Download:
pdf |
pdfUnited States
Department of
Agriculture
Economic
Research
Service
Technical
Bulletin
Number 1934
August 2012
Nonresponse Bias Analysis
of Body Mass Index Data
in the Eating and Health
Module
Karen S. Hamrick
www
.e
sda.gov
u
.
rs
Visit Our Website To Learn More!
www.ers.usda.gov
Recommended citation format for this publication:
Hamrick, Karen S. Nonresponse Bias Analysis of Body Mass Index Data
in the Eating and Health Module, TB-1934, U.S. Department of Agriculture,
Economic Research Service, August 2012.
Photo: Thinkstock.
The U.S. Department of Agriculture (USDA) prohibits discrimination in all its programs
and activities on the basis of race, color, national origin, age, disability, and, where
applicable, sex, marital status, familial status, parental status, religion, sexual
orientation, genetic information, political beliefs, reprisal, or because all or a part of an
individual's income is derived from any public assistance program. (Not all prohibited
bases apply to all programs.) Persons with disabilities who require alternative means
for communication of program information (Braille, large print, audiotape, etc.) should
contact USDA's TARGET Center at (202) 720-2600 (voice and TDD).
To file a complaint of discrimination write to USDA, Director, Office of Civil Rights, 1400
Independence Avenue, S.W., Washington, D.C. 20250-9410 or call (800) 795-3272
(voice) or (202) 720-6382 (TDD). USDA is an equal opportunity provider and employer.
A Report from the Economic Research Service
United States
Department
of Agriculture
www.ers.usda.gov
Technical
Bulletin
Number 1934
August 2012
Nonresponse Bias Analysis of
Body Mass Index Data in the
Eating and Health Module
Karen S. Hamrick, [email protected]
Abstract
The ERS Eating and Health Module, a supplement to the American Time Use Survey
(ATUS), included questions on height and weight so that respondents’ Body Mass Index
(BMI—a measure of body fat based on height and weight) could be calculated and
analyzed with ATUS time-use data in obesity research. Some respondents did not report
height and/or weight, and BMIs could not be calculated for them. Analyses focusing
on correlations between BMIs and time use could be biased if respondents who did not
report height and/or weight differ significantly in other observable characteristics from
the rest of the survey respondents. However, findings reveal that any nonresponse bias
associated with the height and weight data appears to be small and would not affect
future analyses of BMIs and time-use pattern correlations.
Keywords: time use, American Time Use Survey, Eating and Health Module, nonresponse bias, item nonresponse, Body Mass Index, BMI, dissimilarity index, paradata
Acknowledgments
The author wishes to acknowledge the support and assistance of USDA, ERS colleagues
Jayachandran Variyam, Ephraim Leibtag, and Jeremy Weber. Also, the author wishes
to thank Rachel Krantz-Kent, Rose Woods, others on the American Time Use Survey
staff, and an anonymous reviewer for their helpful comments. This report also benefited
from comments from audience members at the 2012 Federal Committee on Statistical
Methodology Research Conference. Special thanks are extended to John Weber and
Wynnice Pointer-Napper for editorial and design assistance.
Contents
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Relevant Literature—Item Nonresponse, Time Use, and BMI . . . . . . . 4
Missing BMIs—Characteristics of Respondents . . . . . . . . . . . . . . . . . . . 7
Missing BMIs—Time-Use Profiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Is a Remedy for Missing BMI Data Needed? . . . . . . . . . . . . . . . . . . . . . 25
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
ii
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Summary
What Is the Issue?
ERS collected data on Americans’ time-use patterns and food-related information in its Eating and Health Module (EH Module), a supplement to the
American Time Use Survey (ATUS). The EH Module also included questions on height and weight so that respondents’ Body Mass Index (BMI—a
measure of body fat based on height and weight) could be calculated and
analyzed with ATUS time-use information. Though the EH Module had a
high rate of cooperation among respondents, just under 5 percent of respondents did not report height and/or weight, and ERS could not calculate
BMIs for these individuals. This raises concerns of bias in the data due to
the missing BMI observations. In this report, ERS examines the BMI data
to determine if the missing values hinder the ability of researchers to use
the data in future analyses. If respondents who did not report height and/or
weight differed significantly in other observable characteristics from the rest
of the survey respondents, then time-use estimates may be higher or lower
than they would be if BMIs were available for all respondents.
What Did the Study Find?
• Respondents who did not report height and/or weight had disproportionately higher indicators of being reluctant or uncooperative survey participants than other respondents. For example, it took more phone calls over
more weeks to obtain a completed interview from these participants. This
suggests that for these respondents, the tendency to not report height and/
or weight had less to do with sensitivity to height and weight questions
and more to do with negative views toward participating in the survey.
• The time-use profiles of the total population and of men with missing
BMIs closely resembled the profiles of respondents who were normal
weight (18.5 ≤ BMI < 25.0).
• The time-use profiles of women with missing BMIs closely resembled the
profiles of women who were overweight (25.0 ≤ BMI < 30.0).
• These findings suggest that those who did not report height and weight
are unlikely to be at either end of the BMI spectrum—underweight
(BMI<18.5) or obese (BMI>30.0)—mitigating any bias in the data.
Based on these findings, any bias in the EH Module height and weight data
stemming from nonresponse appears to be small and would not affect future
analyses of correlations between BMI and time use.
How Was the Study Conducted?
This study used data from the Bureau of Labor Statistics American Time Use
Survey and the ERS Eating and Health Module for 2006-08. Researchers
analyzed demographic characteristics, such as gender and age, of respondents who did not provide weight or height information. Data quality
measures (e.g., completeness or incompleteness of diary reports recording
respondents’ activities) served as indicators of respondents’ reluctance or
iii
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
uncooperativeness toward participating in the ATUS. Multivariate analysis
was performed on respondents’ demographic and survey characteristics.
A measure of dissimilarity was used to compare time profiles across BMI
groups to determine which BMI group most resembled the respondents with
missing BMIs in terms of activities reported in the time diaries.
iv
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Introduction
ERS collected data on Americans’ time-use patterns and food-related information in its Eating and Health Module (EH Module), a supplement to the
American Time Use Survey (ATUS). The module included questions on
height and weight so that respondents’ Body Mass Index (BMI—a measure of
body fat based on height and weight) could be calculated and analyzed with
ATUS time-use data in obesity research. Because some survey respondents
refused to report their height and/or weight, ERS could not calculate BMIs
for these individuals. Does the lack of BMIs for some respondents create bias
in the data? Bias would occur if respondents without corresponding BMIs
had different time-use patterns than other respondents, resulting in under- or
over-estimates of average time spent in various activities. This report presents
an examination of the BMI data to determine if the missing values hinder the
ability to use the data in other analyses.
In examining the potential for bias, ERS provides technical information in
this report that will benefit other researchers using the Eating and Health
Module data. In addition, ERS contributes to the literature by investigating
nonresponse (i.e., missing information for a survey question) in time use
surveys, and the ATUS in particular, and by using paradata (i.e., data about
the process of data collection) in the examination of this nonresponse.
Eating and Health Module
The ATUS, sponsored by the Bureau of Labor Statistics (BLS) and conducted
by the U.S. Census Bureau, has collected time use data nearly every day since
the survey was implemented in 2003. In the survey, one individual age 15 or
older from each sampled household is interviewed about his or her activities
for the 24-hour period from 4 a.m. the day before the interview to 4 a.m. of
the interview day. Survey respondents identify their main activity during any
time period for which they were engaged in more than one activity at a time.
They also report where they were and whom they were with when the activities occurred. All ATUS respondents also participated in the BLS Current
Population Survey (CPS). During the ATUS interview, they were given the
opportunity to update information on household labor force participation that
they provided to the CPS. Thus, the ATUS data include time diary, demographic, labor force participation, and household information.
The EH Module, a supplement to the ATUS developed by ERS and funded
by ERS and the National Cancer Institute, was fielded from January 2006 to
December 2008, producing 3 full years of data. The ATUS collected data
from over 12,000 respondents each year. From 2006 to 2008, the ATUS and
EH Module collected data from 37,832 respondents.1 Weighting factors calculated by the U.S. Census Bureau and applied to the individual respondent data
enabled the ATUS and EH Module to produce nationally representative estimates for an average day over 2006-08.
The EH Module contained questions on eating patterns; height, weight, and
health status; USDA Supplemental Nutrition Assistance Program2 participation and household income; meals obtained at school by household children;
and grocery shopping and meal preparation (see box, “ATUS Eating and
1A small number of respondents
(82, or 0.2 percent of the total sample)
completed the ATUS survey but did
not complete the EH Module.
2As
of October 1, 2008, the Food
Stamp Program was renamed the
Supplemental Nutrition Assistance
Program.
1
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Health Module Questions”). Since the ATUS is a time-use survey, it does not
include food intake information. Nonetheless, the ATUS data provide important information on eating/drinking duration, frequency, and context that
allow for the characterization of eating patterns of different groups.
Data on height and weight collected through the EH Module enable
researchers to examine for correlation between BMI and time-use patterns.
This may be particularly useful in the design of programs targeted at
reducing the incidence of obesity, the most common food and nutritionrelated health problem in the United States. Information on patterns of time
spent in various activities by those who are obese can help researchers understand how behaviors differ among people of different weight status.
The Eating and Health Module microdata can be downloaded from BLS
at http://stats.bls.gov/tus/ehdatafiles.htm; estimates and documentation are
available at www.ers.usda.gov/data-products/eating-and-health-module(atus).aspx
2
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
ATUS Eating and Health Module Questions
day program? This question refers to ONLY
BREAKFASTS prepared at the school or
center—not meals brought from home.
Eating as a secondary activity
Because many Americans eat while engaged in other
activities, such as driving or watching television, information
is needed on eating as both a primary and secondary activity.
This question records when and during which activities the
respondent was eating or drinking beverages.
What about LUNCH? In the past week, did [fill
names of children in the household 18 or under]
eat a LUNCH that was prepared and served at
a school, a paid day care or Head Start Center,
or a summer day program? This question refers
to ONLY LUNCHES prepared at the school or
center—not meals brought from home.
Question:
We’re interested in finding out more about
how people fit meals and snacks into their
schedules. Yesterday, you reported eating or
drinking between [fill times from respondent’s
time diary]. Were there any other times you
were eating yesterday—for example while you
were doing something else? About how long
would you say you were eating while you were
[fill activity]? Not including plain water, were
there any other times yesterday when you were
drinking any beverages? About how long would
you say you were drinking while you were [fill
activity]?
Height, weight, and general health
From this self-reported information, Body Mass Index
(BMI) can be calculated, and time-use patterns, such as
activity levels and eating patterns, can be analyzed by BMI
levels.
