26.12.2012 Views

Current Population Survey Design and Methodology - Census Bureau

Current Population Survey Design and Methodology - Census Bureau

Current Population Survey Design and Methodology - Census Bureau

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Chapter 16.<br />

Quality Indicators of Nonsampling Errors<br />

(Updated coverage ratios, nonresponse rates, <strong>and</strong> other measures of quality can be found by clicking on ‘‘Quality Measures’’ at<br />

.)<br />

INTRODUCTION<br />

Chapter 15 contains a description of the different sources<br />

of nonsampling error in the CPS <strong>and</strong> the procedures<br />

intended to limit those errors. In the present chapter, several<br />

important indicators of potential nonsampling error<br />

are described. Specifically, coverage ratios, response variance,<br />

nonresponse rates, mode of interview, time-insample<br />

biases, <strong>and</strong> proxy reporting rates are discussed. It<br />

is important to emphasize that, unlike sampling error,<br />

these indicators show only the presence of potential nonsampling<br />

error, not an actual degree of nonsampling error<br />

present.<br />

Nonetheless, these indicators of nonsampling error are<br />

regularly used to monitor <strong>and</strong> evaluate data quality. For<br />

example, surveys with high nonresponse rates are judged<br />

to be of low quality, but the actual nonsampling error of<br />

concern is not the nonresponse rate itself, but rather nonresponse<br />

bias, that is, how the respondents differ from the<br />

nonrespondents on the variables of interest. Although it is<br />

possible for a survey with a lower nonresponse rate to<br />

have a larger nonresponse bias than a survey that has a<br />

higher nonresponse rate (if the difference between respondents<br />

<strong>and</strong> nonrespondents is larger in the survey with the<br />

lower nonresponse rate than it is in the survey with the<br />

higher nonresponse rate), one would generally expect that<br />

larger nonresponse indicates a greater potential for bias.<br />

While it is relatively easy to measure nonresponse rates, it<br />

is extremely difficult to measure or even estimate nonresponse<br />

bias. Thus, these indicators are simply a measurement<br />

of the potential presence of nonsampling errors. We<br />

are not able to quantify the effect the nonsampling error<br />

has on the estimates, <strong>and</strong> we do not know the combined<br />

effect of all sources of nonsampling error.<br />

COVERAGE ERRORS<br />

When conducting a sample survey, the primary goal is to<br />

give every person in the target universe a known probability<br />

of being selected for the sample. When this occurs, the<br />

survey is said to have 100 percent coverage. This is rarely<br />

the case, however. Errors can enter the system during<br />

almost any phase of the survey process, from frame creation<br />

to interviewing. A bias in the survey estimates<br />

results when characteristics of people erroneously<br />

included or excluded from the survey differ from those of<br />

individuals correctly included in the survey. Historically in<br />

the CPS, the net effect of coverage errors has been an<br />

<strong>Current</strong> <strong>Population</strong> <strong>Survey</strong> TP66<br />

U.S. <strong>Bureau</strong> of Labor Statistics <strong>and</strong> U.S. <strong>Census</strong> <strong>Bureau</strong><br />

underestimate of the size of the total population for most<br />

major demographic population subgroups before the<br />

population controls are applied (known as undercoverage).<br />

Coverage Ratios<br />

One way to estimate the coverage error present in a survey<br />

is to compute a coverage ratio. A coverage ratio is the<br />

outcome of dividing the estimated number of people in a<br />

specific demographic group from the survey by an independent<br />

population total for that group. The CPS coverage<br />

ratios are computed by dividing a CPS estimate using the<br />

weights after the first-stage ratio adjustment by the independent<br />

population controls used to perform the national<br />

<strong>and</strong> state coverage adjustments <strong>and</strong> the second-stage<br />

ratio adjustment. See Chapter 10 for more information on<br />

computation of weights. <strong>Population</strong> controls are not error<br />

free. A number of approximations or assumptions are<br />

required in deriving them. See Appendix C for details on<br />

how the controls are computed. Chapter 15 highlighted<br />

potential error sources in the population controls. Undercoverage<br />

exists when the coverage ratio is less than 1.0<br />

<strong>and</strong> overcoverage exists when the ratio is greater than<br />

1.0. Figure 16−1 shows the average monthly coverage<br />

ratios for September 2001 through September 2004.<br />

In terms of race, Whites have the highest coverage ratio<br />

(90.7 percent), while Blacks have the lowest (82.2 percent).<br />

Females across all races have higher coverage ratios<br />

than males. Hispanics 1 also have relatively low coverage<br />

rates. Historically, Hispanics <strong>and</strong> Blacks have lower coverage<br />

rates than Whites for each age group, particularly the<br />

20−29 age group. This is by no fault of the interviewers or<br />

the CPS process. These lower coverage rates for minorities<br />

affect labor force estimates because people who are<br />

missed by the CPS are on the average likely to be different<br />

from those who are included. People who are missed are<br />

accounted for in the CPS, but they are given the same<br />

labor force characteristics as those of the people who are<br />

included. This produces bias in the CPS estimates.<br />

This graph, as well as two other graphs of coverage ratios<br />

by race <strong>and</strong> gender, can be found at . (Their<br />

updates, with more current data, will be posted on this<br />

site as they are made available.) The three graphs provide<br />

1 Hispanics may be any race.<br />

Quality Indicators of Nonsampling Errors 16–1

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!