188 
accommodated by the use of an area 
sampling frame derived from map and 
photographic material (Cotter and 
Nealon, 1987). The entire land surface 
area of states are divided into strata 
related to land cover. Multistage 
sampling is then employed to select 
fields. After a selection of fields is 
made, two small plots are randomly 
located within a field. Different plot 
sizes and configurations are used for 
different crops. For example, corn has 
two rows 15 feet in length, whereas, 
wheat has three rows 21.6 inches in 
length (Matthews, 1985). Once sample 
plots have been located, measurements 
are made monthly during the growing 
season to forecast yield. 
The observed variables used to forecast 
yield change with stage of crop 
growth. At the end of the season, when 
biological maturity has occurred, just 
prior to farmer harvest, the surveyor 
harvests the unit to determine gross 
yield. Later, after farmer harvest, 
new plots are randomly located for 
harvest loss determination to obtain 
net yields. 
The survey system is designed to 
produce forecasts and estimates of a 
specified precision at state and 
regional levels. For example, the 1990 
plan calls for 1,540 soybean samples in 
the eleven major soybean states. The 
total area for the eleven states to be 
counted and eventually harvested is 
only 1.5 acres. This small area will 
however provide a yield estimate judged 
to be of sufficient precision by NASS. 
Nearly all the models are "component" 
models. With wheat, for example, the 
number of heads and weight per head are 
forecast separately. Linear regression 
models associate a history of five year 
relationships between the pre-harvest 
measured variable and the resultant 
value at harvest. Then a current year 
observation is made and the model 
provides a forecast result. Large 
numbers of these simple models are used 
during the growing season. They are 
generated by state and stage of growth 
called maturity category. For example, 
wheat is divided into seven maturity 
stages or categories. These are: 
preflag, flag or early boot, late boot 
or flower, milk, soft dough, hard dough 
and ripe. Early in the season for the 
preflag stage only number of stalks are 
counted and used to forecast number of 
heads. Head weight is a historical 
average. The same is true for the flag 
stage. For flowering, a count of 
emerged heads plus heads in late boot 
are used to forecast the total number 
of heads. A count of so called 
"fertile" spikelets per head and a 
historical average are used to forecast 
weight per head. In the milk and early 
dough stage, again, emerged heads plus 
heads in late boot are used to forecast 
total number of heads. Weight per head 
uses total grains per head and actual 
weight of immature heads. For the last 
two, hard dough and ripe, stages there 
is no forecast. For the number of 
heads, actual counts are used, and for 
weight per head, heads are 
micro-threshed, the grain weighted and 
adjusted to a standard moisture. 
The actual number of models is quite 
substantial. Wheat would be in several 
maturity categories in a state during a 
given month, so there would be a model 
for each category. The process varies 
somewhat by crop but the wheat example 
illustrates the general case. 
Forecasts for winter wheat are made 
from May 1 through harvest. 
In this program, particular attention 
is given to standardized training of 
the field survey staff. The program is 
nationally directed so that all 
procedures are conducted in the same 
way in each state. The objectivity of 
data collection can certainly stand up 
to scrutiny. 
However, there is a major flaw in the 
system. A number of validation surveys 
have shown that the process is biased. 
Also, no corrective procedures have 
been found to eliminate this bias. 
Therefore, the Agricultural Statistics 
Board (ASB) of NASS systematically 
adjusts survey results using a ten-year 
average bias. One paper has recently 
shown, mathematically, that the 
forecast system is indeed biased 
(Birkett, 1990). 
Starting during the 1990 forecast 
season, new objective yield regression 
models for corn and soybeans will be 
used. The new models will use the same 
input as the current models but they 
should deliver results of improved 
accuracy. The present system operates 
at the plot level. Measurements from 
each sample plot enter the appropriate 
model and a forecast is generated. 
Then, forecasts of each plot are 
averaged over a state (assuming a 
self-weighting sample). 
The new approach will operate at the 
state level and, in addition, new 
regional level models will be used. 
Individual state models will be 
constrained to produce forecasts that 
are consistent with the regional 
models. The regions are comprised of 
all states in a particular objective 
yield program. For corn, this would be 
ten states. In general, each monthly 
model has one independent variable per 
component and the model form is simple 
linear regression. The dependent 
variable is the ASB yield and the 
independent variables, as stated 
earlier, are the same as the present 
program. The coefficients for the 
simple model are derived from 
historical data. In other words,