procedures could have over-adjusted or under- 
adjusted for commodity production. To address this, 
a second set of variables, known as commodity 
targets, was added to the calibration algorithm. 
These targets were commodity totals from 
administrative sources or from NASS surveys of 
nonfarm populations (e.g. USDA Farm Service 
Agency program data, Agricultural Marketing 
Service market orders, livestock slaughter data, 
cotton ginning data). The introduction of these 
commodity coverage targets strengthened the overall 
adjustment procedure by ensuring that major 
commodity totals remained within reasonable 
bounds of established benchmarks. Commodity 
coverage targets with acceptable ranges were 
established by subject-matter experts for each State, 
with New England treated as a State. 
Each State was calibrated separately. The calibration 
algorithm addressed commodity coverage. The 
algorithm was controlled by the 65 State farm 
operation coverage targets and the State commodity 
coverage targets. To ensure that the calibration 
process converged with so many constraints, it was 
desirable to provide some tolerance ranges for each 
target. Although full calibration to a single point 
estimate would assure that the weighted total among 
census respondents equaled its target for each 
calibration variable in either set, it was not always 
possible to calibrate to such a large number of target 
values while ensuring that farm weights were within 
a reasonable range and not less than one. Because of 
this and because calibration targets are estimates 
themselves subject to uncertainty, NASS allowed 
some tolerance in the determination of the adjusted 
weights. Rather than forcing the total for each 
calibration variable computed using the adjusted 
weights to equal a specific amount, NASS allowed 
the estimated total to fall within a tolerance range. 
This tolerance strategy made it possible for the 
calibration algorithm to produce a set of satisfactory, 
adjusted weights. 
Ranges for the farm operation coverage targets were 
determined differently from the commodity targets. 
The State target for number of farms had no 
tolerance range. The tolerance range for the 64 other 
State farm operation coverage targets was the 
estimated smoothed State total for the variable plus 
or minus one-half of the standard error of the 
capture-recapture estimate. This choice limited the 
2012 Census of Agriculture 
USDA, National Agricultural Statistics Service 
cumulative deviation from the estimated total for a 
variable when State totals were summed to a U.S. 
level total. The commodity target tolerance ranges 
were determined by subject-matter experts, based on 
the amount of confidence in the source, and usually 
were less than plus or minus two percent of the 
target. Ranges were not necessarily symmetric 
around the target value. 
Census data collection was assumed to be complete 
for very large and unique farms with their weight 
being controlled to 1 during the calibration 
adjustment process. For all other farms, adjustment 
weights were obtained using truncated linear 
calibration which forced the final census record 
weights to fall in the interval [1,6]. Adjustments 
began with the nonresponse and misclassification 
adjusted weights. Through calibration, a second 
stage weight that simultaneously satisfied all farm 
operation coverage and commodity coverage 
calibration targets was obtained. Calibration was 
seldom able to adjust weights so that all State targets 
were met. Within the calibration process, the highest 
priority for meeting a target was given to the number 
of farms, total land in farms, and top cash-receipt 
commodities accounting for 80 percent of the State’s 
production. All remaining targets associated with 
commodities and characteristics of farms and farm 
operators had equal priority. If a value within the 
tolerance range of any variable could not be 
achieved in a given State, the variable was removed 
as a target in that State and the calibration algorithm 
was rerun. 
Weight computations in the final algorithms were 
performed to several decimals. Thus, the fully- 
adjusted weights were non-integer numbers. To 
ensure that all subdomains for which NASS 
publishes summed to their grand total, fully-adjusted 
weights were integerized. This eliminated the need 
for rounding individual cell values and ensured that 
marginal totals always added correctly to the grand 
total. As an example of how the integerization 
process worked, assume there were five census 
records in a county with final noninteger coverage 
weights of 2.2, for a total of 1 1 . The integerization 
process randomly selected four of these records and 
rounded their final weight down to 2.0 and rounded 
the fifth record up to 3.0, for a total of 1 1. 
The proportions of selected census data items that 
APPENDIX A A- 13 
