One approach to test for differential associations between plant foods with health uses a scoring approach: foods categorized into animal or 'healthy' plant-based or 'unhealthy' plant-based groups to construct a plant-based diet index (PDI), healthy PDI (hPDI), and unhealthy PDI (uPDI). To evaluate robustness of associations between diet indices and incident coronary heart disease (CHD) risk when recategorizing food groups in indices. Using REasons for Geographic and Racial Differences in Stroke (REGARDS) data, we replicated a published use of the scoring approach. Using Cox proportional hazards regression, we assessed ramifications of the following on associations between diet indices and CHD risk: 1) reconfiguring foods within and among food groups, using potatoes as an example, 2) leave-one-out analysis for each of 12 plant-based food groups, and 3) agnostically redefining each food group as 'healthy' or 'unhealthy'. Over 153,286 person-years of follow-up, there were 868 cases of CHD. Replication analyses did not reach statistical significance. General patterns of magnitude of hazard ratios (HRs) in replication and reconfiguration models were PDI HRs < hPDI HRs < uPDI HRs for women, and hPDI < PDI < uPDI for men. Five models reconfiguring potatoes resulted in small, varied differences in PDI, hPDI, and uPDI associations. Leave-one-out analyses resulted in greater variation of associations between indices and CHD. In agnostic models, each plant-based food group was classified in indices as 'healthy' and 'unhealthy' with statistically significant beneficial or deleterious associations with CHD. Averaged over 4,096 models, HRs' shifts were small when food groups were moved between 'healthy' and 'unhealthy'. Statistically significant associations between hPDI, uPDI, and PDI and incident CHD were not replicated. Small perturbations of the scoring approach had varied impacts on HRs. Agnostically constructing diet indices demonstrated the potential for guilt (or benefit) by association: any of the food groups we studied could be categorized with others in an index showing beneficial or deleterious associations.