Public Opinion Quarterly, Vol. 71, No. 1, Spring 2007, pp. 3–22
HOUSEHOLD TELEPHONE SERVICE AND USAGE PATTERNS IN THE UNITED STATES IN 2004: IMPLICATIONS FOR TELEPHONE SAMPLES CLYDE TUCKER J.MICHAEL BRICK BRIAN MEEKINS
Abstract Changes in the U.S. telephone system, especially the rapid growth in the prevalence and use of cell phones, raise concerns about undercoverage error in random digit dial (RDD) telephone samples. A supplement to the Current Population Survey (CPS) was conducted in 2004 to examine telephone service and usage in U.S. households. This article explores the potential for biases in RDD surveys resulting from the increases in cell phones by presenting estimates of the percentage of households with different types of telephone service, including the percentage of cell-only households, and giving demographic profiles of households by type of telephone service. Logistic regression models examine variables that predict whether households are without a telephone or only have cell phones. These predictors may be used for weighting adjustments to reduce undercoverage biases. We address some additional issues, including the wording of questions for measuring telephone service, that are relevant if telephone-sampling methods are revised to include cell phones. The estimates from the CPS supplement are also used to help understand some of the new sampling and weighting problems associated with selecting samples from cell phone numbers.
CLYDE TUCKER is with the Bureau of Labor Statistics; 2 Massachusetts Avenue, NE; Washington, DC, USA. J. MICHAEL BRICK, Westat; 1650 Research Boulevard; Rockville, MD, USA. BRIAN MEEKINS is with the Bureau of Labor Statistics; 2 Massachusetts Avenue, NE; Washington, DC, USA. The authors would like to thank the Census Bureau for conducting the supplement, those who helped develop the instrument (especially C. Steeh of Georgia State University and E. Cohen of Arbitron who provided questionnaires used in earlier cell phone surveys), and staff at Westat and BLS who tested and evaluated the instrument. Address correspondence to Clyde Tucker; e-mail:
[email protected]. The views expressed in this article are those of the authors and do not necessarily represent the views of the Bureau of Labor Statistics, or the Department of Labor.
doi:10.1093/poq/nfl047 Advanced Access publication March 3, 2007 ß The Author 2007. Published by Oxford University Press on behalf of the American Association for Public Opinion Research. All rights reserved. For permissions, please e-mail:
[email protected].
4
Tucker et al.
Introduction The percentage of households in the US with a telephone has important implications for the undercoverage bias in random digit dial (RDD) telephone surveys and for using landline telephones to collect data in household surveys. In earlier years, undercoverage bias in RDD surveys was the result of households without landline telephones because having a telephone in the household was virtually equivalent to having landline telephone service. As the percentage of households with only cell phones has increased, the difference between not having any telephone service and not having a landline is of interest. Currently, nearly all cell phones are excluded in RDD samples.1 As a result, households that have only cell phones are not covered in these surveys, in addition to households with no landline service. In the last few years, cell phones have become vastly more popular. Figure 1 shows the percentage of households reporting a cell phone bill but no landline bill in the Consumer Expenditure Interview Survey (CEIS). CEIS is an area probability sample conducted by the U.S. Census Bureau for the Bureau of Labor Statistics (BLS) that covers the full civilian population of the U.S. In face-to-face interviews, households are questioned about their expenditures, including those on cell or mobile phone service. The data used in the figure are from the wave 2 interviews (households are in sample 5 consecutive quarters but the first wave is used only for bounding purposes) collected from the first quarter of 1994 to the third quarter of 2004. While the CEIS is a good barometer of the growth of cell only households, it is not well suited to produce precise estimates of levels. The ability to assess the number and characteristics of households accurately by the type of telephone service from existing data sources is limited. The only large-scale area probability sample that identifies households by telephone service is the National Health Interview Survey (NHIS). The National Center for Health Statistics sponsors this survey and began collecting these data in 2004. Blumberg, Luke, and Cynamon (2006) report some early findings from this survey. As discussed in detail later, the 2004 estimates of the percentage of adults living in households with only cell phones from the NHIS are very similar to those reported subsequently from a 2004 supplement to the Current Population Survey (CPS). The NHIS estimates about 5.0 percent of adults lived in cell-only households during 2004 (4.5 percent in the first half of the year and 5.5 percent in the second half ), and the CPS supplement estimates 5.4 percent for 2004. The estimates from the two surveys for demographic and economic subgroups are also very consistent. The only major difference between the estimates for the two surveys are for the percentage of adults in households without any type of 1. In a 2005 RDD survey50.1 percent of all sampled RDD numbers were found to be cell phone numbers after the numbers were cleaned by the sampling vendor.
