BRFSS Technical Notes

  1. Methodology
  2. Data Analysis
  3. Population Density Groups
  4. Questionnaire Design
  5. Types of Questions
  6. Limitations
  7. Quality Control

Sampling
During 1992-1998, the Kansas Behavioral Risk Factor Surveillance System (BRFSS) was conducted using a simple random sampling method. In this method of sampling, each telephone number in the population has an equal probability of being called. The simple random sample is created by combining the known area codes and prefixes in the surveillance area with randomly generated suffixes.

From 1999-2001 & 2003-2008, the Kansas BRFSS was conducted using disproportionate stratified sampling methodology that considers the entire state as a single geographical stratum. This method of probability sampling involved assigning sets of one hundred telephone numbers with the same area code, prefix, and first two digits of the suffix and all possible combinations of the last two digits ("hundred blocks") into two strata. Those hundred blocks that have at least one known household number are designated high density (also called "one-plus blocks"); hundred blocks with no known household numbers are designated low density ("zero blocks"). The high density stratum is sampled at a rate 1.5 times higher than the low density stratum, resulting in greater efficiency.

In 2002, the sampling method was slightly modified. The survey was conducted using disproportionate stratified sampling methodology that considers the entire state as a single geographical stratum as in the earlier years but the probability sampling for assigning set of telephone number consisted of three strata: listed one-plus block numbers, not listed one-plus block numbers, and zero block numbers. Not listed one-plus numbers are sampled at two-thirds the rate of listed numbers; zero block numbers are sampled at one-fifth the rate of listed numbers. The sampling was changed to increase survey efficiency.

Beginning in 2009, the sampling method was modified by implementation of disproportionate stratified sampling methodology that included selection of land line telephone numbers within 10 geographic strata comprised of county grouping instead of random selection of telephone numbers from the entire state as a single geographic stratum. These 10 geographical strata include; Johnson county, Sedgwick county, Shawnee county, Wyandotte county, Northwest public health district, Southwest public health district, North Central public health district, South Central public health district excluding Sedgwick county, Northeast public health district excluding Johnson, Shawnee and Wyandotte counties, and Southeast public health district. The sample that is drawn from each geographical stratum is based on population size within each geographical stratum, the confidence level and the margin of error. This is a methodology that is commonly used to target collection for geographically identifiable su bpopulations, for example people in rural areas. It also increases the accuracy of prevalence estimates for a small subpopulation. This modification in the sampling methodology of the 2009 and future Kansas BRFSS is made to address the need to collect adequate sample to provide local or county level data. These data are needed to determine priority health issues, to identify population subgroups at higher risk of illness, and to monitor the health status of local communities. This goal can be achieved by providing BRFSS data for the individual counties (counties with bigger population sizes) and for bioterrorism regions. As in previous years, this method of probability sampling involved assigning sets of one hundred telephone numbers with the same area code, prefix, and first two digits of the suffix and all possible combinations of the last two digits ("hundred blocks") into two strata. Those hundred blocks that have at least one known household number are designated high density (also called "one-p lus blocks"); hundred blocks with no known household numbers are designated low density ("zero blocks"). The high density stratum is sampled at a rate 1.5 times higher than the low density stratum, resulting in greater efficiency.

Approximately the same number of persons is called each month throughout each calendar year to reduce bias caused by seasonal variation of health risk behaviors. Potential working telephone numbers are dialed during three separate calling periods (daytime, evening, and weekends) for a total of 15 call attempts before being replaced. Upon reaching a valid household number, one household member ages 18 years or older is randomly selected. If the selected respondent is not available, an appointment is made to call at a later time or date. Because respondents are selected at random and no identifying information is solicited, all responses to this survey are anonymous.

In 2010, the landline telephone survey used the survey methodology identical to that of 2009 survey.

Changes in Kansas BRFSS Survey Sampling Methodology: From 2011 onwards, a major change has been made in the sampling methodology of the Kansas BRFSS. This change is instituted to comply with the guidelines provided by the CDC for 2011 survey. From 2011 onwards, a dual frame sampling methodology (landline telephone sample and cellular telephone sample) will be used for Kansas BRFSS instead of single frame methodology (landline telephone sample).

In 2011, the CDC advised all states and territories to implement a dual frame sampling methodology for BRFSS survey and to include both: adults 18 years and older living in private residences with landline telephone service; and adults 18 years and older living in private residences with cellular telephone only service. The states were advised to target at least 20 percent of their total sample of complete interviews to be from cellular telephone only service households. This change in sampling methodology of the BRFSS is made to address the impact of growing number of households with cellular telephone only service and differences in the demographic profile of the people who live in cellular telephone only service households and to maintain representativeness, coverage, and validity of BRFSS data.1 To be in compliance with the current guidelines regarding BRFSS sampling methodology, Kansas BRFSS program has implemented dual frame sampling methodology for 2011 Kansa s BRFSS survey.

