Context: precedents and study concept
The Federal Office for Radiation Protection (German acronym: BfS) is not a nuclear sceptic institution. The current study is the third on the subject (the preceding, so-called Michaelis studies published 1992 and 1998-9), but the first with a positive result. All used the long-term and near-complete database of the German Child Cancer Databank.
The previous two studies confined themselves to compare cancer rates between two areas: within and outside a certain radius around West German nuclear plant sites. The first found a hint of increased leukemia rates in the inner 5 km, but the second (which considered the next five years) failed to confirm that, except for a cluster at a single plant, the Krümmel plant, but the actual cause of that cluster was subject to strong controversy (as I mentioned in my Brunsbüttel, Krümmel (German nuclear controversy) diary).
BfS would have thought the second study was the end of the story, but there was repeated criticism also from experts. Then in 2003 (still under the environment ministry of Jürgen Trittin/Greens), the decision was taken to do a much more detailed study, with the following basic ideas:
- instead of comparing two zones, the distance of every single child cancer case (and control sample child) from the nuclear plant (with its oldest chimney as zero point) shall be determined with a precision of 100 metres;
- detailed interviews shall be conducted with the families of every child in the sample, to look for all kinds of potential carcinogen factors or patterns;
- separately from the study group, a 12-man expert board was named to closely monitor the outlining and conduct of the research, and review & comment the results.
The last included prominent critics of the previous studies. However, its main contribution seems to have come from members sceptical of the emerging positive result, in the form of recommendations for additional statistical significance tests.
The study is titled Epidemologische Studie zum Kinderkrebs und Fehlbildungen in der Umgebung von Kernkraftwerken (KiKK; Epidemological study on Child Cancer and deformities in the environs of Nuclear power plants).
Some basic numbers for scale
The study encompasses a timespan of 24 years: 1980 to 2003 (shorter for plants not yet commissioned/ shut down for more than 5 years).
The area studied was 41 counties: those surrounding 16 of the 18 (+2 research) West German nuclear power plant sites, the closest neighbour or two, and (if not identical) another in the prevailing wind direction.
In these areas and this timespan, more than 350,000 people have been children under 5. The total number of valid cancer cases among these was 1592, of which 593 had leukemia. The control group numbered 4735.
For comparison: the total number of cancer cases in this age group and period was 13,373 for all Germany. Of this, leukemia is a subset of 5,893 cases.
The actually achieved child home - power plant distance precision (difficult to establish for some old cases) was 25 metres.
Headline findings
The authors chose cancer rate increases in the area within the 5-km-radius around the nuclear sites to show the strength of the signal.
The number of all under-five cancer cases in the 24 years was 77, while the expectation would be 48 -- that is, 29 extra cases attributable to the location alone.
Within this, the number of leukemia cases was 37, against an expected 17 (20 extra, two-thirds of all extra cases).
More in-depth details and results
The distance distributions of the cancer case (Fälle) and cancer-free control samples:
An inset only for the statistically interested (and still leaving away most of the details):
Following a common statistical procedure for rare diseases, the control group was selected by matching some factors to the cancer cases. That is, for each single case, the researchers looked for cancer-free children with the same (a) age, (b) sex and (c) closest power plant (d) at the time of the diagnosis (within certain intervals) in randomly selected communities, and then randomly selected up to six from these (using three for analysis).
Such data can be analysed with "conditional logistical regression" (CLR), which was developed specifically to avoid sparse-data biases, and from what I found is primarily used in medial studies. One checks the distribution of the cancer cases against that of the control cases in matched groups. The tested model has the form
Relative Risk = 1 + β*distance-1.42
approximated by
log(Odds Ratio) = β'*distance-1
where the distance function matches a general formula for radiation exposure and β is the parameter to be estimated. 5% low probability margin was used as confidence threshold.
The point for laymen is that if we estimate this number and it is safely greater than zero, we can confidently say that a part of the child cancer rate depends on distance from a nuclear plant.
Now, the result for the full sample is clearly positive: 1.18, with a 95% chance that it is above 0.46, and a mere 0.34% chance that it is zero or below (i.e. that there is no cancer rate - nuclear power plant distance function).
The result in form of a diagram, showing the odds ratio as the function of distance:
Analogous to the older studies, another regression analysis was done according to areas (with the binary variable "inside/outside radius x km" instead of distance). The result:
Limiting
radius | Odds ratio (OR)
(inside/outside) | OR above this limit
with 95% probability | probability of no
real distance trend |
| 5 km | 1.61 | 1.26 | 0.06% |
| 10 km | 1.18 | 1.03 | 2.51% |
| 20 km | 1.06 | 0.96 | 17.8% |
| 30 km | 1.10 | 0.98 | 9.0% |
| 40 km | 1.04 | 0.88 | 36% |
| 50 km | 1.38 | 1.05 | 2.47% |
Separate tests were ran on the lower-numbered sub-samples for specific kinds of cancer, with a significant result only for leukemia:
Cancer
type | β'
(shall be >0) | Standard
deviation | β' above this limit
with 95% probability | probability of no
real distance trend |
| leukemia | 1.75 | 0.67 | 0.65 | 0.44% |
| -lymphatic | 1.63 | 0.75 | 0.39 | 1.53% |
| -myeloic | 1.99 | 1.45 | -0.41 | 8.6% |
| CNS tumor | -1.02 | 1.44 | -3.40 | 76%% |
| embryonic tumor | 0.52 | 0.83 | -0.84 | 26% |
For a limited test of changes over time, power plant times were separated in two: the first at least 11 years and the rest; and a two-parameter regression with time period as binary variable was run. Though the result was a flatter distance function for the second period than for the first, the difference was not statistically significant.
