Precision Standards Guidelines for Reporting MEPS-HC Descriptive Statistics
The following statistical standards are applied for tabular estimates that can be obtained from the MEPS website, and are recommended as general rules for researchers analyzing MEPS public-use file data.
A. Continuous Variables (estimates of means, totals or ratios)
- Number of Sample Persons
- n ≥ 60: Published estimates should be based on an unweighted sample of at least 60 persons for the subgroup of interest.
- Relative Standard Error (RSE = SE/Estimate)
- RSE > .50: Estimate should not be reported or displayed in tables due to extremely large sampling error.
- .30 ≤ RSE ≤ .50: Estimate can be reported but flagged with an * to indicate that its precision is questionable.
- RSE < .30: Estimate is deemed precise and can be reported with no *.
B. Categorical Variables (estimates total counts or percentages)
- Number of Sample Persons
- n ≥ 601: Published estimates should be based on an unweighted sample of at least 60 persons for the subgroup
represented in the total count (e.g. # children without insurance) or in the denominator of a percentage1.
- Relative Standard Error (RSE = SE/Estimate)
- RSE > .50: Do not report estimate if upper bound of 95% confidence interval for the estimate is > 10%. If upper bound is 10% or less, then it is permissible to report with an *2.
- .30 ≤ RSE ≤ .50: : Estimate can be reported but flagged with an * to indicate that its precision is questionable.
- RSE < .30: Estimate is deemed precise and can be reported with no *.
1For example, the denominator when estimating the percentage of children without
health insurance (# children without insurance/total # of children) is the total # of children in the sample.
2The purpose of this exception is to avoid suppressing all estimates of
characteristics with low estimated prevalence (e.g., less than 10%) due to large RSEs attributable to small denominators.
For example, the RSE for an estimate of 2% with an SE of 1 percentage point is 0.5. In this situation, the upper end of a
95% confidence interval would be around 4% which provides useful information that the prevalence of the characteristic is
fairly uncommon (i.e. likely no higher than 4%). However, if the estimate had been 50% with an RSE of .50, then the
95% confidence interval would span from around 0 to 100. That estimate would clearly be too imprecise to provide any
useful information.
|
Back
Email Updates
