People often choose one study or another to substantiate a particular view on radiation risk. Currently, radiation risk is best represented by the Linear, No Threshold (LNT) model. Let's review some basics and see why that is, and how various ongoing studies fit in to the big picture.
The LNT model is concerned with total dose versus the excess risk of cancer incidence (sometimes cancer mortality which is about one-half of incidence).
Cancer incidence is determined by whether or not real people are diagnosed with cancer. So this effect is measured at the organism level:
LNT is a macro-description which represents the totality of all the sub-organism level interactions in regards to excess cancer incidence (as opposed to let's say, heart disease incidence).
Let's recall the evidence pyramid:
In the case of the LNT model we don't have randomized controlled double blind studies. Those types of studies involve intentionally exposing people to radiation in order to perform the study.
The most powerful evidence comes from cohort studies, which is where one group who has been exposed to radiation (not under the control of a researcher) is followed over time and compared with a similar unexposed group. Note that this is an organism level approach (unlike in vitro research) and the organism is a human (unlike animal research) which is measuring excess cancer incidence versus dose directly.
The National Academy Of Sciences (NAS) BEIR VII group (who performed a systematic review published in 2006) was specifically tasked with determining the best model of cancer incidence versus dose. They correctly gave great evidentiary weight to the best ongoing cohort study, which is that of the Japanese atomic bomb survivors. The NAS is particularly noteworthy (as opposed to other study sources) because of its appointment process for reviewers.
Unfortunately, cancer has many contributing factors to its progression so it's statistically impossible to be able to discern with confidence at low excess radiation doses whether or not there is an excess of cancer incidence attributable to that excess dose (randomized controlled double blind studies would provide greater statistical power but are not done as already discussed).
We know this statistical limitation exists from day one. A cohort study will always have a statistical association threshold (SAT). BEIR VII correctly examines "down the pyramid" for other sources of evidence (which will have all sorts of non-linearities) which might suggest a threshold or non-linearity below the SAT for excess cancer risk. They discuss in brief (see Public and Executive Summaries) and in detail (see full text) why they conclude that the best model of cancer incidence versus dose is the LNT model.
Of course various studies representing various pyramid layers are ongoing. When contemplating them, consider the "big picture" as described above. Don't fixate on a small number of lower tier studies just because they reinforce a bias you may have or because the mainstream media has headlined one of them.
UPDATED (05/01/12) - A good discussion on the strengths & weaknesses of LNT and how LNT relates to public policy.
UPDATED (06/21/12) - A couple of other fundamental concepts should be pointed out on the topic of the LNT model & radiation studies. LNT doesn't just apply to radiation and cancer it applies to all genotoxic carcinogens which includes some viruses and some chemicals like N-nitroso-N-methylurea. This is because if the substance can damage DNA, even though DNA can repair itself, it doesn't do so with 100% fidelity. So there isn't any threshold and the dose response has to be inferred from epidemiology. Some carcinogens are non-genotoxic, they don't damage the genome directly but work through other mechanisms like increasing cell proliferation or decreasing apoptosis (planned cell death). The non-genotoxic carcinogens (such as dioxin) typically have thresholds.
In regards to any particular epidemiological study keep in mind the Bradford-Hill criteria for causation. Apply these criteria to the study you're considering and it will give you a sense of the merit of that study compared with others.
UPDATED (06/23/12) - Here is a link to the abstract for the latest study on the Japanese atomic bomb survivors. You should be able to get a pdf version of the entire study if you search for the title. There appears to be an unexplained anomaly in the 0.3-0.7 Gy range with lower than expected risks, but higher than expected risks in the lower dose range.
UPDATED (08/09/12) - Here is a link to a 2003 paper which explains LNT very well. I wish I had recalled it earlier.
UPDATED (04/10/12) - Here is a nice summary of the modern (as of 12/12) view of radiation induced carcinogenesis.