The COVID-19 pandemic, caused by the SARS-CoV-2 virus, has led to significant morbidity and mortality, straining healthcare systems worldwide. Fundamental approaches for controlling viral spread and mitigating its impact are vaccinations and non-pharmaceutical interventions (NPIs). Before vaccines became available, governments relied on NPIs with largely unknown epidemiological and societal impacts. Despite numerous studies, the effectiveness of NPIs on COVID-19 dynamics remained uncertain, especially over multiple pandemic waves. With a gradual roll-out of vaccines, population immunity increased, but this increase in was counteracted by the emergence of immune-escaping variants of concern (VoCs) and waning of both infection- and vaccine-induced immunity. The long-term dynamics of this decline are currently not well characterized, particularly in the context of multiple infections and infections with different VoCs. Given the only partially observed nature of epidemics and their non-linear dynamics, mathematical models are uniquely suited for their analysis. In my thesis, I applied mathematical models to various COVID-19 data, from aggregated population-level data of infections and hospitalizations to antibody (Ab) titers in individuals, with the goal of quantifying the effectiveness of NPIs and vaccines, identifying protective Ab thresholds, and characterizing immunity waning dynamics. Specifically, my first objective was to estimate the effectiveness of NPIs and vaccines in France and explore counterfactual NPI and vaccine implementation scenarios. We developed a population-based mechanistic model, which we fit to epidemiological data in France from March 2020 to October 2021. The model showed a significant reduction in viral transmission by lockdowns, school closures, and curfews, though their effectiveness decreased over time. Simulations demonstrated that vaccines had saved nearly 160k lives over the study period, but an earlier implementation or a faster rollout could have prevented even more deaths. To understand why NPI effectiveness estimates vary across studies, we evaluated two methodologies in my second objective: mechanistic models and a commonly used two-step regression approach. The latter first estimates the reproductive number (Rt) and then regresses it against NPI parameters. Using simulated data of varying complexity, mechanistic models consistently showed minimal bias (0-5%) and high confidence interval (CI) coverage, whereas the two-step regressions had biases up to 20% and much lower CI coverage. The bias stemmed from the depletion of susceptibles and challenges in estimating Rt, indicating that caution is warranted with this method despite its simplicity and speed. Accurate epidemiological models require up-to-date parameters. My third objective was therefore two-fold: 1) to relate SARS-CoV-2 specific Ab levels to the risk of infection and 2) to characterize antibody waning. Using Ab data from over 220k Canadian blood donors between April 2020 and December 2023, we found that both anti-S and anti-N Abs reduced infection risk, with anti-N showing a stronger effect at lower titers. We used biphasic decay models to characterize waning dynamics and estimated that that 51.3% (95% CI 40.6-66.1%) of individuals would drop below detectable anti-N Ab levels within three years after a single infection. The duration of Ab detection increased after subsequent infections. However, antibodies waned within months below thresholds needed to attain substantial protection, even after multiple infections and vaccinations, indicating that continuous vaccine booster doses might be needed to sustain protection. The analyses I conducted in my PhD research highlight the importance of timely interventions and continuous monitoring of immunity to better prepare for future outbreaks. Moreover, I illustrated that mathematical models are a powerful tool to inform public health decision making and strategies.