Household Telephone Service and Usage Patterns
5
10 Cell bill only
9 8 7 6 5 4 3 2 1
20043
20041
20033
20031
20023
20021
20013
20011
20003
20001
19991
19993
19983
19981
19973
19971
19963
19961
19953
19951
19943
19941
0
Figure 1. Cell Phone-only Households in the U.S. from Reports of Telephone Expenditures on the Consumer Expenditure Survey: 1994–2004. telephone, where the NHIS estimate is considerably lower than the CPS estimate. This issue is discussed in some detail subsequently. While the estimates from the NHIS are very valuable, that survey does not ask about how households use their telephone services.2 Because of new developments in telecommunications, data on usage is much needed to better understand the complexities associated with sampling households by telephone service. For example, some households are dropping their landline telephone and only using cell phones, and others are keeping their landline but using their cell phones most of the time. Some of these changes may involve acquiring new telephone numbers while other times the numbers are being transferred or ported from one type of service to another. Piekarski (2003) discusses this environment and its implications for telephone surveys. Needless to say, understanding the trends in the percentages of households with different types of telephone service and their uses of these services is essential for telephone survey researchers. In February 2002, a group of researchers discussed the implications of these emerging trends for survey research, particularly RDD surveys. As a result, the U.S. Census Bureau and the BLS offered to collect data on telephone service in households as a supplement to the CPS. A subgroup of the researchers developed and tested an instrument to assess the types of telephone services and key aspects of the use of these telephone services
2. Blumberg et al. (2006) cite usage estimates from a web panel survey. Tuckel and Daniels (2006) present usage estimates, based on annual surveys of about 2,000 households.
6
Tucker et al.
by household members. The cell phone interview was a supplement to the February 2004 CPS, with responses from over 31,000 households. This article uses the data from the supplement to address key questions about telephone samples. Although the data are from 2004, they provide a basis to speculate about future issues. In the next section, we describe the design of the CPS supplement, details of the questionnaire design process, and some lessons learned from evaluations of the administration of the items. Section ‘‘Effect of Telephone Usage on RDD Samples’’ We then give estimates of the number of households and adults by type of telephone service (landline only, landline and cell, cell only, and no service). These data are used to explore potential biases in RDD samples that exclude cell phones, and to suggest methods that might be used to reduce these biases. In the following section, we discuss issues that might arise if cell phone numbers are included in telephone samples, based on supplement estimates of usage of telephone services. The next to last section addresses the validity and accuracy of the CPS telephone availability question that is included in the standard CPS. The final section speculates on future research needs.
The CPS Supplement The CPS is a national sample of households and persons that is the source of the official government statistics on employment and unemployment. Its probability sample design uses a rotating panel of households that are in the sample for four consecutive months and then interviewed for the same 4 months the next year (U.S. Department of Labor and U.S. Department of Commerce 2002, Chapter 3). Each of these eight rotation groups is a random sample. The 2004 cell phone supplement sampled the households in rotation groups 2, 3, 5, 6, and 7.3 The interview was administered in February 2004. Because a small portion of the sample was inadvertently skipped in the February supplement (those who said they did not have access to a telephone in the regular CPS), a sample of households who reported not having access to a telephone in the November 2004 CPS was selected to complete the supplement. A total of 32,087 households received the supplement in February, and the response rate (including those not completing the CPS itself) was 86.1 percent. The sample included 2,132 households that did not receive the supplement due to an inappropriate skip pattern noted earlier (these households were excluded from the response rate computations). In November 2004, 882 households with responses that caused the erroneous skip in February were sampled, and 680 completed the supplement. This is
3. Replacement households, whole households that replaced households in the sampled unit from the time of the last interview, were also excluded.
Household Telephone Service and Usage Patterns
7
a 77 percent response rate for the supplement, and about a 70 percent overall response rate for this component including nonresponse to the CPS. The primary objective of the supplement was to classify households by the type of telephone service. A secondary goal was to classify members in households with both landline and cell phone service by their pattern of usage. A key restriction was that, since the interview was a supplement to the CPS, the length of the interview could not exceed 2 minutes. Another restriction was that no sampling of persons could be done, so the same respondent used for the regular CPS interview was used (in the CPS any adult household member can respond for the entire household). The supplement items were taken from existing surveys to the extent possible. Drafts of the questionnaire were tested and refined using cognitive interviewing techniques at Westat. Three rounds of interviews, using concurrent probes and vignettes in the debriefing, were conducted over the telephone. During data collection in the CPS, BLS conducted interviewer debriefings and used behavior coding to evaluate the instrument (Esposito 2004). The appendix contains the wording for the items in the supplement, excluding the skip patterns and computer fills. In both the cognitive testing and in the evaluation of the production interviews, parts of the interview were identified as being difficult for either the respondent or the interviewer. In particular, the question (Q3) asking the proportion of all calls to the household received on a cell phone raised the most concerns. The lack of a specific reference period, not having a code for ‘‘half the time,’’ and the difficulty in reporting for other members of the household were noted in all the development and evaluation work for this item. Another issue was that elderly respondents sometimes struggled with the meaning of ‘‘landline.’’ Since respondents often give the number of extensions rather than the number of unique telephone numbers, a verification question was added to address this problem (VER2). There was some confusion about the definition of a ‘‘working’’ cell phone (Q2), as well as how to treat cell phones used primarily for business. In addition, some respondents were unsure of the meaning of ‘‘regular’’ cell phone usage (Q2b). The difference between ‘‘answering’’ and ‘‘using’’ a cell phone (Q2c, Q2d, and Q2e) was unclear to some. To assist the interviewers in overcoming these problems, specific guidelines were included in the training procedures. For Q3, interviewers were to specify a reference period of a ‘‘typical week.’’ The interviewers were to probe if the respondent said ‘‘half the time,’’ asking if that was ‘‘more than half’’ or ‘‘less than half.’’ Respondents having difficulty reporting for other household members were asked to base their answers on their own experience. Detailed definitions of both a ‘‘landline’’ and a ‘‘working’’ cell phone were given to the interviewer. The same was true for the meaning of ‘‘regular’’ cell phone usage and the difference
8
Tucker et al.
