What you have seen closely matches my own observations in some real World tests I carried out at the end of 2014;
Semtech LoRa Transceivers – a KISS approach to Long Range Data Telemetry - January 2015
At the LoRa settings in use, 3dBm was just enough to make the link work, it failed at 2dBm. The key part was this;
Following the 40km\2mW LoRa result, Howie DeFelice produced a link budget calculation
spreadsheet for the LoRa devices.
One anomaly that was clear from the results of Howie’s link budget calculation was that the signal
strength of a 2mW transmitter at 40km calculated as -114dBm.
As 3dBm\2mW was the proven limit of reception, you can conclude that the receiver needed at
least -114dBm of signal to reliably operate. However, the Semtech LoRa calculator application
claims the sensitivity at the bandwidth and spreading factor used was -131dBm, so where has the
missing 17dBm gone?
The sensitivity quoted in the data sheets is probably a figure that cannot be achieved, at least in
non-laboratory situations on Earth, since the real life background noise level is so high. I checked
and a typical background noise level reported by the SX1278 RSSI register during the 40km test
was around -100 to -105dBm. My RF Explorer spectrum analyser reports a similar level. With
noise at that level there may be little value to be gained by the LoRa receiver having a sensitivity
of up to -148dBm as the data sheet specifies.
Where the LoRa device seems to be is getting it’s real world performance from (it is clearly
substantially better than the FSK receiver in the RFM22B) is its ability to receive signals below the
noise level. The acceptable signal to noise ratio (SNR) for the spreading factor used in the 40km
test (8) is quoted in the Semtech data sheets as -10dB, so the receiver should work if the signal
was 10dB below the received noise level. 10dB below a noise level of -100dBm is -110dBm, close
to the predicted signal level that 40km\2mw would have produced (-114dBm).
So whilst the receiver predicted sensitivity (of -131dBm) might suggest far greater range than
2mW for 40km gives in reality, it’s the relationship between local noise level and SNR performance
that appears to predominate.