The spikes are not necessarily disappearing - they are being filtered (as are the positive ones!); you can see this in the severe change in amplitude between the before/after signals. It's just that, if we look closely at the positive spikes in the original signal, we see they are stronger than the negative ones. Specifically, their peaks have a wider duration - particularly the ones that remain after filtering. Namely, the positive spikes have a sequence of several very positive values right next to each other. Apparently the negative spikes do not have this property. Maybe this is something in the data that is trying to tell you something about the source of these spikes? (it's not random)
In other words, the filter appears to be working and your observation about negative spikes is due to your data being different in the positive direction than in the negative direction.
Since these spikes - positive and negative - may be outliers, if you want to remove them, it may be better just to perform a standard deviation calculation on the signal (the "std" command) and then eliminate signals that are more than 3 or 4 standard deviations from the mean. E.g., remove data larger than mean+4*sigma and less than mean-4*sigma. A common practice is to set these outliers to the mean, or even to NaN.
Filtering of outliers just using Butterworth filters (or others) can do very strange things to data as the outlier peak magnitudes are reduced, but the filtering process causes these tall peaks to spread out in time, and thus significantly distort the surrounding data. These distortions are nearly impossible to remove in later steps, so best to try to eliminate the outliers properly.
Another option is to use median filtering - but this can introduce other artifacts into the results.
For experimental work, outlier removal using standard deviations is generally better practice than butterworth filtering or median filtering.
