Background and objective: Survival models fail in accounting for non proportionality in the cost-accumulation process. In case of no censoring, the log-gamma GLM model seems to be a robust alternative. We propose the additive Aalen model as a valid alternative, allowing to: obtain a reasonable fit of the cost-accumulation process, maintaining the additivity of the effects, and obtaining additional information as compared to other models. Methods: To assess the performance of the Aalen's additive regression model in comparison to the Gamma-based estimators, a series of Monte Carlo simulations was performed. From two parametric distributions - Lognormal and Gamma â " we generated four distributions of cost data. We considered two Lognormal data generating mechanisms by setting alternatively the variance of the error equal to 1 or 2 and assuming a null interaction between asset of covariates and the error term. Two data generating mechanisms based on a Gamma distribution were assumed by defining alternatively the shape parameter equal to 0.5 (monotonically declining pdf) or 2 (bell shaped but right-skewed pdf). Seven different sample sizes (k=20, 50, 200, 500, 1000, 2000, 4000) were considered for each distribution, bringing to a total of 28 (4 distributions for 7 sample sizes) simulation scenarios. For each of the 1000 Monte Carlo experiments three models were fitted: a Gamma regression with log link, the Aalen additive model and the linear OLS model for the natural logarithm of y as a reference. At each replication of the simulation experiments, mean and standard deviation of the predicted values, absolute bias of the estimated mean, the Mean Squared Error (MSE) and coverage probability (proportion of 95% confidence intervals for the predicted mean at a mean value of the covariate x containing 1) were computed for the different estimators. Then, the Aalen model has been applied to the data of the COSTAMI trial. Results: The bias for the Aalen model is about 10% for the log-normal generated data and almost negligible for the Gamma case, showing a conservative coverage also for small samples. Both the Gamma and the Aalen model outperformed the OLS estimator, showing a bias from between 28% and 80%. In terms of coverage of the 95% confidence intervals for the predicted mean, both Gamma and Aalen models are in general conservative also for small samples, underestimating the true coverage by about 5%. In the COSTAMI case, the Aalen model allowed the evaluation of each covariate effect also at extreme observed costs. Conclusion: In a non-censoring framework the Aalen additive model performs well and can be a reasonable alternative to the standard Gamma regression models. Still, it offers additional information about the relationships between costs and specific covariates
