(A) Fit results for ordinary least squares regression analyses of the pooled dataset. Note that the red line for the exponential fit is hidden under the linear and quad2 fits. Column 2 shows the p-values for a t-test on each coefficient. In the quadratic case, the linear term was not significant, and thus it was re-fit as Quadratic 2. Column 4 shows the correctness of the fit as the adjusted R2. Each of the fits (with all significant terms) explains similar amounts of the variance, and thus it is hard to differentiate between curve types. However, in each case there is a highly-significant (p<10–14 with associated F-test values similar) increasing trend, indicating the presence of a gradient. (B) To compare anterior and posterior portions of the gradient, we split the dataset at x = 310 and ran the fits on the two portions. x < 310. Results were almost identical to the full fit (A). As before, the linear term in the two-term quadratic was not significant. x > 310. Results were almost identical to the full fit (A), including the exponential curve. Here the quadratic fit is clearly not statistically significant, indicating it is not a mix of behaviors. (C) The exponential curve is based on the linear fit of x with the logarithm of [RA]. Note that the magnitude of differences in [RA] is sufficiently small that the logarithmic change basically becomes a translation. Thus on the semi-translated data lower points in the middle cause the linear fit of the log to have a small slope, making the exponential of the coefficient small enough to make its resulting fit look linear. Thus, given measured differences in relative RA abundance, exponential, linear, and quadratic fits are indistinguishable, even at the tail end of the gradient.