between actually ‘‘answering’’ a cell phone and just ‘‘using’’ a cell phone to make outgoing calls. The February and November data were combined and the Census Bureau produced weights for the combined set that adjusted for nonresponse and matched population controls. The initial weights for the supplement were the base weights (inverse of probability of selection) multiplied by special weighting factors to account for subsampling of the ultimate sampling units in the field. Noninterview adjustments specific to each of the supplements were applied to these weights. In addition, the February supplement was further adjusted because only five of the eight rotation groups were asked the telephone supplement. Finally, the November supplement was weighted to reflect the number of households inadvertently skipped in February. The final data set that contains responses from 31,203 households was used in all the analyses presented here. The standard errors for estimated percentages are calculated using generalized variance parameters provided by the Census Bureau. The Census Bureau does not provide information for users to directly calculate standard errors from the data for the supplement. For estimated means and regression coefficients, general variance parameters were used to compute an average design effect. To be conservative, we inflated this design effect by 10 percent and used it to compute estimated standard errors of the estimates (See U.S. Department of Labor and U.S. Department of Commerce 2002, Chapter 14, and U.S. Census Bureau 2005).
Effect of Telephone Usage on RDD Samples As noted earlier, standard RDD samples exclude households that only have cell phones, resulting in undercoverage bias that is a function of the percent not covered and the differences in the characteristics of the covered and not covered. To address the magnitude of the undercoverage, estimates of the percentage of households that only have cell phones and the differences in characteristics between cell-only households and the households with landlines are given. Data from the supplement are then used to inform the procedures for poststratifying the weights to reduce undercoverage bias for standard RDD surveys. Logistic regression analysis is used for this purpose. Table 1 shows that in 2004 the vast majority of households had a landline (88.6 percent), with slightly more having both landline and cell phone service (46.4 percent) than landline only service (42.2 percent). Only 6.0 percent of all households have only cell phone service. The remaining 5.4 percent of households reported having no phone service, but this is probably an overestimate due to response issues that are discussed later. If the percentage of households without any service is overestimated,
Household Telephone Service and Usage Patterns
9
Table 1. Percent of Households, by Telephone Service Status and Household Characteristics (n ¼ 32,969) Group
Total
Land & cell Land only
Total (%) Region Northeast Midwest South West Metropolitan Central city Metro, noncentral city Non-Meto Household size 1 person More than 1 person Home ownership Own Rent Dwelling unit Single unit Multi-unit
100.0
46.4 (0.3)
100.0 100.0 100.0 100.0
44.4 45.6 45.6 50.2
100.0 100.0 100.0
41.3 (0.5) 52.4 (0.4) 38.1 (0.6)
43.4 (0.5) 38.5 (0.4) 50.0 (0.7)
8.1 (0.3) 5.3 (0.2) 4.9 (0.3)
7.2 (0.3) 3.8 (0.2) 7.0 (0.3)
100.0 100.0
25.7 (0.5) 53.8 (0.3)
58.8 (0.6) 36.2 (0.3)
8.1 (0.3) 5.3 (0.2)
7.3 (0.3) 4.8 (0.1)
100.0 100.0
52.7 (0.3) 31.9 (0.5)
41.2 (0.3) 44.5 (0.5)
3.1 (0.1) 3.1 (0.1) 12.8 (0.4) 10.9 (0.3)
100.0 100.0
51.3 (0.3) 30.9 (0.6)
40.8 (0.3) 46.5 (0.6)
4.0 (0.1) 3.9 (0.1) 12.6 (0.4) 10.1 (0.4)
(0.7) (0.6) (0.5) (0.6)
Cell only
None
42.2 (0.3)
6.0 (0.1)
5.4 (0.1)
47.1 43.2 40.6 39.3
3.9 6.4 6.8 6.9
4.6 4.8 6.9 4.4
(0.7) (0.6) (0.5) (0.6)
(0.3) (0.3) (0.2) (0.3)
(0.3) (0.3) (0.2) (0.3)
NOTE.—Standard errors in parentheses.
then it is likely that the percentage of cell-phone only households is underestimated. Table 1 also presents the telephone service distributions by key household characteristics. The table shows that the percentage of households that are cell-only varies by the characteristics of the households. The percentage is small in the northeast (3.9 percent), and large in central cities (8.1 percent). The northeast region had the lowest percentage of households that were cellonly within each of the categories of metropolitan status, although not all the differences were statistically significant. The percentage of one-person households that are cell-only (8.1 percent) is greater than that of larger households (5.3 percent). The contrasts are sharpest by home ownership and type of dwelling unit. Cell-only households constitute a much larger percentage of rented units (12.8 percent) than of owned units (3.1 percent), and a much larger percentage of multi-unit dwellings (12.6 percent) than single-unit dwellings (4.0 percent). The percentage of cell-only households and the differences in the table can be combined to approximate the undercoverage bias estimate for simple means from RDD surveys. For example, assuming only coverage errors the
10
Tucker et al.
