26.12.2014 Views

Clinical Biochemistry of Domestic Animals (Sixth Edition) - UMK ...

Clinical Biochemistry of Domestic Animals (Sixth Edition) - UMK ...

Clinical Biochemistry of Domestic Animals (Sixth Edition) - UMK ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

24<br />

Chapter | 1 Concepts <strong>of</strong> Normality in <strong>Clinical</strong> <strong>Biochemistry</strong><br />

differing significantly from 1 would indicate that MS d/a<br />

was estimating something additional to σ e<br />

2<br />

, namely, n σ d/a 2 ,<br />

where n is the number <strong>of</strong> replications per sample. Based<br />

on the assumed data, the hypothesis σ d/a<br />

2<br />

0 is rejected.<br />

σ d/a<br />

2<br />

in this design is the true intradog variability for this<br />

response. The test result that the day effect is significant<br />

means that this intradog variability is larger than zero.<br />

The second test is that the variance component for<br />

“ animals ” is equal to zero or σ a 2 0. It can be determined<br />

by considering the expected mean squares <strong>of</strong> Table 1-14<br />

that the test for no significant variability among animals<br />

( H o : σ a<br />

2<br />

0) is made using MS d/a in the denominator <strong>of</strong> the<br />

F -statistic. This test is also significant. This means that σ a<br />

2<br />

is<br />

not equal to zero, indicating a significant source <strong>of</strong> variability<br />

among dogs in the plasma cortisol levels recorded in the<br />

experiment.<br />

1 . Estimating Variance Components<br />

Once significance has been established for one or more <strong>of</strong><br />

the variance components in an experimental design, interest<br />

focuses on estimating the variance component(s). Estimates<br />

are readily obtainable using the appropriate expected mean<br />

squares in conjunction with the mean squares obtained<br />

from the data. For example, Table 1-7 shows that an estimate<br />

<strong>of</strong> σ e<br />

2<br />

is MS e (σˆ e<br />

2<br />

0.0096). An estimate <strong>of</strong> σ d/a<br />

2<br />

can<br />

be obtained by noting that MS d/a estimates σ e<br />

2<br />

n σ d/a<br />

2<br />

;<br />

solving for σ d/a 2 followed by substitution <strong>of</strong> MS e for σ e<br />

2<br />

yields ( MS d/a MS e )/n as the estimate. Based on the data<br />

used in the example, σˆ d/a<br />

2<br />

(1.0958 0.0096)/2 0.5431.<br />

An estimate <strong>of</strong> σ a<br />

2<br />

, obtained in a similar manner, is<br />

( MS MS )/ nn<br />

where n d is the number <strong>of</strong> days samples were taken (3).<br />

Based on the data an estimate <strong>of</strong> σ a<br />

2<br />

is 0.4387. The term<br />

σ a<br />

2<br />

is the estimate <strong>of</strong> the interanimal variability, whereas<br />

σ d/a<br />

2<br />

estimates the intra-animal variability in plasma cortisol<br />

level for the underlying population <strong>of</strong> dogs. Interval estimates<br />

for these variance components can be obtained using<br />

these point estimates by methods described elsewhere<br />

( Harter and Lum, 1955 ; Mickey et al. , 2004 ; Satterthwaite,<br />

1941).<br />

d/a<br />

2 . Estimating the Variance <strong>of</strong> the Grand<br />

Mean Response<br />

Another way to visualize the importance <strong>of</strong> these variance<br />

components is to analyze their impact on the estimate<br />

<strong>of</strong> the variance <strong>of</strong> interest. In some applications, there<br />

would be interest in estimating the grand mean ( μ ) <strong>of</strong> the<br />

response. In the present example, this would involve estimating<br />

the mean plasma cortisol level taking into account<br />

any random animal effect and day effect. The variance <strong>of</strong><br />

μˆ , Var( μˆ ), is given as σ a<br />

2<br />

/ n a σ d/a<br />

2<br />

/ n a n d σ e<br />

2<br />

/ n a n d n ( Little<br />

d<br />

et al. , 1991 ; Neter et al. , 1996 ). In the present example,<br />

Var( μˆ ) is estimated as 0.4387/20 0.5431/60 0.0096/<br />

120 0.0219 0.0091 0.0001 0.0311, and by far<br />

the greatest contribution to this variance is that due to the<br />

variability among animals in their response. The intradog<br />

variance component, although slightly larger than the interdog<br />

variance component, makes a considerably smaller<br />

impact on Var( μˆ ).<br />

3 . Estimating the Total Variability <strong>of</strong> a<br />

Single Response<br />

In other applications, interest centers about the total variability<br />

( σ total 2 ) associated with a single response. A single<br />

response is a linear combination <strong>of</strong> the terms in the response<br />

model and, using the assumption <strong>of</strong> the independence <strong>of</strong><br />

terms in the model, has a variance equal simply to the sum <strong>of</strong><br />

the variance components. Specifically, σ total 2 σ a<br />

2<br />

σ d/a<br />

2<br />

<br />

σ e<br />

2<br />

( Kringle, 1994 ), which in this example is estimated as<br />

0.4387 0.5431 0.0096 0.9914. Here the total variability<br />

<strong>of</strong> a single response is divided nearly equally between<br />

“animals ” and “days nested within animals. ”<br />

Other possible designs could be considered. What has<br />

been demonstrated is that the method <strong>of</strong> analysis <strong>of</strong> variance<br />

in conjunction with experimental design can be useful<br />

in answering a variety <strong>of</strong> questions.<br />

Nested designs are frequently used to assess sources <strong>of</strong><br />

variability in an assay. For example, several laboratories<br />

could be involved in doing a particular assay, with several<br />

autoanalyzers in each laboratory and multiple technicians<br />

running these autoanalyzers. Inference in this context centers<br />

around being able to identify if there are significant<br />

sources <strong>of</strong> variation among the laboratories, among autoanalyzers<br />

within a given laboratory, and among technicians<br />

operating a given autoanalyzer. The goal <strong>of</strong> analyses <strong>of</strong><br />

this sort is to identify large sources <strong>of</strong> variability. Once the<br />

larger sources <strong>of</strong> variability have been identified, changes<br />

are made in the system in an effort to reduce the variability<br />

associated with each source. The long-term objective<br />

is to have an assay with sources <strong>of</strong> variability that are as<br />

small as possible. <strong>Clinical</strong> analysts conventionally divide<br />

the square root <strong>of</strong> the estimates <strong>of</strong> the variance components<br />

(the sample standard deviations) resulting from such<br />

assay experiments by the grand mean to obtain coefficients<br />

<strong>of</strong> variation for each source <strong>of</strong> variability ( Kringle, 1994 ).<br />

These coefficients <strong>of</strong> variability should be much smaller<br />

than those derived as intra-animal and interanimal variability.<br />

Interested readers are strongly encouraged to consult<br />

texts written on experimental design and ANOVA ( Mickey<br />

et al. , 2004 ; Neter et al. , 1996 ).<br />

REFERENCES<br />

Björkhem , I. , Bergman , A. , and Falk , O. ( 1981 ). Clin. Chem. 27 ,<br />

733 − 735 .

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!