Towards understanding replication of software engineering experiments

Summary form only given. To consolidate a body of knowledge built upon evidence, experimental results have to be extensively verified. Experiments need replication at other times and under other conditions before they can produce an established piece of knowledge. Several replications need to be run to strengthen the evidence. Most SE experiments have not been replicated. If an experiment is not replicated, there is no way to distinguish whether results were produced by chance (the observed event occurred accidentally), results are artifactual (the event occurred because of the experimental configuration but does not exist in reality) or results conform to a pattern existing in reality. The immaturity of experimental SE knowledge has been an obstacle to replication. Context differences usually oblige SE experimenters to adapt experiments for replication. As key experimental conditions are yet unknown, slight changes in replications have led to differences in the results that prevent verification. There are still many uncertainties about how to proceed with replications of SE experiments. Should replicators reuse the baseline experiment materials? How much liaison should there be among the original and replicating experimenters, if any? What elements of the experimental configuration can be changed for the experiment to be considered a replication rather than a new experiment? The aim of replication is to verify results, but different types of replication serve special verification purposes and afford different degrees of change. Each replication type helps to discover particular experimental conditions that might influence the results. We need to learn which types of replications are feasible in SE as well as the acceptable changes for each type and the level of verification provided.