Calibration for the Census of Agriculture

0226-AgCensus-Calibration for the Census of Agriculture.pdf

2017 Census of Agriculture

Calibration for the Census of Agriculture

OMB: 0535-0226

Document [pdf]
Download: pdf | pdf
JSM2015 - Survey Research Methods Section

Calibration for the Census of Agriculture
Andrea Lamas1, Kelly Toppin1, Matthew Williams2
Linda J Young1, Cliff Spiegelman3

National Agricultural Statistics Service, USDA, 1400 Independence Ave, SW,
Washington, DC 20250
2
Department of Health and Human Services, 1 Choke Cherry Road, Rockville, MD
20857
3
Texas A&M University, College Station, TX 77843
1

Abstract

The National Agricultural Statistics Service (NASS) conducts a Census of Agriculture
every 5 years, in years ending in 2 and 7. For the 2012 Census of Agriculture, NASS used
capture-recapture methods to adjust the Census for under-coverage, non-response, and
misclassification of farms/non-farms. After these adjustments, the weights were calibrated
and integerized. Calibration was conducted to ensure that state and national totals were
unbiased for variables where administrative data were available. The integerization process
rounded weights but did not change marginal totals. NASS researched alternative
calibration methods applied to the Census. Here the constraints and limitations of those
methods are discussed.

Key Words: Capture-recapture, Calibration, Non-response, Under-coverage

1. Introduction
The USDA’s National Agricultural Statistics Service (NASS) conducts hundreds of
surveys and prepares reports that cover every aspect of U.S. agriculture. The majority of
the reports provide estimates that impact U.S. markets and price of commodities. Some
examples of these include corn, soybeans, wheat and upland cotton. The largest survey that
NASS conducts is its Census of Agriculture. The census provides information on
characteristics of U.S. farms and ranches and people who operate them. It is used by
federal, state and local governments and others who provide services to farms and rural
communities. Its estimates are produced at the national, state and county levels. The
estimates impact community planning, availability of operational loans and other funding,
location and staffing of service centers, and farm programs and policies.
Estimates produced by the Census of Agriculture are adjusted in two ways. First the
estimates are adjusted through capture-recapture. Second, the estimates are adjusted
through calibration. This ensures census estimates are consistent with available
information on commodity production. However, through the current calibration
methodology, all targets are rarely met. Other issues also arise through this process. The
purpose of this work is to discuss the constraints and limitations of the current calibration
methodology and to propose an alternative methodology.

1803

JSM2015 - Survey Research Methods Section

2. Census of Agriculture
NASS conducts the Census of Agriculture every five years (years ending in 2 and 7). The
census is a count of U.S. farms and ranches and the people who operate them. As
established by Congress in 1974, a farm is any place from which $1,000 or more of
agricultural products were produced and sold or normally would have been sold during the
year. During the census, data are collected on land use and ownership, operator
characteristics, production practices, income and expenditures, and numerous other
characteristics. The census provides the most uniform, comprehensive agricultural data for
every county in the nation. It is a list-based survey; the Census Mail List (CML) is the list
of all operations mailed a census questionnaire.

2.1 Census Estimates

Several sources of error are known to exist on the census of agriculture. The CML contains
agricultural operations that are farms and agricultural operations that are non-farms. Some
farming operations are not on the CML, due to incompleteness of the list. Because of this,
there is list under-coverage on the census. Also, not all agricultural operations on the CML
respond, resulting in non-response. Lastly, misclassification occurs on the census due to
errors in census reporting. This occurs when, based on their response to the census
questionnaire, some non-farms are classified as farms, or when some farms are classified
as non-farms. To adjust for these sources of error, the census estimates are adjusted through
capture-recapture.
NASS also obtains administrative data on commodity production. After the census
estimates are adjusted through capture-recapture, they are then calibrated to ensure the
estimate are consistent with the administrative data.

