We point out the pitfall of using spatial patterns of correlation coefficients obtained from filtered data to infer physical mechanisms. Filtering is often used to emphasize a signal of interest by removing unrelated variability, but it alters subsequent correlation maps when the removed variance contains spatial structure, as is most often the case in geophysical applications. This then results in misleading patterns of correlation. In the case where filtering involves the removal of independent signals, we show that the resulting increase in correlation is entirely due to the removed signal. We include a short discussion of alternative methods that can be used to generate more consistent maps of statistical relationships.