undercoverage, nonresponse, and misclassification 
as well as the total percent adjustment for selected 
items are displayed in Tables A and C. 
MEASURED ERRORS IN THE CENSUS 
PROCESS 
Although the census of agriculture does not 
inherently rely on a sample, it uses statistical 
procedures in compiling the CML, in its data 
collection procedures, in data editing and processing, 
and in compiling the final data. Additionally, it uses 
statistical procedures to both measure errors in the 
various processes and in making adjustments for 
those errors in the final data. One example is the 
statistical process used to account for undercoverage, 
nonresponse of farms on the CML, and 
misclassification of responses to the census. The 
basis of the undercoverage adjustment is the capture- 
recapture procedure that uses the area sample 
enumeration from the June Agricultural Survey. The 
largest contribution to error in the census estimates is 
due to the adjustments for nonresponse, 
undercoverage, misclassification, calibration and 
integerization. 
Variability in Census Estimates due to 
Statistical Adjustment 
In conducting the 2012 Census of Agriculture, 
efforts were initiated to measure error associated 
with the adjustments for farm operations that were 
not on the CML, for farm operations that were on the 
CML but did not respond to the census report form , 
for farms and nonfarms that were misclassified as 
nonfarms and farms, respectively, for calibration, 
and for integerization. These error measurements 
were developed from the standard error of the 
estimates at the national, State, and county levels and 
were expressed as coefficients of variation (CVs) at 
the national and State levels and as generalized 
coefficients of variation (GCVs) at the county levels. 
The standard error of an estimate is an estimate of 
the standard deviation of the sampling distribution of 
the estimator. Because Texas and Alaska were 
modeled separately from the other States, the 
variances of a national-level data item for these two 
States were computed separately and added to the 
A- 16 APPENDIX A 
variance of that data item for the rest of the U.S. The 
standard error was then the square root of the total 
variance. In each case, standard errors were 
computed using the group jackknife approach. To 
conduct the jackknifing, k mutually exclusive and 
exhaustive groups of JAS segments were formed. 
The groups were selected using a stratified random 
design so that each group reflected the survey 
design, including State and agricultural strata within 
a State. In turn, each group, j — 1, 2, ..., k, was 
deleted and the capture-recapture estimate CRP was 
computed for each data item i at the specified 
geographical level, such as nation, State, or county, 
using the remaining (k— 1) groups. Estimates of the 
variance and standard error associated with the 
capture-recapture estimate CR t are then, respectively, 
S ( CR ' j) - CR i ) 2 - SE(CR > ) = V? 
k j= \ 
Increasing k improves the estimate of the variance 
but, as k increases, the observations become too 
sparse to reflect the survey design and to provide 
country- wide coverage. Based on 2007 data, k - 10 
was determined to be the largest number of groups 
that could be formed and still have each group 
provide adequate coverage within all States and 
agricultural strata. Thus, 10 jackknife groups were 
used to provide standard errors for 2012 State and 
national estimates. To capture the additional 
variability from calibration and integerization, the 
standard errors were computed using the calibrated, 
integerized capture-recapture estimates from the 
jackknife groups. For the estimate of the number of 
farms with a given set of characteristics, only the 
CML records with those characteristics were used to 
obtain the overall estimate as well as the estimates 
from each jackknife group. 
When the constraints of the calibration process 
produced an artificially small standard error, the 
more conservative capture-recapture standard error 
was used. Note that the jackknife groups must only 
be constructed once, and different subsets of the 
records were used to compute estimates and standard 
errors for the data items. 
The CV is a measure of the relative amount of error 
2012 Census of Agriculture 
USDA, National Agricultural Statistics Service 