The dual frame sampling methodology for 2011 survey included two components: 1) Landline telephone service survey component; and 2) Cellular telephone only service component. The landline telephone survey component of this dual frame sampling method remained identical to the sampling method for 2009 and 2010 surveys. It comprised of implementation of disproportionate stratified sampling methodology that included selection of landline telephone numbers within 10 geographic strata comprised of county grouping instead of random selection of telephone numbers from the entire state as a single geographic stratum. These 10 geographical strata include; Johnson county, Sedgwick county, Shawnee county, Wyandotte county, Northwest public health district, Southwest public health district, North Central public health district, South Central public health district excluding Sedgwick county, Northeast public health district excluding Johnson, Shawnee and Wyandotte counties, and Southeast public health district. The sample that is drawn from each geographical stratum is based on population size within each geographical stratum, the confidence level and the margin of error. The landline telephone component sampling was designed to reach non-institutionalized adults ages 18 years and older living in the private residences in Kansas. As in previous years, this method of probability sampling involved assigning sets of one hundred telephone numbers with the same area code, prefix, and first two digits of the suffix and all possible combinations of the last two digits ("hundred blocks") into two strata. Those hundred blocks that have at least one known household number are designated high density (also called "one-plus blocks"); hundred blocks with no known household numbers are designated low density ("zero blocks"). The high density stratum is sampled at a rate 1.5 times higher than the low density stratum, resulting in greater efficiency. The cellular telephone survey component of this dual frame sampling method included the sampling frame comprised of all 1000-series blocks dedicated to cellular devices serving the state with a nonzero chance of inclusion. The cellular telephone survey component sampling was designed to reach non-institutionalized adults ages 18 years and older living in the private residences with cellular telephone only service in Kansas.

In 2012, the CDC further advised all states to make two additional changes in the 2012 BRFSS methodology. These additional changes included: 1) inclusion of respondents living in households with both cell phone and landline service but receiving 90 percent or more of their calls on cell phones (cellular telephone mostly households) in cell phone survey sample thus addressing the impact of increased use of cell phones in households with dual telephone service (cell phone mostly households) in addition to the impact of growing number of households with cellular telephone only service and differences in the demographic profile of the people who live in cellular telephone only service households to further maintain representativeness, coverage, and validity of BRFSS data; and 2) inclusion of residents living in college housing with landline and/or cellular telephone service in both landline and cell phone samples. These changes in the BRFSS survey sampling methodology will allow inclusion of respondents from the cellular telephone mostly households, as well as respondents living in the college housing, thus making the survey sample more representative of the general population.

Also, in 2012 survey, for landline telephone sample, landline service over the internet was counted as landline service. This included Vonage, Magic Jack and other home-based phone services. Besides these changes, sampling methodology for landline and cellular phone components for 2012 survey was same as 2011 survey.

Sample Size
From 2000-2003 Kansas BRFSS survey sample size was about 4,000 respondents and from 2004-2008 it was about 8,000 respondents.

The target sample size in odd numbered years beginning in 2009 is 16,000 complete interviews. The target sample size in even numbered years will remain 8,000. The target sample for 2009 survey was 16,000 complete interviews; and for 2010 survey was 8,000 complete interviews.

For 2011 Kansas BRFSS survey, the target total (combined landline and cell phone sample) sample size was about 19,200 respondents with a target of 16,000 respondents for the landline telephone survey component and 3,200 respondents for the cellular telephone survey component.

For 2012 Kansas BRFSS survey, the target total (combined landline and cell phone sample) sample size was about 10,000 respondents with a target of 8,000 respondents for the landline telephone survey component and 2,000 respondents for the cellular telephone survey component (20% of the state's total combined landline and cell phone sample).

Weighting Procedure
Data weighting is an important statistical process that attempts to remove bias in the sample. It corrects for differences in the probability of selection due to non-response and non-coverage errors. It adjusts variables of age and gender between the sample and the entire population. Data weighting also allows the generalization of findings to the whole population, not just those who respond to the survey.

Once BRFSS data are collected, statistical procedures are undertaken to make sure the estimates of health indicators generated by the analysis of survey data are representative of the population for each state and/or local area.

This weighting process of BRFSS data includes calculation of design weight as one of its components: In BRFSS survey, the design factors that affect weighting include; number of residential telephones in household, number of adults in household and geographic or density stratification. The formula for calculation of design weight is:

Design weight = STRWT * 1 OVER IMPNPH * NUMADULT

Weighting process of BRFSS also involves adjustment for the distribution of the sample data so that it reflects more accurately the total population of the sampled area. The method used to for this adjustment till 2010 is called the post-stratification methods. This method involves calculation of post-stratification factor by computing the ratio of the age, race, and sex distribution of the state population divided by that of the sample.

This post stratification factor is then multiplied by the design weight to compute an adjusted, final weight variable. Thus the weighting process adjusts not only for variation in selection and sampling probability but also for demographic characteristics so that projections can be made from the sample to the general population. The computational formula below is intended to reflect all the possible factors that could be taken into account in weighting a state's data till 2010. If a factor does not apply, its value is set to one.

The formula for weighting using post-stratification method:

FINALWT = Design WT * POSTSTR

or

FINALWT = STRWT * 1 OVER IMPNPH * NUMADULT * POSTSTR.

