.032 -12.101 0.000 UREGION Region U_USBORN Foreign- or U.S.-born UBCPSED Education level, CPS REFUSAL YEAR Some of the five-year NSAF call attempt trends (1997 to 2002) are shown in table 3-2. In the five years between the first and third rounds of the NSAF, the average number of call attempts needed to complete the screener interview almost doubled, while the number of calls needed to complete the extended interview declined. The increase in call attempts needed to complete the screener is consistent with the literature in what was happening to most RDD telephone studies during the same period. It is less clear, however, why there was a decline in the average call attempts needed to complete the NSAF extended interview. Some of the change could be explained by the use of a more efficient calling strategy (Cunningham et al. 2005). But that strategy involved how to reach households that do not answer the phone, which should have a greater impact on the screener interview. More likely the decline was a combination of two things: more people who are difficult to reach refused to complete the screener, thus not being eligible for the screener; and, since more call screener attempts were made, more information 3- 6 about when to call back to complete an extended was available. As shown in table 3-2, over 90 percent of households were screened by the fifth call attempt in rounds 1 and 2, whereas barely more than 70 percent of household were screened after five call attempts in round 3. On the bright side, there seems to have been some benefit to doing additional work on completing screener interviewers, in that it helped informed when to make the calls to complete the extended interviews. Table 3-2. Across-Round Call Attempt Comparisons Round 1, 1997 Mean number of call attempts Screener Extended Median number of call attempts Screener Extended Response rates Screener Extended (adults) Extended (child) % completed on first call attempt Screener Extended % completed on third call attempt Screener Extended % completed on fifth call attempt Screener Extended % completed on tenth call attempt Screener Extended Round 2, 1999 Round 3, 2002 2.61 10.91 2.21 11.12 4.95 9.58 2.00 7.00 1.00 7.00 3.00 6.00 77.4% 79.9% 84.1% 76.7% 77.5% 81.4% 66.0% 78.7% 83.5% 49.9% 0.7% 59.2% 0.0% 25.3% 0.0% 79.3% 19.7% 85.0% 24.6% 55.6% 27.3% 90.0% 38.1% 92.9% 41.3% 71.3% 46.5% 97.3% 65.2% 98.3% 65.3% 88.0% 70.5% 3- 7 4. EVALUATION OF THE EFFORTS TO CONVERT REFUSALS Refusal conversion in telephone surveys is a standard practice at most survey organizations and accounts for a significant percentage of the final sample. The rationale for refusal conversion is to increase response rate and hence reliability. But, in achieving this goal, we must also be alert to potential unintended effects on data quality. There have been analyses of how reluctant responders differ from others on the distribution of their answers to substantive questions, as well as how these two types of respondents compare demographically. An analysis of differences between responder groups in a large study reported by Lavrakas and others (1992) found some demographic differences. These lines of research, however, do not address the question of whether reluctant respondents may have other response behaviors that bear on the quality of the data obtained from them. Forty years ago, Cannell and Fowler (1963) found that reluctant respondents provided less accurate data. They attributed this effect mainly to lower respondent motivation. Citing this result some years later, Bradburn (1984) states the issue more generally, suggesting a possible effect of interviewer persistence on response behaviors. He asserted, “There are... a number of people who end up responding because they have given up trying to fend off the interviewer...[and]…go through the interview quickly—in other words do it but don’t work hard.” Of course, it may also be that these respondents who are reluctant to participate also simply have less interest in the survey topic (Groves et al. 2004). While it would be difficult to disentangle these possible effects, both are likely be in the direction of a decreasing effort on the part of respondents in answering questions. The type of question—for example, a simple yes-no item versus an open-ended question—may affect the amount of cognitive effort required. Effort may also vary by recall task, such as a question that asks about a simple attribute such as the respondent’s age versus asking for the respondent’s detailed medical history. In addition to these factors, effort may be affected simply by how motivated the respondent is to provide an answer. This reduced effort may be stated in terms of cognitive strategies respondents use. One cognitive strategy that a respondent may use is to provide the minimum response that will satisfy the interviewer and allow the interview to proceed, with the hope of ending it as quickly as possible. Krosnick and Alwin (1987) have termed this general behavior for minimizing cognitive effort “satisficing.” In a survey interview, this could result in such respondent behaviors as increased item refusals or “don’t know” responses, more primacy and recency effects in selecting from a list of response categories, and reduced completeness of answers to open-ended questions. The concern that respondents who initially refused the NSAF may provide less accurate data is fortunately less of a concern for the 2002 study than the prior two rounds. As seen in chart 4a, the overall percentage of completed interviews that were a result of converted refusals declined in 2002. This may seem odd given that the number of people refusing to complete the NSAF is much higher in 2002, but this can at in least in part be explained by the change in how monetary incentives were used. In 2002, the majority of households were sent a $2 bill in advance, instead of the prior-round strategy of sending or offering $5 only to people who refused the screener. So the expectation was that the advance payment would reduce refusals and those who do refuse would be more difficult to convert. 