Question:
I’m going to switch topics and ask you a few
final questions about your physical health that
might affect how you use your time. In general,
would you say that your health is Excellent,
Very Good, Good, Fair, or Poor? How tall are
you without shoes? How much do you weigh
without shoes?
Grocery shopping and meal preparation
Question:
I’d like to ask a couple of questions about food
preparation. Are you the person who usually
does the grocery shopping in your household?
Are you the person who usually prepares the
meals in your household?
Household income
This question asks if total household income before
taxes was above or below a certain amount. The ATUS
Computer Assisted Telephone Interviewing software
automatically calculated the dollar amount of 185 and 130
percent of the poverty threshold based on the respondent’s
household composition. These income thresholds—185
percent and 130 percent—determine income eligibility
for food assistance programs.
Food Stamp Program participation
This information allows analysis of the time-use patterns
of food stamp recipients versus others, and in particular,
low-income persons who are not participating in the
program.
Question:
In the past 30 days, did you or anyone in your
household get food stamp benefits?
Question:
Last month, was your total household income
before taxes more or less than [fill 185 percent
of poverty threshold] per month?
Breakfast and lunch obtained at school
Question:
Please think back over the past week starting
last Monday up to today, Monday. In the past
week, did [fill names of children in the household 18 or under] eat a BREAKFAST that
was prepared and served at a school, a paid
day care or Head Start Center, or a summer
If answer was LESS:
Was it more or less than [fill 130 percent of
poverty threshold] per month?
A text version of the Eating & Health Module questionnaire
is available at www.bls.gov/tus/ehmquestionnaire.pdf
3
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Relevant Literature—Item Nonresponse,
Time Use, and BMI
Considerable research has been conducted on the effects of unit nonresponse in household surveys. Unit nonresponse refers to sampled individuals
or households who choose not to participate in a survey. In such cases, the
concern is that estimates based on survey data affected by unit nonresponse
may have an increased variance and a bias. A bias would result if those who
do not participate have different characteristics related to the survey topic
than those who do respond to the survey, and so an estimate calculated from
survey data would be different than the actual population value. The literature in this area is extensive (see, for example, Groves (2006), Groves et al.
(2002), and Bethlehem et al. (2011)).
Abraham et al. (2006, p. 676) investigated unit nonresponse in the ATUS and
concluded, “We find little support for the hypothesis that busy people are less
likely to respond to the ATUS but find considerable support for the hypothesis that people who are weakly integrated into their communities are less
likely to be contacted.” However, reweighting the estimates to account for this
effect produced only modest differences in time use.
Abraham et al. (2008) also looked at unit nonresponse in the ATUS in relation to estimates of time spent in volunteer work. They investigated whether
the social processes that lead one to participate in a survey also lead one
to volunteer, which would result in an overrepresentation of volunteers in
the ATUS. Their findings reveal an association between the factors that
encourage volunteering and those that encourage survey participation,
meaning that time-use estimates would contain bias, in this case an overestimate, in time spent in volunteering and other pro-social activities. They do
not recommend an adjustment for the bias; however, they suggest strategies to
improve sample weighting to account for this nonresponse.
O’Neill and Sincavage (2004) conducted a response analysis survey on ATUS
nonrespondents to better understand the differences between respondents and
those who refused to respond. One-third of the ATUS nonrespondents cited
CPS-related survey fatigue—their decision not to participate in the ATUS
was based on their having previously participated in the CPS, which is an
8-month survey over a 16-month period. Several nonrespondents stated that
they were tired of the Census Bureau calling them, and they felt that participating in the CPS more than satisfied their requirement.
Item nonresponse occurs when an individual or household participates in a
survey but does not provide information for one or more questions. Several
factors may account for item nonresponse: The respondent may think that
the information requested by a question is too sensitive to reveal, such as
income;3 the respondent may think that a question is not directly related to
the survey topic; or the respondent may not know the answer to a question
and may not want to spend time researching the answer. It is also possible
that the respondent does not know the answer and cannot obtain the answer.
The research on item nonresponse is not as extensive as that on unit nonresponse; however, thorough overviews exist (see, for example, Bethlehem et al.
(2011, chapter 14), Dixon (2005), and Mason et al. (2002)).
3Tourangeau
et al. (2000, chapter 9)
provide a good discussion of sensitive
questions and nonresponse.
4
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Fricker and Tourangeau (2010, p. 941) investigated nonresponse, reluctant
respondents, and data quality in the CPS and in the ATUS. Although their
focus was unit nonresponse and data quality, they studied item nonresponse
as well. Their data quality indicators for ATUS were “(1) total number of
diary activities reported; (2) missing diary reports of basic daily activities; (3)
round values for activity durations; and (4) item nonresponse on ATUS laborforce questions.” They found that respondents with high nonresponse propensity scores had fewer activities reported in their time diaries. In addition, they
examined respondents who were refusal conversions, that is, they originally
refused to participate in the ATUS but later agreed to be interviewed. They
found that data quality for refusal conversion respondents was worse than for
other nonrespondents.
Other areas of survey-related research include late respondents and data
quality. Triplett et al. (1996) studied late respondents (converted-refusal
cases) to the 1993-94 time diary study conducted by the U.S. Environmental
Protection Agency. They concluded that the converted-refusal respondents
consistently provided less information than other respondents. Friedman et
al. (2003, p. 992) investigated whether the characteristics of late respondents
to the Health Care Survey of Department of Defense Beneficiaries were
different from those of other respondents. They used the “continuum of resistance” model that posits that “…individuals who require the most contacts
before participating in a survey are also the most resistant to being interviewed, and the more resistant a respondent, the more similar he or she is to
the most resistant individuals in the population—the nonrespondents.” They
studied these late respondents—those who needed the most callbacks before
participating—as a way of gaining insight into nonrespondents. They found
that late respondents have different characteristics than nonlate respondents
and are also more likely to have “don’t know,” “not applicable,” or just blank
responses to survey questions. As a consequence, their responses may be less
reliable than those of nonlate respondents.
Some survey methodologists recommend that researchers adjust the data
to avoid bias from nonresponse. Graham (2009) and Schafer and Graham
(2002) provide thorough overviews of dealing with missing data. Both
discuss the nature of “missingness” and review various approaches avoiding
bias. One such approach is to re-calculate the sampling weights to account
for the nonresponse. Estimates using the nonresponse-adjusted weights may
be used as population estimates or may be used to estimate the extent of any
nonresponse bias. See Abraham et al. (2006) and Dixon (2008) for examples
of weight recalculation. Another approach to dealing with missing data is to
impute values that are missing due to item nonresponse. Kyureghian et al.
(2011) used a parametric Bayesian model for multiple imputation methods,
and Andridge and Little (2010) discuss hot deck imputation methods.
The survey methodology literature on BMI focuses on bias of self-reported
height and weight measures. Representative research includes Danubio et
al. (2008), Kuczmarski et al. (2001), and Hill and Roberts (1998). Research
on item nonresponse on BMI is limited, and all of it pertains to children.
Wagstaff et al. (2009) imputed BMI values for individuals age 2-15 in the
1999-2000 National Health and Nutrition Examination Survey (NHANES).
BMI values are missing for 8 percent of the children/youth sampled. Wagstaff
et al. compared active imputation of BMI (imputing BMI directly) with
5
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
passive imputation of BMI (imputing height and weight). Using 1999-2000
NHANES data for non-Hispanic White youths age 2-18, they found little
difference between the active and passive BMI imputation methods. Nandram
and Choi (2010) used the NHANES III data to impute missing values for
those respondents age 2-19 with missing BMI. They modeled population
means of small domains of age, race, and sex within counties. They focused
on analysis of BMI of children/youth and created BMI “growth curves” by
age. Tiggemann (2006) studied missing BMI using data from the Flinders
University (Australia) survey of teenage boys and girls in South Australia that
investigated the role of media and adolescent self-image. Over one-fourth of
the sample had missing BMI values. Tiggemann concludes that nonreporting
of height and/or weight was not random and consistent with the “motivated
nonreporting hypothesis,” and, consequently, analysis without the missing
values or imputation with a mean BMI would produce biased estimates.
She also concludes that “…the current study provides an illustration of how
treating missing values as meaningful data can provide some useful information and raise some interesting questions” (p. 349).
6
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Missing BMIs—Characteristics
of Respondents
Obesity is the most common food and nutrition-related health problem in
America. Data from the ATUS and the EH Module enable researchers to
identify the types of activities and eating patterns that are associated with
obesity and those that are associated with healthy weight, overall health, and
well-being.
As mentioned earlier, researchers used the height and weight information gathered from EH Module respondents to calculate BMIs.4 They then
analyzed BMIs in conjunction with time diary, demographic, and labor
force information to better understand associations among BMI status and
time-use patterns and demographic and labor force characteristics. Although
BMIs were calculated from self-reported height and weight in the ATUS,
researchers found that differences between self-reported BMIs and measured
BMIs are small and, as a result, self-reported BMIs are acceptable for use
in analysis of data on nonelderly adults (see Cawley and Burkhauser, 2006;
Kuczmarski et al., 2001; and Danubio et al., 2008). In addition, the expected
underreporting of BMIs (through underreporting of weight and overreporting
of height) in the EH Module data does not appear to be large (Pinkston and
Stewart, 2009). So while the EH Module BMIs should not be used to obtain
an official measure of obesity in the United States, the data are suitable for
analyzing time-use behavior as it relates to BMI.5
Missing values for height and/or weight
EH Module data collected over 2006-08 reveal that only 1,848 respondents,
or 4.88 percent of 37,832 completed interviews, did not report height and/or
weight, and have missing BMI values (tables 1, 2). 6, 7 ERS did not impute
BMIs for any of the missing values. An additional 347 respondents (0.92
percent) told interviewers they were pregnant, and these individuals were not
asked about their weight. As a consequence, these respondents have missing
BMIs; however, they are not included in this analysis.
Item nonresponse is considered a source of nonsampling error.8 Respondents
who are willing to participate in the ATUS may be unwilling to answer
sensitive questions. In response to concerns about sensitivity, ERS placed
questions about general health, height, weight, and income at the end of
the survey.9 Respondents with missing BMIs may have declined to report
height and/or weight, perceiving this information as irrelevant to a time-use
survey.10 It is also possible that some individuals, such as growing teenagers
and elderly individuals with age-related weight loss, may not know their
current height and/or weight. Because the ATUS interviews are conducted
via computer-assisted telephone calls, respondents may be reluctant to pause
the interview to measure themselves.