Table 2. Percent of Adults, by Telephone Service Status and Adult Characteristics (n ¼ 84,062) Total Land & cell Land only Cell only Total (%) Age 18–24 25–34 35–54 55þ Race Hispanic Black non-Hispanic All other races (including White, non-Hispanic) Marital status Married Not married Education level Less than high school High school diploma Some college Bachelor’s degree Masters or more
100.0
51.0 (0.2)
38.6 (0.2)
100.0 100.0 100.0 100.0
38.9 54.0 60.9 39.3
28.4 29.8 30.3 54.6
100.0 100.0 100.0
38.7 (0.6) 39.9 (0.7) 54.6 (0.3)
43.8 (0.6) 45.6 (0.7) 36.8 (0.3)
100.0 100.0
55.3 (0.3) 40.0 (0.4)
37.6 (0.3) 3.3 (0.1) 41.2 (0.4) 11.1 (0.3)
100.0 100.0 100.0 100.0 100.0
24.7 45.4 57.5 66.8 67.7
57.5 43.9 33.1 26.3 27.9
(0.1) (0.6) (0.3) (0.4)
(0.5) (0.4) (0.4) (0.5) (0.7)
5.4 (0.1)
None 4.9 (0.1)
(0.9) 20.1 (0.8) 12.6 (0.6) (0.5) 10.1 (0.4) 6.1 (0.3) (0.3) 4.7 (0.1) 4.1 (0.1) (0.4) 1.8 (0.1) 4.3 (0.2)
(0.6) (0.4) (0.4) (0.5) (0.7)
7.3 (0.3) 10.2 (0.4) 5.7 (0.3) 8.8 (0.4) 5.1 (0.1) 3.5 (0.1)
5.9 5.5 6.1 5.0 2.6
3.9 (0.1) 7.7 (0.2)
(0.3) 11.9 (0.4) (0.2) 5.3 (0.2) (0.2) 3.3 (0.2) (0.2) 1.9 (0.2) (0.2) 1.9 (0.2)
NOTE.—Standard errors in parentheses.
percent of single-unit dwellings would be overestimated by about 2 percent in an unadjusted RDD sample.4 The undercoverage biases are small, but the bias is likely to increase as the percentage of households with only cell phones increases. Table 2 gives characteristics of adults (persons 18 and older) by the telephone status of the household. It shows that while only 5.4 percent of adults live in cell-only households, 20 percent of young adults (aged 18–24 years) are in cell-only households. The percent of older adults in cell-only households are much lower (1.8 percent). The percentage distribution by race and ethnicity shows more Hispanic adults in cell-only households (7.3 percent) than other race and ethnic groups. Adults who are not married are notably more likely to live in a cell-only household (11.1 percent) than their married counterparts (3.3 percent). The distribution by education 4. Let the percent of single-unit dwellings be P, the overall coverage rate C, and the coverage rate for single-unit dwellings C1. The undercoverage bias is P(1 C1/C). In the example, P(1 C1/C) ¼ P(1 0.958/0.94) ¼ 1.9 percent P.
Household Telephone Service and Usage Patterns
11
level is not highly variable, with a range from 2.6 percent to 6.1 percent. Age and marital status are the two variables in the table that are subject to the largest undercoverage biases in RDD surveys, if no weighting adjustments are made. The Blumberg et al. (2006) findings from the NHIS for these characteristics are very similar, even though they cover January 2004 through June 2005. The NHIS estimates by region, home ownership, number of persons in a household, and age, all show the same differences discussed earlier for the CPS. Like the CPS estimates, the NHIS estimates large differences by home ownership, the number of adults in a household, and age. Even more subtle differences, such as the percentage of adults who are cell-only being lower in the northeast region than other regions, are consistent for the two surveys. To investigate the potential for undercoverage bias in a substantive variable in an RDD survey, we consider the employment status of household members. The unemployment rate (not seasonally adjusted or composited) for all households in February 2004 was 5.9 percent. The rate for cell-only households and households with no service was 7.9 percent. If the CPS excluded cell-only households and those with no telephone service, the estimated unemployment rate would have been 5.7 percent, a 0.2 percent underestimate. As a way to explore approaches to reduce undercoverage bias in RDD surveys, logistic regression models using household and person level characteristics were examined. One model predicted living in a cell-only household and a second predicted living in a household without a landline (combining cell-only and no telephone status). These exploratory models were computed using forward stepwise logistic regression, using the final poststratified weights normalized to sum to the effective sample size.5 The parameters of the final model for predicting cell-only status are shown in table 3, and for predicting no landline status in table 4. The significant predictors in the regression models are consistent with the descriptive results given earlier. The household level variables that are significant predictors for both models are type of unit (single, multi-unit), home ownership, presence of a child in the household, and region. At the adult level, age, education level, and marital status are significant predictors in both models. For predicting cell-only status, race/ethnicity, and employment status are important predictors, while for predicting no landline phone, metropolitan status, and number of household members are significant. The interactions for the two models differ, in that, only interactions between age and employment status predict cell-only households, but a number of interactions are involved in determining whether a household has a landline. 5. The normalized weights may not account for the full clustering effect of the sample design in the CPS.
12
Tucker et al.