Present-day low dose ionising radiation models calculate with an OR of 0.5 for 1 Gray/year. However, those models are limited to specific forms of radiation, and adults (even adult males), with insufficient research data to say something about children (who have systematically different types of cancer). A recently released study of leukemia in under-six children near Chernobyl found an OR of 1.46 for doses of 1-5 milligray. So the effect now found seems four-three orders of magnitude stronger.
Further statistical tests
What about a model with a different function of distance? On a general level, two further methods (the so-called 'fractional polynom method' and the 'Box-Tidwell Model') were used as back-up test. They brought no significant improvement, but the best fits had the form
Odds Ratio = β*distance-½
resp.
log(Odds Ratio) = β*distance0.55
The main difference in layman terms: higher cancer rates at larger distances and lower ones in the innermost 2 kilometres than with the original model.
For leukemia, a quadratic relation was suggested by other authors, so that was fitted, too. However, the Akaike information criterion was almost the same (1633, vs. 1631 for the linear).
What if one site is responsible for the signal, e.g. as the Krümmel plant in the previous study? Upon the expert board's recommendation, the regression analysis was re-done for each reduced dataset (all sites minus one site), but each time the signal was there. But the Krümmel plant's region does push leukemia higher: without it, β reduces from 1.75 to 1.39.
There was one circumstance that could lead to strong selection effects: if lack of proper cooperation in recruiting the control group candidates is a function of distance. Altogether, recruitment failure was 10.5%, and indeed recruitment failure was higher by some 5 percentage points in the inner 5 kilometres. Again upon the expert board's recommendation, two more regression analyses were carried out, first leaving out match groups of cancer cases in fully uncooperative communities (9% of all), then those in partially cooperative ones too (a fifth in all; leaving 1310 cancer cases). β' falls somewhat from 1.18 to 1.11 resp. 1.01, and the probability that there is no real distance trend grows to a still minute 1.58%.
Yet another expert board advised confidence check went the opposite way: including all control group candidates (up to six controls per cancer case), not just the originally intended max. 3. The result was almost identical, only with somewhat reduced error margins. This proves that the original choice of three controls from samples of up to sixdidn't spoil the data.
During the interview part of the study, the suspicion came up that some 5% of the control group has the wrong address in the database (chiefly: families moved in later than the date in community records). What effect that may have had, was tested with a simulation and an in-depth check on a sub-sample. The former showed that even if the 5% erroneous sample is concentrated in one region, the basic result (distance trend with strong confidence) doesn't change. A check of some addresses with various methods showed non-systematic fluctuations, and regression analyses again showed no significantly different parameters.
Could migration movements be correlated with nuclear plant construction? No such effect could be found (in all plant regions, the strongest migration phenomenon was German Reunification).
Detailed interviews
A subset of the cancer cases and roughly twice as many controls were interviewed to look for other potential factors causing child cancer. Questions about issues like birth data, disease history, vaccinations, medicines taken by child & mother, allergies, family situation, economic situation, kindergarden or at home, etc.
The families' cooperation was much lower than for the simple distance study, with significant tendencies (distance from power plant [less in the closest region], child's age, class, size of settlement), thus the study authors stress that a distance statistic of the interviews is of questionable validity (this part of the study was much less useful than planned) -- but the expert board requested it anyway (and got no significant skew).
The few correlations found: increased cancer risk for non-vaccinated children and those with few social contacts; increased leukemia and non-Hodgkin risk for low class and high birth weight; less central nervous system tumors for children who have many contacts with living things (animals, kindergarden, many adults in household). All of these fit with results of other studies.
Another test planned with the interview subjects was to check their address history, to establish an average distance from plants from birth to cancer diagnosis. But there was again no confidence in distance statistics.
Studies by others
The KiKK was emphatically NOT the first study concluding that there is a statistically significant increase in cancer rates. It is rather the case that the predecessors of this study were a main reason for the issue to remain unresolved and controversial (and a reason for many to dismiss earlier studies with positive results). The two preceding studies themselves were inspired by late eighties British studies finding increased leukemia near British sites.
The press release mentions two meta-studies from 2006 and 2007 that concluded that there is a weak child leukemia effect (unfortunately witout naming them, and I didn't find a reference in the study itself).
Comments by the expert board
The expert board commentary is generally approving, but one senses friction rather clearly. They hint that the study group didn't let them check on their data handling and request an external audit (which is underway, until the end of this year).
The experts' most significant disapproval concerns giving only the increased cancer rate in the inner 5 km: while the effect is lower beyond, the area (and population) much bigger: say, within the 50 km radius, the extra cancer cases are "at least 121-275". In fact I omitted a calculation of overall cancer risk from the KiKK study, which simply divided the 5 km data with all under-5 cancer cases in Germany - definitely invalid.
The expert board also criticise the claim that the effect is way beyond wat one would expect from the radiation exposure, stating that knowledge of the latter is rather vague, though IMO they gloss over nuance in ther study text (either that or the text was modified after the expert board opinion was written).