Because some respondents did not report height, some did not report weight,
and some did not report either, the sum of missing height and missing weight
is greater than the total number of missing BMI (see table 3). For height and
especially for weight, more respondents refused to report values than those
who answered that they did not know their height and/or weight. However,
4Height and weight are bottom- and
top-coded for confidentiality. The EH
Module included a screening question
for pregnancy as pregnant women were
not asked their weight and thus have
missing BMIs.
5This
research does not address
whether BMI should be adjusted to
correct for underreporting of weight
and overreporting of height due to
self-reporting as the focus here is
on missing BMIs. See Pinkston and
Steward (2009) and Danubio et al.
(2008) for discussions of BMI bias
correction.
6All
estimates presented in
this report were calculated using
American Time Use Survey (ATUS)
data and Eating and Health Module
(EH Module) data for 2006-08.
The ATUS Respondent, Activity,
Activity Summary, Roster, Who, and
ATUS-Current Population Survey
files were used, along with the ATUS
Methodology Call History and Case
History files, and the EH Module
Respondent and Replicate Weights
files. The dataset used contains 37,832
respondents, with a total of 753,604
activities. Excluding respondents
who reported being pregnant results
in a total of 37,485 respondents. All
estimates and unweighted counts are
for individuals age 15 and over. All
calculations were done using SAS 9.2.
Estimation procedures outlined in the
ATUS User’s Guide (http://stats.bls.
gov/tus/atususersguide.pdf) and the
EH Module User’s Guide (http://ers.
usda.gov/Publications/AP/AP047/)
were followed. Averages were calculated as the mean unless otherwise
stated. Standard errors were calculated according to Section 7.5 of the
ATUS User’s Guide. The EH Module
Replicate Weights were used to calculate standard errors.
7Body mass index is calculated as
weight (lb) / [height (in)] 2 x 703. Adult
BMI groups are underweight (BMI
< 18.5), normal weight (18.5 ≤ BMI
< 25), overweight (25 ≤ BMI < 30),
and obese (30 ≤ BMI). For purposes
of interpreting BMI, the Centers for
Disease Control (CDC) defines adults
as those age 20 and over and uses a
different interpretation for youth and
teens. However, here these adult groupings are for convenience of exposition
used for all respondents age 15 and
7
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Continued on page 8
Table 1
Continued from page 7
ATUS/EH Module respondents by Body Mass Index (BMI) status
and by sex
older as the focus is on the missing
BMIs for the entire dataset and not on
analyzing time use patterns of those
who are, say, overweight. In analysis
by BMI, ERS uses the CDC adult
and youth/teen definitions. For more
information on BMI, see www.cdc.
gov/healthyweight/assessing/bmi/
index.html
Item
Men
Women
Total
Number
Missing BMI
Underweight (BMI<18.5)
Normal weight (18.5≤BMI<25.0)
Overweight (25.0≤BMI<30.0)
Obese (30.0≤BMI)
Total
389
148
4,669
6,814
4,407
16,427
1,459
471
8,498
5,649
4,981
21,058
1,848
619
13,167
12,463
9,388
37,485
Note: Cell counts (unweighted), age 15 and over, pregnant women excluded.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey (ATUS) and Eating and Health (EH) Module.
Table 2
ATUS/EH Module respondents with missing Body Mass Index (BMI)
by year
Item
Respondents missing BMI (number)
Share of total (%)
Total respondents
2006
2007
2008
Total
588
4.6
12,764
593
4.9
12,108
667
5.3
12,613
1,848
4.9
37,485
Note: Cell counts (unweighted), age 15 and over, pregnant women excluded.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey (ATUS) and Eating and Health (EH) Module.
Table 3
Missing height and weight respondents by type of nonresponse
Respondent type
Missing height
Missing weight
Don’t know
381
466
Refused
8ERS and the Bureau of Labor
Statistics concluded that including
height and weight in the EH Module
would not lead to unit nonresponse
because the ATUS is not a health
survey, and individuals would not
expect questions about height and
weight. For a discussion of nonresponse and sensitive questions, see
Tourangeau et al. (2010).
9See Vernon (2005) for a discussion
of pretesting the EH Module. A text
version of the questionnaire is at http://
stats.bls.gov/tus/ehmquestionnaire.pdf
10Dixon
(2002) finds that some
respondents in the Current Population
Survey respond to labor force questions—indicating that they agree with
the purpose of the survey—but not to
demographic questions that they may
perceive as irrelevant.
Total
Number
478
1,083
859
1,549
Note: Cell counts (unweighted), age 15 and over, pregnant women excluded.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey (ATUS) and Eating and Health (EH) Module.
whereas a “refusal” is a refusal to answer the question, a “don’t know”
response may be a soft refusal, so it is unclear how to interpret the “don’t
know” responses.
Demographic characteristics of BMI nonrespondents
In analyzing BMI nonresponse, it is useful to look at the basic demographic
characteristics of those with missing BMI. As shown in table 1, the majority
of nonrespondents were women (79 and 75 percent of unweighted and
weighted counts, respectively). The share of respondents missing BMIs varies
slightly by age group; however, the estimates are not statistically different
from each other, so there is essentially no difference in nonresponse across
age groups (table 4).
8
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
11Each sampled designated person is
assigned a diary day of the week, and
the interview is conducted on the next
day (e.g., if Monday is the diary day,
Tuesday is the interview day). If the
designated person is not reached or the
interview cannot be completed, call
attempts are made on the interview day
for up to 8 weeks.
Those who were employed at the time of the ATUS interview were more
likely to report height and weight, which is reflected by the lower share with
missing BMIs for that category of respondents (table 5).
Survey characteristics of BMI nonrespondents
Another way to analyze BMI nonresponse is by looking at ATUS respondent
characteristics. Perhaps the respondents with missing BMIs are reluctant to
be interviewed (i.e., they are “uncooperative respondents).” An indicator of an
uncooperative respondent is the number of phone calls made by the Census
Bureau to obtain a completed interview.11, 12 Based on the ATUS call history
data from one of the ATUS survey methodology files, the average number of
call attempts appears to be higher for those respondents with missing BMIs
(7.1) than for the total population (6.8); however, these averages are not statistically different at the 90-percent level, so the number of call attempts made
to those with missing BMIs is about the same as the number of calls made
to all others (table 6). A related characteristic of ATUS respondents is the
number of weeks (1-8) that were needed to successfully complete the interview. The number of weeks required for respondents with missing BMIs was
12The
number of attempted calls also
includes times that the respondent’s
case file was opened. So, an actual
call may not have been made if the file
was viewed or queued up for a call.
Thus, cases with high numbers of call
attempts likely have fewer actual calls
made. Information from Mary Dorinda
Allard, Director, BLS Division of
Labor Force Statistics, in a discussion
on March 11, 2011.
Table 4
Respondents with missing Body Mass Index (BMI) by age group
Missing BMI
Unweighted
Number
Total population, nonmissing BMI
UnStandard Weighted,
Weighted
weighted
error
90% CI
––————— Percent —————––
Unweighted
Number
UnStandard Weighted,
Weighted
weighted
error
90% CI
––————— Percent —————––
Total
population
1,848
4.9
4.9
0.14
±0.24
35,637
95.1
95.1
0.14
±0.24
Age 15-19
129
4.8
4.4
0.46
±0.75
2,588
95.2
95.6
0.46
±0.75
Age 20-39
600
5.2
5.1
0.27
±0.45
11,016
94.8
94.9
0.27
±0.45
Age 40-64
812
4.9
4.9
0.21
±0.34
15,736
96.1
95.1
0.21
±0.34
Age 65+
307
4.6
4.7
0.35
±0.57
6,297
95.4
95.3
0.35
±0.57
Note: Age 15 and over, pregnant women excluded.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use Survey and Eating and Health Module.
Table 5
Missing Body Mass Index (BMI) by employment status, population age 15 and over
Missing BMI
Unweighted
Number
Unweighted
Weighted
Total population, nonmissing BMI
Standard Weighted,
error
90% CI
––————— Percent —————––
Unweighted
Number
Unweighted
Weighted
Standard Weighted,
error
90% CI
––————— Percent —————––
Employed
1,112
4.6
4.6
0.18
±0.30
22,988
95.4
95.4
0.18
±0.30
Not
employed
736
5.5
5.4
0.26
±0.43
12,649
94.5
94.6
0.26
±0.30
Note: Age 15 and over, pregnant women excluded.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use Survey and Eating and Health Module.
9
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 6
Call attempts and weeks called by Body Mass Index (BMI) group
Item
Total
population
Missing
BMI
Underweight
Normal
weight
Overweight
Obese
Number of call attempts
Mean
(unweighted)
6.7
7.0
6.5
6.8
6.8
6.5
6.8
(0.06)
[±0.09]
7.1
(0.20)
[±0.33
6.8
(0.44)
[±0.72
6.9
(0.09)
[±0.15]
6.8
(0.10)
[±0.16]
6.6
(0.10)
[±0.17]
Median
(unweighted)
4
4
4
4
4
4
Median
(weighted)
4
5
4
4
4
4
75th percentile
(unweighted)
9
9
8
9
9
8
75th percentile
(weighted)
9
10
9
9
9
9
Minimum
1
1
1
1
1
1
Maximum
94
72
42
94
78
86
Mean
(weighted)
Number of weeks call attempts made
Mean
(unweighted)
2.2
2.4
2.2
2.2
2.2
2.2
2.2
(0.01)
[±0.02]
2.4
(0.05)
[±0.09]
2.3
(0.12)
[±0.19]
2.2
(0.02)
[±0.03]
2.3
(0.02)
[±0.04]
2.2
(0.02)
[±0.04]
Median
(unweighted)
1
2
1
1
1
1
Median
(weighted)
1
2
1
1
1
1
75th percentile
(unweighted)
3
3
3
3
3
3
75th percentile
(weighted)
3
3
3
3
3
3
1
1
1
1
1
1
8
8
8
8
8
8
Mean
(weighted)
Minimum
Maximum
Note: Age 15 and over, pregnant women excluded. Variables TUATTMPTNO and
TUATTMWEEK from ATUS call history file used. Standard errors are in parentheses,
90-percent confidence intervals in brackets.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey (ATUS) and Eating and Health Module
10
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
higher (2.4 weeks—weighted mean) and statistically different from that for
the total population (2.2 weeks). The median for those with missing BMIs
was higher as well—2 weeks versus 1 week for the other groups. The higher
average number of weeks needed to complete an interview may indicate that
these respondents were reluctant to be interviewed.
Characteristics of late respondents are likely to be similar to those of nonrespondents. Since the 75th percentile for the number of weeks that call
attempts were made is 3 weeks, late respondents are defined as those who
needed to be called for 4 or more weeks to participate in the ATUS. About 24
percent (weighted) of respondents with missing BMIs were late respondents,
a higher share than for late respondents not missing BMI values (20 percent)
(table 7). So, the missing BMI group has a higher share of those reluctant to
be interviewed than the nonmissing BMI group.