Final weight variable is the use for analysis of survey data to generate estimates of health indicators for general population.

Additional facts about data weighting are:

  • Weighting consists of a lot more than post-stratification.
  • Weighting for design factors has more of an effect on final results than does post-stratification.
  • Weighting for design factors is also more important conceptually.
  • Weighting affects both the point estimate (bias) and confidence intervals (precision). 

Application of the weighting process allows comparability of data. However weighting can only be performed when the sampling methodology is carefully controlled.

New Weighting Methodology: Iterative Proportional Fitting or Raking

Since 1980s, as mentioned above the CDC has used "post stratification statistical method" to weight BRFSS survey data to simultaneously adjust survey respondent data to known proportions of age, race and ethnicity, gender, geographic region, or other known characteristics of a population. This type of weighting is important because it makes the sample more representative of the population and adjusts for non-response bias. In 2006, in accordance with the recommendations of the expert panel of survey methodologists, CDC began testing a more sophisticated weighting method called "iterative proportional fitting", or "raking".

Raking method adjust the data so that groups which are underrepresented in the sample can be accurately represented in the final dataset. Raking allows for the incorporation of cell phone survey data, permits the introduction of additional demographic characteristics and more accurately matches sample distributions to known demographic characteristics of populations. The use of raking method reduces non-response bias and has been shown to reduce error within estimates.

Raking has several advantages over post stratification. First, it allows the introduction of more demographic variables suggested by the BRFSS expert panel such as education level, marital status, and home ownership into the statistical weighting process than would have been possible with post stratification. This advantage further reduces the potential for bias and increases the representativeness of estimates. Second, raking allows for the incorporation of a now crucial variable telephone source (landline or cellular telephone) into the BRFSS weighting methodology.

Beginning with the 2011 dataset, the CDC has adopted raking or method in place of post stratification weighting procedure as the sole BRFSS statistical weighting method.

The new BRFSS weighting methodology is comprised of two components:

  • Design Weight
  • Raking Adjustment

Design Weight: Design Weight is calculated by using computational formula:
Design Weight = _STRWT * (1/NUMPHON2) * NUMADULT 

  • The stratum weight (_STRWT) is calculated using:
    • Number of available records (NRECSTR) and the number of records selected (NRECSEL) within each geographic strata (_GEOSTR) and density strata (_DENSTR);
    • Geographic strata (entire state, counties, census tracts, etc.); and
    • Density strata (1=listed numbers, 2=not listed numbers).
    • Within each _GEOSTR *_DENSTR combination: The stratum weight (_STRWT)is calculated from the average of the NRECSTR and the sum of all sample records used to produce the NRECSEL.

The computational formula for stratum weight:
STRWT = NRECSTR / NRECSEL 

  • 1/ NUMPHON2 is the inverse of the number of residential telephone numbers in the respondent's household.
  • NUMADULT is the number of adults 18 years and older in the respondent's household.

Final Weight is calculated for analysis of survey data to generate estimates for health indicators that are representative of the general population. 

The computational formula for Final weight:

Final Weight = Design Weight * Raking Adjustment

Raking adjustment: Raking adjusts estimates within each state by using: 

  • Telephone source,
  • Detailed race and ethnicity,
  • Regions within state,
  • Education level,
  • Marital status,
  • Age group by gender,
  • Gender by race and ethnicity,
  • Age group by race and ethnicity, and
  • Renter/homeowner status.

Raking is completed by adjusting for one demographic variable (or dimension) at a time. For example, when weighting by age and gender, weights would first be adjusted for gender groups, then those estimates would be adjusted by age groups. This procedure would continue in an iterative process until all group proportions in the sample approach those of the population, or after 75 iterations.

Weighted data analysis techniques are used to analyze BRFSS survey to generate population based estimates of health indicators. The Final weight variable is used in these analyses. 

Weight Trimming in Raking

Weight trimming is used to increase the value of extremely low weights and decrease the value of extremely high weights. The objective of weight trimming is to reduce errors in the outcome estimates caused by unusually high or low weights in some categories.

Source: Above description (language) on "New Weighting Methodology" is provided to the state BRFSS programs through the factsheets titled Behavioral Risk Factor surveillance System (BRFSS) Fact Sheet: Raking and Behavioral Risk Factor Surveillance System Improving Survey Methodology prepared by the Public Health Surveillance Program Office and Division of Behavioral Surveillance, Office of Surveillance, Epidemiology and Laboratory Services, Centers of Disease Control and Prevention.

The new survey methodology, including dual frame sampling and the iterative proportional fitting (IPF) or raking method, was used starting with the 2011 data. Therefore, DO NOT COMPARE 2011 to present data with previous years.

Data Reliability
Telephone interviewing has been demonstrated to be a reliable method for collecting behavioral risk data and can cost three to four times less than other interviewing methods such as mail-in interviews or face-to-face interviews. The BRFSS methodology has been utilized and evaluated by the CDC and other participating states since 1984. Content of survey questions, questionnaire design, data collection procedures, surveying techniques, and editing procedures have been thoroughly evaluated to maintain overall data quality and to lessen the potential for bias within the population sample.