Mixing data for multivariate statistical study of groundwater quality

In the present paper, a multivariate statistical modeling study of water quality data from different places of Kozhikode Gity, Kerala, India, has been conducted applying multiple linear regression (MLR), structural equation modeling (SEM), and adaptive neuro-fuzzy inference system (ANFIS) modeling. First, we combined water quality data from different places in the study area over different time periods to obtain a unified multiple linear regression (MLR) model. By mixing three data sets from different places and time periods in four different ways, different regression models were formed with total dissolved solids (TDS) as the dependent variable and calcium, magnesium, nitrate, sodium, chloride, potassium, total hardness, and sulfate as independent variables. The effectiveness of each model was then tested against a data set, which corresponded to a different period and location. One unmixed model and three mixed models showed similar performance. An SEM was developed for the data set, which is obtained by mixing all the three data sets. The same regression coefficients are found for the SEM and the corresponding MLR. An improvement in the sample size as a result of mixing of data sets could be thought of as the reason for this phenomenon. We thus selected the MLR obtained by mixing all three data sets as our unified model. For the mixed data set, we then developed an ANFIS model with calcium, magnesium, nitrate, sodium, chloride, potassium, total hardness, and sulfate as input variables and TDS as the output variable. On the external data set, the ANFIS model showed a better performance than the MLR model.