Exposure assessment is one of the major tasks when identifying occupational risk factors. At present, most of the exposure assessment methodologies that rely exclusively on monitoring data are costly and time consuming. In contrast, the job exposure matrix (JEM) is a useful alternative assessment, particularly when occupational data are predominantly limited to occupational history. JEMs typically provide information on the exposure intensity and prevalence of selected stressors (e.g., chemical, physical, psychosocial) for defined exposed groups. JEMs are commonly constructed from measurements and/or expert judgement. Currently, there is no nationwide JEM available in Switzerland and only a few exposure measurements for selected chemicals and occupations have been collected. However, many countries with similar working conditions have performed exposure measurement campaigns that have led to the generation of exposure databases and JEMs. In this work, we propose an approach to develop a JEM with limited available occupational exposure data. The first part of this thesis focuses on utilising multiple imputation to address the problem of missing data, which is commonly found in occupational exposure datasets. We tested the performance of multiple imputation under various missing patterns, mimicking plausible missing scenarios observed in occupational exposure databases. By varying the magnitude and the patterns of the missingness of the data, the plausibility and uncertainty of the imputed data was assessed. The results showed that multiple imputation was able to appropriately impute missing values for various categorical (occupations, industries) and continuous (exposure measurements) variables in exposure datasets that contained up to 95% missing values. In the second part of this work, hierarchical Bayesian modelling for the prediction of JEM components with limited exposure data was explored. The framework provided by the Bayesian analysis for the development of a JEM is appealing primarily because various sources of knowledge can easily be taken into account during modelling. As such, the existing information constituted by exposure measurements, the literature and expert judgment can be combined into the model. Moreover, the iterative process allows updating the posterior distribution with new information as it becomes available. In the context of limited exposure data in Switzerland and because a large body of knowledge on occupational exposure is available from other countries with similar working conditions, we evaluated the use of the Bayesian framework for the development of a JEM. The results showed that existing information can modulate the posterior probability and that the multilevel structure of the exposure dataset can be implemented in the Bayesian model. Hence, in this work, multiple imputation coupled to hierarchical Bayesian modelling was demonstrated to be an appealing approach for the construction of a JEM under the circumstances of limited exposure data.