The spectral characteristics of multiple variables were regularly extracted for concentration inversion in the quantitative analysis of heavy metal with LIBS. However, overlapped spectral information might be contained among variables and the complexity of the regression model also would be increased. To extract effective feature variables, the variable selection method based on PLS was studied. PLS model established with the concentration of the element under test used as the dependent variable and multiple LIBS spectrum characteristic variables used as independent variables. The optimal variable subset was extracted on the basis of the original variable importance projection index for variable selection. The results showed that the optimal variable subset for Pb was composed of the Pb Ⅰ 405.78 nm peak, the spectral values before Pb Ⅰ 405.78 nm peak, the intensity corrected by the internal standard element and the signal to background in lake water, and the correlation coefficient square of the training set was 0.912. The optimal variable subset was used for PLS regression analysis, and RSD and RE of the test set were 10.2% and 7.9%, respectively, significantly better than the predictive results of the internal standard. It also showed that variable screening results for different elements and different water samples was applicative to some extent, but the internal standard element failed in correction for different water samples. The results provided high quality characteristic data for the LIBS quantitative analysis, and the methods also provided a reference for other quantitative analysis involving variable selection.