Table 3. Logistic Regression Model for Predicting Cell-only Telephone Status (n ¼ 28,955) Predictors of cell-only telephone status Region: Northeast Region: Midwest Region: South Region: West (reference cell) Age: 18–24 Age: 25–34 Age: 35–54 Age: 55þ (reference cell) White Black All other ethnicities (reference cell) Less than high school High school diploma Some college Bachelor’s degree Masters or more (reference cell) Owner occupied dwelling Single-unit dwelling Married Not employed No child in household Interactions 18–24* Not employed 25–34* Not employed 35–54* Not employed Max-rescaled R2 Likelihood Ratio 2 Prob( 2)
Estimate
SE
Odds ratio
0.32 0.23 0.36
0.09 0.08 0.07
0.7 1.3 1.4
1.71 1.23 0.59
0.13 0.12 0.11
5.6 3.4 1.8
0.09 0.31
0.11 0.13
1.1 0.7
0.70 0.65 0.65 0.44
0.14 0.12 0.12 0.13
2.0 1.9 1.9 1.5
0.84 0.44 0.51 1.00 0.49
0.07 0.07 0.06 0.15 0.06
0.4 0.6 0.6 0.4 1.6
0.58 0.74 1.01
0.21 0.20 0.18
1.8 2.1 2.7
0.18 829.3 50.01
The variance explained by these logistic regressions is rather small, as is typical when estimating a rare characteristic. Nevertheless, some of these variables such as home ownership, age, type of dwelling unit, and marital status should be considered for adjusting the weights of RDD surveys that exclude cell phones. Keeter (2006) shows that simply adjusting for age largely eliminated the bias due to cell-only households in estimating the vote in the 2004 presidential election. In addition to undercoverage bias, nonresponse bias in RDD surveys might be expected if some households with both landlines and cell phones primarily use their cell phones and infrequently answer their landlines. To provide some insight into this possibility, the households with both landlines and cell
Household Telephone Service and Usage Patterns
13
Table 4. Logistic Regression Model for Predicting No Landline Telephone Status (n ¼ 30,642) Predictors of no landline telephone status Region: Northeast Region: Midwest Region: South Region: West (reference cell) Age: 18–24 Age: 25–34 Age: 35–54 Age: 55þ (reference cell) Less than high school High school diploma Some college Bachelor’s degree Masters or more (reference cell) Central city Metro, noncentral city Non-Metro (reference cell) 1 member household 2 member household 3–4 member household 5þ member household (reference cell) Owner occupied dwelling Single-unit dwelling Married No child in household Interactions: Northeast *married Midwest *married South*married 5 H.S.*single-unit H.S. diploma*single-unit Some college*single-unit Bachelor’s*single-unit No child in hh*2 in hh No child in hh*3–4 in hh Single-unit*owned Max-rescaled R2 Likelihood Ratio 2 Prob( 2)
Estimate
SE
Odds ratio
0.28 0.34 0.41
0.09 0.08 0.07
0.8 1.4 1.5
1.49 1.04 0.61
0.08 0.07 0.06
4.4 2.8 1.8
0.87 0.70 0.55 0.34
0.15 0.14 0.14 0.15
2.4 2.0 1.7 1.4
0.10 0.25
0.06 0.06
0.9 0.8
0.04 0.25 0.23
0.15 0.12 0.08
1.0 1.3 0.8
0.59 0.51 0.28 0.30
0.09 0.19 0.09 0.16
0.6 0.6 0.8 1.4
0.27 0.30 0.05 0.56 0.19 0.07 0.01 0.44 0.10 0.34
0.13 0.12 0.11 0.20 0.19 0.20 0.21 0.19 0.18 0.11
1.3 0.7 1.1 1.8 1.2 1.1 1.0 0.6 1.1 0.7
0.16 1143.1 50.01
14
Tucker et al.
Table 5. Percent of Households with Both Landlines and Cell Phones, by Calls Received on Cell Phone (n ¼ 14,960) Total Total (%) All or almost all More than half Less than half Very few or none
100.0 8.8 24.4 35.8 31.0
(0.2) (0.4) (0.4) (0.4)
Single unit
Multi-unit
100.0 7.1 23.5 36.5 33.0
100.0 18.0 29.1 31.9 20.9
(0.2) (0.4) (0.5) (0.4)
(0.8) (1.0) (1.0) (0.9)
NOTE.—Standard errors in parentheses.
phones in the supplement were asked how often they received calls on their cell phones. Table 5 shows that nearly 9 percent of these households receive all or almost all of their calls on cell phones. Suppose households who receive all or almost all their calls on their cells do not respond to surveys from their landline, then RDD surveys would not have data from the 6 percent of households without a landline plus the 4 percent that did not answer their landline phone (9 percent of the 46.2 percent of households with both types of service). In all, 10 percent of all households would not be represented. As table 5 shows the problem is more severe for some types of households such as households in multi-unit dwellings.6 Even though the weighting adjustments may reduce these potential biases, undercoverage problems are going to become more important as the prevalence of cell phones continues to grow. One approach to explore the types of households with both landline and cell phones that might be underestimated due to nonresponse if cell phone numbers are not included in the sampling frame is to use logistic regression to predict adults who take more than half of their calls on a cell phone. A logistic regression model with this as the outcome variable was estimated in the same fashion as the early models, and the results are shown in table 6. The odds ratios for receiving at least half their calls on a cell phone are less than unity for households in the northeast and married households, while they are greater than unity for young-adult households and urban households.
Issues Associated with Sampling Cell Phones In this section we identify and suggest methods that might be useful in dealing with issues that must be addressed if cell phone numbers are 6. A Rao–Scott 2 test shows the distributions in table 5 are significantly different (p-value 5.01).