Another indicator of an uncooperative respondent is the number of activities
that the respondent reported in the time diary. The mean number of activities
reported across the total population was 19.7 (weighted), with a median of 19,
a minimum of 5, and a maximum of 9113 (table 8). Based on the number of
activities in the time diaries, the missing BMI group has fewer activities as
measured by mean, median, and maximum value than the total population.
The figure 1 box plot shows the different distributions of the number of diary
activities for the BMI groups. The missing BMI group has a lower box than
the other BMI groups, indicating a distribution with fewer diary activities
not only for the mean and median but also for the 25th and 75th percentiles.
Interestingly, men overall have fewer average reported diary activities than
women. This finding suggests that these respondents may have decided to
participate in the ATUS but did not report detailed information for their time
diaries and were perhaps reluctant to provide answers for the questionnaire
portions of the survey as well.
13Note that BLS ATUS excludes
interviews that have fewer than 5
reported activities and/or reported
activities that do not cover at least
21 hours of the diary day. There
is no constraint on the maximum
number of diary activities. Email
correspondence from Rachel KrantzKent, Manager, American Time Use
Survey, October 7, 2010.
Another indicator of respondent cooperation is the degree to which respondents answered other sensitive questions. The final questions in the survey
instrument ask for information on general health, height, weight, and income.
General health information is related in content to the height and weight questions. Of the 504 respondents who did not report general health, most (92.5
percent) did not report height and/or weight and had missing BMIs. This
Table 7
Missing Body Mass Index (BMI) by late respondent status—call attempts for 4+ weeks before interview
Missing BMI
Unweighted
Number
Unweighted
Weighted
Total population, nonmissing BMI
Standard Weighted,
UnUnerror
90% CI weighted weighted
––————— Percent —————––
Number
Weighted
Standard Weighted,
error
90% CI
––————— Percent —————––
Late
respondent
(4+ weeks)
405
21.9
24.1
1.26
±2.08
6,762
19.0
19.7
0.30
±0.49
Not late
respondent
1,443
78.1
75.9
1.26
±2.08
28,875
81.0
80.3
0.30
±0.49
Note: Age 15 and over, pregnant women excluded. CI = confidence interval.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use Survey and Eating and Health Module.
11
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 8
Number of diary activities by Body Mass Index (BMI) group
Item
Total
population
Missing
BMI
Underweight
Normal
weight
Overweight
Obese
Number of activities in diary
Mean
(unweighted)
19.9
18.6
19.8
20.7
19.7
19.3
19.7
(0.05)
[±0.08]
18.7
(0.22)
[±0.35]
19.2
(0.46)
[±0.76]
20.3
(0.08)
[±0.13]
19.5
(0.09)
[±0.15]
19.3
(0.10)
[±0.17]
Median
(unweighted)
19
17
19
19
19
18
Median
(weighted)
19
17
18
19
18
18
75th percentile
(unweighted)
24
23
24
25
24
23
75th percentile
(weighted)
24
23
23
24
24
24
Minimum
5
5
6
5
5
5
Maximum
91
58
91
75
81
85
Mean
(weighted)
Number of activities in diary, men only
Mean
(unweighted)
18.1
15.2
17.0
18.5
18.3
17.8
18.0
(0.06)
[±0.10]
15.4
(0.41)
[±0.67
17.2
(0.77)
[±1.26]
18.2
(0.13)
[±0.21]
18.2
(0.10)
[±0.17]
17.9
(0.13)
[±0.22]
Median
(unweighted)
17
14
16
18
17
17
Median
(weighted)
17
14
15
17
17
17
75th percentile
(unweighted)
22
19
20
22
22
22
75th percentile
(weighted)
22
19
20
22
22
22
5
5
6
5
5
5
64
40
44
64
57
54
Mean
(weighted)
Minimum
Maximum
—continued
share, however, accounts for only one-fourth of the missing BMI respondents.
Perhaps some individuals with missing BMIs are in poor health and may be
reluctant to disclose their height and weight. Some respondents may consider
this area of questioning to be too intrusive or irrelevant to the survey. The
three-fourths of respondents with missing BMIs who did report general health
were found to be less likely to report their health as “Excellent” or “Very
good,” more likely to report it as “Good “or “Fair,” and equally likely to report
it as “Poor” than those with a BMI value (table 9). So, the distribution of
12
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 8
Number of diary activities by Body Mass Index (BMI) group—Continued
Total
population
Item
Missing
BMI
Underweight
Normal
weight
Overweight
Obese
Number of activities in diary, women only
Mean
(unweighted)
21.3
19.5
20.7
21.9
21.4
20.7
21.3
(0.07)
[±0.12]
19.8
(0.24)
[±0.39]
20.1
(0.47)
[±0.78]
21.7
(0.10)
[±0.16]
21.4
(0.16)
[±0.26]
20.9
(0.15)
[±0.25]
Median
(unweighted)
20
18
20
20
20
19
Median
(weighted)
20
18
19
20
20
20
75th percentile
(unweighted)
26
24
25
26
26
25
75th percentile
(weighted)
26
24
24
26
26
25
Minimum
5
5
6
5
5
5
Maximum
91
58
91
75
81
85
Mean
(weighted)
Note: Age 15 and over, pregnant women excluded. Standard errors are in parentheses,
90-percent confidence intervals in brackets.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey and Eating and Health Module.
Figure 1
Number of diary activities by Body Mass Index group
Number of activities in diary
25.0
22.5
20.0
17.5
15.0
12.5
Missing BMI Underweight Normal weight Overweight Obese
Note: Extreme values omitted, 1.5 clip factor used. Age 15 and over, pregnant women
excluded.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey and Eating and Health Module.
13
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 9
Respondents with missing Body Mass Index (BMI) but reported
general health
Item
Missing BMI count
(number)
Total population,
nonmissing BMI
count (number)
Missing BMI
(row percent
unweighted)
Total population,
nonmissing BMI
(row percent
unweighted)
Missing BMI (row
percent weighted)
Total population,
nonmissing BMI
(row percent
weighted)
Reported health status
Excellent
Very good
Good
Fair
Poor
Total
138
409
527
247
61
1,382
6,806
12,258
10,566
4,385
1,584
35,599
10.0
29.6
38.1
17.9
4.4
100
19.1
34.4
29.7
12.3
4.5
100
10.9
(1.34)
[±2.20]
29.2
(1.67)
[±2.75]
38.8
(1.78)
[±2.93]
17.5
(1.37)
[±2.26]
3.6
(0.60)
[±0.99]
100
19.1
(0.28)
[±0.46]
34.7
(0.33)
[±0.55]
30.2
(0.37)
[±0.60]
12.0
(0.22)
[±0.36]
4.0
(0.13)
[±0.22]
100
Note: Age 15 and over, pregnant women excluded. A total of 466 respondents had missing
BMI and missing general health, and an additional 38 respondents had BMI but missing
general health. Standard errors in parentheses, 90-percent confidence intervals in brackets.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey and Eating and Health Module.
missing BMI respondents’ health status appears slightly different from that of
the rest of the respondents.
Questions about income can be problematic in household surveys, as some
respondents may be highly sensitive to requests for this information. The
EH Module benefited from several rounds of cognitive pre-testing in which
ERS was able to craft income questions that produced a high response
rate.14 The first question asked if household income was above or below 185
percent of the poverty threshold for the respondent’s household size. If the
respondent answered “below” or gave a “don’t know” or “refused” answer,
then a followup question asked whether the household’s income was above
or below 130 percent of the poverty threshold for the respondent’s household
size. Among the respondents who did not report household income for the
185-percent question,15 43 percent were missing BMIs (table 10). This share
is considerably larger than and statistically different from the 10 percent of
respondents who reported height and weight but did not report income.
A final indicator is diary quality. Time diary quality, as reported by the
Census interviewer and accessible through the ATUS case history data, can
be used to evaluate whether the missing BMI observations were from uncooperative respondents. After each ATUS interview is completed, the Census
interviewer answers two data quality questions: “Is there any reason the
information from this interview should NOT be used?” and “Why do you
think the data should not be used?”16
14The
EH Module income questions
had a nonresponse rate of 10 percent,
which is lower than the CPS income
nonresponse rate of 13 percent. Using
household earnings, ERS and BLS
imputed income for some respondents,
yielding only 6 percent with missing
income in the released EH Module
data. See Hamrick (2010) for more
information.
15Used here is the original missing
values for variable EEINCOME1;
that is, EEINCOME1 without the
values that were imputed for those
who did not report income. The flag
EXINCOME1 was used to remove
imputed values. EUINCOME2 (more/
less than 130 percent of poverty
threshold) was not used here.
16See
ATUS Questionnaire, June
2010, http://stats.bls.gov/tus/tuquestionnaire.pdf.
14
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 10
Respondents with missing Body Mass Index (BMI) by reporting
or not reporting income
Item
Income reported
Missing BMI count (number)
Income not
reported
1,103
745
32,718
2,919
Missing BMI (row percent unweighted)
59.7
40.3
Total population, nonmissing BMI (row
percent unweighted)
91.8
8.2
Missing BMI (row percent weighted)
57.2
(1.67), [±2.75]
42.8
(1.67), [±2.75]
Total population, nonmissing BMI (row
percent weighted)
89.6
(0.20), [±0.32]
10.4
(0.20), [±0.32]
Total population, nonmissing BMI count
(number)
Note: Age 15 and over, excludes pregnant women. Standard errors are in parentheses,
90-percent confidence intervals in brackets.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey and Eating and Health Module.
In only 275 of the 37,484 completed interviews did the Census interviewer
think that the respondent’s time diary was not of good quality (table 11). The
unweighted share of those with missing BMIs and no diary quality issues
(97.7 percent) is less than the share for those with a BMI value (99.4 percent),
as expected. The weighted share of those with missing BMIs and no diary
quality issue is 98.2 percent, statistically different from the 99.3 percent for
those with a BMI value and no diary quality issue. Although the number and
share of respondents with diary quality issues is very small, these results
contribute to understanding the factors accounting for the missing BMIs for
some respondents.
Multivariate analysis
A probit model was used to systematically analyze the influence that personal
characteristics and indicators of uncooperativeness have on the probability
that an observation will have missing BMIs (see table 12). In the model, the
dependent variable is missing BMI. The model allows for testing the hypothesis that respondents with missing BMIs are uncooperative respondents.
