Noise modeling in power line communications has recently drawn the attention of researchers. However, when characterizing the noise process in narrowband communications, previous works have only focused on small-scale phenomena involving fine-grained details. Nevertheless, the communication link's reliability is also affected by long-term noise phenomena that might affect transfer rates at higher layers as well. This paper addresses the problem of long-term noise characterization for narrowband power line communications and provides a statistical analysis of the long-term trends affecting the noise levels. We present a statistical description of the noise process in the time and frequency domains based on real field measurements in the FCC band (10 kHz - 490 kHz). The collected data comprises more than 1.8 billion samples taken from three different locations over a time period of approximately 10 days. The noise samples have been statistically analyzed by considering stationarity, autocorrelation, and independence. Although our results -- being unprecedented -- are interesting per se, they improve the noise pattern knowledge, thus paving the way for the design and implementation of more robust PLC protocols.