4- 1 As it turned out, providing $2 at the initial attempt to complete the screener works about as well on response rates as a $5 treatment at refusal conversion (Cantor et al. 2003). A benefit of the incentive is that it drew attention to the advance materials, as some post-interview questions found that people who received the $2 in advance were more likely to have read the advanced letter describing the project. It is unclear how much of the effect of the incentive adds to the perceived benefits of participating on the survey, but by having reduced the number of initial refusals, it is likely that the prepaid incentive used in round 3 increased the accuracy of respondents’ answers. How did refusal conversion alter the final NSAF demographic distribution? The extended refusal conversions efforts in all three rounds increased the percentage of male respondents relative to female respondents. While the refusal conversion of those who refused the screener had practically no impact on the final distribution of race, the extended refusal conversion efforts did increase the final number of black families interviewed in all three rounds. In the first two rounds of the NSAF, extend refusal conversion efforts helped increased the percentage of low-income families interviews, though attempt to convert screened refusals had no impact on the percentage of families found to be low income. In the third round, both the screener and extended refusal conversion efforts had little impact on the final percentage of low-income families interviewed. In all three rounds, the extended refusal conversions did increase the percentage of interviews done with families that did not have any children under 18. This was important despite the 4- 2 emphasis the survey placed on collecting information about children, since response rates were lower among adult-only households. As in the previous section that looks at call attempts, there is expected correlation between demographic groups. For instance, refusal conversion increased the percentage of male respondents, which is more likely to occur in interviews conducted in households without children, therefore the percentage of interviews with families without children also increased as a result of refusal conversion. An ordinary least squares regression model similar to the model run in the previous section was run for the combined 1997 to 2002 NSAF data. The dependant variable this time was previous refusal status, and the independent variables were the demographic variables immigration status, marital status, census region, foreign-born status, age, gender, education, ethnicity, race, labor force status, poverty level, hours worked, and number family of various age groups. Two other independent variables were the survey variables years of survey and year of survey. Means were substituted for missing data, and there was no missing data for total number of call attempts. The regression results are shown in table 4-1. The regression analysis shows that converted refusals were less likely to be families that had children and more likely to be families that had proportionately more older adults. The regression also indicates that refusal conversion actually did not help in increasing the number of respondents who work or the respondents who reported having never been married. In addition, refusal conversion did not effectively increase the number of non-U.S. citizens interviewed. Refusal conversion was much more effective with lesseducated respondents. While refusal conversion was more effective among the non-Hispanic population, refusal conversion did not have a differential impact for low-income or black families. 4- 3 Table 4-1. Initial Refusal Regression Output Dependent variable: refusal status (0=never refused, 1=refused screener, 2=refused extended, 3 = refused both) Sample: all three rounds of the NSAF, 1997–2002 (R SQUARE=.146) INDEPENDENT VARIABLES B Std Error (Constant) 38.67 1.96 UIMMSTAT Immigration status -0.068 0.013 UMARSTAT New marital status -0.007 UREGION Region Beta T Sig T 19.75 0.00 -0.045 -5.172 0.000 0.001 -0.020 -5.828 0.000 -0.006 0.002 -0.008 -3.122 0.002 U_USBORN Foreign- or U.S.-born 0.048 0.022 0.019 2.194 0.028 AGE Age 0.001 0.000 0.013 2.960 0.003 SEX Gender -0.027 0.005 -0.016 -5.706 0.000 UBCPSED Education level, CPS -0.031 0.004 -0.025 -8.552 0.000 UBETH Hispanic 0.070 0.008 0.029 9.326 0.000 UBRACE Race (2 category) 0.005 0.006 0.002 0.805 0.421 -0.016 0.003 -0.018 -6.204 0.000 U_SOCPOV Social family income as % of poverty 0.002 0.002 0.003 1.037 0.300 U_USHRS Hours worked per week this year 0.000 0.000 -0.004 -1.536 0.125 UAGE1 No. family members age 0-5 -0.003 0.003 -0.003 -0.909 0.364 UAGE2 No. family members age 6–17 -0.004 0.002 -0.005 -1.703 0.089 UAGE3 No. family members age 18–24 0.018 0.004 0.014 4.617 0.000 UAGE4 No. family members age 25–34 0.023 0.005 0.022 4.974 0.000 UAGE5 No. family members age 35–44 0.047 0.004 0.047 10.523 0.000 UAGE6 No. family members age 45–54 0.066 0.005 0.059 14.567 0.000 UAGE7 No. family members age 55–64 0.098 0.006 0.064 16.714 0.000 UAGE8 No. family members age 65+ 0.108 0.007 0.043 15.633 0.000 -0.019 0.001 -0.052 -19.690 0.000 0.027 0.000 0.368 137.230 U_LFSR Labor force status recode YEAR TNC (total number of call) 4- 4 0.000 5. SOCIODEMOGRAPHIC STUDY OF SURVEY NONRESPONDENTS The impact of nonresponse on survey estimates depends on the size or number of nonrespondents and the extent to which nonrespondents differ from respondents on survey items of interest (Gray et al. 1996; Groves 1989; Keeter et al. 2000). Whether survey estimates are unbiased depends on the assumption that nonrespondents are missing at random (MAR). When nonrespondents differ from respondents, the MAR assumption does not hold and nonresponse error may be present in some of the survey estimates. Little empirical evidence exists to evaluate the MAR assumption because any information about nonrespondents must come from administrative records or other auxiliary sources. This section of the nonresponse report compares respondents and nonrespondents from the Urban Institute’s 2002 National Survey of America’s Families using auxiliary data derived from the 2000 U.S. biennial census. The address information that was obtained for both NSAF respondents and nonrespondents enabled merging of census block group level data. A census block is the smallest geographical area for which census data are collected. A clustering of blocks forms the geographically larger census block group and, at the next level, several block groups combine to form a census tract. Although the Census Bureau’s goal is for each block group to contain 400 housing units, block groups generally vary between 250 and 550 housing units. For confidentiality reasons, census data is available publicly at the block group level (U.S. Census Bureau 2003). Linking census block group data to the NSAF allows comparisons of block group characteristics of respondents to those of nonrespondents. For this research, neighborhood classifications are defined as being above or below the national averages on a set of characteristics. For