Modeling Risk-Factor Trajectories When Measurement Tools Change Sequentially During Follow-up in Cohort Studies: Application to Dietary Habits in Prodromal Dementia

Abstract

Modeling risk-factor trajectories is critical to understanding the natural history of diseases, yet the measurement tools used to assess risk factors often evolve during follow-up in cohorts, and such change prevents longitudinal analyses using standard models. We addressed this issue with a latent process model. Trajectories of average intakes of 5 food families (fish, meat, fruits, vegetables, and carbohydrate-rich foods) were described in prodromal dementia during the 10 years prior to diagnosis of cases and compared with those of controls, using a case-control sample nested within the Three-City Study, Bordeaux, France (1999-2012). Food intakes were measured by 2 or 3 different subquestionnaires across 5 repeated food frequency questionnaires. The sample comprised 205 incident cases and 410 controls matched for age, sex, education, and number of repeated food frequency questionnaires. Intakes of fish, fruits, and vegetables decreased at the approach of diagnosis among cases, suggesting reverse causation. This study demonstrated that the latent process model approach constitutes a powerful framework for modeling risk-factor trajectories, even when measurement tools change sequentially over time. Coupled with a case-control approach to contrast trajectories in prodromal disease versus healthy status, it can help us to understand the dynamic, causal relationships between risk factors and diseases.

Association of Midlife Diet With Subsequent Risk for Dementia

“…In this study the association was not statistically significant when the AHEI was assessed in late midlife, and the median follow-up was less than 15 years. As also reported in another study,35 the present study suggested a slight decrease in diet quality in the years preceding dementia diagnosis, compatible with the hypothesis that change in diet quality is a feature, among others, of preclinical dementia. This observation of a possible signal could be explored in further studies.…”

Intake of Meat, Fish, Fruits, and Vegetables and Long-Term Risk of Dementia and Alzheimer’s Disease

“…Indeed, as AD pathophysiological process begins 10 to 20 years before the first clinical signs [18][19][20], the nutritional habits in the years preceding the diagnosis could be affected by the subclinical expression of prodromal dementia. This hypothesis was confirmed by a recent study that showed a reduction in the consumption frequency of fish, fruits and vegetables in the years preceding the diagnosis of dementia [21].…”

“…For instance, preparation of fish-based dishes could require greater autonomy to buy and cook fresh products, and thus fish consumption frequency might be reduced in subjects in the prodromal phase of AD, as recently shown [21]. In agreement, two long-term studies (the Rotterdam cohort: 9·6 years of follow-up and 5395 subjects at baseline; and the Framingham Heart study: 9·1 years of follow-up and 899 subjects at baseline) did not find any association with fish consumption [12,14].…”

“…Thus, the associations could depend on the nutritional habits of the participants in terms of the type of vegetables eaten. As for fish, a previous publication [21] suggested that prodromal dementia could affect the capacity to buy fruits and vegetables, or to cook them, leading to associations that can be explained by reverse causality. Therefore, a protopathic bias could also partly explain the divergence between our findings and previous literature data.…”

Are latent variable models preferable to composite score approaches when assessing risk factors of change? Evaluation of type-I error and statistical power in longitudinal cognitive studies

“…4,8,10 Theoretically, the multivariate latent process approach should be preferred because it directly models the assumed construct measured by the repeated psychometric tests. 9 As such, it uses all the information available, does not require that all the tests be measured at the same time, 11,12 and can account for differential test functioning, 3,10,13 invariance across time and reliability of the tests. 3,10 By contrast, the composite score can only be estimated for psychometric tests measured at the same time, and the model assumes that the psychometric tests measure the same construct in a predefined manner without any differential test functioning or reliability.…”