driven algorithm to define strata. Certain American 
Indian farms were treated as a separate group, 
effectively having their own donor pool. 
In response to each donor request issued by the edit, 
a dedicated system process would search the 
appropriate stratum and respond with the most 
similar donor, while giving preference to more 
recent donors. In relatively rare instances where it 
was unable to provide a donor, the donor selection 
process issued an appropriate failure message to the 
edit. Imputation failures occurred for several 
different reasons. The requirement that an imputed 
value be positive could have ruled out all available 
donors, as could have the necessity for the donor 
record to satisfy a particular constraint - say, that the 
donor record has cattle, but no milk cows. In 
general, an imputation failure occurred if there was 
no satisfactory donor in the same profile as the report 
being edited. Records with imputation failures were 
either held until more records were available in the 
donor pool or referred to an analyst. In addition, 
when such a failure occurred in finding a donor for 
expenditure data, a program provided values from a 
table of donor pool averages in lieu of values from 
an individual donor, wherever possible. This 
‘failover’ utility was new for the 2012 census 
imputation process, and significantly reduced the 
number of imputation failures among the 
expenditure and labor variables. During the early 
stages of editing, records requiring imputation for 
production (and hence yields) of field crops or hay, 
land values, or certain expenditure variables were set 
aside or “parked.” These records were edited when 
the donor pools contained only 2012 records, 
ensuring that 2012 data were used in imputations for 
these variables. 
After receiving a donor's data, the edit substituted 
the values into the edited record. In many cases, the 
donor record's data value was scaled using another 
data field specified in the edit logic. In such cases, 
the size of the auxiliary field's value in the edited 
record, relative to its value in the donor record, was 
used to inflate or reduce the donor record's value for 
the imputed field. The imputed data were then 
validated by the same edit logic to which reported 
data were subject. Since imputation was conducted 
independently for each occurrence, reports requiring 
multiple imputations may have drawn from multiple 
donors. 
2012 Census of Agriculture 
USDA, National Agricultural Statistics Service 
Data Analysis 
The complex edit ensured the full internal 
consistency of the record. Successfully completing 
the edit did not provide insight as to whether the 
report was reasonable compared to other reports in 
the county. Analysts were provided an additional set 
of tools, in the form of listings and graphs, to review 
record-level data across farms. These examinations 
revealed extreme outliers, large and small, or unique 
data distribution patterns that were possibly a result 
of reporting, recording, or handling errors. Potential 
problems were researched and, when necessary, 
corrections were made and the record interactively 
edited again. 
When NASS summarizes the census of agriculture, it 
assigns the data from an individual report to the 
“principal” county. The principal county is based on 
the operator’s response to a census question and is 
the one county in which the majority of agricultural 
products are produced. Because some large 
operations have significant production in multiple 
counties, some reports were broken up into multiple 
source counties, to more accurately allocate the data. 
Similarly, large farms operating in more than one 
State were treated as distinct, state-specific 
operations. A separate report form was completed 
for each county or State and a separate record was 
added. 
ACCOUNTING FOR UNDERCOVERAGE, 
NONRESPONSE, AND MISCLASSIFICATION 
Although much effort was expended making the 
CML as complete as possible, the CML did not 
include all U.S. farms, resulting in list 
undercoverage. Some farm operators who were on 
the CML did not respond to the census, despite 
numerous attempts to contact them. In addition, 
although each operation was classified as a farm or a 
nonfarm based on the responses to the census report 
form, some were misclassified; that is, some 
nonfarms were classified as farms and some farms 
were classified as nonfarms. NASS’s goal was to 
produce agricultural census totals for publication that 
were fully adjusted for list undercoverage, 
nonresponse and misclassification at the county 
level. 
APPENDIX A A- 9 