Household Telephone Service and Usage Patterns
15
Table 6. Logistic Regression Model for Predicting Whether More than Half of Incoming Calls Received on Cell Phone for Households with Cell and Landline Service (n ¼ 14,013) Predictors
Estimate
Region: Northeast 0.49 Region: Midwest 0.24 Region: South 0.04 Region: West (reference cell) Central city 0.43 Metro, noncentral city 0.51 Non-Metro (reference cell) Single-unit dwelling 0.34 Age: 18–24 0.74 Age: 25–34 0.46 Age: 35–54 0.25 Age 55þ (reference cell) White 0.22 Black 0.46 All other ethnicities (reference cell) Married 0.36 No child in household 0.09 Not employed 0.46 Interactions Age: 18–24*No child 0.66 Age: 25–34*No child 0.82 Age: 35–54*No child 0.35 Max-rescaled R2 Likelihood Ratio X2 Prob(X2)
SE
Odds ratio
0.06 0.06 0.05
0.6 0.8 1.0
0.07 0.06
1.5 1.6
0.06 0.23 0.17 0.16
0.7 2.1 1.6 1.3
0.08 0.10
0.8 0.6
0.05 0.16 0.05
0.7 0.9 0.6
0.26 0.19 0.18
1.9 2.3 1.4
0.09 469.5 50.01
to be included in the sampling frame for telephone surveys. For example, table 5 shows that about 30 percent of all households with both landlines and cell phones report receiving very few or none of their calls on their cell phones. Even though there are potential reporting problems for this item, the data suggest that a substantial percentage of households receive very few calls on their cell phone. It is likely that these households might respond at a lower rate to their cell phones than others, potentially leading to nonresponse bias. In a dual frame survey of landlines and cell phones, Brick et al. (2006) found evidence of this type of nonresponse. Another important question is whether the sampling unit for telephone samples should change from the household to the person as we begin to interview persons on cell phones. We assume that the unit of analysis is an adult, not a household as is typical in RDD surveys. If a cell phone is
16
Tucker et al.
Table 7. Percent of Multiple-Adult Households with Cell Phones in which the Cell is Answered by more than One Person in the Household (n ¼ 11,931) Group
Total
Not shared
Shared
Total (%) Both landline and cell Cell only
100.0 100.0 100.0
49.8 (0.5) 50.9 (0.5) 33.9 (1.8)
50.2 (0.5) 49.1 (0.5) 66.1 (1.8)
NOTE.—Standard errors in parentheses.
a personal device that can be uniquely linked to the person who answers it, then the sampling unit should be the person who answers the phone. A consequence of doing this is that adults in cell-only households who do not have a cell phone are undercovered. An alternative is the traditional approach to RDD sampling with the household as the sampling unit. In this scheme, the person answering the cell phone identifies all the adults in the household and an adult is sampled. The sampled adult then must be interviewed and this might involve calling another telephone number. Table 7 gives estimates from the supplement that show the link between cell phones and individuals in 2004 which, is far from unique. These estimates are based on responses to item Q2e (item Q2c if there is only one cell phone in the household) that asked if more than one member of the household answers this cell phone. The estimates are for households that have at least one cell phone and have more than one adult. In households with both cell phones and landlines that have more than one adult, 49.1 percent of the households said that more than one person answers this cell phone. The percentage of households in which this cell phone is answered by others is even larger for cell-only households (66.1 percent). Figure 2 shows that as the number of cell phones in the household increases, the percentage of households that have more than one person answering this cell phone decreases somewhat, except when there is only one cell phone in the household.7 The estimates from the CPS supplement suggest that shifting the sampling unit for cell phones to the person may cause some sampling problems, at least in the environment existing in 2004. Brick et al. (2007), in this issue of POQ, address this matter more completely based on a survey that sampled both landlines and cell phones.
7. The amount of ‘‘sharing’’ of cells may be less than these estimates show due to the wording of the question or the fact that it is asked at the household level. Even if the estimates are considerable overestimates of sharing, the key point is that many cell phones may be answered by more than one person.
Household Telephone Service and Usage Patterns
17
70 60
Percent
50 40 30 20 10 0 1
2
3
4
Numberof cell phones in household
Figure 2. Percent of households with more than one adult that share at least one cell phone, by the number of cell phones in the household. If the household remains the sampling unit, another issue is accurately assessing the probability of selecting a household. A household could be sampled on any of its cell and landline phone numbers, and ascertaining the probability associated with these multiple phone numbers is essential for producing unbiased estimates using a weighting adjustment for ‘‘multiplicity.’’ The ordinary RDD type of multiplicity adjustment (e.g., Massey and Botman 1988) divides the sampling weight by the number of lines in the household. This may be problematic in households with both types of service. First, all the different land and cell lines have to be identified, and this may result in response errors. Second, the varying weights might increase the variance of the estimates and create an inefficiency in the sample design. Third, the adjustment procedure in this situation is rather complex. Other approaches are possible and might offer some benefits. For example, if a dual frame approach is used, then only the number of lines of the type of service from which the number is sampled is needed (e.g., see Brick et al. 2007 in this issue of POQ). Table 8 shows the CPS supplement estimates of the number of households by the type of telephone service, the number of landlines and the number of adults in the household. Overall, the supplement estimates about 7 percent of households with a landline have more than one landline number, which is consistent with our experience in RDD surveys. The mean number of landlines for landline-only households is 1.02, and for households with both landlines and cells it is 1.12. Table 9 shows the same types of estimates for cell phones. Overall, about 49.5 percent of households with cell phones have more than one cell phone. The mean number of cell phones in households with both landlines and cells is 1.68 and in cell-only households is 1.44, both much larger than the means for landlines. With such a large number of cell phones there may be more
18
Tucker et al.