Number of diary activities, poor quality time diary, number of weeks that
calls were made (1-8), and missing income information were used as indicators of uncooperativeness.17 Demographic, economic, geographic, and household controls are included to see if respondents with various characteristics
are likely to have missing BMIs.
As expected and consistent with the descriptive analysis presented earlier,
the probability of missing BMI is higher for female respondents. The probability of missing BMI is also higher for those with less than a high school
education, noncitizens, employed persons, and those with household income
less than 185 percent of the poverty threshold. Among all respondents, probability of missing BMI was lower for teenagers, Asians, those of mixed race,
and those that have any health status except “Good.” Region and household
composition seem to have little association with missing BMI. Interestingly,
17Missing general health could not
be included as it is closely correlated
with missing income and so resulted
in a near singular matrix; likewise for
number of call attempts and weeks call
was made.
15
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 11
Missing Body Mass Index (BMI) respondents by diary quality
Item
Missing BMI count
(number)
No diary
quality issue
Diary quality problems
Intentionally
wrong
Could not
remember
Deliberately
long durations
Total
Other
reason
1,806
4
19
11
8
1,848
35,404
32
85
77
39
35,637
Missing BMI (row
percent unweighted)
97.7
0.2
1.0
0.6
0.5
100
Total population,
nonmissing BMI (row
percent unweighted)
99.4
0.1
0.2
0.2
0.1
100
Missing BMI (row
percent weighted)
98.2
(0.37), [±0.61]
0.2
(0.16), [±0.26]
0.9
(0.26), [±0.43]
0.4
(0.15), [±0.24]
0.3
(0.12), [±0.20]
100
Total population,
nonmissing BMI (row
percent weighted)
99.3
(0.06), [±0.10]
0.1
(0.02), [±0.03]
0.2
(0.04), [±0.06]
0.3
(0.04), [±0.07]
0.1
(0.03), [±0.05]
Total population,
nonmissing BMI count
(number)
100
Note: Age 15 and over, excludes pregnant women. Standard errors are in parentheses, 90-percent confidence intervals in brackets.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use Survey and Eating and Health Module.
in this multivariate analysis that controls for other factors, those who were
employed have a higher probability of missing BMI than other respondents,
whereas in the descriptive statistics, the same group had a lower share of
missing BMI. Employed respondents may have a higher probability of
missing BMI because those who are overweight but not obese have, on
average, longer paid work hours than other respondent types.18
18Hamrick
The general health variables were all negative and significant, indicating that
those reporting general health of ”Excellent,” “Very good,” “Fair,” or “Poor”
had a lower probability of missing BMI than those reporting general health of
“Good.” However, because missing general health cannot be included in the
model as it is closely related with missing income, these general health coefficients may also be indicating that these probabilities are less than for those
with missing general health.
Among measures of cooperativeness, all coefficients were significant, and
coefficient signs were in the expected directions. For all respondents, the
more activities reported in the time diary, the lower the probability of a
missing BMI. And having a time diary flagged as a poor quality diary
increases the probability of having a missing BMI, consistent with the possibility that missing BMI observations are from uncooperative respondents.
The more weeks that call attempts were made to interview a respondent, the
more likely the respondent’s BMI is missing. Likewise, not reporting household income increased the probability of missing BMI.
16
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
et al. (2011), app. table 6.
Table 12
Probit model, dependent variable = missing Body Mass Index (BMI)
Missing
BMI probit
coefficient
-2.0836
Item
Marginal
probability
Standard Significance
error
level
-0.1320
0.0824
***
0.6139
-0.7448
0.0389
-0.0472
0.0375
0.0763
***
***
0.1029
-0.0447
0.0065
-0.0028
0.0378
0.0606
***
0.2501
0.0158
0.0502
***
Some college
0.0115
College degree
0.0704
Advanced degree
0.1094
Noncitizen
0.4787
African American
-0.0362
Asian
-0.3326
Mixed race
-0.3200
Labor force and household income:
Employed
0.0935
0.0007
0.0045
0.0069
0.0303
-0.0023
-0.0211
-0.0203
0.0421
0.0534
0.0589
0.0548
0.0476
0.0950
0.1295
0.0059
0.0378
**
Intercept
Demographic variables:
Female
Age 15-19
Age 40-64
Age 65 and over
Less than high school
education
*
***
***
**
Income below 185% poverty
threshold
0.0887
0.0056
0.0465
*
Income missing (measure of
respondent cooperation)
1.1028
0.0698
0.0488
***
-0.0027
0.0010
0.0017
-0.0011
0.0572
0.0530
0.0572
0.0400
-0.0005
-0.0031
0.0378
0.0422
-0.0356
-0.0230
-0.0173
-0.0281
0.0691
0.0437
0.0535
0.0903
***
***
***
***
-0.0005
0.0197
0.0020
0.1274
***
**
0.0011
0.0084
**
Region:
Midwest
-0.0424
South
0.0164
West
0.0265
Nonmetro
-0.0174
Household composition:
One-adult household
-0.0078
No children in household
-0.0493
General health:
Excellent health
-0.5621
Very good health
-0.3638
Fair health
-0.2725
Poor health
-0.4436
Measures of respondent cooperation:
Number of diary activities
-0.0078
Poor quality time diary
0.3116
Week call made (1-8)
0.0170
Likelihood ratio test, probability
> ChiSq
Number of observations
<.0001
37,485, 4.9% observations have missing BMI.
Note: Population (weighted), age 15 and over, pregnant women excluded.
*** indicates significance at the 1-percent level, ** indicates significance at the 5-percent
level, and * indicates significance at the 10-percent level. Excluded group is male, age 20-39,
high school diploma, citizen, White, not employed, income above 185 percent of the poverty
threshold, Northeast, Metro, two-adult household with children, and good health.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey and Eating and Health Module.
17
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Missing BMIs—Time-Use Profiles
The ATUS and the EH Module facilitate analysis of Americans’ time spent
in various activities. Having an understanding of the characteristics of survey
respondents with missing BMIs is useful to time-use research; however, it may
be more important to understand whether missing BMIs cause bias in timeuse estimates. Are the respondents who did not report height and/or weight
different with respect to their time-use patterns than other respondents? On
average, those who did not report height and/or weight reported fewer activities
in their diaries, but what were the time-duration differences of these activities?
For time spent per day in major activities19 (table 13), the most striking difference between the missing BMI group and the other BMI groups is the long
amount of time spent by the missing BMI group in Personal Care (which includes
Sleeping)—586 minutes (9.8 hours). The missing BMI group, of which the
majority are women, also had a higher average time spent per day in household activities and caring for household members than the other BMI groups.
Interestingly, the missing BMI group also had the highest average time spent in
Other Activities. Other Activities include gaps in the time diary that the ATUS
interviewer was not able to code due to insufficient detail, respondent refusal,
or the inability of the respondent to recall activity. The higher average Other
Activities time is consistent with the concept of uncooperative respondents.
19For definitions of the major activity
groups, see ATUS User’s Guide
Appendix H: Bridge between published
tables activity categories and ATUS
coding lexicon activity categories, http://
stats.bls.gov/tus/atususersguide.pdf
For major time-use activities by gender, the male missing BMI group spent
about the same amount of time, on average, on Personal Care than underweight males (table 14a). The average time males with missing BMIs spent
in Eating and Drinking was the shortest among all BMI groups. The female
missing BMI group had a relatively high average time spent in Personal Care,
the lowest average time spent in Eating and Drinking, and the most time
spent in the Other category (table 14b).
The missing BMI group, both men and women, had a long-duration average
time spent in Personal Care. The two main activities in Personal Care are
Sleeping and Grooming; the missing BMI groups and the total population
group differed more in time spent Sleeping than in time spent Grooming (table
15). The nonresponse bias as a result would be that minutes spent in Sleeping
are underreported in analysis of BMI groups excluding the missing BMI
respondents. However, it is unclear whether sleep duration is underreported for
any specific BMI group, as the underreporting occurs across all BMI groups.
Based on the weighted absolute-deviation index (WADI),20 a measure of dissimilarity of time use “activity profiles” across groups, the difference between the
missing BMI group and the other BMI groups can be measured systematically
over the 17 major time-use activities. Stewart (2006) defines WADI as:
WADI = ∑ ik=1
ai − bi ai + bi
a −b
= ∑ ik=1 i i
2880
ai + bi 2880
20Note that weighting with respect
to the WADI means applying the share
of total time spent on an activity to
each difference, whereas weighted
elsewhere in this report indicates that
sample weights were used to produce
national estimates.