example, a neighborhood that has a Hispanic population above the national average of 12.5 percent is characterized as a Hispanic neighborhood. Learning the extent to which nonrespondents’ neighborhood characteristics differ from those of respondents sheds light on the appropriateness of the assumption that nonrespondents are missing at random and the presence of nonresponse bias in the sample. Census block groups for each NSAF household (respondent and nonrespondent households) were obtained using reverse directory call back services. These services identify a household’s address through its telephone number. Of the initial NSAF telephone sample of over a halfmillion telephone numbers, 214,181 were identified as residential. Of these residential telephone numbers 75.7 percent (or 162,157) were matched to an address using reverse directory call back services and subsequently matched to the census block group in which the address is found. Future research may explore the remaining households by matching them to census data at the zip code level,3 but the current research only examines nonresponse in the block group matched NSAF households. 3 Using telephone exchange information, the most populace zip code within an exchange could be assigned to each household, then either use zip code–level census data or assign the most populace block group within that zip code. 5- 1 Once matched to census data, NSAF respondent and nonrespondent households were characterized as located in block groups above or below the national average on the indicators below: v Racial composition v Linguistic isolation, foreign-born and U.S. citizenship v Poverty rate v Receipt of public assistance v Education v Urbanization indicator v Employment v Time spent living in neighborhood v Home ownership Neighborhood characteristics are explored for respondents and nonrespondents at the screener interview and extended interview levels. Nonrespondents at both levels are broken down into refusals and other nonresponse categories. The latter category includes passive refusals and households that did not respond because they were difficult to contact. Households selected for the extended interview are a subset of those who completed the screener interview. Criteria for selection into the extended interview sample include the presence of at least one householder under the age of 65, the presence of at least one child under the age of 17, and family income below 200 percent of the federal poverty level. Due to these sampling criteria, analysis of nonresponse at the extended interview is less comparable to other national telephone surveys of the general population than analysis of nonresponse at the screener level. Comparison of Screener Respondents versus Nonrespondents Most telephone nonresponse occurs before the first question of a survey is even asked. Similar to other national telephone surveys, the NSAF experiences its highest rates of nonresponse when attempting to complete the screener survey (Curtin et al. 2005). Table 5-1 shows the variation in household screener completion, refusal, and other nonresponse rates by neighborhood type. The above-average and below-average subcolumns indicate the rate of survey response, refusal, or other nonresponse for each neighborhood characteristic. Given the large sample size, small differences found in this table are significant. 5- 2 Table 5-1. Household Screener Response, Refusal, and Other Nonresponse Rates by Neighborhood Characteristics Neighborhood characteristics Urban Rural Screener Completes Overall Rate= 65.9% (n = 106,804) Screener Refusals Overall Rate=26.6% (n = 43,134) Above average 64.4% 70.3 Above average 27.3% 24.9 Below average 24.6% 27.1 Below average 70.3% 64.6 Screener Other Nonresponse Overall Rate = 7.5% (n = 12,219) Above Below average average 8.4% 5.1% 4.9 8.3 White Black Asian Other race Hispanic 65.9 66.0 62.3 66.2 64.5 65.8 65.8 67.0 65.8 66.2 27.3 25.5 27.6 24.5 25.8 24.9 26.9 26.3 26.9 26.8 6.9 8.5 10.1 9.3 9.7 9.3 7.3 6.7 7.3 7.0 Spanish language isolation Asian language isolation Linguistically isolated: any language 64.4 62.6 66.2 66.7 25.9 27.3 26.7 26.4 9.7 10.2 7.1 6.9 62.6 66.8 26.7 26.6 10.7 6.6 Foreign born Noncitizen 61.3 61.8 67.6 67.2 27.9 27.2 26.1 26.4 10.8 11.0 6.3 6.4 Same house in 1995 Different house in 1995 66.2 65.1 65.4 66.4 26.9 26.5 26.2 26.7 6.9 8.4 8.4 6.9 High school degree only College degree or higher 67.7 62.9 64.5 68.0 25.9 28.7 27.2 25.1 6.5 8.4 8.3 6.9 Employed—all Employed—females 65.1 65.4 66.5 66.2 27.3 27.0 26.0 26.3 7.6 7.6 7.5 7.5 Receives public assistance 66.7 65.8 24.4 26.8 8.9 7.4 Less than 50% of FPL Less than 100% of FPL Less than 200% of FPL 66.5 67.2 68.0 65.6 65.3 64.8 25.1 24.8 24.4 27.2 27.4 27.8 8.4 8.1 7.6 7.2 7.3 7.48 Owner occupied Renter occupied 66.6 64.3 64.7 66.7 27.0 26.2 25.9 26.8 6.5 9.5 9.4 6.5 Rural neighborhoods tend to have higher household screener completion rates than urban neighborhoods, although the reason appears to be the result of having a lower other nonresponse rate. This suggests that people from rural areas are almost as likely to refuse a survey but have higher response rates since they are much easier to contact. Over half of the 5.7 percentage point increase in the rural completion rate can be accounted for by the 3.4 percentage point decrease in the other nonresponse rate. Maybe more telling is that the other nonresponse rate for urban neighborhoods is almost twice as high as it is in rural neighborhoods. This implies that nonresponse in urban neighborhoods is not only a problem of high refusal rates but also of making contact with respondents. 5- 3 There were no differences found in screener completion rates between white and black neighborhoods. This is surprising given that NSAF weighting adjustments give greater weights to black non-Hispanics. This anomaly could mean that increasing the share of black nonHispanics through post-stratification might create a slight overrepresentation of black respondents living in nonblack neighborhoods, which could affect estimates if their answers are significantly different than black respondents living in black neighborhoods. This potential problem is similar to the 1997 NSAF nonresponse analysis (Groves and Wissoker 1999), where there was some evidence that by treating black respondents and nonrespondents the same, the weight may have overstated the number of black low-income households. Though this potential problem was investigated, the report did not find evidence that this in fact occurred. While completion rates are roughly the same in black and white neighborhoods, the source of