The Behavioral Risk Factor Surveillance System (BRFSS) is an annual survey conducted by the Centers for Disease Control (CDC). The central focus of the survey is to measure preventative behaviors and practices and associated health outcomes for US adults. The survey has been administered since 1984 and currently over half a million interviews are conducted per year.

We pool the surveys for the years 2006-2010 which covers the years in which the life satisfaction question was included in the survey. The total sample size of respondents who answered the central life satisfaction question is 1,988,275. The full data including all respondents from the 2006-2010 BRFSS (only the subset variables from the full survey that we use), FIPS code county location for each respondent and county level characteristics can be downloaded here.

Since we wish to explain variation in satisfaction by respondents’ demographics, we filter the data to delete missings for variables used in the analysis. We eliminate respondents with missing values for demographics and county locations. We also only include responses for White, Black and Hispanic respondents. Finally, since we want to control for location effects we also eliminate respondents from counties where we observe 50 or fewer respondents. R code for filtering the data can be obtained here.

We use the following variables from the BRFSS:

Table 1. Variable Definitions and Sample Selection Rules.

BRFSS Variable	Notation	Description	Selection Rule
INCOME2	Income	Annual Income	Removed missing values
AGE	Age	Age in years	Removed missing values
MARITAL	Marital	Marital Status	Removed missing values
SEX	Gender	Gender of respondent
CHILDREN	Children	Number of children in household	Removed missing values
EMPLOY	Employment	Employment Status	Removed missing values
GENHLTH	Health	General Health	Removed missing values
RACE2	Race	Race	Include only White, Black and Hispanic respondents
CTYCODE	County	County FIPS CODE	Removed observations from unknown counties, counties with less than 50 respondents and counties where county level demographics were not available

These selection rules lowers the sample size from 1,988,275 to 1,390,826 (the filtered data can be downloaded from here). The effect of each variables selection rule is shown in the left figure below. The biggest reduction is due to missing income observations and removing observations from counties with a low number of respondents. However, the sample selection doesn’t bias the average response of the key outcome variable: The proportions for each level of the satisfaction measure is not affected by the selection rules.

Figure: Effect of Sample Selection rules on Sample Size and Outcome Measure. The left chart shows the effect of successive sample selection rules on final sample size. The right chart compares the distributions of the satisfaction measure for the full and final (filtered) sample.