Data not collected for research purposes, such as those from electronic health records, are increasingly used for making causal inference about exposure effects. However, these data do not rely on a planned study design that ensures balance in patients’ characteristics across exposure groups. Furthermore, the mere fact of being observed in those data is often strongly related to personal characteristics. When ignored, imbalances associated with those features can bias the estimation of the marginal effect of an exposure on a longitudinal outcome. We propose and demonstrate a methodology to account for those features in observational study settings, focusing on a setting where we aim to estimate the causal marginal effect of a continuous exposure on a longitudinal, ordinal outcome. We apply the methodology to assess the marginal effect of the weekly number of hours spent playing video games, on the number of suicide attempts in youth from the Add Health study in the US.