A to Z Index  |  FAQs  |  About BLS  |  Contact Us

Error Measurement

Definition

In the process of collecting data, some error occurs. BLS tends to categorize these errors in to two types: sampling error and nonsampling error.

Sampling error affects survey-based data. Sampling error occurs when the results of the survey differ from the “true” population result.

Nonsampling error can affect any collected data. Types of nonsampling error include keypunch errors, errors in the collection or processing of data, misclassification of data, and nonresponse from survey members.

Stylized example of error measurement

Janet asks 100 women to report their height. She averages their responses and finds that the average height of the group to be 70 inches (or 5 feet 10 inches). After doing some research, she finds that according to the Centers for Disease Control and Prevention, the average height for women is 63.8 inches (or nearly 5 feet 4 inches).  Her sampling error is 6.2 inches.

In further research, Janet looks at the individual responses and finds that two responses that should have been keyed as 59 inches were instead keyed as 95 inches (7 feet 11 inches!). This type of keypunch error would be considered nonsampling error. From this example, the nonsampling error contributed to the sampling error. In practice, nonsampling error can be difficult to identify and quantify.

Understanding confidence intervals

To help quantify sampling error. BLS provides standard error estimates. Standard errors are used to construct confidence intervals around an estimate, which essentially means that BLS believes with 90 percent certainty that the “true” estimate is within the band surrounding the estimate.

For example, for the month of June wages are reported as $15.15 and the standard error is 0.15. To construct a 90-percent confidence interval, multiply the standard error (0.15) by 1.65 (which is the critical z value for measuring a 90-percent interval). 0.15×1.65=0.2475 Next we add this product to our wage level to get the upper bound. Then subtract it from our wage level to get our lower band.$15.15+0.2475=$15.40 (rounded upper limit)$15.15-0.2475=$15.13 (rounded lower limit) 90-percent confidence interval=$15.13 to $15.40 Now, let’s imagine that for July, the new wage estimate was$15.25. There is a $0.10 rise in the wage, but it is not statistically different from$15.15 because $15.25 falls within the 90-percent confidence band surrounding the$15.15 wage estimate.

Error measurement at BLS

Users may find more information about error measurement for each program in the Calculation section of the Handbook of Methods.