BRFSS Data Quality, Validity, and Reliability
Overview of Behavioral Risk Factor Surveillance System (BRFSS) 2004
Expert Panel Review and Recommendations
In May 2002, the Behavioral Risk Factor Surveillance System (BRFSS) held
its first expert panel review. The second review meeting was held in
November 2004. At both meetings, approximately twenty leading survey
statisticians, methodologists, and operational experts gathered to discuss the challenges facing the field of survey
research and implications for the BRFSS. The goal of these meetings was to
develop options and prioritize recommendations for maintaining data quality
in the face of societal and technological changes, managing an increasingly
complex surveillance system, and meeting the demand for more local-level
data and expanded analysis capabilities.
The meetings began with 3–4
overview presentations on challenges facing survey research generally in the
areas of statistics, operations, estimation, and system infrastructure. The
larger working group was then divided into smaller discussion panels. The
2002 meeting panels focused on technological, methodological, and system
challenges. The 2004 breakout groups were organized around statistical and
operational issues. Each group held open discussion and deliberation of
issues, then drafted recommendations for research and system improvements
which were presented at a plenary session. These recommendations were then
prioritized by the larger working group in terms of importance and
The May 2002 BRFSS expert panel meeting generated 42 recommendations, of
which 37 were implemented or are in progress (see Morbidity and Mortality
Weekly Report, May 23, 2003, Vol. 52, No. RR-9). The November 2004 panel
made 38 specific recommendations, including use of more complex weighting and
imputation procedures, experiments with incentives, multiple modes of data
collection and alternative sampling frames, and assessment of processes for
reaching language-isolated households (see appendix for list of specific
recommendations). These recommendations will help shape BRFSS research and
process improvement activities for the next 2–3
Participants at the 2004 meeting were Lina Balluz (CDC), Michael Battaglia (Abt
Associates), Paul Biemer (University of North Carolina at Chapel Hill and
RTI International), Stephen Blumberg (CDC), Barbara Bowman (CDC), J. Michael
Brick (Westat), Donna Brogan (Emory University), Bonnie Davis (Public Health
Institute), Michael Elliott (University of Pennsylvania School of Medicine),
Amy Ferketich (Ohio State University), Earl Ford (CDC), Martin Frankel
(Baruch College at CUNY and Abt Associates), William Garvin (CDC), Gary
Gentry (Public Health Institute), Wayne Giles (CDC), Ziya Gizlice (North
Carolina Department of Health and Human Services), Virginia Bales Harris
(CDC), Ruth Jiles (CDC), William Kalsbeek (University of North Carolina at
Chapel Hill), Jim Lepkowski (University of Michigan), Paul Levy (RTI
International), Michael Link (CDC), Ali Mokdad (CDC), Cynthia Nelson
(Northern Illinois University), Sarah Nusser (Iowa State University), Colm
O’Muircheartaigh (National Opinion Research Center), Charlie Palit
(University of Wisconsin-Madison), Nagi Salem (Minnesota Department of
Health), Fritz Scheuren (National Opinion Research Center), and Donna Stroup
2004 BRFSS Expert Panel Recommendations
High priority/Short timeframe:
- State-specific nonresponse analysis: examine disposition codes (past
and future) for use in response rate improvement and model-assisted
- Nonresponse simulation analysis: use past years' BRFSS disposition
codes to assess how simulated variations in response rates may affect
estimates (i.e., simulate 30, 40, and 50 percent response rates based
on call history information).
- Evaluate call center “house effects” by estimating intra-interviewer
correlation (inter-interviewer variance), which could be contributing to
some of the cross-state variances.
- Evaluate modifications to current weighting and poststratification
adjustments: examine coherence of weighting procedures across states;
consider the effects of alternative weighting adjustments, such as race,
education, geography, and telephone interruption; consider use of
raking and weight trimming; develop diagnostic testing of weights;
and consider the use of model-assisted sampling.
