You're right, Peter.

Your results would mean Va is negative, which it  is impossible. I would check out the input file for possible errors....

 I agree with Pablo's. You should probably review your data. There must be something wrong when you typed it or when you conducted the different steps of analysis. 

Thanks for the comments. In that case Va is -0.16331. Anyway the data set seems okay, as results of other groupings and other analyses are okay and in a "possible" range (and it's not my first analysis using Arlequin, but the first with weird results). Input file was created by GeneAlex.

You can try checking with L Excoffier. Perhaps you have discovered a bug in Arlequin.

The genetic differentiation components measured in AMOVA are really covariances among alleles, not variances, and negative values should be interpreted in that context.  That said, such strongly negative values for between-population differentiation are problematic.  However, obligately outcrossing organisms, or those that suffer from severely reduced viability due to inbreeding, can have substantially negative values for the within-population inbreeding coefficient.  If your analysis isn't estimating the inbreeding coefficient separately, or is estimating it incorrectly (e.g. inflated estimates of inbreeding due to the presence of null alleles), I wonder if these effects could be masquerading as negative allelic covariances between populations in the absence of population structure.

Also, you don't explain the hierarchy of structuring among regions vs. host plants.  If this hierarchy were set up incorrectly in terms of the way populations are actually structured, it seems like it might be possible that one could get compensating biased estimates for each of these components.  I haven't really thought this through, though . . .

The variance (or covariance) among groups is Fct=Va/Vt. The within-population inbreeding coefficient could be either Fsc=Vb/(vb+Vc) or Fsc=Vb/(Vb+Vc+Vd) depending whether or not the within-individual level is taken into account.

I have interpreted that Peter made two separate analyses: one grouped by hosts and the other grouped by regions. The graphical interface of Arlequin is not error-prone, and in case any error  is made, it should be easy to detect.

Thanks for all the input.

Pablo is right, I did two separate analyses to test possible structuring according to host plant and region, respectively. Though an analysis with combined effects I should do too. Have to check more intensively null alleles, though a first microchecker analysis showed inconspicuous results.