Review of the risk of cancer following low and moderate doses of sparsely ionising radiation received in early life in groups with individually estimated doses

Background: The detrimental health effects associated with the receipt of moderate (0.1–1 Gy) and high (>1 Gy) acute doses of sparsely ionising radiation are well established from human epidemiological studies. There is accumulating direct evidence of excess risk of cancer in a number of populations exposed at lower acute doses or doses received over a protracted period. There is evidence that relative risks are generally higher after radiation exposures in utero or in childhood. Methods and findings: We reviewed and summarised evidence from 60 studies of cancer or benign neoplasms following low- or moderate-level exposure in utero or in childhood from medical and environmental sources. In most of the populations studied the exposure was predominantly to sparsely ionising radiation, such as X-rays and gamma-rays. There were significant (p < 0.001) excess risks for all cancers, and particularly large excess relative risks were observed for brain/CNS tumours, thyroid cancer (including nodules) and leukaemia. Conclusions: Overall, the totality of this large body of data relating to in utero and childhood exposure provides support for the existence of excess cancer and benign neoplasm risk associated with radiation doses < 0.1 Gy, and for certain groups exposed to natural background radiation, to fallout and medical X-rays in utero, at about 0.02 Gy.


Introduction
Although moderate and high doses of sparsely ionising radiation (such as X-rays and gamma-rays), when received at a high dose-rate, are known to be associated with elevated cancer risks (Armstrong et al. 2012; Committee to Assess Health Risks from EXposure to Low Levels of Ionizing Radiation 2006; International Commission on Radiological Protection (ICRP) 2007; United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 2008), less is known about any risks arising from exposures at lower doses and dose rates.
There is growing evidence in the Japanese atomic bomb survivors ) and in groups receiving medical diagnostic exposures or radiation therapy of an excess risk of cancer following lower levels of exposure to radiation, particularly among those exposed in childhood (Little et al. 2018b;Lubin et al. 2017). The pioneering case-control study of Stewart et al (Bithell and Stewart 1975;Stewart et al. 1956;Stewart et al. 1958), which became known as the OXford Survey of Childhood Cancers (OSCC), suggested that there might be excess risk of most types of childhood cancer associated with antenatal exposure to doses of about 0.01-0.03 Gy of X-rays; however, the interpretation of the association found by this and similar case-control studies has been controversial, with potential for recall and selection biases and for confounding (Brent 2014; International Commission on Radiological Protection (ICRP) 2003; United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 2008). More recently there have been a number of studies of childhood cancer associated with natural background exposure to gamma radiation (with cumulative doses generally in the range of at most a few tens of mGy), some (but not all) of which have observed excess cancer risks (Kendall et al. 2021;Mazzei-Abba et al. 2020).
By far the largest part (>80%) of man-made radiation exposure (apart from patients receiving radiotherapy) to children in the US is from computed tomography (CT) scan use, comprising a collective effective dose of about 18,000 person Sv in 2016, with much smaller contributions from conventional radiography (∼1300 person Sv), fluoroscopy (∼700 person Sv), nuclear imaging (∼700 person Sv), and image-guided interventions (∼300 person Sv) (National Council on Radiation Protection and Measurements (NCRP) 2019). Both in the US and Canada rates of CT scan use in children have stabilised since the early 2000s, with some signs of reduction since about 2006 (Smith-Bindman et al. 2019), in contrast to trends in the adult population in the US and elsewhere where rates have carried on increasing (National Council on Radiation Protection and Measurements (NCRP) 2009;Smith-Bindman et al. 2019). There have been a number of studies evaluating risks of cancer after CT scanning in childhood, many of which have indicated some excess risk (Berrington de Gonzalez et al. 2016;Journy et al. 2015;Journy et al. 2016;Krille et al. 2015;Mathews et al. 2013;Meulepas et al. 2019;Pearce et al. 2012) although the interpretation of these findings is not straightforward (Boice 2015; Walsh et al. 2014).
There have been a number of recent reviews of this low and moderate dose literature, in particular by the National Council on Radiation Protection and Measurements (NCRP) (National Council on Radiation Protection and Measurements (NCRP) 2018; Shore et al. 2018;Shore et al. 2019) and by a large group of collaborators coordinated by the National Cancer Institute (NCI) Daniels et al. 2020;Gilbert et al. 2020;Hauptmann et al. 2020;Linet et al. 2020;Schubauer-Berigan et al. 2020), although most studies surveyed in both cases related to exposure in adulthood.
Childhood cancers have never been common diseases and recent decades have seen great improvements in therapy (Stiller 2007). Nevertheless, they remain diseases of concern, particularly due to the temporal and spatial variations in incidence ("clusters") that have been observed in, for example, childhood leukaemia (Steinmaus et al. 2004). In this paper we shall review cancer risks following exposure to sparsely ionising (low linear energy transfer (LET)) radiation exposure early in life (in utero and in childhood) in many of these low and moderate dose studies, and by a meta-analysis quantitatively assess the degree of compatibility of relative risk estimates derived from the main studies. As such, the focus of this review is quite distinct from the recent reviews of low dose risk by the NCRP (National Council on Radiation Protection and Measurements (NCRP) 2018) and by the NCI (Berrington de Daniels et al. 2020;Gilbert et al. 2020;Hauptmann et al. 2020;Linet et al. 2020;Schubauer-Berigan et al. 2020), which concentrated on all-age exposure with the former conducted in the context of radiological protection, and specifically did not assess in utero exposure in any depth; however, the issue of risks after radiation exposure in childhood has been dealt with in a number of other reviews (Linet et al. 2009; United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 2013). In the Discussion we shall also briefly review the radiobiology to assess the biological plausibility of these epidemiological associations.