- Evaluate methods for imputing missing data: consider the use of
imputation for weighting variables, other descriptive variables,
and substantive variables
High priority / Longer timeframe:
- Continue alternative and mixed-mode studies looking at mode
telephone, mail, Web, and in-person interviewing; appropriateness of question wording for self-administered
and interviewer-administered questionnaires; and utility as a refusal
- Investigate a predominantly mail-mode BRFSS, focusing on
evaluation of within-household selection techniques, use of
substitution for nonrespondents, and follow-up of refusals and noncontacts in subsequent months.
- Conduct assessment of how Spanish-language interviewing is
handled, including techniques for translation and assessment of
cultural comparability with English version. Also assess cultural
implications of contact attempts – trust issues, legitimacy, etc.
- Assess extent of language isolation: evaluate and if necessary
modify disposition codes to capture more information about what language
is potentially spoken, whether the language is other than English
or Spanish, and whether the language is Spanish, and a bilingual
interviewer is not available at the time.
- Conduct feasibility, utility, and quality assessment of Language
Line Services for on-the-phone translations into languages other
than English or Spanish.
- Nonresponse adjustment: collect basic demographic information
(e.g., sex, age, race) of the selected but unavailable respondent from
the proxy household member who answers the respondent selection
- Assess noncoverage and nonresponse bias in selected counties
(selected from SMART-BRFSS counties) comparing RDD sample to an area
probability sample. Start with address-based sample. Conduct
interviews initially by telephone, then follow up with in-person
interviews. Goal may be to justify adequacy of low response rates,
not to adjust the estimates per se.
- Continue and expand research on external validation for quality.
- Ensure cross-state standardization for CATI programming.
- Evaluate tailored and modified introductions highlighting local
- Sample release: conduct analysis to evaluate the
statistical, operational, and analytic (including weighting)
pros and cons of releasing sample weekly or biweekly.
- Field period: consider extending current one-month field
period to six or eight weeks to improve response rates.
- Evaluate partial interviews and terminations to determine
their feasibility for use in nonresponse adjustment and
imputation of missing data.
- Evaluate current within-household randomization procedure to
determine if it is optimal, given the changes in household
composition over time and underrepresentation of particular
- Evaluate cell-phone-only coverage bias and nonresponse bias
with multiframe, multimode experiments. Study design
components using a sample from a landline frame and a cellular
frame, conducting a survey initially by telephone, and applying
follow-up techniques to complete nonrespondent interviews
(consider use of an abbreviated mail questionnaire). Analysis
would consist of comparing characteristics of cellular
respondents with land line respondents, comparing
characteristics of telephone respondents with responses obtained
in the follow-up, and comparing results with external sources
- Evaluate nonresponse bias and response propensity by seeding
the sample with persons of known characteristics and by applying frames
of other surveys.
- As part of the reporting process, identify methods of
presenting quality declarations (i.e., information on the
quality of the data) to explain how to interpret the quality
measures associated with the study.
- Consider the use of an abbreviated questionnaire for nonresponse
follow-up and bias assessment.
- Conduct cognitive and language-level assessment of the English
version of questionnaire to ensure that it is correctly
comprehended by all.
- Consider the use of specialized, stand-alone surveys to reach
important hard-to-reach or hard-to-interview populations (such
as institutionalized adults, mobile-only households, and Native
- Incentives: consider experiments with incentives,
including up-front vs. promised incentives and
incentives for use with nonrespondents.
- Examine the impact of imputation on sampling error.
- Use NHIS as a standard for evaluating the validity of
telephone, Web, and mail surveys.
- Use NHIS to evaluate within-household correlation on
health estimates (if correlations are low, then surveying
multiple household members by mail may be more cost-effective than interviewing only one selected household
- Consider the effectiveness of quota
sampling with backend weighting adjustments to determine if
better estimates are produced at a lower cost.
Back to Data Quality, Validity, and