nonresponse varied for these two groups. White neighborhoods had a higher refusal rate, while black neighborhoods suffered from a higher other nonresponse rate. This indicates that potential respondents in black neighborhoods are more difficult to reach but once reached they can be more cooperative. Asian and Hispanic neighborhoods both had lower completion rates than white and black neighborhoods. These differences may be explained by the low completion rates found in neighborhoods with high levels of linguistically isolated households and by the low completion rates of neighborhoods with above-average numbers of foreign-born and noncitizen residents. In fact, the screener completion, refusal, and other nonresponse rates of Hispanic and Asian neighborhoods are mirrored by the language isolation measures for those groups. The screener completion rate in Spanish-language linguistically isolated neighborhoods is higher than in Asian-language linguistically isolated neighborhoods but may be attributable to the availability of a Spanish translation of the NSAF survey. The NSAF is available in English and Spanish only. Presumably, households that are linguistically isolated in a language other than Spanish would be unable to respond or even to refuse the survey request, explaining the higher other nonresponse rate and lower refusal rate of Asian-language linguistically isolated neighborhoods. Transient households, those that had moved to the current neighborhood in the past five years, did not differ much in their willingness to complete the NSAF screener from more established households. However, neighborhoods with above-average numbers of transient household had a higher nonresponse rate than neighborhoods where people had lived for more than five years. Here the small finding is fairly intuitive, since more transient respondents are often thought to be more difficult to contact and interview. While this research supports the notion that nonresponse increases in transient neighborhoods, the difference of 1.1 percentage points is modest. Education tended to have a negative impact on response rates. The data show that screener completion rates in less-educated neighborhoods where respondents tended to have only a high school degree or equivalent are nearly 5 percentage points higher than in neighborhoods with above-average numbers of college graduates. Highly educated respondents tend to not only refuse the survey request at a higher rate than less-educated respondents but are also more difficult to contact, as evidenced by their higher refusal and other nonresponse rates. Neighborhoods with higher overall employment and higher female employment had only slightly lower screener completion rates than the overall average for the survey. This is a positive finding, given that previous NSAF nonresponse analysis found that employment negatively impacted response rates (Black and Safir 2000; Triplett et al. 2002). Both the long field period (nine months) and the selection of any adult to complete the screener probably offset the 5- 4 difficulty that most national telephone surveys experience in attempting to reach employed respondents. Neighborhoods with more than the national average number of households receiving public assistance as well as poor neighborhoods had slightly higher completion rates than the overall average screener completion rate for the survey. This finding seems somewhat counterintuitive, but there may be topic interest effect among poorer households that increases their propensity to respond to the NSAF. Research has demonstrated that respondent interest in the topic of a survey can affect survey participation decisions (Groves et al. 2004). Finally, neighborhoods that have a high concentration of homeowners tend to have higher screener completion rates than high rental occupancy neighborhoods. This difference may be explained by the higher other nonresponse rate in homeowner neighborhoods and indicate that renters are more difficult to contact. This finding is similar to the trend seen among respondents in transient neighborhoods where home rental rather than home ownership is likely to be more common. Comparison of Extended Interview Respondents versus Nonrespondents Households that completed an extended interview are a subset of the households that completed the screener interview. The research on extended interview nonresponse on the NSAF is less comparable to other household surveys since not all high-income and childless households are asked to complete an extended interview. In addition, the respondent was often the most knowledgeable adult for a child in the household. However, the findings in this section should be useful for those studies collecting data on children or low-income populations. Table 5-2 shows variation in response rates by neighborhood characteristics at the extended interview level. As before, the above- and below-average columns indicate the rates of response, refusal, or other nonresponse to the extended interview by neighborhood type. The sample sizes are still large, so again even the small differences found in this table are significant. 5- 5 Table 5-2. Household Extended Response, Refusal, and Other Nonresponse Rates by Neighborhood Characteristics Extended Completes Overall Rate= 84.3% (n=33,919) Neighborhood characteristics Urban Rural Above average 83.4% 86.9 Extended Refusals Overall Rate=10.4% (n = 4,179) Below Above average average 86.6% 10.6% 83.4 9.7 Below average 9.7% 10.6 Extended Other Nonresponse Overall Rate = 5.3% (n = 2,137) Above Below average average 6.0% 3.7% 3.5 6.0 White Black Asian Other race Hispanic 85.4 82.2 82.1 82.3 82.4 81.9 84.9 85.0 84.7 84.8 10.4 10.9 10.9 9.0 9.3 10.4 10.2 10.2 10.6 10.7 4.2 6.9 7.0 8.8 8.3 7.7 4.9 4.8 4.7 4.5 Spanish language isolation Asian language isolation Linguistically isolated: any language 82.3 82.3 84.8 84.8 9.3 10.2 10.7 10.4 8.5 7.5 4.6 4.8 81.7 85.1 9.8 10.6 8.5 4.4 Foreign born Noncitizen 80.8 81.1 85.6 85.3 10.7 10.3 10.3 10.4 8.6 8.6 4.2 4.3 Same house in 1995 Different house in 1995 84.1 84.6 84.5 84.1 11.0 9.6 9.6 11.0 4.9 5.9 5.9 4.9 High school degree only College degree or higher 85.1 83.6 83.7 84.7 10.2 11.5 10.5 9.8 4.7 4.9 5.8 5.5 Employed—all Employed—females 84.8 85.0 83.9 83.7 10.6 10.4 10.2 10.4 4.5 4.6 5.9 5.9 Receives public assistance 84.0 84.3 9.0 10.5 7.0 5.2 Less than 50% of FPL Less than 100% of FPL Less than 200% of FPL 83.8 84.1 84.5 84.5 84.4 84.1 9.5 9.5 9.3 10.8 10.8 11.1 6.7 6.4 6.2 4.7 4.8 4.7 Owner occupied Renter occupied 84.9 83.2 83.3 84.9 10.7 10.0 10.0 10.6 4.4 6.9 6.8 4.5 A comparison of the screener