Evaluation of diagnostic scores with adjustment for covariates

Abstract

Diagnostic tests yield measurements on very different types of scales. Quantitative scales may consist of non-negative integers, either unbounded or bounded, with a fixed number of different values, or they may consist of continuous or percentage values. Remembering a different threshold value for each diagnostic variable would be cumbersome, in particular if covariates have to be taken into account. As a convenient way to overcome such problems we propose to compute z-scores for all measurements. They will be adjusted for covariates so that any individual can be judged on any test result on one single scale with an appropriate standard normal quantile as threshold. Two issues need to be addressed: Selection of covariates in the regression model which delivers the adjustment and normality of the residuals. The first will be treated by cross-validation and the latter by applying an appropriate transformation. We apply this methodology to neuropsychological tests and adjust for age, length of education and sex. Normality of residuals is needed on the diagnostically relevant side only. This allows to use parametric transformations, which can be easily implemented, e.g. in database systems. Since we have measurements at baseline and at follow-up we also analyze change values in a similar manner. For ease of interpretation, we transform the resulting z-scores back to the original scale.

Progression of cognitive impairments in idiopathic REM sleep behaviour disorder

“…The raw scores on the test results were transformed into z-scores demographically adjusted for age, gender and education to analyse results in a more clinically meaningful manner 30. The z-scores below −1.5 were categorised as a pathological or impaired score, and the frequency of cases with pathological scores was assessed for each cognitive measure 27…”

Considering the base rates of low performance in cognitively healthy older adults improves the accuracy to identify neurocognitive impairment with the Consortium to Establish a Registry for Alzheimer’s Disease-Neuropsychological Assessment Battery (CERAD-NAB)

“…A description of these tests is provided in Table 2. Altogether, ten raw scores and demographically adjusted for age, gender, and education z -scores were derived (see [25, 33]). Because education is a strong predictor for premorbid cognitive performance, the number of years of education was used as its surrogate [13, 34, 35].…”

“…To account for the well-known influence of demographic variables, we used standardized z -scores adjusted for age, gender, and education [25] in all analyses. The base rates of low scores in the normative sample were calculated for multiple cut-off scores for the entire battery.…”

Impact of APOE status on cognitive maintenance in healthy elderly persons

“…Testretest change scores were analyzed by subtracting age, gender, and education adjusted z-scores at follow-up (T 1 ) from z-scores at baseline (T 0 ) for each CVLT and CERAD-NAB variable. Since linear regression models were to be used for this purpose, the most appropriate models were selected by applying Predicted Residual Sum of Squares (PRESS) statistics, a procedure that is described in detail elsewhere (Berres et al, 2008;Zehnder et al, 2007). The PRESS statistics procedure was applied also to the z-score differences (T 1 À T 0 ) of each CVLT and CERAD-NAB variable.…”