To determine the loss cost within property and casualty (P&C or general, non-life) insurance, actuaries make use of predictive modeling techniques. It is of utmost importance that these techniques are able to accurately capture the structure of the data. This holds particularly true for a workers' compensation product, which is characterized by an inherent hierarchical structure of industry sectors and branches. Due to hierarchical clustering, there will be within-cluster dependence and between-cluster heterogeneity. In addition to the hierarchical structure, company-level rating factors can serve as an additional source of information in the ratemaking process. In general, these are obtained from the insurer's historical data which can be supplemented by data from an external source. We develop a general framework for predictive regression models suitable for hierarchically structured data. We argue that this framework encompasses models such as the hierarchical credibility model of Jewell, the extension thereof by combining it with generalized linear models by Ohlsson and (generalized) linear mixed models. Using a workers' compensation data set from a Belgian insurer, we examine the added value of company-level rating factors in a data-driven way. Part of the set of company-level rating factors are provided by the insurer and another part is extracted from an external, financial database. Additionally, we examine and compare the predictive performance of the aforementioned three methods as well as the effect of different distributional assumptions on model performance.