Literature review
A literature search of PubMed was last performed on 16th May 2021 using the search terms given in the Supplementary Methods. Additionally, recent UNSCEAR reports (United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 2008;2018) were scanned to assess additional literature, as well as recent review articles (Kendall et al. 2021;Linet et al. 2009;Linet et al. 2012;Wakeford and Bithell 2021). We restricted d.

e.
Preston et al (Preston et al. 2007) brain/CNS and breast cancer data restricted to < 0.1 Gy; f. Cahoon et al (Cahoon et al. 2017b) lung cancer data restricted to < 0.1 Gy.
These we term the "lower dose risks". These we contrasted with using instead: a. the full dose range thyroid nodule data of Hatch et al (Hatch et al. 2019); b. the full dose range thyroid cancer data of Kopecky et al (Kopecky et al. 2006);

c.
Cardis et al (Cardis et al. 2005) thyroid cancer data using a linear model restricted to < 2 Gy;

e.
Preston et al (Preston et al. 2007) brain/CNS and breast cancer data restricted to < 1 Gy; and f. Cahoon et al (Cahoon et al. 2017b) lung cancer data restricted to < 1 Gy.
These we term the "higher dose risks".
All statistical models were fitted using the metafor package (Viechtbauer 2010; in R (R Project version 3.6.1 2019). Further details of the statistical methods are given in the Supplementary Methods.

Results
The second stage of the literature review yielded 49 studies in which radiation exposure has been quantitatively assessed, whether in utero or in childhood (Tables 1-5). Supplementary   Table S6 gives details of a further 11 studies of natural background radiation that are not included in Table 1, generally because they were not informative or because their results were effectively subsumed within larger studies; also given in Supplementary Table S6, for completeness, are details of those 6 studies that are also included in Table 1.

Risks of in utero exposure
There are strong estimates of excess risk of cancer in childhood in the OSCC study of Bithell and Stiller (Bithell and Stiller 1988) and of Bithell (Bithell 1993) ( Table 2,  Supplementary Table S7), at doses that likely do not exceed 0.03 Gy, and borderline significant indications of excess risk for lympho-haematopoietic malignancies in the range of attained ages up to 61 years in the Southern Urals study of Schüz et al (Schüz et al. 2017) and for all thyroid nodules (mainly benign) at an attained age of 25-30 years in the Ukraine 131 I-exposed cohort of Hatch et al (Hatch et al. 2019) (Table 3, Supplementary Table S7).
There are weaker indications of excess brain tumour risk in a case-control study of medical diagnostic exposures (Pasqual et al. 2020) ( Table 2, Supplementary Table S7), for solid cancer in the in utero exposed Japanese atomic bomb survivors in the incidence study of Preston et al (Preston et al. 2008) at attained age of 12-55 years and in the mortality study of Sugiyama et al (Sugiyama et al. 2021) at attained age of 5-67 years (Table 4, Supplementary  Table S7), and for lymphoma, leukaemia and solid cancer in the offspring of US radiologic technologists (Johnson et al. 2008) at attained age up to 20 years (Table 3, Supplementary  Table S7).

