Background Most research on dietary patterns and health outcomes does not include longitudinal exposure data. We used an innovative technique to capture dietary pattern trajectories and their association with haemoglobin A1c (HbA1c), homeostasis model of insulin resistance (HOMA-IR) and prevalence of newly diagnosed diabetes.

Methods We included 4096 adults with 3–6 waves of diet data (1991–2006) and biomarkers measured in 2009 from the China Health and Nutrition Survey. Diet was assessed with three 24-h recalls and a household food inventory. We used a dietary pattern previously identified with reduced rank regression that positively predicted diabetes in 2006 (high in wheat products and soy milk and low in rice, legumes, poultry, eggs and fish). We estimated a score for this dietary pattern for each subject at each wave. Using latent class trajectory analysis, we grouped subjects with similar dietary pattern score trajectories over time into five classes.

Results Three trajectory classes were stable over time, and in two classes the diet became unhealthier over time (upward trend in dietary pattern score). Among two classes with similar scores in 2006, the one with the lower (healthier) initial score had an HbA1c 1.64% lower (−1.64 (95% CI −3.17 to −0.11)) and non-significantly a HOMA-IR 6.47% lower (−6.47 (−17.37 to 4.42)) and lower odds of diabetes (0.86 (0.44 to 1.67)).

Conclusions Our findings suggest that dietary pattern trajectories with healthier scores longitudinally had a lower HbA1c compared with those with unhealthier scores, even when the trajectories had similar scores in the end point.