2.2 June Agricultural Survey

To adjust for errors due to under-coverage, non-response, and misclassification, using
capture-recapture, two independent surveys are required. The census of agriculture is the
first survey and the June Agricultural Survey (JAS) is used as the second survey. The June
Agricultural Survey (JAS) has an area frame and is conducted annually. It collects
information on U.S. crops, livestock, grain storage capacity and type and size of farms.
Because the distribution of crops and livestock can vary widely across a state in the U.S.,
land is divided, in preparation for sampling, into homogeneous groups or strata, such as
intensively cultivated land, urban areas and range land. The general strata definitions are
similar from state to state; however, minor definitional adjustments may be made
depending on the specific needs of a state. Each land-use stratum is further divided into
substrata by grouping areas that are agriculturally similar. This yields greater precision for
state-level estimates of individual commodities. Within each substratum, the land is
divided into primary sampling units (PSUs). A sample of PSUs is selected and smaller,
similar-sized segments of land are delineated within these selected PSUs. Finally, one
segment is randomly selected from each selected PSU to be fully enumerated. Through inperson canvassing, field interviewers divide all of the land in the selected segments into
tracts, where each tract represents a unique land-operating arrangement. Each tract is
screened and classified as agricultural or non-agricultural. Non-agricultural tracts belong
to one of three categories: (1) non-agricultural with potential, (2) non-agricultural with
unknown potential, or (3) non-agricultural with no potential. A tract is considered
agricultural if it has qualifying agricultural activity either inside or outside the segment.
Otherwise, it is non-agricultural. An agricultural tract will subsequently be classified as a

1804

JSM2015 - Survey Research Methods Section

farm if its entire operation (land operated both inside and outside the segment) qualifies
with at least $1,000 in sales or potential sales. All non-agricultural tracts and agricultural
tracts with less than $1,000 in sales are classified as non-farms.

2.3 Capture-Recapture Weight

For a farm to be captured by the census, the farm must first be on the CML, respond to the
census, and be classified as a farm based on the response to the questionnaire. Therefore,
the probability of capture is
πœ‹πΆ = πœ‹(𝐢𝑀𝐿, π‘Ÿπ‘’π‘ π‘π‘œπ‘›π‘‘π‘’π‘‘, π‘“π‘Žπ‘Ÿπ‘š π‘œπ‘› 𝑐𝑒𝑛𝑠𝑒𝑠|π‘“π‘Žπ‘Ÿπ‘š)
= πœ‹(𝐢𝑀𝐿|πΉπ‘Žπ‘Ÿπ‘š) πœ‹(π‘…π‘’π‘ π‘π‘œπ‘›π‘‘π‘’π‘‘|𝐢𝑀𝐿, πΉπ‘Žπ‘Ÿπ‘š) πœ‹(𝐢𝑒𝑛𝑠𝑒𝑠 πΉπ‘Žπ‘Ÿπ‘š|𝐢𝑀𝐿, π‘…π‘’π‘ π‘π‘œπ‘›π‘‘π‘’π‘‘, πΉπ‘Žπ‘Ÿπ‘š)
where
c = capture.
This probability of capture accounts for under-coverage, non-response and
misclassification of non-farms as farms. However, the misclassification of farms as nonfarms, or correct census farm classification is not included in the probability of capture.
Therefore, the probability of correct census farm classification is
where

πœ‹πΆπΆπΉπΆ = πœ‹(πΉπ‘Žπ‘Ÿπ‘š|πΉπ‘Žπ‘Ÿπ‘š π‘œπ‘› 𝑐𝑒𝑛𝑠𝑒𝑠)
CCFC = correct census farm classification.

A matched dataset of CML and JAS records is created and logistic regression models are
developed for each probability. Therefore, the capture-recapture weight is:
πœ‹Μ‚πΆπΆπΉπΆ
πœ‹Μ‚πΆ
and the capture-recapture estimate is
𝐢𝑅 = βˆ‘

where

π‘–βˆˆπΉ

πœ‹Μ‚πΆπΆπΉπΆπ‘–
πœ‹Μ‚πΆπ‘–

F = set of all CML records classified as a farm based on their responses to the
Census questionnaire.