Risks of radiation exposure in childhood
3.2.1. Risks associated with environmental radiation exposure 3.2.1.1. Naturally occurring environmental exposures.: A large number of studies of natural background radiation and childhood cancer have been conducted, as shown in Table 1 and Supplementary Table S6. The maximum doses are generally very low, in no case exceeding 0.05 Gy (Table 1). Among the main studies listed in Table 1, all are of natural background gamma radiation, but some also include assessment of the risks of radon exposure (Berlivet et al. 2021;Berlivet et al. 2020;Demoury et al. 2017;Kendall et al. 2013). All these studies, and two studies that assessed only gamma radiation (Nikkilä et al. 2016;Spycher et al. 2015) are of European national populations, and are register based.
Most studies do not yield significant excess risks, the only exceptions being the British study of Kendall et al (Kendall et al. 2013) and the Swiss study of Spycher et al (Spycher et al. 2015). We note that since the original database search was conducted an updated Swiss study has been published (Mazzei-Abba et al. 2021), which reported very similar relative risk estimates to those of Spycher et al (Spycher et al. 2015) and we judge that the meta-analysis would be little affected. The small Finnish study of Nikkilä et al (Nikkiläet al. 2016) is probably of limited statistical power, but the French studies are of much larger populations (Berlivet et al. 2021;Berlivet et al. 2020;Demoury et al. 2017). The small size of most of the studies in Supplementary Table S6 means that they have little realistic chance of detecting an effect of radiation exposure.

Exposures due to man-made environmental contamination.:
The Chernobyl nuclear accident in northern Ukraine in 1986 resulted in large releases of radioisotopes of iodine and caesium, and other radionuclides to a lesser extent, resulting in a miXture of internal and external exposure, particularly to parts of the populations of Ukraine, Belarus and Russia (United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 2011). One of the main health consequences has been the markedly elevated incidence of thyroid cancer, largely due to intakes of radioiodine by children (Brenner et al. 2011). However, as noted in the Methods, the thyroid doses in most of these studies include both high doses and high dose rates; an exception in this respect is the Belarus and Russia thyroid cancer case-control study of Cardis et al (Cardis et al. 2005), which did consider risks over lower dose (and dose rate) ranges (Table 5).
More relevant to the present review are two case-control studies of childhood leukaemia among those exposed in utero or under the age of 6 years while living in heavily contaminated areas of the former USSR at the time of the Chernobyl accident; the mean dose to the active bone marrow (ABM) was around 10 mGy (Davis et al. 2006;Noshchenko et al. 2010). There are modest excess risks in the Russian and Belarusian components of the study of Davis et al (Davis et al. 2006) and the Ukrainian study of Noshchenko et al (Noshchenko et al. 2010), only the second of these significant (Table 5). Raised, albeit non-significant, risks were observed in a study of childhood leukaemia in Utah and fallout from the Nevada Test Site (Stevens et al. 1990) (Table 5). A cohort study in Utah, Nevada and Arizona of thyroid disease and childhood exposure to fallout from the Nevada Test Site found a positive association for thyroid neoplasms (Lyon et al. 2006) (Table 5). By contrast, a cohort study of persons exposed in early life to 131 I releases from the Hanford nuclear site (Davis et al. 2004) yielded virtually no excess risks of thyroid cancer or benign thyroid nodules (Table 5), but this study probably had only limited power to detect the increased risks predicted by studies of external exposure in childhood (Table 4).

