associated with the sample estimate: 
CV = — - ' 100",. 
CR, 
where SE(CRi) is the standard error of the capture- 
recapture estimate for data item i. This relative 
measure allows the reliability of a range of estimates 
to be compared. For example, the standard error is 
often larger for large population estimates than for 
small population estimates, but the large population 
estimates may have a smaller CY, indicating a more 
reliable estimate. For county-level estimates, a 
generalized coefficient of variation (GCVs) was 
determined for each estimate within a State. A 
generalized variance function relates a function of 
the variance of an estimator to a function of the 
estimator. Within a State, the standard error of an 
estimate for a data item was often found to be 
linearly related to the estimate of that item with an 
intercept of zero. Based on this modeled relationship, 
the GCV is the slope of the line relating the standard 
error to the estimate, multiplied times 100 to 
represent the GCV as a percentage. 
The standard error is the product of the CY (or GCV 
for county estimates) and the estimate divided by 
100. As an example, if the GCV for a State is 25 
percent and a county’s estimate is 4, then the 
standard error is 25 (4)/ 100 = 1. The standard error of 
an estimated data item from the census provides a 
measure of the error variation in the value of that 
estimated data item based on the possible outcomes 
of the census collection, including variants as to who 
was on the CML, who returned a census form, who 
was misclassified either as a farm or as a nonfarm, 
and the uncertainty associated with calibration and 
integerization. With 95 percent confidence, an 
estimate is within two standard errors of the true 
value being estimated. For this example, with 95 
percent confidence, the estimate of 4 is within 2(1) = 
2 of the true county value. 
Table B presents the fully adjusted estimates with 
the coefficient of variation for selected items. 
NONMEASURED ERRORS IN THE CENSUS 
PROCESS 
As noted in the previous section, sampling errors can 
2012 Census of Agriculture 
USDA, National Agricultural Statistics Service 
be introduced from the coverage, nonresponse and 
misclassification adjustment procedures. This error 
is measureable. However, nonsampling errors are 
imbedded in the census process that cannot be 
directly measured as part of the design of the census 
but must be contained to ensure an accurate count. 
Extensive efforts were made to compile a complete 
and accurate mail list for the census, to elicit 
response to the census, to design an understandable 
report form with clear instructions, to minimize 
processing errors through the use of quality control 
measures, to reduce matching error associated with 
the capture-recapture estimation process, and to 
minimize error associated with identification of a 
respondent as a farm operation (referred to as 
classification error). The weight adjustment and 
tabulation processes recognize the presence of 
nonsampling errors; however, it is assumed that 
these errors are small and that, in total, the net effect 
is zero. In other words, the positive errors cancel the 
negative errors. 
Respondent and Enumerator Error 
Incorrect or incomplete responses to the census 
report form or to the questions posed by an 
enumerator can introduce error into the census data. 
Steps were taken in the design and execution of the 
census of agriculture to reduce errors from 
respondent reporting. Poor instructions and 
ambiguous definitions lead to misreporting. 
Respondents may not remember accurately, may 
give rounded numbers, or may record an item in the 
wrong cell. To reduce reporting and recording errors, 
the report form was tested prior to the census using 
industry accepted cognitive testing procedures. 
Detailed instructions for completing the report form 
were provided to each respondent. Questions were 
phrased as clearly as possible based on previous tests 
of the report form. Computer-assisted telephone 
interviewing software included immediate integrity 
checks of recorded responses so suspect data could 
be verified or corrected. In addition, each 
respondent’s answers were checked for completeness 
and consistency by the complex edit and imputation 
system. 
Processing Error 
Processing of each census report form was another 
potential source of nonsampling error. All mail 
APPENDIX A A- 17 