3. Calibration
3.1 Census Calibration

After the capture-recapture adjustment, calibration is conducted to ensure that census
estimates are consistent with administrative data on commodity production. NASS obtains
information on most commodities from administrative sources or from NASS surveys of
non-farm populations. Some examples are USDA Farm Service Agency program data,
Agriculture Marketing Service market orders, livestock slaughter data and cotton ginning
data.
The targets used in the census calibration are the commodity administrative data and 65 of
the capture-recapture estimates. The 65 capture-recapture estimates are estimated for each
state. They are the number of farms, land in farms, and the number of farms by the

1805

JSM2015 - Survey Research Methods Section

following characteristics: 8 categories of value of agricultural sales, age of farm operator,
female operators, race of farm operator, Hispanic origin of the principal farm operator, 10
major commodities by their 4 sales categories and 7 farm type categories.

3.2 Current Calibration Methodology

The current calibration methodology is truncated linear calibration with weights between
1 and 6 (Fetter, 2009). Each state is calibrated separately. High priority targets are
calibrated first and are treated as hard (fixed) targets. Within the set of priority targets, the
target having its estimate furthest from the target value is included in calibration first. Once
that target is hit, the next target with the estimate furthest from the target value is included.
If a target cannot be hit, it is removed from the list for targets, and the next target with the
estimate furthest from the target value is included. Once this process is complete with the
high priority targets, a stepwise algorithm is used to calibrate the remaining. In the stepwise
algorithm, all variables are treated with equal priority and as hard targets. However, once
a target has been entered and has been hit, it is then treated as a soft target (within an
interval) as other variables are entered in the stepwise algorithm. Each soft target is
calibrated within a pre-specified tolerance range (generally less than 2% of the target).
Output weights from calibration are to several decimals but census results are published at
the integer level. Therefore weights are integerized (or rounded) to ensure all tables and
breakdowns are summed to correct totals.
Not all records are treated similarly during calibration. For large and unique farms, census
data collection was assumed to be complete. Weights are controlled to be one during the
calibration adjustment process for these records. Specialty operations have weight
restriction of the interval [1, 3]. For all other farms, calibration adjustments begin with the
capture-recapture adjusted weights but are truncated to [1, 6].
With this methodology, all targets are rarely met through calibration. The fact that all
variables are treated as hard targets when being entered into the stepwise algorithm
constrains feasible solutions. Also, after calibration, all estimates are rounded to integers
in a manner that preserves farm totals. This rounding for large farm producers becomes
problematic.

3.3 Proposed Calibration Methodology

A new calibration methodology was developed. It is similar to the current methodology in
that it first hits high priority targets as hard targets. However, it then includes all targets at
once as soft targets using the LASSO (least absolute shrinkage and selection operator)
penalty. Also, in the proposed calibration methodology, the weight restriction scheme and
truncation of DSE weights input into calibration were evaluated. The LASSO methodology
was run with and without weight restrictions and with DSE weights input and output from
calibration between [1, 6] or [0.9, 6]. The calibration methodologies, weight restrictions
schemes and weight truncations were compared based on the number of targets missed.
Results for Michigan, North Carolina, and Texas were obtained using 2012 Census of
Agriculture data.