and extended response and nonresponse rates shows that completion rates are much higher for the extended interview, even though the extended interview often takes more than 45 minutes to complete. This is evidence that length of the questionnaire has little impact on response rate. Additionally, other nonresponse makes up over half the overall extended nonresponse but only 28 percent of the overall screener nonresponse. This is likely a result of the added difficulty of scheduling a long interview with a chosen respondent, whereas the screener interview was completed by any householder 18 years of age or older. While differences in the extended completion rate by neighborhood characteristics were generally smaller than the differences found at the screener level, they tend to follow the same 5- 6 patterns as the screener interview completion rates. There were a few exceptions, such as a decline in the completion rates of black neighborhoods and very poor neighborhoods (income below 50 percent of federal poverty level). However, neighborhoods with high employment rates maintained response rates above the overall rate for the survey, while at the screener level the opposite is true. This finding may be attributable to the NSAF data collection design of administering the survey to the adult in the family who is most knowledgeable about the sampled child(ren). In addition to providing information about the child(ren), the most knowledgeable adult provides information about him- or herself and any spouse or partner in the household, thereby relieving the need for scheduling time to administer the survey to the spouse or partner. In our descriptive analysis of neighborhood characteristics, most of the findings have shown some relationship between the neighborhood type and the propensity to responds. To further the research, a regression model is used to answer the question, “Does not knowing the percentage of neighbor that has a certain characteristic help us predict a household’s likelihood of participating?” The results of this model are shown in table 5-3. The results indicate that knowing any of the block group characteristics used in this model would provide some insight in predicting nonresponse. These results still show that rural areas predict higher response but this factor seems smaller compared with some other predictors of higher response (below poverty, owner occupancy, and receipt of public assistance). The reason households in rural areas are more likely to respond may have more to do with characteristics about rural areas than with the household simply being located in a rural area. In this model, having a greater percentage of either more college graduates or more high school–educated respondents predicts lower response rates. Thus, the only real surprise is that the sign of the high school only coefficient is the opposite of what one would have guessed based on the descriptive analysis. Finally, the descriptive analysis did not find much association between being black or employed. The regression confirms this in that these two characteristics have only slight predictive powers. 5- 7 Table 5-3. Linear Regression Dependent variable: screener disposition (0=nonrespondent, 1=respondent) Independent variables: percentage of block with that characteristic (values 0 to 100) (Constant) Black Rural Hispanic Different house in 1995 Less than 100% of FPL Other race Owner occupied High school degree only Receives public assistance College degree or higher Employed—all Unstandardized Coefficients Std. B error .567 .016 -.034 .007 .044 .004 -.115 .015 .058 .010 .124 .020 .151 .030 .098 .007 -.123 .022 .299 .110 -.212 .014 .053 .012 Standardized Coefficients Beta -.012 .028 -.034 .015 .023 .022 .043 -.022 .008 -.066 .011 t 34.758 -4.670 11.493 -7.587 5.615 6.247 5.022 13.193 -5.594 2.725 -15.171 4.280 Sig. .000 .000 .000 .000 .000 .000 .000 .000 .000 .006 .000 .000 95% Confidence Interval for B Lower Upper bound bound .535 .598 -.049 -.020 .036 .051 -.144 -.085 .038 .079 .085 .163 .092 .211 .083 .112 -.166 -.080 .084 .513 -.239 -.185 .029 .077 Conclusions Most of the findings only show modest differences in completion rates by neighborhood type. However, even differences of a couple percentage points could introduce enough bias into some survey estimates to affect the overall findings. In addition, the differences in completion rates by neighborhood characteristics are probably somewhat conservative in that the households that were not matched to block group data are probably more transient and therefore would tend to not respond to the survey. Some of the potential bias associated with different response by neighborhood type is dealt with by the undercoverage and post-stratification weights developed for the NSAF (Bethlehem 2002). The NSAF methods and response rate evaluation report did some evaluation of the potential bias owing to nonresponse (Brick et al. 2003). However, this report was fairly vague on the degree of how much increasing nonresponse rates are biasing survey estimates. This paper goes one step further in that it identifies types of neighborhoods that have differing response propensity (rural, Asian, Hispanic, educated, owner occupancy rates). A natural next step would be to look at the completed cases block group characteristics using the survey weights. Unfortunately, most nonresponse for the screener occurs where post-stratification information is unavailable. However, post-stratification is not as effective if nonrespondents differ from respondents, which does appear to be at least partially true. 5- 8 6. SUMMARY Survey research methodologies have improved over the past two decades, bolstered by both technological advances and learning through expanded experience. Telephone surveys such as the National Survey of America’s Families have been very popular because they often yield high response rates and less item nonresponse, provide more control of the question ordering, and allow surveyors to use skip patterns and recall information during the interview. Higher cooperation rates and the ability to reach people by phone had been two major advantages of telephone surveys, but both these advantages are now on the decline, the primary reason the NSAF response rate in 2002 was 10 percentage point lower than it was in 1997. Still, telephone surveys remain a very viable mode of collecting survey data, though changing technology, such as answering machines, cell phones, and call screening has made it more difficult to achieve high response rates on telephone surveys. While there is no such thing as an “official” acceptable