These analyses are based on Eq. 1 with a multiplicative main effects model for the variable A, B, S, L, IG, and F. The effect of dose is represented with an exponential relative risk function or an additive excess relative risk (ERR) function. The main effects model with exponential relative risk is expressed as

where , **A** = (age
-52.5)/100, and **D** is external dose in Sv. In Eq. 3 B, **S, L, IG**,
and **F** are factors and **A** and **D** are continuous
variates.
Score [38]
test statistics for external dose are presented for selected cause of
death categories using all ten dose groups. Additional summary
results---parameter estimates, standard errors, and likelihood ratio
test (LRT) statistics---are given for each cause of death category
with the highest dose group deleted (low dose analysis). The score
test for **D** in Eq. 3
is identical [39] to that obtained for
**D** variable in Eq. 4 below, and can be compared to the standard
normal distribution to evaluate the strength of the dose-response
relation. These analyses are then repeated using adjusted doses. The
score test and low dose exponential relative risk are used as a
screening procedure to identify cause-of-death categories that may
show a strong association with dose.

Most summary statistics (estimates and SEs) for relative risk parameters are expressed in log percent (L%) units, i.e. they are given in logarithmic units multiplied by 100---see [13] Chapter 22,[15] and the Appendix. For the ERR estimates likelihood based intervals are given---see [16,17].

More detailed results are presented for several cause of death categories using a main effects model with the additive excess relative risk (ERR) function to describe the dose-response relation for external radiation, e.g.

The main effects model provides an overall descriptive summary of the
the effects of each stratification variables on cause specific
mortality. Thus inclusion of these potential confounding variables
(**A,B,S,L**) and exposure variables (**F** and **IG**) provides a broader context
in which to evaluate the relative importance of the estimated effect
of external radiation. A saturated model for the confounding
variables **AG,B,S,** and **L** was also considered and was found to have
little effect on the dose parameter estimate. Detailed results are
given for all cancer in the Appendix (see Table AVII).
The score statistic for a linear dose term for the
main effects model and the saturated model (i.e. stratified analysis)
were routinely calculated and no important difference for any cause of
death categories were found.

A detailed analysis for all cancer mortality that uses the Akaike Information Criteria (AIC) to contrast the effectiveness of several exponential and ERR models is presented [18,19]. For Poisson data the AIC = Deviance + (number of parameters), combines a measure of the discrepancy between the fitted values and the data (as measured by the deviance) and the simplicity of the model as reflected in the number of parameters. McCullagh and Nelder---see Sec 3.9--- [20] discuss a general approach to model selection which includes the AIC and note situations when = 2 (used here) provides a reasonable choice. Ostrouchov and Frome [21] discuss various approaches to model selection based on AIC and related criteria where the objective is to find a subset of models that adequately describe a large ADS. See global modal search at (Computational Statistics.)

- Back To ORMS Home Page
- Last Modified 3Jul97 FromeEL@ornl.gov(touches: 202097 )