4. Results and Conclusions
For all states (Michigan, North Carolina and Texas), four variations of calibration were
compared. The first is the current methodology used by NASS, where truncated linear
methodology is used, no weight restriction changes were made, and DSE input and output

1806

JSM2015 - Survey Research Methods Section

weights were between [1, 6]. The second approach uses the proposed LASSO
methodology, however keeps all weight restrictions and truncations the same. The third
again uses the LASSO and changes the weight restriction scheme. Instead of large and
unique farms whose census data collection was assumed to be complete having a weight
of one, only records who do not have a non-response, under-coverage and misclassification
adjustment receive a weight of one. Lastly, the fourth approach uses the LASSO, does not
change the weight restriction scheme. However, DSE input weights to calibration are
changed from [1, 6] to [0.9 to 6]. Output weights of calibration are also changed from [1,
6] to [0.9 to 6]. These changes allow the input weights to calibration and output weights
from calibration to be less than 1.
The results for Michigan, which has 175 targets, are in Table 1 below. The truncated
linear methodology missed 9 targets. Alternatively, the LASSO methodology missed 6
targets. Changing the weight restriction scheme (only records who do not have a nonresponse, under-coverage and misclassification adjustment receive a weight of one) and
using the LASSO methodology, still missed 6 targets. Using LASSO and allowing the
DSE input weights to calibration and output weights from calibration to range from [0.9,
6] reduced the number of missed targets to 4.
Calibration
Methodology
Truncated
Linear
LASSO
LASSO
LASSO

Table 1: Michigan Calibration Results
Weight
DSE Input
Output Weights Targets
Restriction Weight to
from Calibration Missed
Change
Calibration
No
[1,6]
[1,6]
9
No
Yes
No

[1,6]
[1,6]
[0.9,6]

[1,6]
[1,6]
[0.9,6]

6
6
4

North Carolina has 184 targets and the results for this state are in Table 2. The truncated
linear methodology missed 4 targets. Alternatively, the LASSO methodology missed 3
targets. Changing the weight restriction scheme (to only records who do not have a nonresponse, under-coverage and misclassification adjustment receive a weight of one) in the
LASSO methodology, reduced the number of missed targets to 1. Using LASSO, and
allowing the DSE input weights to calibration and output weights from calibration to
range from [0.9, 6] resulted in 3 missed targets.
Table 2: North Carolina Calibration Results
Weight
DSE Input
Output
Targets
Restriction Weight to
Weights
Missed
Change
Calibration of Calibration
Truncated
No
[1,6]
[1,6]
4
Linear
LASSO
No
[1,6]
[1,6]
3
LASSO
Yes
[1,6]
[1,6]
1
LASSO
No
[0.9,6]
[0.9,6]
3

Calibration
Methodology

The results for Texas with 346 targets are in Table 3 below. The truncated linear
methodology missed 14 targets. Alternatively, the LASSO methodology missed 12

1807

JSM2015 - Survey Research Methods Section

targets. Changing the weight restriction scheme (only records who do not have a nonresponse, under-coverage and misclassification adjustment receive a weight of one) in the
LASSO methodology, reduced the number of missed targets to 5. Using LASSO and
allowing the DSE input weights to calibration and output weights from calibration to
range from [0.9, 6] resulted in 11 missed targets.
Calibration
Methodology
Truncated
Linear
LASSO
LASSO
LASSO

Table 3: Texas Calibration Results
Weight
DSE Input
Output Weights
Targets
Restriction Weight to
of Calibration
Missed
Change
Calibration
No
[1,6]
[1,6]
14
No
Yes
No

[1,6]
[1,6]
[0.9,6]

[1,6]
[1,6]
[0.9,6]

12
5
11

In conclusion, the proposed code does as well or better than the operational code. Changing
the weight restriction scheme decreases the number of targets missed. Changing the input
and output weights from calibration decreases the number of targets missed, but this
requires changes to the methods used for rounding. Therefore, how best to round the values
is the next research project.

References
Fetter, Matthew J. 2009. An Overview of Coverage Adjustment for the 2007 Census of
Agriculture. Proceedings of the Government Statistics Section, JSM 2009. Pp 32283236.
Singh, A. C., and C. A. Mohl (1996). Understanding Calibration Estimators in Survey
Sampling, Survey Methodology 22, 107–115.

1808


File Typeapplication/pdf
File TitleCalibration for the Census of Agriculture
File Modified2015-11-25
File Created2015-10-02

Β© 2024 OMB.report | Privacy Policy