Definitive diagnosis of Alzheimer’s disease (AD) is often unavailable, so clinical diagnoses with some degree of inaccuracy are often used in research instead. When researchers test methods that may improve clinical accuracy, the error in initial diagnosis can penalize predictions that are more accurate to true diagnoses but differ from clinical diagnoses. To address this challenge, the current study investigated the use of a simple bias adjustment for use in logistic regression that accounts for known inaccuracy in initial diagnoses.

A Bayesian logistic regression model was developed to predict unobserved/true diagnostic status given the sensitivity and specificity of an imperfect reference. This model considers cases as a mixture of true (with rate = sensitivity) and false positives (rate = 1 - specificity) while controls are mixtures of true (rate = specificity) and false negatives (rate = 1 - sensitivity). This bias adjustment was tested using Monte Carlo simulations over four conditions that varied the accuracy of clinical diagnoses. Conditions utilized 1000 iterations each generating a random dataset of n = 1000 based on a true logistic model with an intercept and three arbitrary predictors. Coefficients for parameters were randomly selected in each iteration and used to produce a set of two diagnoses: true diagnoses and observed diagnoses with imperfect accuracy. Sensitivity and specificity of the simulated clinical diagnosis varied with each of the four conditions (C): C1 = (0.77, 0.60), C2 = (0.87, 0.44), C3 = (0.71, 0.71), and C4 = (0.83, 0.55), which are derived from published values for clinical AD diagnoses against autopsy-confirmed pathology. Unadjusted and bias-adjusted logistic regressions were then fit to the simulated data to determine the models’ accuracy in estimating regression parameters and prediction of true diagnosis.

Under all conditions, the bias-adjusted logistic regression model outperformed its unadjusted counterpart. Root mean square error (the variability of estimated coefficients around their true parameter values) ranged from 0.23 to 0.79 for the unadjusted model versus 0.24 to 0.29 for the bias-adjusted model. The empirical coverage rate (the proportion of 95% credible intervals that include their true parameter) ranged from 0.00 to 0.47 for the unadjusted model versus 0.95 to 0.96 for the bias-adjusted model. Finally, the bias-adjusted model produced the best overall diagnostic accuracy with correct classification of true diagnostic values about 78% of the time versus 62-72% without adjustment.

Results of this simulation study, which used published AD sensitivity and specificity statistics, provide evidence that bias-adjustments to logistic regression models are needed when research involves diagnoses from an imperfect standard. Results showed that unadjusted methods rarely identified true effects with credible intervals for coefficients including the true value anywhere from never to less than half of the time. Additional simulations are needed to examine the bias-adjusted model’s performance under additional conditions. Future research is needed to extend the bias adjustment to multinomial logistic regressions and to scenarios where the rate of misdiagnosis is unknown. Such methods may be valuable for improving detection of other neurological disorders with greater diagnostic error as well.

Learning curve patterns on list-learning tasks can help clinicians determine the nature of memory difficulties, as an “impaired” score may actually reflect attention and/or executive difficulties rather than a true memory impairment. Though such pattern analysis is often qualitative, there are quantitative methods to assess these concepts that have been generally underutilized. This study aimed to develop a model that decomposes learning over repeated trials into separate cognitive processes and then include other testing data to predict performance at each trial as a function of general cognitive functioning.

Data for CVLT-II learning trials were obtained from an outpatient neuropsychology service within an academic medical center referred for clinical reasons. Participants with a cognitive diagnosis of non-demented (ND) or probable Alzheimer’s disease (AD) were included. The final sample consisted of 323 ND [Mage = 58.6 (14.8); Medu = 15.4 (2.7); 55.7% female] and 915 AD [Mage = 72.6 (9.0); Medu = 14.2 (3.1); 60.1% female cases. A Bayesian non-linear beta-binomial multilevel model was used, which uses three parameters to predict CVLT-II recall-by-trial: verbal attention span (VAS), maximal learning potential (MLP), and learning rate (LR). Briefly, VAS predicts expected first trial performance while MLP, conversely, predicts the expected best performance as trials are repeated, and LR weights the influence of VAS versus MLR over repeated trials. Predictors of these parameters included age, education, sex, race, and clinical diagnosis, in addition to raw scores on Trail Making Test Parts A and B, phonemic (FAS) fluency, animal fluency, Boston Naming Test, Wisconsin Card Sorting Test (WCST) Categories Completed, and then age-adjusted scaled scores from WAIS-IV Digit Span, Block Design, Vocabulary, and Coding. Random intercepts were included for each parameter and extracted for comparison of residual differences by diagnosis.

The model explained 84% of the variance in CVLT-II raw scores. VAS reduced with age and time-to-complete Trails B but improved with both verbal fluencies and confrontation naming. MLP increased as a function of WAIS Digit Span, animal fluency, confrontation naming, and WCST categories completed. Finally, LR was greater for females and WAIS-IV Coding and Vocabulary performances but reduced with age. Participants with AD had lower estimates of all three parameters: Cohen’s d = 2.49 (VAS) - 3.48 (LR), though including demographic and neuropsychological tests attenuated differences, Cohen’s d = 0.34 (LR) - 0.95 (MLP).

The resulting model highlights how non-memory neuropsychological deficits affect list-learning test performance. At the same time, the model demonstrated that memory patterns on the CVLT-II can still be identified beyond other confounding deficits since having AD affected all parameters independent of other cognitive impairments. The modeling approach can generate conditional learning curves for individual patient data, and when multiple diagnoses are included in the model, a person-fit statistic can be computed to return the mostly likely diagnosis for an individual. The model can also be used in research to quantify or adjust for the effect of other patient data (e.g., neuroimaging, biomarkers, medications).

Evidence-based diagnostic methods have clinical and research applications in neuropsychology. A flexible Bayesian model was developed to yield diagnostic posttest probabilities from a single person’s neuropsychological score profile by utilizing sample descriptive statistics of the test battery across diagnostic populations of interest.

Three studies examined the model’s performance. One simulation examined estimation accuracy of true z-scores. A diagnostic accuracy simulation utilized descriptive statistics from two popular neuropsychological tests, the Wechsler Adult Intelligence Scale–IV (WAIS-IV) and Repeatable Battery for the Assessment of Neuropsychological Status (RBANS). The final simulation examined posterior predictive accuracy of scores to those reported in the WAIS manual.

The model produced minimally biased z-score estimates (root mean square errors: .02–.18) with appropriate credible intervals (95% credible interval empirical coverage rates: .94–1.00). The model correctly classified 80.87% of simulated normal, mild cognitive impairment, and Alzheimer’s disease cases using a four subtest WAIS-IV and the RBANS compared to accuracies of 60.67–65.60% from alternative methods. The posterior predictions of raw scores closely aligned to percentile estimates published in the WAIS-IV manual.

This model permits estimation of posttest probabilities for various combinations of neuropsychological tests across any number of clinical populations with the principal limitation being the accessibility of applicable reference samples. The model produced minimally biased estimates of true z-scores, high diagnostic classification rates, and accurate predictions of multiple reported percentiles while using only simple descriptive statistics from reference samples. Future nonsimulation research on clinical data is needed to fully explore the utility of such diagnostic prediction models.