OBJECTIVE To assess the cost-effectiveness of add-on therapy drugs within the formulary of the Mexican Institute of Social Security for partial refractory epilepsy. METHODS Cost-effectiveness study that used costs information from a retrospective cohort of a multicentric study. Efficacy (>50% seizure reduction) and adverse events of oxcarbazepine, topiramate, levetiracetam and lamotrigine were obtained from a meta-analysis. Experts suggested excluding vigabatrin from the analysis, and gabapentin was also eliminated because IMSS formulary recommends its use only for neuropathic pain. Study perspective was institutional with 1 year time horizon, and no discount rate was used. Costs were estimated from financial information from IMSS, and are reported in US 2006 dollars. A decision tree with a Bayesian approach included efficacy and adverse events. Mean cost-effectiveness and incremental ratios, net health benefits and net economic benefits were calculated. ICER confidence interval was estimated with ellipse method. Sensitivity analysis included threshold, scenarios, one-way and probabilistic Monte Carlo simulation. RESULTS Levetiracetam had the lowest mean cost-effectiveness ratio, \$6238. Incremental analysis showed that topiramate was dominated by levetiracetam, while the ICER for lamotrigine and oxcarbazepine was \$1938 and \$2156 compared with not providing add-on therapy. Acceptability curves showed that lamotrigine was the most cost-effective option with a WTP between \$955 and \$1476. Levetiracetam was the most cost-effective option when WTP was above $1476, and the components analysis confirmed this result. CONCLUSIONS Levetiracetam was the alternative with lowest cost per controlled patient and provided the largest health benefit compared with using standard therapy alone (not add-on). Topiramate was dominated by levetiracetam. Standard therapy was the cheapest alternative; however, due to its low effectiveness, it had more health care costs per controlled patient. Sensitivity analysis showed that the base study case was robust.