response rate, response rates are the industry’s standard which people judge the quality of a survey. Surveys that achieve a response rate of 70 percent or higher are generally thought of as high-quality surveys, and nonresponse is not usually a concern. Studies such as the NSAF that have response rates between 50 and 70 percent often benefit from some nonresponse weighting adjustment to reduce potential nonresponse bias. Nonresponse adjustments involves adding a weighting adjustment that when used in analysis will increase the overall the impact of the data that was collected from people who have characteristics similar to the nonrespondents. Determining the characteristics of nonrespondents involves comparing respondents to census estimates as well as looking at the characteristics of respondents who were the most difficult to obtain interviews from. While nonresponse is primarily a function of people refusing to participate, the difficult to reach respondents also contribute to the over nonresponse rate. These two problems make it imperative for a quality telephone survey to include both a refusal avoidance and refusal conversion strategy, and to schedule enough time to make multiple calls at varying times and days of the week in order to contact difficult to reach respondents. The lower your response rate, the more vulnerable your study is to nonresponse bias. Unlike sampling error, it is very difficult to quantify the effect nonresponse error has on the quality of your survey. Nonresponse error is not easy to quantify because survey researchers do not know whether the nonrespondents differ from the respondents in terms of how they would have responded to a survey. While survey researchers will never know for sure how a respondent would have responded to questions asked, they can predict how a respondent’s absence from the study may affect survey estimates by studying how the characteristics of nonrespondents differ from respondents. To discover who nonrespondents are, it is very important to understand that people who refuse have different characteristics than people who are simply very difficult to contact. The number of call attempts to finalize a case is a common measure of how difficult it is to reach a respondent. Checking to see if anyone in the household ever refused the survey is the common measure used to measure the willingness of respondents to participate. Hence, this reports looks separately at the difficult-to-contact respondents and the respondents who initially refused to complete the survey in order to better understand the nature of NSAF nonrespondents. 6- 1 The hard-to-reach and less cooperative respondents shared many characteristics: for instance, they were more likely to be foreign born, never married, Hispanic, and work full time. But there were striking differences, especially with education. For the NSAF, less-educated respondents were harder to reach, while higher-educated respondents were easier to reach but less cooperative. It was much easier to reach rural households, but while rural families traditionally have been more cooperative, in the 2002 NSAF the difference was small and not significant. Similarly, larger households were easier to contact but slightly less cooperative. Somewhat surprising and certainly good news for the NSAF study was that while low-income families were more difficult to reach, they were more cooperative though not significantly more so. Finally, only modest differences were found when census track information was used to compare the 2002 nonrespondents with the respondents. This is encouraging, and given that the survey weights includes some post-stratification to match key census control totals, these small differences are not likely to have significant impact on the survey estimates. Nevertheless, this census comparison suggests there likely will be less precision for the survey estimates for certain subpopulations, such as Asian, Hispanic, urban, and higher-educated respondents. 6- 2 References Abi-Habib, Natalie, Tamara Black, Simon Pratt, Adam Safir, Rebecca Steinback, Timothy Triplett, Kevin Wang, et al. 2005. 2002 NSAF Collection of Papers. Methodology Report No. 6. Washington DC: The Urban Institute. Abi-Habib, Natalie, Adam Safir, and Timothy Triplett. 2004. 2002 NSAF Survey Methods and Data Reliability. Methodology Report No. 1. Washington DC: The Urban Institute. ———. 2004. 2002 NSAF Public Use File Data Documentation Guide. Methodology Report No. 11. Washington DC: The Urban Institute. Abi-Habib, Natalie, Adam Safir, Timothy Triplett, and Pat Cunningham. 2003. 2002 NSAF Questionnaire. Methodology Report No. 12. Washington DC: The Urban Institute. Alexander, Charles H. 1988. “Cutoff Rules for Secondary Calling in a Random Digit Dialing Survey.” In Telephone Survey Methodology, edited by Robert M. Groves, R. M. Kahn, L. Lyberg, James T. Massey, Joseph Waksberg, and William L. Nicholls II (113–26). New York: Wiley. Beaumont, Jean-Francois. 2000. “An Estimation Method for Nonignorable Nonresponse.” Survey Methodology 26(2). Bethlehem, Jelke G. 2002. “Weighting Nonresponse Adjustments Based on Auxiliary Information.” In Survey Nonresponse, edited by Robert M. Groves, Don Dillman, John L. Eltinge, and Roderick J.A. Little (275–87). New York: Wiley. Black, Tamara, and Adam Safir. 2000. “Assessing Nonresponse Bias in the National Survey of America’s Families.” In Proceedings of the Survey Research Methods Section (816–21). Alexandria, VA: American Statistical Association. Blair, Johnny, and Young I. Chun. 1992. “Quality of Data from Converted Refusers in Telephone Surveys.” Paper presented at the American Association for Public Opinion Research, St. Petersburg, Florida. Blair, Johnny, and D. O’Rourke. 1985. “Alternative Callback Rules in Telephone Surveys: A Methods Research Note.” Paper presented at the American Association for Public Opinion Research, St. Petersburg, Florida. Brick, Michael J., Bruce Allen, Pat Cunningham, and David Maklan. 1996. “Outcomes of a Calling Protocol in a Telephone Survey.” In Proceedings of the Survey Research Methods Section (142–48). Alexandria, VA: American Statistical Association. Brick, Michael J., Pam Broene, David Cantor, David Ferraro, Tom Hankins, Carin Rauch, and Teresa Strickler. 2000. 1999 NSAF Response Rates and Methods Evaluation. Methodology Report No. 8. Washington, DC: The Urban Institute. Brick, Michael J., Ismael Flores Cervantes, and David Cantor. 1998. 1997 NSAF Response Rates and Methods Evaluation. Methodology Report No. 8. Washington, DC: The Urban Institute. Brick, Michael J., David Ferraro, Carin Rauch, and Teresa Strickler. 