A common approach for performance monitoring and diagnosis in wireless sensor networks (WSNs) is to send meta-data to a sink node to process. WSN radio constraints limit the amount of this meta-data that can be sent. Logging it to the node's onboard storage can also aid in long-term performance debugging. Using the stored data it is possible for nodes to do statistical analysis for the detection of performance anomalies in the network, rather than at the sink. In this paper we compare the cost and accuracy of performing anomaly detection in the network and at the sink.