Table 8. Percent of Household with Landlines, by Number of Landlines, Type of Service and Number of Adults (n ¼ 28,694) Number of landlines Type of service Total Land & cell Land only 1-adult households Land & cell Land only 2-adult households Land & cell Land only 3þ adult households Land & cell Land only
Total %
1
2
3þ
100.0 100.0
89.7 (0.3) 97.4 (0.1)
9.2 (0.2) 2.4 (0.1)
1.1 (0.1) 0.3 (0.05)
100.0 100.0
96.1 (0.6) 99.0 (0.2)
3.6 (0.6) 0.8 (0.2)
0.3 (0.2) 0.2 (0.1)
100.0 100.0
90.0 (0.6) 97.4 (0.4)
9.1 (0.6) 2.6 (0.4)
1.0 (0.2) 0.2 (0.1)
100.0 100.0
82.4 (1.2) 92.5 (1.2)
15.1 (1.1) 7.0 (1.1)
2.5 (0.5) 0.5 (0.3)
NOTE.—Standard errors in parentheses.
difficulties in getting respondents to accurately report the number of cell phones. Clearly, some of the shortcut methods used in standard RDD surveys such as only determining if the household has more than one line may be inadequate if cell phone numbers are sampled. Furthermore, samples from cell phone numbers will inherently have a larger design effect due to the multiple selection probabilities.
The Standard CPS Telephone Availability Questions Concerns about the reliability of questions on the supplement were discussed earlier. Unfortunately, the problems are not limited to the supplement items. At the end of the first CPS interview, respondents are asked if there is a telephone in the household (see the Appendix for the item). If the household does not have a telephone, the respondent is asked whether a phone is available elsewhere. These questions were intended to gather information for recontacting the household in the subsequent months, not to measure telephone service. Yet, they are often used for classifying households by telephone service. Table 10 shows that only about 38 percent of the households reporting no phone in the household in the first interview reported no phone in the supplement, and 21 percent of these households reported in the supplement that they only had a cell phone. For households reporting in the CPS that they
Household Telephone Service and Usage Patterns
19
Table 9. Percent of Household with Cells, by Number of Cells, Type of Service, and Number of Adults (n ¼ 16,807) Number of cells Type of service Total % All households Land & cell 100.0 Cell only 100.0 1-adult households Land & cell 100.0 Cell only 100.0 2-adult households Land & cell 100.0 Cell only 100.0 3þ adult households Land & cell 100.0 Cell only 100.0
1
2
3
4
5þ
48.4 (0.4) 39.4 (0.4) 65.7 (1.1) 26.3 (1.1)
8.6 (0.2) 6.7 (0.6)
2.9 (0.1) 0.6 (0.1) 1.1 (0.2) 0.1 (0.1)
93.3 (0.8) 97.3 (1.0)
5.5 (0.7) 2.5 (0.9)
1.0 (0.3) 0.2 (0.3)
0.1 (0.1) 0.1 (0.1) – –
40.3 (0.9) 52.2 (1.0) 48.2 (3.5) 46.2 (3.5)
6.1 (0.5) 5.0 (1.5)
1.3 (0.2) 0.2 (0.1) 0.6 (0.5) 0.0 (0.1)
31.5 (1.5) 31.8 (1.5) 24.0 (1.4) 10.3 (1.0) 2.4 (0.5) 32.4 (6.0) 31.1 (5.9) 29.0 (5.8) 6.3 (3.1) 1.2 (1.4)
NOTE.—Standard errors in parentheses.
Table 10. Percent of Households, by Telephone Service in Supplement and Standard CPS (n ¼ 32,969) Standard CPS telephone status Supplement telephone status Telephone in household No telephone in household Total (%) Both landline and cell Landline only Cell only No phone
100.0 48.6 43.4 5.0 3.0
(0.3) (0.3) (0.1) (0.1)
100.0 15.6 25.3 20.7 38.4
(0.8) (1.0) (0.9) (1.1)
NOTE.—Standard errors in parentheses.
had neither a phone in the household nor could be reached by phone elsewhere, 60 percent reported they had telephone service in the supplement, with 42 percent of these having a landline or both a landline and cell. These results raise questions about the validity of respondent reports about telephone service in the standard CPS questions. It is quite likely that the proportion of households reporting no telephone service in the CPS is overestimated. Both the 2000 Census and the NHIS (Blumberg et al. 2006) reported 52 percent of households without any telephone service, while the CPS estimate is over 5 percent. One possible reason for the overstatement
20
Tucker et al.
is that the CPS question is prefaced by the interviewer stating that they would like to conduct future interviews by telephone. After hearing this, some respondents may be reluctant to tell the interviewer they have a telephone in the home or could be reached by phone elsewhere. Recently, the first question was modified to tell respondents to include cell phones, but it is not clear that this change will reduce the overstatement.
