Inverted parameters estimates

[benjamin]

Has anyone come across parameter estimates that have inverted their sign? That is, I have two primary site-level covariates (AVGMIN and AVGPRECIP) that I am trying to model psi. The two predictor covariates do seem to be somewhat correlated (r=-0.40), but the correlation is not strictly linear. When I have two models estimating the relationships between these covariates with psi (Model 1: AVGMIN + AVGPRECIP, Model 2: AVGPRECIP), I get the following beta estimates for AVGPRECIP:

Model1: psiAVG_PRECIP 0.264371 (0.064708)
Model2: psiAVG_PRECIP -0.509655 (0.053829)

The sampling covariates used in modeling detection probability estimates are the same.

When I plot the estimates of psi against AVGPRECIP they seem to better support Model2, but the more well-supported model (lowest AIC) is the one that includes both covariates.

Any suggestions in terms of interpretation? Thank you.