The non-Gaussian cold spot detected in wavelet space in the WMAP 1–year data , is detected again in the coadded WMAP 3–year data at the same position ( b = -57 ^ { \circ } ,l = 209 ^ { \circ } ) and size in the sky ( \approx 10 \arcdeg ) .
The present analysis is based on several statistical methods : kurtosis , maximum absolute temperature , number of pixels below a given threshold , volume and Higher Criticism .
All these methods detect deviations from Gaussianity in the 3–year data set at a slightly higher confidence level than in the WMAP 1–year data .
These small differences are mainly due to the new foreground reduction technique and not to the reduction of the noise level , which is negligible at the scale of the spot .
In order to avoid a posteriori analyses , we recalculate for the WMAP 3–year data the significance of the deviation in the kurtosis .
The skewness and kurtosis tests were the first tests performed with wavelets for the WMAP data .
We obtain that the probability of finding an at least as high deviation in Gaussian simulations is 1.85 \% .
The frequency dependence of the spot is shown to be extremely flat .
Galactic foreground emissions are not likely to be responsible for the detected deviation from Gaussianity .