Nn

EY./X,
= SLOPE
vi

/

/

The approximate confidence interval for R, (assuming R, 1s normally
distributed) is

Ry + 2NV'(R,) ,
where z_

(3)

is the value of the standard normal deviate corresponding to

the appropriate probability a.

The assumption of normality is reason-

ably good_if the coefficients of variation of both ¥ and x, sy/(/n Y)
and 8y/(/n X), respectively, are less than 0.1 and the number of obser-

vations is greater than 30 (Cochran, 1963, p. 164).

If these conditions

do not apply, V'(R,) (equation 2) will tend to be ton small.
Cochran
(1963, p. 164) gives an alternative method for computing confidence
limits of Ry that should be more accurate than the above method
(equation 3), but is also more difficult to compute. When the sample
size is small, the t distribution rather than the standard normal is
recommended for calculating the confidence interval.

When the straight line is restricted to pass through the origin, as in
our case to estimate an average ratio, and both X and Y¥ are normally
distributed, then the maximum likelihood estimate of the slope (common

X
FIGURE 5.

ratio} is also Ry = Y/X (Creasy, 1956).

Ratio Estimation When Var(¥) is Proportional to X?,

Recall that these estimators are based on the assumption that X is
known with little error.
In that case, each estimate (R,, R,, R)
is statistically unbiased and the choice between them depends on
whether the variance of Y is constant, proportional to X, or
proportional to X* (Snedecor and Cochran, 1967 pp. 166-170).
But
what 1f X is not "constant", the usual case in transuranic field

studies?

For example, Y and X might be 229-2"0py and 24! 4m con-

centrations, respectively, in the same aliquot of soil.
Cochran
(1963, p. 161) shows that the ratio estimate R, = Y/X, even though
Y and KX are both random variables, is an unbiased estimate of the
ratio of the true means if the regression of Y on X is a straight
line through the origin.Hence, regardless of whether X is constant
or variable, if the data forms a straight line through the origin,
Y/X is an unbiased estimate of

the ratio of true means.

However,

when X is a variable, the vartance of R, is only approximate and
equation (1) may not be appropriate for estimating Var(R,).
Cochran's (1963, p. 31) approximate variance for R, is

VICR,) =

DT CEH R)X,)?/n(n-1) x?
i=l
n

—

.

The variance of the observa-

tions from the regression line (line with slope Rj) was approximated
by Creasy (1956) as
n

(¥,-Y)? -

n

n

Y

¥°x

x

n-1

n

.

(X,-X) ¥
(4).

The variance of Ro for this case has not been completely derived, but

Ricker (1973) recommends

n

V"(Ry) = ex/ (X,-%)?
i=l

(5)

(where 32.
is computed using equation 4) as being as good an approxi~
mation a’ She computationally involved method proposed by Creasy.
The
confidence interval is then given by

(2)
R2 + t vv" (Ro)

608
609

Select target paragraph3