Small Area Estimation for the Tobacco Use Supplement to the Current Population Survey Benmei Liu Statistical Research and Applications Branch/ SRP/DCCPS May 10 2018 TCRB Scientific Staff Meeting Outline
Overview of Small Area Estimation (SAE) Research goals SAE models and implementation Model-based estimates for maps Summary and discussion Why Small-Area Estimation? The TUS-CPS is designed to produce reliable estimates at the national and state levels. Policy makers, cancer control planners and researchers
often need county level data for tobacco related measures in order to better evaluate tobacco control programs, monitor progress, and conduct tobacco-related research. The standard direct estimation (design-based) methods for TUS data cannot provide reliable estimates at county level due to small (or zero) sample size Model-based methods that combine information from multiple related sources are needed to increase precision
Overview of the Model-based SAE Techniques Borrowing strength from relevant sources (Census/ Administrative information, related surveys) Borrowed strength comes from covariates, and from other counties with similar characteristics Methods of combining Information - Choose good small area model - Use good statistical methodology Mixed models (fixed effects + random effects) at area level or
unit level have been popularly used in the small area estimation literature (Rao 2003, Jiang and Lahiri 2006). Among the many models developed in the SAE literature, the most prominent approach is the Fay-Herriot area- level model, originally developed to estimate per-capita income for U.S. areas with populations of less than 1,000. Statistical Inferences Using Mixed Models The final estimates are combinations of the direct estimates with the synthetic estimates.
Fully Bayesian approach or empirical best prediction approach (analytic formulas) can be used for the estimation. Application of SAE Techniques in Estimating Proportions Estimate cancer risk factors & screening behaviors for states and counties by combining data from Behavior Risk factor Surveillance System (BRFSS) and National Health Interview
Survey (NHIS) (http://sae.cancer.gov/) Estimate poverty rates for states, counties, and school districts in the Census Bureaus Small Area Income and Poverty (SAIPE) program (http://www.census.gov/did/www/saipe/) Estimate substance rates for states with data from the National Survey on Drug Use and Health (NSDUH) ( http://www.samhsa.gov/data/NSDUH/2k11State/NSDUHsae2011/i ndex.aspx ) Estimate proportions at the lowest level of literacy for states and
counties with data from the National Assessment of Adult Literacy (NAAL) (http://nces.ed.gov/naal/estimates/overview.aspx) Research Goals Produce model-based, county level estimates (n=3,137) for the following key measures (2010/2011 TUS, age 18+): 1. Percent of population currently smoking 2. Percent of population that has ever smoked 3. Percent of population that has quit for 24+ hours,
among those who have smoked within the past year 4. Percent of population governed by a smoke-free workplace policy 5. Percent of population governed by a smokefree home rule Involved collaboration among NCI, the Census Bureau, and the University of Maryland Parameters of interest
The population proportions: where is a binary response for unit k in county is the population total in county . Let denote the sample size, denote a vector of auxiliary variables. Direct Estimates of and Associated Variances Direct estimates (design-unbiased):
Variances of the direct estimates: =, . Where is the design effect reflecting the complex design (Kish 1965). Problem of : Variance too large (imprecise estimates) for small sample sizes
Small area estimation techniques to address imprecise estimates Commonly Used Area Level Model: Fay-Herriot Model The well known Fay-Herriot model (Fay & Herriot 1979): Sampling model: - is the sampling variance and is assumed known Linking model:
where ; This is equivalent to: where , ; Several transformations on the direct estimates are proposed to stabilize sampling variance
Fay-Herriot Model with C&R Arcsin Transformation Let ; (Carter & Rolph, 1974 JASA) o Sampling model: o Linking model: where Goal: To estimate Model was chosen based on an extensive simulation
study Hyper parameters to be estimated are , Estimate the Design Effects The design effect (or DEFF) is the ratio of the actual variance of a sample to the variance of a simple random sample of the same number of elements Multiple ways can be used to estimate DEFF. We used Kishs traditional design effect formula
given the clustering design of TUS-CPS, and estimated the state level design effects. We then used the state level DEFF to estimate the county level DEFF. Auxiliary Variables The pool of auxiliary variables include: - 30 county-level demographic & socio-economic variables obtained from ACS 2005-2009, 2008-2012, Census 2000 & 2010, and other administrative records;
- 5 state level tobacco policy data (cigarette taxes, clean air laws, tobacco control funding, Medicaid Coverage for Tobacco-Related Treatment, year in which Quitline service was established) Classical model selection procedures are applied to reduce the number of auxiliary variables for each outcome Tested forcing in several strong unit level covariates: only worked for current smoking and smoking cessation. Statistical Inference and Model Diagnosis
Hiearchical Bayesian approach through Markov Chain Monte Carlo (MCMC) methods were used to estimate the parameters of the statistical models. Extensive model selection and model diagnosis procedures are used to select the final models and assess the goodness of fit for each model. Modeled estimates were compared to the available direct estimates. The ratio of the two is expected to converge to 1 as the sample size gets larger.
Ratio of the Direct Over the Modeled Estimates for the Current Smoking Prevalence Model-based vs Design-based Estimates for Current Smoking Prevalence Maryland 2010/11 Model-based Estimates for Percent of Population Currently Smoking Among Age 18+: TUS-CPS 10/11 Model-based Estimates for Percent of Population Ever Smoked
Among Age 18+: TUS-CPS 10/11 Model-based Estimates for Percent of Population Live in SmokeFree Home Among Age 18+: TUS-CPS 10/11 Model-based Estimates for Percent of Population Attempt Quit Smoking for 24+ Hours Among Age 18+: TUS-CPS 10/11 Model-based Estimates for Percent of Population Governed by a Smoke-free Workplace Policy* Among Age 18+: TUS-CPS 10/11 Individual Self-Reported
*Workplace has an official smoking policy: Smoking Not allowed in ANY public areas and work areas https://sae.cancer.gov/tus-cps/ Law Legislations Applications of the SAE estimates and maps
Other Applications of SAE at NCI Small Area Estimates for Cancer Risk Factors and Screening Behaviors by Combining multiple Surveys - Multiple years of state and county level estimates are produced for 7 variables - Utilize data from both the National Health Interview Survey and the Behavior Risk Factor Surveillance System Small area estimates using the NCI-sponsored Health Information national Trends Survey (HINTS)
State level estimates are produced for 15 cancer-related knowledge variables Spatio-temporal Models for Cancer Burden Mapping To estimate age-standardized mortality rates and incidence rates by US county from a number of cancers and map the estimates to identify patterns and outliers The SAE website
24 The State Cancer Profiles Website 25 Summary and Discussion More details and results are available at https://sae.cancer.gov/tus-cps/. County level estimates of current and ever smoking
prevalence derived from this project are available upon request. Similar estimates were derived from the combining NHIS/BRFSS project and released at https://sae.cancer.gov/nhis-brfss/. Model-based SAE techniques represent a promising means of generating estimates where there is small (or zero) state or county sample. The SAE results provide a useful resource for the cancer surveillance, evaluation, and research communities. We are currently working on the TUS-CPS SAE estimates
for the 2014/2015 data cycle. Future works include model improvements and estimates for county by Race/ethnicity groups Any Questions? Thank you! Contact info: Benmei Liu [email protected]