where
i = activity
k = number of activities
ai = time in minutes spent in activity i by group a
bi = time in minutes spent in activity i by group b
18
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 13
Time spent (minutes) in major activities by Body Mass Index (BMI) group
Total
population
Missing
BMI
562.1
(0.86)
[±1.41]
111.6
(0.89)
[±1.46]
Caring for household members
Underweight
Normal
weight
Overweight
Obese
586.2
(4.86)
[±7.99]
120.9
(4.29)
[±7.05]
591.5
(7.33)
[±12.06]
100.1
(6.10)
[±10.04]
567.2
(1.66)
[±2.73]
112.8
(1.38)
[±2.28]
554.1
(1.48)
[±2.44]
111.8
(1.46)
[±2.41]
558.8
(1.99)
[±3.27]
108.8
(1.97)
[±3.24]
31.2
(0.42)
[±0.69]
41.8
(2.68)
[±4.41]
27.5
(4.14)
[±6.81]
33.0
(0.69)
[±1.14]
28.2
(0.72)
[±1.19]
30.8
(1.01)
[±1.67]
Caring for
nonhousehold
members
12.8
(0.35)
[±0.57]
12.7
(1.55)
[±2.56]
13.5
(3.27)
[±5.38]
11.6
(0.54)
[±0.88]
13.6
(0.70)
[±1.15]
13.4
(0.70)
[±1.15]
Paid work
226.1
(1.77)
[±2.91]
198.8
(7.97)
[±13.12]
153.7
(12.12)
[±19.93]
209.3
(3.18)
[±5.24]
245.7
(3.09)
[±5.08]
233.8
(3.56)
[±5.86]
Educational
27.8
(0.72)
[±1.18]
26.0
(3.68)
[±6.05]
83.6
(9.17)
[±15.08]
44.2
(1.66)
[±2.73]
17.0
(1.24)
[±2.04]
15.7
(1.22)
[±2.01]
Purchasing goods
38.0
(0.46)
[±0.76]
40.9
(2.05)
[±3.38]
43.5
(4.48)
[±7.37]
38.1
(0.77)
[±1.26]
37.5
(0.89)
[±1.46]
37.5
(0.97)
[±1.60]
Purchasing
services
7.3
(0.23)
[±0.38]
8.4
(0.95)
[±1.57]
10.7
(2.58)
[±4.24]
7.0
(0.38)
[±0.62]
7.4
(0.40)
[±0.66]
7.3
(0.40)
[±0.66]
1.3
(0.08)
[±0.14]
0.7
(0.07)
[±0.12]
0.9
(0.25)
[±0.41]
0.5
(0.29)
[±0.47]
0.3
(0.22)
[±0.37]
1.5
(0.68)
[±1.12]
1.3
(0.15)
[±0.25]
0.5
(0.11)
[±0.18]
1.3
(0.15)
[±0.25]
0.6
(0.11)
[±0.18]
1.5
(0.20)
[±0.33]
0.9
(0.20)
[±0.33]
Eating and drinking
73.9
(0.40)
[±0.66]
65.5
(1.63)
[±2.67]
71.8
(3.03)
[±4.98]
75.6
(0.75)
[±1.23]
75.5
(0.74)
[±1.21]
71.4
(0.72)
[±1.19]
Leisure
285.7
(1.38)
[±2.28]
284.1
(5.86)
[±9.64]
284.2
(11.57)
[±19.04]
270.3
(2.33)
[±3.83]
286.2
(2.49)
[±4.09]
307.3
(2.78)
[±4.57]
Sports
22.6
(0.46)
[±0.76]
13.2
(1.39)
[±2.28]
19.9
(3.34)
[±5.50]
27.4
(0.86)
[±1.42]
23.6
(0.92)
[±1.51]
16.6
(0.84)
[±1.38]
Religious
9.5
(0.22)
[±0.36]
10.5
(1.04)
[±1.71]
7.8
(1.75)
[±2.89]
9.2
(0.39)
[±0.63]
9.4
(0.40)
[±0.65]
10.2
(0.43)
[±0.71]
Volunteer
10.0
(0.35)
[±0.57]
7.2
(0.96)
[±1.58]
6.7
(1.68)
[±2.76]
10.7
(0.58)
[±0.96]
10.4
(0.60)
[±0.99]
9.2
(0.64)
[±1.05]
Phone and mail
7.1
(0.19)
[±0.32]
6.4
(0.84)
[±1.38]
10.0
(1.33)
[±2.20]
8.3
(0.35)
[±0.58]
6.3
(0.28)
[±0.45]
6.5
(0.37)
[±0.61]
Item
Personal care
Household
activities
Purchasing household services
Government
and civic
—continued
19
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 13
Time spent (minutes) in major activities by Body Mass Index (BMI)
group—Continued
Item
Total
population
Missing
BMI
Underweight
Normal
weight
Overweight
Obese
Other
12.2
(0.34)
[±0.57]
15.8
(1.51)
[±2.48]
13.5
(3.26)
[±5.36]
13.7
(0.66)
[±1.08]
11.5
(0.63)
[±1.04]
10.3
(0.45)
[±0.74]
Total
1,440.0
1,440.0
1,440.0
1,440.0
1,440.0
1,440.0
Note: Population (weighted), age 15 and over, pregnant women excluded. Activities listed
are ATUS major activity groups. Travel time included with each activity. Standard errors in
parentheses, 90-percent confidence intervals in brackets.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey (ATUS) and Eating and Health Module.
Table 14a
Time spent (minutes) in major activity groups, men only
Activity
Total
population
Missing
BMI
Underweight
Normal
weight
Overweight
Obese
Personal care
551.5
(1.35)
[±2.22]
592.4
(10.99)
[±18.08]
592.8
(11.28)
[±18.56]
563.9
(2.92)
[±4.81]
545.3
(2.12)
[±3.48]
541.8
(2.62)
[±4.32]
Household
activities
85.0
(1.17)
[±1.93]
60.2
(5.85)
[±9.63]
61.7
(10.44)
[±17.17]
81.6
(2.21)
[±3.63]
91.0
(1.91)
[±3.14]
82.9
(2.32)
[±3.82]
Caring for household members
20.3
(0.50)
[±0.82]
18.7
(3.41)
[±5.60]
12.6
(4.09)
[±6.72]
17.2
(0.78)
[±1.29]
20.6
(0.63)
[±1.04]
23.5
(1.31)
[±2.16]
Caring for
nonhousehold
members
10.7
(0.47)
[±0.77]
6.6
(1.76)
[±2.90]
10.6
(4.42)
[±7.28]
9.9
(0.71)
[±1.17]
10.9
(0.80)
[±1.31]
11.5
(0.98)
[±1.62]
Paid work
271.1
(2.51)
[±4.13]
254.0
(19.92)
[±32.77]
159.1
(25.46)
[±41.87]
244.0
(5.26)
[±8.65]
285.4
(4.75)
[±7.81]
284.8
(5.31)
[±8.74]
Educational
25.5
(1.02)
[±1.67]
37.3
(9.16)
[±15.07]
105.3
(19.40)
[±31.91]
45.8
(2.45)
[±4.02]
14.6
(0.29)
[±2.12]
16.1
(1.69)
[±2.78]
Purchasing goods
30.3
(0.68)
[±1.12]
26.1
(4.10)
[±6.74]
29.7
(7.78)
[±12.80]
27.00
(1.10)
[±1.81]
31.9
(1.13)
[±1.85]
32.0
(1.22)
[±2.01]
Purchasing
services
5.2
(0.27)
[±0.45]
6.2
(1.81)
[±2.98]
5.0
(0.56)
[±0.93]
5.3
(0.44)
[±0.72]
4.9
(0.46)
[±0.76]
Purchasing household services
1.6
(0.14)
[±0.24]
0.4
(0.20)
[±0.33]
8.8
(3.79)
[±6.23]
0
(0)
[0]
1.4
(0.31)
[±0.51]
1.7
(0.24)
[±0.39]
1.8
(0.34)
[±0.56]
Government
and civic
0.7
(0.12)
[±0.20]
1.3
(1.06)
[±1.75]
0.6
(0.44)
[±0.72]
0.5
(0.14),
[±0.24]
0.6
(0.16)
[±0.26]
0.9
(0.32)
[±0.53]
—continued
20
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 14a
Time spent (minutes) in major activity groups, men only—Continued
Total
population
Missing
BMI
Underweight
Normal
weight
Overweight
Obese
Eating and drinking
76.1
(0.58)
[±0.96]
59.8
(2.39)
[±3.94]
67.5
(5.41)
[±8.90]
77.7
(1.20)
[±1.98]
77.3
(0.98)
[±1.61]
74.2
(1.06)
[±1.74]
Leisure
299.8
(2.20)
[±3.62]
331.4
(15.09)
[±24.82]
343.7
(23.90)
[±39.31]
295.5
(3.98)
[±6.54]
291.3
(3.60)
[±5.93]
312.9
(4.06)
[±6.68]
Sports
29.8
(0.78)
[±1.29]
15.1
(3.08)
[±5.06]
19.1
(5.50)
[±9.04]
35.8
(1.49)
[±2.45]
30.4
(1.44)
[±2.36]
24.2
(1.47)
[±2.42]
Religious
7.9
(0.31)
[±0.50]
11.5
(2.64)
[±4.35]
6.2
(2.32)
[±3.82]
7.7
(0.56)
[±0.93]
7.8
(0.44)
[±0.73]
7.8
(0.53)
[±0.87]
Volunteer
9.2
(0.49)
[±0.81]
3.5
(1.17)
[±1.93]
5.7
(3.14)
[±5.17]
8.7
(0.76)
[±1.25]
10.3
(0.85)
[±1.40]
9.0
(0.99)
[±1.62]
Phone and mail
3.9
(0.20)
[±0.32]
3.7
(1.18)
[±1.94]
6.4
(2.42),
[±3.98]
5.0
(0.43),
[±0.70]
3.8
(0.32)
[±0.52]
2.8
(0.27)
[±0.45]
Other
11.4
(0.51)
[±0.84]
12.0
(2.69)
[±4.42]
10.2
(4.54)
[±7.47]
13.2
(1.11)
[±1.83]
11.8
(0.97)
[±1.59]
8.8
(0.59)
[±0.97]
Activity
Total
1,440.00 1,440.00 1,440.00 1,440.00 1,440.00 1,440.00
Note: BMI = Body Mass Index. Population (weighted), age 15 and over. Activities listed
are ATUS major activity groups. Travel time included with each activity. Standard errors in
parentheses, 90-percent confidence intervals in brackets.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey (ATUS) and Eating and Health Module.