2003. 2002 NSAF Response Rates and Methods Evaluation. Methodology Report No. 8. Washington, DC: The Urban Institute. R-1 Brick, Michael J., David Ferraro, Teresa Strickler, and Benmei Liu. 2003. 2002 NSAF Sample Design. Methodology Report No. 2. Washington DC: The Urban Institute. Brick, Michael J., Jill Montaquila, and Fritz Scheuren. 2002. “Estimating Residency Rates for Undetermined Telephone Numbers.” Public Opinion Quarterly 66(1): 18–39. Cannell, C. F., and F. J. Fowler. 1963. “Comparison of a Self-Enumerative Procedure and a Personal Interview: A Validity Study.” Public Opinion Quarterly 27(2): 250–64. Cantor, David, Bruce Allen, Pat Cunningham, Michael J. Brick, P. Slobasky, and Genevieve Kenney. 1997. “Promised Incentives on a Random Digit Dial Survey.” In Nonresponse in Survey Research, Proceedings of the Eighth International Workshop on Household Survey Nonresponse, edited by A. Koch and R. Porst (219–28). Manheimm, Germany. Cantor, David, Pat Cunningham, Timothy Triplett, and Rebecca Steinback. 2003. “Comparing Incentives at Initial and Refusal Conversion Stages on a Screening Interview for a Random Digit Dial Survey.” In Proceedings of the Survey Research Methods Section. Alexandria, VA: American Statistical Association. Cantor, David, Kevin Wang, and Natalie Abi-Habib. 2003. “Comparing Promised and Pre-Paid Incentives for an Extended Interview on a Random Digit Dial Survey.” In Proceedings of the Survey Research Methods Section. Alexandria, VA: American Statistical Association. Converse, Nate, Adam Safir, and Fritz Scheuren. 2001. 1999 NSAF Public Use File Documentation. Methodology Report No. 11. Washington, DC: The Urban Institute. Curtin, Richard, Stanley Presser, and Eleanor Singer. 2000. “The Effect of Response Rate on the Index of Consumer Sentiment.” Public Opinion Quarterly 64(4): 413–28. ———. 2005. “Changes in Telephone Survey Nonresponse over the Past Quarter-Century.” Public Opinion Quarterly 69(1): 87–98. Dillman, Don, John L. Eltinge, Robert M. Groves, and Roderick J. A. Little. 2002. “Survey Nonresponse in Design Data Collection and Analysis.” In Survey Nonresponse, edited by Groves et al. (3–26). New York: Wiley. Dunkleberg, William C., and George Day. 1973. “Nonresponse Bias and Call backs in Sample Surveys.” Journal of Marketing Research 10 (May): 160–68. Elliott, M. R., and Roderick J. A. Little. 2000. “Subsampling Callbacks to Improve Survey Efficiency.” Journal of American Statistical Association 95: 730–38. Goldstein, Kenneth M., and M. Kent Jennings. 2002. “The Effect of Advance Letters on Cooperation in a List Sample Telephone Survey.” Public Opinion Quarterly 66: 608–17. Greenberg, B. S., and S. L. Stokes. 1990. “Developing an Optimal Call Scheduling Strategy for a Telephone Survey.” Journal of Official Statistics 6: 421–35. Groves, Robert M. 1989. Survey Costs and Survey Errors. New York: Wiley. Groves, Robert M., and Mick P. Couper. 1998. Nonresponse in Household Interview Surveys. New York: Wiley. Groves, Robert M., Don Dillman, John L. Eltinge, and Roderick J.A. Little, eds. 2001. Survey Nonresponse. New York: Wiley. R-2 Groves, Robert M., F. J. Fowler, Mick P. Couper, James M. Lepkowski, Eleanor Singer, and Roger Tourangeau, eds. 2004. Survey Methodology. New York: Wiley. Groves, Robert M., and R. M. Kahn. 1979. Surveys by Telephone: A National Comparison with Personal Interviews. New York: Academic Press. Groves, Robert M., and L. Lyberg. 1988. “An Overview of Nonresponse Issues in Telephone Surveys.” In Telephone Survey Methodology, edited by Groves et al. (91–211). New York: Wiley. Groves, Robert M., Stanley Presser, and S. Dipko. 2004. “The Role of Topic Interest in Survey Participation Decisions.” Public Opinion Quarterly 68(1): 2–31. Groves, Robert M., Eleanor Singer, and Amy D. Corning. 2000. “Leverage-Saliency Theory of Survey Participation.” Public Opinion Quarterly 64(3): 299–308. Groves, Robert M., Eleanor Singer, Amy D. Corning, and Ashley Bowers. 1999. “A Laboratory Approach to Measuring the Effects on Survey Participation of Interview Length, Incentives, Differential Incentives, and Refusal Conversion.” Journal of Official Statistics 15: 251–68. Groves, Robert M., and Douglas Wissoker. 1998. Early Nonresponse Studies of the 1997 National Survey of America’s Families. Methodology Report No. 7. Washington, DC: The Urban Institute. Hansen, Morris H., and William N. Hurwitz. 1946. “The Problem of Non-Response in Sample Surveys.” Journal of American Statistical Association 41(236): 517–29. Keeter, Scott, Carolyn Miller, Andrew Kohut, Robert M. Groves, and Stanley Presser. 2000. “Consequences of Reducing Nonresponse in a Large National Telephone Survey.” Public Opinion Quarterly 64(2): 125–48. Kish, Leslie. 1965. Survey Sampling. New York: Wiley. Lavrakas, Paul J., Sandra L. Bauman, and Daniel M. Merkle. 1992. “Refusal Report Forms (RRFs), Refusal Conversions and Non-response Bias.” Paper presented at the American Association for Public Opinion Research conference, St Petersburg, Florida. Len, I-Fen, and Nora Shaeffer. 1995. “Using Survey Participants to Estimate the Impact of Nonparticipation.” Public Opinion Quarterly 59(2): 236–58. Link, Michael W., and Robert W. Oldendick. 1999. “Call Screening: Is It Really a Problem for Survey Research?” Public Opinion Quarterly 63(4): 577–89. Losch, Mary E., Maitland Aaron, Gene Lutz, Peter Mariolis, and Gleason Steven C. 2002. “The Effect of Time of Year of Data Collection on Sample Efficiency: An Analysis of Behavioral Risk Factor Surveillance Survey Data.” Public Opinion Quarterly 66: 594– 607. Lynn, Peter, Paul Clarke, Jean Martin, and Paul Sturgis. 2002. “The Effects of Extended Interviewer Efforts on Nonresponse Bias.” In Survey Nonresponse, edited by Groves et al. (135–47). New York: Wiley. R-3 Massey, James T., P. R. Barker, and Hsiung S. 1981. “An Investigation of Response in a Telephone Survey.” In Proceedings of the Survey Research Methods Section (63–72). Chicago, IL: American Statistical Association. Massey, James T., Charles Wolter, Siu Chong Wan, and Karen Liu. 1996. “Optimum Calling Patterns for Random Digit Dialed Telephone Surveys.” In Proceedings of the Survey Research Methods Section (426–31). Alexandria, VA: American Statistical Association. Merkle, Daniel M., Sandra L. Bauman, and Paul J. Lavrakas. 1993. “The Impact of Callbacks on Survey Estimates in an Annual RDD Survey.” In Proceedings of the Survey Research Methods Section (1070–75). Alexandria, VA: American Statistical Association. Newcomer, Kathryn E., and Timothy Triplett. 