3.2.2.
Risks associated with medical diagnostic exposure and other low and moderate dose radiation exposure-There have been a number of studies evaluating risks of cancer after diagnostic CT scan exposure in childhood, some of which have yielded excess risks of various cancers (Berrington de Gonzalez et al. 2016;Journy et al. 2015;Journy et al. 2016;Kojimahara et al. 2020;Krille et al. 2015;Mathews et al. 2013;Meulepas et al. 2019;Nikkilä et al. 2018;Pearce et al. 2012), and summarised in Table 2. Most studies only assessed risk of leukaemia and brain tumours (Table 2). Only in the Finnish case-control study (Nikkilä et al. 2018), the Australian study (Mathews et al. 2013) and the UK study (Pearce et al. 2012), the low dose part of which is subsumed in the study of Little et al (Little et al. 2018b), was the excess risk of leukaemia statistically significant. In the UK study (Berrington de Gonzalez et al. 2016;Pearce et al. 2012), the Australian study (Mathews et al. 2013), the German study (Krille et al. 2015) and the Dutch study (Meulepas et al. 2019) there were significant excess risks of brain tumours (Table 2). Table 2 shows leukaemia risks in the < 0.1 Gy pooled dataset of Little et al (Little et al. 2018b), also thyroid cancer incidence following external exposure in childhood at thyroid doses of < 0.2 Gy and < 0.1 Gy (Lubin et al. 2017). Risks in both studies are statistically significantly elevated (Table 2). Table 4 shows the paediatrically exposed moderate dose range brain and breast cancer data of Preston et al (Preston et al. 2007), the lung cancer data of Cahoon et al (Cahoon et al. 2017b) and the solid cancer data of Grant et al , all from the LSS cohort. The risks in these datasets are generally compatible with those in the medical diagnostic studies, although the central estimates of risk tend to be higher in the LSS (Tables 2, 4).

Meta-analysis of cancer risks associated with radiation exposure in early life
3.3.1. Restricted maximum likelihood (REML) and maximum likelihood analyses-The REML analysis of Table 6 in relation to postnatal exposure suggests that, overall, there were significantly elevated risks of all cancers and brain/CNS tumours (p < 0.05), with marginally significant (p = 0.108) elevation of risk for leukaemia. Risks for lymphoma and the remainder category (cancers other than leukaemia, lymphoma, brain/CNS tumours) were markedly lower, and were not statistically significant (p > 0.5). In relation to in utero exposure there was (at least using the lower dose analysis) significant excess risk postnatal exposure endpoints, a number above 40%, implying that a material proportion of the variance was due to inter-study heterogeneity; however, this was generally not the case for in utero exposure. Analysis of the data for in utero exposed subjects using adjustments to the ERR for attained age yielded much larger risks (Supplementary Table S4) than those using the unadjusted data (Table 6), but only for thyroid nodules was there a significant positive trend when using the lower dose estimates (p = 0.033).
Further maximum likelihood analyses were performed in order to assess the significance of certain contrasts between lower and higher dose meta-analyses, exhibited in Table 7. These suggest that for thyroid cancer (including thyroid nodules) there was significantly lower risk (p = 0.033) associated with postnatal exposure compared with in utero exposure when using the lower dose estimates. Lung cancer and thyroid (including nodules) risk were significantly higher at lower levels of dose (p = 0.031, p = 0.001 respectively) using the lower dose set of risk estimates. Lung cancer risk was significantly higher for low dose-rate exposures than for moderate and high dose-rate exposures (p = 0.031), although information was only available for the lower dose set of risk estimates, whereas for thyroid (including nodules) risks were significantly higher (p < 0.001) for the moderate and high dose-rate exposures than for lower dose-rate exposure. However, the problems of convergence with all the thyroid cancer/thyroid nodules model fits complicate interpretation of all these findings for this endpoint and caution in interpretation is required. There was significant (p < 0.05) inter-study heterogeneity for certain endpoints for certain of these contrasts, a generally consistent feature of the analyses of leukaemia, brain/CNS tumours and thyroid cancer (including nodules) ( Table 7). The I 2 statistic was somewhat variable, generally near 0, but for brain/CNS tumours consistently above 50%, implying that a relatively large amount of the variance for this endpoint was accounted for by inter-study heterogeneity.
The analysis of Supplementary Tables S3 and S5, in which inverse-variance weighted linear models were employed to refit thyroid nodule data < 0.799 Gy of Hatch et al (Hatch et al. 2019), leukaemia data of Stevens et al (Stevens et al. 1990), and thyroid cancer data < 0.284 Gy of Kopecky et al (Kopecky et al. 2006) yielded generally similar findings, although the postnatal risk of leukaemia reduced but became (at least when using the lower dose set of risk estimates) highly significantly increasing (p < 0.001) (Supplementary Table S3). Additional analysis suggested that when using the lower dose study estimates there was highly significant heterogeneity in risk by endpoint (using the 6-endpoint split of Tables 6 and 7) overall (combining postnatal and in utero exposures) (p < 0.001), and also considering only studies of post-natal exposure (p < 0.001), but not for studies only of in utero exposure (p = 0.194) (results not shown). When using the higher dose study estimates there was highly significant heterogeneity in risk by endpoint overall (combining postnatal and in utero exposures) (p < 0.001), but not when considering only studies of postnatal exposure (p = 0.461), or for studies only of in utero exposure (p = 0.266) (results not shown). A complication with the analyses overall and for postnatal exposure only is that there were indications of non-convergence, whether using the lower or the higher risk estimates.  Table S8 implies that for many endpoints, in particular all postnatal endpoints, leukaemia, thyroid cancer (including nodules) and all cancer other than thyroid cancer (including nodules) there is significant selection bias (p < 0.05). However, the analysis of Supplementary Table S8 also demonstrates that adjusting for selection bias using the trim-and-fill method of Duval and Tweedie (Duval and Tweedie 2000) does not in general lead to marked changes in the central estimates, whether using the lower or higher set of risk estimates -only for lymphoma does the adjustment for selection bias lead to a marked reduction in ERR, although for the lower dose risk estimates the risk for all post-natal studies substantially increases after adjustment.

