OBJECTIVES Despite significant patient heterogeneity and complex treatment pathways, averages are commonly relied upon when defining patient populations and treatment effects within type 2 diabetes modeling. As a result, clinicians may struggle to relate results to the clinical setting. This study compares outcomes when using patient-level and average cohort inputs within a published simulation model, based on the UKPDS68 outcomes equations. METHODS UK patient data (2,251 patients initiating dual therapy) were obtained from The Health Improvement Network (THIN). Simulations, performed over a medium-term horizon of 20 years, utilised either patient-level data, collating outputs over all replications, or average cohort data. The outputs (total costs, benefits and complication rates) were then compared. RESULTS Average baseline characteristics were: age: 63.36 (±11.14) years; HbA1c: 8.39% (±1.23); total cholesterol: 4.18 (±0.92) mmol/L; systolic blood pressure: 135.07 (±14.76) mmHg; weight: 89.85 (±19.01) kg. The mean treatment effect was a reduction in HbA1c of 1.01 (±1.23) %. Over 20 years, fewer macrovascular and microvascular events (-82/1,000 patients) and higher all-cause mortality (+17/1,000 patients) were predicted when using patient-level data compared to the average profile. Differences in the frequency and timing of deaths were driven primarily by variation in age and led to fewer estimated life-years (-0.66), quality-adjusted life-years (QALYs; -0.59) and costs (-£551) per patient. Patients estimated to have lower costs and higher QALYs than those associated with the average profile were younger, with higher HbA1c and cholesterol but lower blood pressure at baseline.  CONCLUSIONS Modelling results differ depending on the use of patient-level or average cohort model inputs. Patient-level data may provide insight into the type of patients in whom therapy is likely to be most beneficial. Furthermore, it enables the accurate simulation of correlation between patient characteristics and treatment effect, which are rarely accounted for as part of a standard probabilistic sensitivity analysis.