The efficiency of mid-infrared selection methods for finding obscured AGN is investigated using data in the Chandra Deep Field North .
It is shown that samples of AGN candidates compiled on the basis of mid-infrared colours only suffer substantial contamination from normal galaxies .
X-ray stacking analysis reveals a soft mean X-ray spectrum for these sources , consistent with \Gamma \approx 2.1 .
This suggests that star-forming galaxies and not obscured AGN dominate the stacked signal .
In contrast AGN selection methods that combine mid-infrared with optical criteria are more successful in finding heavily obscured AGN candidates .
A method similar to the one proposed by Fiore et al .
( 2008 ) is adopted to select extremely red objects ( R - [ 3.6 ] > 3.7 mag ) with high 24 \mu m to optical flux ratio ( f _ { 24 \mu m } / f _ { R } > 1000 ) .
About 80 % of these sources are not detected at X-ray wavelengths .
Stacking the X-ray photons at the positions of these sources shows a flat mean X-ray spectrum ( \Gamma \approx 0.8 ) , which suggests Compton-thick sources , low-luminosity and moderately obscured ( N _ { H } \sim 8 \times 10 ^ { 22 } cm ^ { -2 } ~ { } ) AGN , or a combination of the two .
The mid-infrared colours and luminosities of these sources are consistent with ULIRGs at z \approx 2 , while HST /ACS images , available for the optically brighter of these sources , show disturbed optical morphologies in many of them .
The evidence above suggests that this population includes systems in the process of formation .