2004. “Using Surveys.” In Handbook of Practical Program Evaluation, 2nd edition, edited by Joseph S. Wholly, Harry P. Hatry, and Kathryn E. Newcomer (chapter 9). San Francisco: Jossey-Bass. Neyman, J. 1938. “Contribution to the Theory of Sampling Human Populations.” Journal of American Statistical Association 33(201): 101–16. Oldendick, Robert W. 1993. “The Effect of Answering Machines on the Representativeness of Samples in Telephone Surveys.” Journal of Official Statistics 9(3): 663–72. Oldendick, Robert W., and Michael W. Link. 1994. “The Answering Machine Generation.” Public Opinion Quarterly 58(2): 264–73. Piazza, Tom. 1993. “Meeting the Challenges of Answering Machines.” Public Opinion Quarterly 57(2): 219–31. Rao, P. S. R. S. 1983. “Callbacks, Follow-ups, and Repeated Telephone Calls.” In Incomplete Data in Sample Surveys, vol. 1, edited by William G. Madow, I. Olkin, and D.B. Rubin (173–204). New York: Academic Press. Roth, Shelly B., Jill Montaquila, and Michael J. Brick. 2001. “Effects of Telephone Technologies and Call Screening Services on Sampling, Weighting and Cooperation in a Random Digit Dial (RDD) Survey.” In Proceedings of the Survey Research Methods Section. Alexandria, VA: American Statistical Association. Safir, Adam, Rebecca Steinbach, Timothy Triplett, and Kevin Wang. 2002. “Effects on Survey Estimates from Reducing Nonresponse.” In Proceedings of the Survey Research Methods Section. Alexandria, VA: American Statistical Association. Scheuren, Fritz. 2000. “Quality Assessment of Quality Assessment.” In American Statistical Association 2000 Proceedings of the Section on Government Statistics and Section on Social Statistics (44–51). Indianapolis, IN: American Statistical Association. Sebold, Janice. 1988. “Survey Period Length, Unanswered Numbers and Nonresponse in Telephone Surveys.” In Telephone Survey Methodology, edited by Groves et al. (247– 56). New York: Wiley. Shaiko, R. G., Dwyre D., M. O’Gorman, J. M. Stonecash, and J. Vike. 1991. “Pre-election Political Polling and Non-Response Bias Issue.” International Journal of Public Opinion Research 3: 86–99. R-4 Singer, Eleanor. 2002. “The Use of Incentives to Reduce Nonresponse in Household Surveys.” In Survey Nonresponse, edited by Groves et al. (163–77). New York: Wiley. Singer, Eleanor, John Van Hoewyk, and Mary Maher. 2000. “Experiments with Incentives for Survey Participation on the Survey of Consumer Attitudes.” Public Opinion Quarterly 64(2): 171–88. Smith, Tom. 1983. “The Hidden 25 Percent: An Analysis of Nonresponse on the 1980 General Social Survey.” Public Opinion Quarterly 47(3): 386–404. Stasny, Elizabeth A. 1988. “Modeling Nonignorable Nonresponse in Categorical Panel Data with an Example in Estimating Gross Labor Force Flows.” Journal of Business and Economic Statistics 6(2): 207–19. Steeh, C. G. 1981. “Trends in Nonresponse Rates, 1952–1979.” Public Opinion Quarterly 45(1): 40–57. Steeh, C. G., Robert M. Groves, R. Comment, and Hansmire R. 1983. “Report on the Survey Research Center’s Survey of Consumer Attitudes.” In Incomplete Data in Sample Surveys, vol. 2, edited by William G. Madow, I. Olkin, and D. B. Rubin (377–81). New York: Academic Press. Steeh, C. G., Nicole Kirgis, Brian Cannon, and Jeff DeWitt. 2001. “Are They Really as Bad as They Seem? Nonresponse Rates at the End of the 20th Century.” Journal of Official Statistics 17: 227–47. Teitler, Julien O., Nancy E. Reichman, and Susan Sprachman. 2003. “Costs and Benefits of Improving Response Rates.” Public Opinion Quarterly 67(1): 126–38. Traugott, Michael W. 1987. “The Importance of Persistence in Respondent Selection for Preelection Surveys.” Public Opinion Quarterly 51(1): 48–57. Triplett, Timothy. 2002. “What Is Gained from Additional Call Attempts and Refusal Conversion and What Are the Cost Implications?” Washington, DC: The Urban Institute. http://mywebpages.comcast.net/ttriplett13/tncpap.pdf. Triplett, Timothy, and Natalie Abi-Habib. 2005. “Determining the Probability of Selection for a Telephone Household in a Random Digit Dial Sample Design Is Becoming Increasingly More Difficult.” In 2002 NSAF Collection of Papers (1-11–1-17). Washington DC: The Urban Institute. Triplett, Timothy, Johnny Blair, Teresa Hamilton, and Yun Chiao Kang. 1996. “Initial Cooperators vs. Converted Refusers: Are There Response Behavior Differences?” In Proceedings of the Survey Research Methods Section (1038–41). Alexandria, VA: American Statistical Association. Triplett, Timothy, Adam Safir, Kevin Wang, Rebecca Steinbach, and Simon Pratt. 2002. “Using a Short Follow-up Survey to Compare Respondents and Nonrespondents.” In Proceedings of the Survey Research Methods Section. Alexandria, VA: American Statistical Association. Triplett, Timothy, Julie Scheib, and Johnny Blair. 2001. “How Long Should You Wait Before Attempting to Convert a Refusal?” In Proceedings of the Survey Research Methods Section. Alexandria, VA: American Statistical Association. R-5 Tuckel, Peter, and Harry O’Neil. 1996. “New Technology and Non-response Bias in RDD Surveys.” In Proceedings of the Survey Research Methods Section. (889–94). Alexandria, VA: American Statistical Association. Tucker, Clyde, James M. Lepkowski, and Linda Pierkarski. 2002. “List-Assisted Telephone Sampling Design Efficiency.” Public Opinion Quarterly 66(3): 321–38. U.S. Census Bureau. 2002. 2000 Census of Population and Housing, Summary File 3: Technical Documentation. Washington, DC: U.S. Census Bureau. ———. 2003. Census 2000 Geographic Terms and Concepts. Washington, DC: U.S. Census Bureau. Vaden-Kiernan, Nancy, David Cantor, Pat Cunningham, Sarah Dipko, Karen Molloy, and Patricia Warren. 1999. 1997 NSAF Telephone Survey Methods. Methodology Report No. 9. Washington DC: The Urban Institute. Waksberg, Joseph. 1978. “Sampling Methods for Random-Digit Dialing.” Journal of American Statistical Association 73(361): 40–46. Ward, James C., Betrand Russick, and Redelius William. 1985. “A Test of Reducing Callbacks and not at Homes Bias in Personal Interviews by Weighting at Home Respondents.” Journal of Marketing Research 22: 66–73. Warren, Patricia, and Pat Cunningham. 2003. 2002 NSAF Telephone Survey Methods. Methodology Report No. 9. Washington DC: The Urban Institute. Weeks, M. F. 1988. “Call Scheduling with CATI: Current Capabilities and Methods.” In Telephone Survey Methodology, edited by Groves et al. (403–20). New York: Wiley. Weeks, M. F., B. L. Jones, R. E. Folsom Jr., and C. H. Benrud. 1980. “Optimal Times to Contact Sample Households.” Public Opinion Quarterly 44(1): 101–14. Weeks, Michael F., Richard A. Kulka, and Stephanie A. Pierson. 1987. “Optimal Call Scheduling for a Telephone Survey.” Public Opinion Quarterly 51(4): 540–49. R-6