A database of cancer mortality and arsenic concentrations in village wells in an arseniasis-endemic area of southwestern Taiwan has been the predominant source of information for risk assessments of U.S. Environmental Protection Agency and two National Research Council reports on arsenic and drinking water. A limitation of the data, however, is that exposure is ecological, that is, cancer mortality cannot be matched with arsenic exposure on an individual basis, just grouped by village. The resultant potential for bias is examined by comparing dose-response analyses of villages divided into two groups, those with well concentrations in a narrow range and the remainder. The narrow range group suggests a flat or downward change in risk in the low dose range, whereas the other group indicates increasing risk. More disturbingly, however, the dose-response data are highly dispersed for both groups and the dose-response curve predicts background bladder/lung cancer levels much higher than a southwestern Taiwan comparison population. This may be due to a large variability between villages of the study area in bladder/lung cancer not directly attributable to arsenic. Including the comparison population in the dose-response analysis artificially anchors the dose-response curve at a background level inconsistent with the study population and likely just biases the slope factor upward.