7rc = 7i(CML, Responded, Farm on Census|Farm)= 
7i( C M L| F arm )7t (Responded) CML, Farm)7t(Farm on 
Census |CML, Responded, Farm) 
The probability of capturing a farm depends on the 
characteristics of the farm. Using five-fold cross- 
validation, three logistic models were developed 
based on the matched dataset. The first model 
estimated the probability of a farm being on the 
CML. The second model estimated the probability 
that a farm on the CML responded to the census 
report form. The final model estimated the 
probability that a farm that was on the CML and 
responded to the census was identified as a farm 
based on its response. The probability that a farm is 
captured by the census of agriculture is then the 
product of the three conditional probabilities that a 
farm is on the CML, responds, and is identified as a 
farm. 
Note 1: Responses were required for Must cases. 
These operations were only included in modeling the 
probability of a farm being on the CML. 
Consequently, the weight associated with a Must 
record was the reciprocal of the probability of a farm 
being on the CML. 
Note 2: Two sets of models were created. One set 
estimated the probability of capture for Texas farms. 
The other set provided estimated capture 
probabilities for farms in the remaining States, 
except for Alaska. 
Note 3: Because Alaska is not included in the JAS 
and thus has no area frame, the Alaskan agricultural 
operations were not included in the capture-recapture 
process. No adjustments were made for 
undercoverage or misclassification. To account for 
nonresponse, the CML records were divided into 
three groups: (1) the Must records, (2) the Criteria 
Records, and (3) the remaining CML records. The 
must records received a weight of one, thereby 
receiving no adjustment for nonresponse. The 
probability of response for each of the other two 
groups was the proportion of responders within the 
group. Each record within the group was then given 
a weight equal to the reciprocal of the probability of 
response. 
A -12 APPENDIX A 
Misclassification 
An operation is misclassified if (1) it meets the 
definition of a farm, but is classified as a nonfarm on 
the census or (2) it does not meet the definition of a 
farm, but is classified as a farm on the census. The 
first type of misclassification is accounted for when 
modeling the probability of capture. An adjustment 
is still needed for the misclassification of nonfarms 
as farms. As with farm status and capture, the 
probability of this misclassification depends on an 
operation’s characteristics. Thus, a final logistic 
model was developed. Given that an operation was 
classified as a farm on the CML, the probability of 
its being a farm was modeled based on its 
characteristics. Five-fold cross-validation was used 
to ensure that the model was not over- fitted. 
CALIBRATION 
Each operation identified as being in-scope on the 
CML was given a weight equal to the probability of 
misclassification divided by the probability of 
capture. This weight accounted for undercoverage, 
nonresponse, and both types of misclassification. 
The record weighting processes were initially 
applied at the State level to produce adjusted 
estimates of farm numbers and land in farms for 63 
different categories of 8 characteristics of the farm 
operation or the farm operator — value of agricultural 
sales (8); age (2); female; race (4); Hispanic origin of 
principal farm operator ; 4 sales categories for each 
of 10 major commodities (40); and farm type groups 
(7). The State-level number of farms and land in 
farms were two additional adjusted estimates, 
resulting in 65 categories. To reduce the intercensal 
variation at the State level, the State targets were 
smoothed by averaging the 2012 estimates from 
capture-recapture and the published 2007 state 
estimates with the restrictions that the smoothed 
targets were within one standard error of the capture- 
recapture estimates. The smoothed State targets were 
rescaled so that they summed to the national capture- 
recapture estimates. 
These State estimates were general purpose in that 
they did not provide any control over expected levels 
of commodity production of the individual farm 
operation. As a result of this limitation, the 
2012 Census of Agriculture 
USDA, National Agricultural Statistics Service 