General remarks
While understanding of the development of cancer at a cellular and sub-cellular level is very important and steadily increasing, and can inform risk extrapolation, epidemiology provides the human evidence most directly relevant to estimate radiation-related cancer risk.
Because we are interested in the effects of low doses, large study populations are required to achieve sufficient statistical power. Studies of cancer after exposure in childhood have decisive advantages over studies of adult cancers: the frequency of cancer at young ages is relatively low and the ERR per unit dose is generally substantially higher than for exposure in adulthood ; United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 2008; 2013). Most of the risk factors apart from radiation for the common childhood cancers are familial or genetic (Roman et al. 2018); there are few other risk factors that substantially modify risk at a population level in childhood . However, about 20% of childhood leukaemia could be attributed to natural background radiation on the basis of conventional risk estimates (Wakeford 2004;Wakeford et al. 2009).
The meta-analyses that we have undertaken combine studies that consider different ranges of attained ages. Many are of childhood cancers, but some extend into adulthood. We have attempted to allow for this in the analyses, in particular the analyses of exposures in utero (Supplementary Tables S4, S7), but this point should be borne in mind. In addition, as noted above, for many postnatal exposure endpoints (but not for studies of in utero exposure) a large proportion of the total variance is accounted for by inter-study heterogeneity, which complicates interpretation of the findings. Nevertheless, the meta-analysis of Table 6 using REML models suggests that there are significant risks for all cancers in relation to postnatal exposure (p < 0.01), with particularly large relative risks for brain/CNS tumours and leukaemia, whereas in relation to in utero exposure the strongest evidence of excess is for thyroid cancer (including nodules) (p = 0.033), although this is based on one study (Hatch et al. 2019). The meta-analysis using maximum-likelihood fitted models in Table 7 implies that for thyroid cancer (including nodules) there is significantly higher relative risk associated with in utero exposure compared with postnatal exposure. There are significant variations by dose and dose rate for certain endpoints, so that lung cancer relative risk appears to be highest for low dose-rate exposures, although the data are limited. There are suggestions of significant variation in relative risk between the cancer endpoints overall, but not when postnatal and in utero exposures are considered separately. The findings of selection bias in the data are somewhat troubling. Nevertheless, although statistically significant for a few endpoints (in particular leukaemia and thyroid), adjustment for such bias does not generally change the ERR estimates (Supplementary Table S8).
It is important to recognise that although relative risks for childhood exposure are relatively high, the absolute risks associated with exposure are quite low. For example, in the pooled analysis of leukaemia after childhood exposure the excess cases or deaths after 0.1 Gy were in the range 0.1-0.4 per 10,000 person-years of follow up (Little et al. 2018b). It is also important to recognize that there has been evidence for some time that relative risks associated with exposure in utero and in childhood are very likely not constant with attained age (Little et al. 1991). We have attempted to adjust for this at least in relation to the studies of in utero exposure (Supplementary Tables S5, S7), from which can be seen the substantial difference that is made. It is possible that such attained age effects could explain part of the heterogeneity observed in studies of postnatal exposure; unfortunately, it was less easy to get useful information on attained age in all of these studies.