Discussion and Future Research The collection of data on cell phone penetration and usage in 2004 provides a benchmark as the cell-only households are becoming a more serious issue for telephone surveys. The NHIS estimates for 2005 given in Blumberg et al. (2006) support our belief that cell-only households will become even more prevalent in the future. We also assume the percentage of households without any telephone service will remain very small. While it is likely that the percentage of cell-only households will plateau, there is no evidence that plateau is near-at-hand. The results from the CPS and the CEIS indicate that survey researchers will need to develop methods to include cell-only households in their samples to combat the potential bias due to undercoverage. In addition to developing new sampling methodologies, operational methods for actually contacting cell phone owners and convincing them to participate are needed. Conducting surveys by cell phones also raises a host of legal and other issues. Some efforts to identify and resolve these issues have already begun (e.g., Steeh 2004, Brick et al. 2007, and Lavrakas and Shuttles 2005), and more will surely follow. For those interested in estimating telephone service and usage in the future, we suggest some caution. We have identified some problems with both the intent and meaning of some terms used in the CPS supplement that need to be resolved. In addition, the supplement enabled us to clearly identify some problems with the basic CPS telephone service questions.
Appendix TELEPHONE SUPPLEMENT QUESTIONS
Q1.
First I would like to ask about any regular, landline telephone numbers in your household. These numbers are for phones plugged into the wall of your home and they can be used for different reasons, including making or receiving calls, for computer lines or for a fax machine. How many different landline telephone numbers does your household have?
Household Telephone Service and Usage Patterns
21
VER1.
I’d like to verify the information you just provided. I believe you indicated that your household has NO LANDLINE TELEPHONE service for incoming and outgoing calls: Is that correct? VER2. I just want to verify that your household has (Q1) distinct telephone NUMBERS: Is that correct? Q1a. Excluding any numbers used only for faxes and computers, how many of these (Q1) landline telephone numbers are used for incoming calls? Q1b. Excluding a number used only for a fax or computer, do you or any other member of your household take incoming calls on a landline number? Q2. {Excluding students living away at school,} Do you or any other members of your household have a working cellular phone number? Q2a. {Excluding students living away at school,} How many different cell phone numbers do the members of your household have? Q2b. How many of the (Q2a) cell phone numbers do you or any other members of your household use regularly? Q2c. How many of the (Q2a) cell phone numbers are answered by more than one household member? Q2d. Do you regularly answer this cell phone number? Q2e. Is this cell phone number answered by more than one household member? Q3. Of all the phone calls that you or any other members of your household receive, about how many are received on a cell phone? Would you say . . . All or almost all calls, More than half, Less than half, or Very few or none?
1 2 3 4
STANDARD CPS TELEPHONE AVAILABILITY QUESTIONS
Since households included in this survey are interviewed during each of the next 3 months, we attempt to conduct the follow-up interviews by telephone. Is there a telephone in this house/apartment? (If No) Is there a telephone elsewhere on which people in this household can be reached?
References Blumberg, Stephen, Julian Luke, and Marcie Cynamon. 2006. ‘‘Telephone Coverage and Health Survey Estimates: Evaluating Concern about Wireless Substitution.’’ American Journal of Public Health 96:926–31.
22
Tucker et al.
Brick, Michael, Pat Brick, Sarah Dipko, Stanley Presser, Clyde Tucker, and YangYang Yuan. 2007. ‘‘Cell Phone Survey Feasibility in the U.S.: Sampling and Calling Cell Numbers Versus Landline Numbers.’’ Public Opinion Quarterly 71. Brick, Michael, Sarah Dipko, Stanley Presser, Clyde Tucker, and YangYang Yuan. 2006. ‘‘Nonresponse Bias in a Dual Frame Sample of Cell and Landline Numbers.’’ Public Opinion Quarterly 70:780–93. Esposito, James. 2004. An Evaluation of the CPS Cell-Phone_Use Supplement: Composite Report. Unpublished final report. Washington, DC: Bureau of Labor Statistics. Keeter, Scott. 2006. ‘‘The Impact of Cell Phone Noncoverage on Polling in the 2004 Presidential Election.’’ Public Opinion Quarterly 70:88–98. Lavrakas, Paul, and Charles Shuttles. 2005 ‘‘Cell Phone Sampling Summit II Statements on Accounting for Cell Phones in Telephone Survey Research in the U.S.’’ Avaliable at http:// www.nielsenmedia.com/cellphonesummit/statements.html. Massey, James, and Steven Botman. 1988. ‘‘Weighting Adjustments for Random Digit Dialed Surveys.’’ In Telephone Survey Methods, eds. Groves, Biemer, Lyberg, and Massey pp. 143–160. New York: John Wiley & Sons. Piekarski, Linda. 2003. ‘‘Cellular Phones: Challenges and Opportunities.’’ Survey Research, pp. 34(2). Steeh, Charlotte. 2004. ‘‘A New Era for Telephone Surveys.’’ Paper presented at the Annual Conference of the American Association for Public Opinion Research, Phoenix, AZ. Tuckel, Peter and Sally Daniels. 2006. Usage Patterns of the Cell Phone: 2000–2006. Paper presented at the annual conference of the American Association for Public Opinion Research, Montreal, Canada. U.S. Department of Labor [Bureau of Labor Statistics] and U.S. Department of Commerce [Bureau of the Census]. 2002. Current Population Survey Technical Paper 63 (Revised) Design and Methodology. U.S. Census Bureau. 2005. ‘‘Current Population Survey, 2004: Cell Phone Use Supplement Technical Documentation.’’ Avaliable at http://www.census.gov/apsd/techdoc/cps/ cpscel04.pdf.
