OBJECTIVES: The majority of survival analyses assume proportional hazards but provide no justification for the choice of model and no assessment of external validity. This study aimed to define and compare non-parametric (Kaplan Meier), semi-parametric (Cox) and parametric (Accelerated Failure Time) methods that can be used in persistence analyses. METHODS: Each method was implemented on persistence data calculated on French patients with dispensations of long-acting parenteral (LAP) antipsychotic (paliperidone palmitate, risperidone and haloperidol decanoate) in retail pharmacies from the IMS Longitudinal Treatment Dynamics^TMpanel. Three cohorts of patients initiating LAP antipsychotic treatment between March and December 2013 were created. Treatment discontinuation was defined as five consecutive months without dispensing of the initiated treatment. The survival analyses tested different characteristics (age, co-treatments, previous antipsychotic treatments and initiated treatment) to explain a better persistence between patients. Models were compared on their quality of fit, predictive capabilities and other criteria like results interpretation. RESULTS: Of 4,492 patients included, the persistence at one year was better for patients initiating paliperidone (64%) compared to those initiating haloperidol decanoate (46%) or risperidone (35%). This difference was significant (p<0,0001) for non-parametric (Kaplan Meier) and semi-parametric (Cox) tests. Hazard ratios for paliperidone versus haloperidol decanoate and risperidone were respectively 1.171 and 1.435. CONCLUSIONS: Non-parametric and semi-parametric modeling leads to the same conclusion regarding the influence of the initiated product on discontinuation risk. On one hand, the Cox model enables multivariate modeling, which means possibility to adjust on and test other variables that may affect discontinuation risk. On the other hand, it assumes the proportional hazards hypothesis for all variables introduced in the model. Parametric models can overcome this constraint, under the condition appropriate distributions can be found for the model variables.