Studies of atomic bomb survivors
The survivors of the Japanese atomic bombings offer a unique cohort which can throw light on low-dose effects as well as effects at higher doses -around two-thirds of the survivors received doses < 0.1 Gy. Cancer relative risks in the LSS (Table 4) are not very dissimilar in magnitude to those in diagnostically exposed groups (Table 2). There are suggestions that in utero relative risks in the atomic bomb survivors (Table 4) may be lower than those in some other groups, in particular, for endpoints such as lymphoma and brain/CNS tumours (Tables  2, 3, Supplementary Table S7), as is also indicated by the results of the meta-analysis (Table  7). It should be noted that there are limited numbers of cases and deaths among the in utero exposed occurring in childhood (age < 15 years), specifically one death from liver cancer (Delongchamp et al. 1997) and a non-fatal case of Wilms' tumour (Yoshimoto et al. 1988). The first case of leukaemia, also the first death from this cause, occurred at age 18 years (Delongchamp et al. 1997;Yoshimoto et al. 1988).
For leukaemia there are indications of lower relative risk among those exposed in utero during the atomic bombings, with an average dose of ∼ 0.12 Gy (Sugiyama et al. 2021), than in the OSCC (Table 2) and in other case-control studies of intrauterine medical diagnostic exposure. This may be because of elevated sensitivity of the active bone marrow (ABM) to the competing effects of moderate doses of acutely delivered radiation in utero. It is notable that there are no leukaemia cases or deaths observed in childhood among the in utero exposed cohort, although the expected numbers are small (Delongchamp et al. 1997;Wakeford and Little 2003;Yoshimoto et al. 1988). The study of Ohtaki et al (Ohtaki et al. 2004) may have some bearing on this, as it suggested that stable chromosome translocations among the in utero exposed survivors exhibited a biphasic response, steeply increasing below about 0.1 Gy then involuting above that dose, indicating that haematopoietic cells may be damaged irreparably by moderate doses and replaced by viable cells.
significantly increased radiation incidence risk of uterine corpus cancers is found only for exposure during the mid-pubertal period preceding menarche (Utada et al. 2019). During puberty, rapid stem cell proliferation of the terminal end buds mediates development of the mammary gland into a highly branched epithelial network (Scheele et al. 2017). Likewise, a dramatic increase in uterine volume and endometrial thickness occur during puberty prior to menarche (Hagen et al. 2015). A narrow age at exposure window for cancer radiosensitivity in the breast and uterine corpus may be related to radiation exposure during a period of such increased cell proliferation, and have significant implications for radiological protection.

Studies of in utero irradiation
The association between cancer in childhood and a prior radiographic examination of the abdomen of the pregnant mother identified by case-control studies such as those of Stewart et al (Bithell and Stewart 1975;Stewart et al. 1956) and many others (Wakeford 2008) (see Little et al (Little et al. 2021b, submitted)) provides epidemiological evidence that externally delivered doses of ionising radiation of the order of 0.005-0.030 Gy of X-rays increase the risk of cancer (United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 2008). This level of dose is somewhat lower than the lowest doses producing significantly increased risks of cancer in all other epidemiological studies, apart from the natural background radiation studies that we discuss below (United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 2008). Case-control studies have often been used in this setting, which can be subject to a number of biases, in particular selection, participation and recall biases, which can make them a poor choice for studies of medical exposure; however, with care, large case-control studies have been used in a medical setting without appreciable bias (MacMahon 1962). Nevertheless, the interpretation of the associations in this group of studies has been controversial. Doll and Wakeford (Doll and Wakeford 1997), after reviewing the available evidence, concluded that there are strong grounds for a causal interpretation of the association, and although this is still not universally accepted (International Commission on Radiological Protection (ICRP) 2003; National Council on Radiation Protection and Measurements (NCRP) 2013) there has been a degree of consensus in recent years (Armstrong et al. 2012;Wakeford and Little 2003) that this association may represent a cause-and-effect relationship.
As well as the medical diagnostic studies without assessed doses (which are reviewed in a separate paper (Little et al. 2021b, submitted)), there is information in a large number of studies of various exposed groups in which individual dose estimates are available (Tables  2, 3, Supplementary Table S7). One of the more intriguing findings of the meta-analysis are the indications of difference for certain endpoints in the magnitude of relative risk for exposures in utero and in the postnatal period (Tables 6, 7). In particular the analysis of Table 7 highlights the significant difference between these two types of exposure for thyroid cancer (including nodules), with relative risks tending to be higher for in utero than for postnatal exposure. Although not formally statistically significant there are also indications of much higher risks of lymphoma and brain/CNS associated with exposure in utero than for postnatal exposure (Table 7).

