Estimating the health impact of the disaster is very difficult because we don't have a robust model for the damage that low radiation exposure causes.
Ultimately, it's a problem of statistical power. High doses - large effect sizes - are relatively easy to work with and we can see effects without too much issue. But small effect sizes require large samples for suitable statistical power to avoid type II error, and perhaps more importantly, observational studies become extremely prone to even small confounding factors. In the case of studying radiation, or any number of drugs, diets, or environmental risks, you can't ethically perform experiments. "Natural" experiments like Chernobyl, Hiroshima, etc. are still beset by confounding factors.
So the best we can do is take the data from higher exposures and extrapolate that data back to the origin. But then you run into the issue of whether the regression comes back to meet the origin (linear no threshold - LNT), curves and sits along the x-axis for a while (threshold effects), or even dips below (hormesis). The truth is that we don't really know what happens at lower levels of exposure. The conservative approach is to assume the worst case, and thus LNT tends to get the most play in the literature. Hormesis has also picked up an air of woo and psuedoscience over time, but the idea isn't altogether unreasonable.
Calculations of 'how bad' Chernobyl was all depend on what model you assume, so there's plenty of room for variation in that, in addition to the spotty historical evidence.