Table 14b
Time spent (minutes) in major activity groups, women only
Total
population
Missing
BMI
Personal care
572.2
(1.25)
[±2.05]
Household
activities
Underweight
Normal
weight
Overweight
Obese
584.1
(5.11)
[±8.41]
591.0
(8.93)
[±14.69]
569.4
(1.81)
[±2.97]
566.8
(2.48)
[±4.08]
578.0
(2.95)
[±4.86]
137.1
(1.36)
[±2.24]
141.4
(5.31)
[±8.74]
115.2
(7.71)
[±12.69]
133.9
(1.93)
[±3.17]
142.2
(2.63)
[±4.33]
138.1
(2.72)
[±4.48]
Caring for household members
41.6
(0.63)
[±1.04]
49.6
(3.39)
[±5.57]
33.5
(5.26)
[±8.65]
43.7
(1.01)
[±1.66]
39.2
(1.38)
[±2.27]
39.1
(1.54)
[±2.53]
Caring for
nonhousehold
members
14.8
(0.54)
[±0.88]
14.7
(2.02)
[±3.32]
14.7
(4.10)
[±6.74]
12.8
(0.80)
[±1.32]
17.5
(1.17)
[±1.93]
15.5
(1.00)
[±1.64]
Activity
—continued
21
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 14b
Time spent (minutes) in major activity groups, women only—Continued
Activity
Total
population
Missing
BMI
Underweight
Normal
weight
Overweight
Obese
Paid work
182.9
(2.28)
[±3.75]
180.3
(8.78)
[±14.44]
151.6
(16.02)
[±26.35]
185.8
(3.71)
[±6.11]
187.6
(4.51)
[±7.41]
176.3
(4.69)
[±7.72]
Educational
30.0
(1.03)
[±1.69]
22.2
(4.15)
[±6.82]
75.0
(10.65)
[±17.51]
43.0
(2.22)
[±3.65]
20.6
(2.30)
[±3.78]
15.2
(1.83)
[±3.02]
Purchasing goods
45.3
(0.66)
[±1.09]
45.8
(2.49)
[±4.09]
48.9
(5.50)
[±9.05]
45.6
(1.08)
[±1.78]
45.7
(1.29)
[±2.13]
43.8
(1.48)
[±2.44]
Purchasing
services
9.4
(0.37)
[±0.61]
9.2
(1.10)
[±1.81]
11.5
(3.30)
[±5.43]
8.4
(0.51)
[±0.84]
10.5
(0.77)
[±1.27]
10.0
(0.73)
[±1.20]
Purchasing household services
1.1
(0.08)
[±0.13]
1.1
(0.32)
[±0.53]
0.4
(0.31)
[±0.51]
1.3
(0.13)
[±0.22]
0.7
(0.10)
[±0.17]
1.2
(0.20)
[±0.34]
Government
and civic
0.6
(0.09)
[±0.15]
0.3
(0.14)
[±0.23]
1.8
(0.92)
[±1.51]
0.5
(0.14)
[±0.23]
0.5
(0.15)
[±0.25]
1.0
(0.25)
[±0.42]
Eating and drinking
71.9
(0.54)
[±0.89]
67.4
(2.01)
[±3.30]
73.5
(3.52)
[±5.80]
74.1
(0.90)
[±1.48]
72.8
(1.12)
[±1.84]
68.1
(1.01)
[±1.66]
Leisure
272.1
(1.81)
[±2.98]
268.2
(6.24)
[±10.27]
260.8
(13.11)
[±21.57]
253.2
(2.84)
[±4.67]
278.7
(3.40)
[±5.60]
300.9
(3.97)
[±6.52]
Sports
15.8
(0.50)
[±0.82]
12.6
(1.57)
[±2.58]
20.2
(4.06)
[±6.68]
21.7
(0.92)
[±1.52]
13.8
(0.95)
[±1.56]
8.0
(0.57)
[±0.93]
Religious
11.1
(0.33)
[±0.55]
10.2
(1.08)
[±1.78]
8.5
(2.11)
[±3.46]
10.2
(0.51)
[±0.84]
11.7
(0.68)
[±1.13]
12.9
(0.68)
[±1.11]
Volunteer
10.7
(0.47)
[±0.77]
8.5
(1.18)
[±1.94]
7.1
(1.87)
[±3.08]
12.1
(0.82)
[±1.34]
10.5
(0.84)
[±1.38]
9.5
(0.85)
[±1.40]
Phone and mail
10.2
(0.31)
[±0.52]
7.3
(1.06)
[±1.74]
11.5
(1.65)
[±2.71]
10.5
(0.50)
[±0.83]
10.1
(0.51)
[±0.84]
10.7
(0.69)
[±1.14]
Other
13.0
(0.43)
[±0.70]
17.2
(1.84)
[±3.03]
14.8
(3.58)
[±5.89]
14.0
(0.83)
[±1.36]
11.0
(0.66)
[±1.08]
11.9
(0.73)
[±1.21]
Total
1,440.00 1,440.00 1,440.00 1,440.00 1,440.00 1,440.00
Note: BMI = Body Mass Index. Population (weighted), age 15 and over, pregnant women
excluded. Activities listed are ATUS major activity groups. Travel time included with each
activity. Standard errors in parentheses.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey (ATUS) and Eating and Health Module.
22
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 15
Time spent (minutes) in Personal Care activities by Body Mass Index
(BMI) group
Activity
Total
population
Missing
BMI
Underweight
Normal
weight
Overweight
Obese
Men
Sleeping
509.7
(1.25)
[±2.05]
553.9
(11.29)
[±18.58]
557.3
(12.08)
[±19.88]
522.9
(2.75)
[±4.53]
502.1
(1.99)
[±3.27]
500.7
(2.55)
[±4.20]
Grooming
32.9
(0.31)
[±0.50]
31.3
(2.13)
[±3.51]
28.5
(3.70)
[±6.09]
33.1
(0.50)
[±0.83]
33.2
(0.49)
[±0.81]
32.7
(0.54)
[±0.88]
Sleeping
513.4
(1.22)
[±2.00]
528.1
(4.58)
[±7.53]
531.8
(9.64)
[±15.86]
510.9
(1.67)
[±2.76]
507.2
(2.39)
[±3.93]
518.5
(2.83)
[±4.66]
Grooming
47.0
(0.32)
[±0.53]
43.3
(1.06)
[±1.75]
51.3
(2.95)
[±4.85]
48.7
(0.53)
[±0.88]
47.1
(0.64)
[±1.06]
44.5
(0.72)
[±1.18]
Women
Note: Population (weighted), age 15 and over. Sleeping is ATUS activity code 010101, and
Grooming is 010201. Travel time included. Standard errors in parentheses, 90-percent confidence intervals in brackets.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey (ATUS) and Eating and Health Module.
The absolute difference in time spent is divided by the total time spent in
activity i by the two groups, then that is weighted by the fraction of activity i’s
time by total available time to groups a and b (1,440 + 1,440 = 2,880, 1,440 is
the total minutes in a day). These terms are then summed over all activities.
The smaller the WADI, the smaller the difference between the two groups,
and a WADI=0 would indicate no difference in activity profiles. Stewart
(2006) recommends using a WADI index over other dissimilarity indices as
it is not sensitive to the level of aggregation of activities (e.g., ATUS major
groups of two-digit activity codes versus ATUS four-digit activity code
groups) and short-duration activities receive little weight. The index value is
“equal to the average proportional difference in the time spent in all activities” (Stewart, 2006, p. 59). Looking specifically at Personal Care (ATUS
codes 01xxxx), an index of the absolute deviation (ADI), not the WADI, can
be used as it is only one activity so weighting is not needed.
Overall, the WADIs are small, indicating little difference between the
activity profiles of the missing BMI group and those of other BMI groups21
(table 16). Among the BMI groups, the missing BMI group had the lowest
WADI with the normal-weight group for the total population and for men.
This means the time use profile of those with missing BMIs most resembled
that of normal-weight individuals. For women, the lowest WADI was with
the overweight group, which indicates that the activity profile of women
with missing BMIs is closest to that of overweight women. Looking only at
the absolute deviation index for Personal Care activities, those with missing
BMI had Personal Care durations most like the underweight group for the
total population and for men and the obese group for women.
21The
mean BMI for those who
reported height and weight age 15 and
older is 27.13; for men, it is 27.60, and
for women, it is 26.66.
23
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Table 16
Measures of dissimilarity by Body Mass Index (BMI) group
Total
population
Missing
BMI
Underweight
Normal
weight
Overweight
Obese
WADI
0.0356
na
0.0600
0.0407
0.0513
0.0494
ADI, personal
care only
0.0205
na
0.0045
0.0165
0.0282
0.0239
WADI
0.0626
na
0.0758
0.0566
0.0801
0.0692
ADI, personal
care only
0.0358
na
0.0003
0.0246
0.0413
0.0446
WADI
0.0190
na
0.0586
0.0343
0.0250
0.0298
ADI, personal
care only
0.0095
na
0.0059
0.0128
0.0150
0.0053
Item
Total population
Men
Women
Note: na = not applicable WADI = weighted absolute deviation index. Population (weighted),
age 15 and over. Shading indicates smallest index among the BMI groups.
Source: USDA, Economic Research Service using data from 2006-08 American Time Use
Survey and Eating and Health Module.
Perhaps the long sleep times are due to unobserved characteristics more
specific than general health. For example, the missing BMI respondents
could be more likely to be suffering from depression or other illnesses at the
time of the ATUS interview, which could result in longer sleep times, poorer
health, and fewer activities in the time diary. It does appear that the missing
BMI individuals may be in slightly worse health than others and, as a consequence, may sleep more.
However, the longer average times engaged in sleeping may also indicate that
those in the missing BMI group are uncooperative respondents; that is, they
cannot remember; they do not want to make the effort to remember; or they
do not want to disclose their activities in detail. Reporting longer sleep times
allows respondents to cover a large block of time with one activity. This is
consistent with the earlier analysis on indicators of uncooperativeness. As a
consequence, any nonresponse bias of underreported sleep time may be one
of sleep time as reported and not actual sleep time, since the missing BMI
group has a disproportionate share of uncooperative respondents.
24
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Is a Remedy for Missing BMI Data Needed?
Several factors make it unnecessary to devise a remedy for missing BMI
data in the EH Module: less than 5 percent of respondents are missing BMIs;
respondents with missing BMIs have indicators of being uncooperative
respondents with poor data quality; and the index of dissimilarity shows that
the missing BMI group has time use patterns similar to normal-weight men
and overweight women. The BMI data in the EH Module are suitable for
research with the ATUS time diary data. Researchers can use case deletion
in their analysis to exclude respondents with missing BMIs and calculate estimates using the responses with BMI values.
Researchers who want to increase the number of observations, or who want
to fully account for any possible bias, may undertake one of two approaches.
First, sample weights could be recalculated for BMI nonresponse (see
Abraham et al. (2006) for analysis of unit nonresponse in the ATUS). The
ATUS final sample weights control for a variety of factors, including unit
nonresponse and interview day of week. A large number of technical adjustments are included as well.22 A researcher considering this remedy may want
to consider the extensive computational requirement of recalculating weights
with the expected reduction of bias.
22See ATUS User’s Guide section
7.2, and BLS and Census (2006),
Current Population Survey: Design
and Methodology, chapter 10.
The other approach would be to impute BMI, or impute height and weight,
for the missing BMI values. One difficulty in imputing BMI from the ATUS
and EH Module data is that there are no anthropomorphic or medical information available other than the self-reported general health. The studies cited
above that imputed BMI used NHANES data and had medical history and
waist circumference information for the respondents with missing BMIs.
Having anthropomorphic and/or medical information would make for more
informed imputations of BMI. Without this information, BMI imputations
would essentially be cell averages of the demographic and labor force participation group of missing BMI respondents. Another possibility is to use the
NHANES data and probabilistic matching to match the missing BMI respondents with NHANES respondents and then apply the NHANES BMI value to
the ATUS/EH Module respondents.
25
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Conclusions
The 4.9 percent of EH Module respondents who did not report height and/or
weight had disproportionately higher indicators of being reluctant or uncooperative respondents. It took more call attempts over more weeks to interview
these respondents, indicating that they were reluctant to participate in the
survey. They were also more likely to have time diaries with little detail as
measured by the number of activities in the diary, and they were more likely to
have poor quality time diaries. They were less likely to answer other sensitive
questions on the survey. These findings indicate that the respondents’ lack of
reporting height and/or weight had less to do with the height and weight questions and more to do with the respondents’ views of participating in the survey.
