Substituting values for censored data from Texas, USA, reservoirs inflated and obscured trends in analyses commonly used for water quality target development.

Abstract

Listen

Translate article

We calculated four median datasets (chlorophyll a, Chl a; total phosphorus, TP; and transparency) using multiple approaches to handling censored observations, including substituting fractions of the quantification limit (QL; dataset 1 = 1QL, dataset 2 = 0.5QL) and statistical methods for censored datasets (datasets 3-4) for approximately 100 Texas, USA reservoirs. Trend analyses of differences between dataset 1 and 3 medians indicated percent difference increased linearly above thresholds in percent censored data (%Cen). This relationship was extrapolated to estimate medians for site-parameter combinations with %Cen > 80%, which were combined with dataset 3 as dataset 4. Changepoint analysis of Chl a- and transparency-TP relationships indicated threshold differences up to 50% between datasets. Recursive analysis identified secondary thresholds in dataset 4. Threshold differences show that information introduced via substitution or missing due to limitations of statistical methods biased values, underestimated error, and inflated the strength of TP thresholds identified in datasets 1-3. Analysis of covariance identified differences in linear regression models relating transparency-TP between datasets 1, 2, and the more statistically robust datasets 3-4. Study findings identify high-risk scenarios for biased analytical outcomes when using substitution. These include high probability of median overestimation when %Cen > 50-60% for a single QL, or when %Cen is as low 16% for multiple QL's. Changepoint analysis was uniquely vulnerable to substitution effects when using medians from sites with %Cen > 50%. Linear regression analysis was less sensitive to substitution and missing data effects, but differences in model parameters for transparency cannot be discounted and could be magnified by log-transformation of the variables.