Statistical Analyses

A chi-square test was used to assess differences in gender across diagnoses. Fisher’s exact test was used to compare APOE genotype across diagnoses. One-way ANOVAs were used to compare age, education, and OAS score (assessed at the time of olfactory testing) across diagnoses (CN, SCC, MCI, and AD). A one-way ANOVA was also used for the analysis of Trails B time, which was transformed using the Box-Cox procedure (21) to help satisfy the normality assumption of ANOVA. The Kruskal-Wallis test was used to analyze CDR-SOB, MMSE, and BNT scores across diagnoses, while simple logistic regression was used to compare the proportions of correct responses in the OPID-10, OPID-20, OD, and POEM tests across diagnoses. For pairwise comparisons between diagnoses, we used Holm’s stepdown method (22) to correct p-values for multiple comparisons. A subgroup analysis of CN participants was run to compare the performance in the POEM test for correctly versus incorrectly identified OPID-10 odors. A logistic mixed effects model was used to estimate the probability of having a “yes” response on the POEM test, while controlling for the effect of correctly identifying each OPID-10 odor (yes versus no) and accounting for the correlation among POEM responses within each participant.

Primary analyses of the proportions of correct responses in the OPID-20, OD, and POEM tests consisted of multiple logistic regression models that controlled for the following effects: diagnosis, age, gender, education level, and interactions between diagnosis and the covariates (age, gender, and education level). We then examined the interaction with the largest p-value (based on a global F test) and determined whether that p-value was less than 0.1. If not, this interaction was removed and the model was re-fit. We continued in this manner until we had a final model that controlled for the effects of diagnosis, age, gender, education level, and interactions whose p-values (based on a global F test) were less than 0.1.

For each proportion of correct responses in the OPID-20, OD, and POEM tests, we then conducted secondary analyses in two parts. In the first part, we fit a multiple logistic regression model for each proportion that controlled for the effects of: diagnosis, MMSE score (to control for overall cognition), the interaction between diagnosis and MMSE score, age, gender, education level, and the significant or marginally significant diagnosis interactions (with age, gender, or education level) that were identified in the primary analysis of that proportion. We then used a global F test to determine whether the interaction between diagnosis and MMSE score was significant at the 0.05 level or marginally significant at the 0.1 level. If not, this interaction was removed and the model was re-fit. Using the same modeling approach, we then fit additional multiple logistic regression models for each proportion that controlled for the effects of each of the following five AD measures, and their interactions with diagnosis, in separate models: BNT score (to control for a naming deficit), OAS score (to control for olfactory awareness), CDR-SOB score (to control for overall function), Trails B time (to control for executive function required to perform the task), and APOE genotype (to control for genetic risk), which is a well-established risk factor for Alzheimer’s and has been reported to influence odor identification performance. Since no distributional assumptions are required for Trails B time as a covariate, it was not transformed in any of these models

In the second part of our secondary analyses, we fit a multiple logistic regression model for each proportion that controlled for the effects of: diagnosis, the AD measures and their interactions with diagnosis that had significant or marginally significant effects in the first part of our secondary analysis.

For significant main or interaction effects involving diagnosis in both the primary and secondary models for each proportion of correct responses in the OPID-20, OD, and POEM tests, we used Holm’s stepdown method to adjust p-values and the Bonferroni correction to adjust 95% confidence intervals (CIs) for multiple comparisons when assessing which pairs of diagnoses significantly differed from one another.

A subgroup analysis of CN participants was run to compare the performance in the POEM test for correctly versus incorrectly identified OPID-10 odors. A logistic mixed effects model was used to estimate the probability of having a “yes” response on the POEM test, while controlling for the effect of correctly identifying each OPID-10 odor (yes versus no) and accounting for the correlation among POEM responses within each participant.