The time-use profiles of the total population and of men with missing BMI
closely resembled those of respondents with normal weight. For women, the
missing BMI time profile closely resembled that of women who were overweight. This suggests that those who did not report height and weight are
unlikely to be at either end of the BMI spectrum; that is, they are unlikely to be
severely underweight or morbidly obese, mitigating any bias. Since the missing
BMI respondents have time profiles close to those in the middle BMI groups
(normal weight and overweight), excluding their time diaries in the analysis
is unlikely to produce bias in time-use estimates. As a consequence of these
findings, any item nonresponse bias in the EH Module height and weight data
appears to be small, allowing for future analysis of time use by BMI.
26
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
References
Abraham, Katharine G., Sara E. Helms, and Stanley Presser. 2008.
“How Social Processes Distort Measurement: The Impact of Survey
Nonresponse on Estimates of Volunteer Work in the United States.”
NBER Working Paper Series, Working Paper 14076, National Bureau of
Economic Research.
Abraham, Katharine G., Aaron Maitland, and Suzanne M. Bianchi. 2006.
“Nonresponse in the American Time Use Survey: Who Is Missing From
the Data and How Much Does It Matter?” Public Opinion Quarterly
70(5): 676-703.
Andridge, R.R., and R.J.A. Little. 2010. “A Review of Hot Deck Imputation
for Survey Non-response.” International Statistical Review 78(1): 40-64.
Bethlehem, Jelke, Fannie Cobben, and Barry Schouten. 2011. Handbook of
Nonresponse in Household Surveys. Hoboken, NJ: John Wiley & Sons, Inc.
Bose, Jonaki. 2001. “Nonresponse Bias Analyses at the National
Center for Education Statistics,” Proceedings of Statistics Canada
Symposium, Statistics Canada, www.statcan.gc.ca/bsolc/olc-cel/
olc-cel?catno=11-522-X20010016269&lang=eng.
Cawley, John, and Richard V. Burkhauser. 2006. “Beyond BMI: The Value
of More Accurate Measures of Fatness and Obesity in Social Science
Research,” National Bureau of Economic Research Working Paper 12291.
Danubio, M.E., G. Miranda, M.G. Vinciguerra, E. Vecchi, and F. Rufo.
2008. “Comparison of Self-Reported and Measured Height and Weight:
Implications for Obesity Research Among Young Adults.” Economics &
Human Biology 6(1): 181-190.
Dixon, John. 2002. “The Effects of Item and Unit Nonresponse on Estimates
of Labor Force Participation,” Office of Survey Methods, Bureau of Labor
Statistics, www.bls.gov/osmr/abstract/st/st020240.htm.
________. 2005. “Comparison of Item Nonresponse and Unit Nonresponse
in Household Surveys,” Office of Survey Methods, Bureau of Labor
Statistics, www.stats.bls.gov/osmr/abstract/st/st050080.htm.
________. 2008. “Sensitivity Analysis of Nonresponse Bias in the Current
Population Survey,” Office of Survey Methods, Bureau of Labor Statistics,
www.bls.gov/osmr/abstract/st/st080160.htm.
Frazis, H., and J. Stewart. 2010. “How to Think About Time-Use Data:
What Inferences Can We Make About Long- and Short-Run Time
Use from Time Diaries?” IZA Discussion Paper Series DP No. 5306,
Forschungsinstitut zur Zukunft der Arbeit (Institute for the Study of
Labor), http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1708752.
27
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Fricker, Scott, and Roger Tourangeau. 2010. “Examining the Relationship
Between Nonresponse Propensity and Data Quality in Two National
Household Surveys,” Public Opinion Quarterly, 74(5), 934-955.
Friedman, E.M., N.A. Clusen, and M. Hartzell. 2003. Better Late?
Characteristics of Late Respondents to a Health Care Survey. Survey
Research Methods Section, American Statistical Association meetings,
August 3-7, San Francisco, CA. Available at: www.amstat.org/Sections/
Srms/Proceedings/y2003f.html
Graham, J.W. 2009. “Missing Data Analysis: Making It Work in the Real
World.” Annual Review of Psychology 60(1): 549-576.
Groves, Robert M. 2006. “Nonresponse Rates and Nonresponse Bias in
Household Surveys.” Public Opinion Quarterly 70(5): 646-675.
Groves, Robert M., Don A. Dillman, John L. Eltinge, and Roderick J.A. Little
(eds). 2002. Survey Nonresponse. Wiley Series in Survey Methodology.
New York: John Wiley & Sons, Inc.
Hamrick, Karen. 2010. Eating and Health Module User’s Guide,
Administrative Publication No. 047, USDA, Economic Research Service.
Available at: www.ers.usda.gov/publications/ap-administrative-publication/ap-047.aspx
Hamrick, K.S., M. Andrews, J. Guthrie, D. Hopkins, and K. McClelland.
2011. How Much Time Do Americans Spend on Food? U.S. Department
of Agriculture, Economic Research Service, Economic Information
Bulletin No. 86, November. Available at: www.ers.usda.gov/publications/
eib-economic-information-bulletin/eib86.aspx
Hennessy, C.H., D.G. Moriarty, M.M. Zack, P.A. Scherr, and R. Brackbill.
1994. “Measuring health-related quality of life for public health surveillance,” Public Health Reports, 109(5): 665-672. Available at: www.ncbi.
nlm.nih.gov/pmc/articles/PMC1403555/.
Hill, A., and J. Roberts. 1998. “Body Mass Index: A Comparison Between
Self-Reported and Measured Height and Weight.” Journal of Public
Health Medicine 20(2): 206-210.
Kuczmarski, M.F., R.J. Kuczmarski, and M. Najjar. 2001 “Effects of Age on
Validity of Self-Reported Height, Weight, and Body Mass Index: Findings
from the Third National Health and Nutrition Examination Survey, 19881994.” Journal of The American Dietetic Association, 101(1): pp. 28-34.
Kyureghian, G., O. Capps, Jr., and R.M. Nayga, Jr. 2011. “General Remedies
to Local Problems: An Applied Researcher’s Manual to Multiple
Imputation.” 2011 Annual Meeting, Agricultural and Applied Economics
Association, Pittsburgh, PA, July 24-26. Available at http://purl.umn.
edu/108266
Mason, Robert, Virginia Lesser, and Michael W. Traugott. 2002. “Effect
of Item Nonresponse on Nonresponse Error and Inference,” Survey
28
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
Nonresponse. Robert M. Groves, John L. Eltinge, and Roderick J.A. Little
(eds.). New York: John Wiley & Sons, Inc., chapter 10: 149-161.
Mertz, W., J. C. Tsui et al. 1991. “What Are People Really Eating? The
Relation Between Energy Intake Derived From Estimated Diet Records
and Intake Determined to Maintain Body Weight.” American Society for
Clinical Nutrition, 54: 291-295.
Nandram, B., and J.W. Choi. 2010. “A Bayesian Analysis of Body Mass
Index Data From Small Domains Under Nonignorable Nonresponse and
Selection.” Journal of the American Statistical Association 105(489):
120-135.
O’Neill, G., and J. Sincavage. 2004. “Response Analysis Survey: A
Qualitative Look at Response and Nonresponse in the American Time Use
Survey, Bureau of Labor Statistics, Office of Survey Methods Research.”
Bureau of Labor Statistics Office of Survey Methods Research Statistical
Survey Paper, available at: www.bls.gov/osmr/abstract/st/st040140.htm
Pinkston, Josh, and Jay Stewart. 2009. “How Does Time Use Affect the
Probability of Becoming Obese?” Paper presented at the American
Time Use Research Conference, College Park, MD, June 22-24.
Available at: www.popcenter.umd.edu/research/sponsored-events/
atus-conf-workshop-2009.
Schafer, J.L., and J.W. Graham. 2002. “Missing data: Our View of the State
of the Art.” Psychological Methods 7(2): 147-177.
Stewart, Jay. 2006. “Assessing Alternative Dissimilarity Indexes for
Comparing Activity Profiles.” Electronic International Journal of Time
Use Research, 3(1): 49-59.
Tiggemann, M. 2006. “Nonreporting of Body Mass Index: A Research Note
on the Interpretation of Missing Data.” International Journal of Eating
Disorders 39(4): 346-349.
Tourangeau, R., R.M. Groves, and C. Redline. 2010. “Sensitive Topics and
Reluctant Respondents: Demonstrating a Link Between Nonresponse Bias
and Measurement Error.” Public Opinion Quarterly 74(3): 413-432.
Tourangeau, R., L.J. Rips, and K. Rasinski. 2000. The Psychology of Survey
Response. New York: Cambridge University Press.
Triplett, T., J. Blair, T. Hamilton, U.C. Kang. 1996. Initial Cooperators
vs. Converted Refusers: Are There Response Behavior Differences?
American Statistical Association Meetings, Survey Research Methods
Section, August 4-7, Chicago, IL. Available at: www.amstat.org/sections/
srms/proceedings/y1996f.html
U.S. Bureau of Labor Statistics and U.S. Census Bureau. 2006. Current
Population Survey: Design and Methodology. Technical Paper TP66,
October. Available at: www.census.gov/apsd/techdoc/cps/cps-main.html
29
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
U.S. Department of Labor, Bureau of Labor Statistics. 2010. American Time
Use Survey User’s Guide: Understanding ATUS 2003 to 2009. July 2010.
Available at: http://stats.bls.gov/tus/atususersguide.pdf.
Vernon, Margaret K. 2005. “Pre-testing Sensitive Questions: Perceived
Sensitivity, Comprehension, and Order Effects of Questions About Income
and Weight, Bureau of Labor Statistics Working Paper, August. Available
at: http://stats.bls.gov/ore/pdf/st050090.pdf.
Wagstaff, D.A., S. Kranz, and O. Harel. 2009. “A Preliminary Study of
Active Compared With Passive Imputation of Missing Body Mass Index
Values Among Non-Hispanic White Youths.” American Journal of
Clinical Nutrition 89(4): 1025-1030.
Yan, Ting, and Roger Tourangeau 2004. “When Less Is More: Are Reluctant
Respondents Poor Reporters?” Survey Research Methods Section,
American Statistical Association Meetings, Toronto, Canada, August 8-12.
Available at: www.amstat.org/sections/srms/proceedings/y2004f.html.
Websites
American Time Use Survey: stats.bls.gov/tus/
Eating and Health Module: ers.usda.gov/Data/ATUS/
Centers for Disease Control BMI information: http://www.cdc.gov/healthyweight/assessing/bmi/index.html
30
Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module / TB-1934
Economic Research Service/USDA
| File Type | application/pdf |
| File Title | Nonresponse Bias Analysis of Body Mass Index Data in the Eating and Health Module |
| Subject | Time use, American Time Use Survey, Eating and Health Module, nonresponse bias, item nonresponse, Body Mass Index, BMI, dissimil |
| Author | Karen S. Hamrick |
| File Modified | 2013-12-18 |
| File Created | 2012-08-15 |