Studies of natural background radiation
As discussed by Kendall et al (Kendall et al. 2021), many of the early studies concentrated on effects of exposure from inhaled radon (an alpha-particle emitting noble gas) and its radioactive progeny, but as time went on, there was an increasing focus on studies of gamma radiation as well as, or instead of, radon, possibly driven by the realisation that doses from penetrating gamma radiation almost certainly accounted for most of the predicted radiation-related absolute risk of cancers in childhood (Kendall et al. 2021). A large number of studies of natural background radiation and childhood cancer have been conducted (Supplementary Table S6). The early studies tended to be ecological, but as time went on there was a preference for the more reliable case-control design. Insufficient statistical power is a particular problem for case-control studies, as realistically sized interview-based studies, with usually at most a few thousand cases, can never have high (>80%) statistical power to detect realistic excess risks (Little et al. 2010). As shown by Land (Land 1980) if a low power study produces a statistically significant positive trend it is almost always bound to be upwardly biased. Studies reporting positive associations are also more likely than negative ones to be written up and published, leading to a reporting bias. However, Hauptmann et al (Hauptmann et al. 2020) judged that most studies that are set up to study cancer and are of reasonable size would be published, whether null or not. Kendall et al (Kendall et al. 2021) discuss various other issues, in particular the problems of selection bias that may affect case-control studies. Largely in response to these two problems, of lack of statistical power and potential bias, a number of register-based national studies have been conducted over the last decade or so, and the advantages and disadvantages of these are discussed by Kendall et al (Kendall et al. 2021). Although free of selection, participation and recall bias, inevitably registry-based studies will lack individual dose estimates and many other individual covariates that can be collected by an interview-based case-control study. However, surrogates for some of these, for example indicators of socioeconomic status, are included in many datasets, and as noted above there are few large risks factors that operate at a population level for most childhood cancers. Therefore, this loss of information is perhaps not a large concern. There have been a number of European national register-based studies (Berlivet et al. 2021;Berlivet et al. 2020;Demoury et al. 2017;Kendall et al. 2013;Mazzei-Abba et al. 2021;Nikkilä et al. 2016;SpiX et al. 2017;Spycher et al. 2015), recently reviewed by Mazzei-Abba et al (Mazzei-Abba et al. 2020) and by Kendall et al (Kendall et al. 2021). Most of these studies remain underpowered; only the British study (Kendall et al. 2013) and the French study (Berlivet et al. 2021;Demoury et al. 2017) have reasonable power, of 50% or more, to detect the predicted excess risk.

Studies of exposures due to man-made environmental contamination
In terms of population exposure, the Chernobyl reactor accident in northern Ukraine in 1986 was by far the largest nuclear accident, in particular, leading to significant intakes of internally deposited radio-active iodine among the populations of the former USSR (United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 2011). As noted in the Supplementary Methods, many thyroid doses and dose rates were very high, so that the screening studies of these populations, recently reviewed by Hatch and Cardis (Hatch and Cardis 2017) are outside the scope of the present review. The mean ABM dose from Chernobyl exposure, mainly external gamma radiation from deposited radiocaesium, was just 10 mGy and the study of childhood leukaemia in the Chernobylexposed populations has been problematical (Davis et al. 2006): there were concerns about the representativeness of the controls in the Ukrainian part of the study (Davis et al. 2006;Moysich et al. 2011), which led to a further case-control study being conducted in Ukraine that produced a substantially lower risk estimate (Noshchenko et al. 2010) ( Table 5). The positive findings for thyroid cancer and benign neoplasms in the Nevada fallout study (Lyon et al. 2006) are also consistent with studies of exposures at slightly higher (but still low) levels of external dose (Lubin et al. 2017) (Tables 2, 5). They are also consistent with risks in the Chernobyl screening studies, although doses and dose rates here are high (Brenner et al. 2011;Little et al. 2014;Little et al. 2015). Both comparisons suggest that internal exposure or low dose rate exposure are unlikely to be the reason for the absence of thyroid cancer risks in the Hanford study (Davis et al. 2004). It is possible that the statistical power in the Hanford study is low.
The Fukushima Dai-ichi nuclear accident in Japan in 2011 also released radioactive iodine, but an order of magnitude less than the release during the Chernobyl accident, and thyroid doses were estimated to be much less than those received by children living around Chernobyl (United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 2014; 2021b). As yet, no studies of populations exposed from Fukushima have incorporated individual dosimetry; in particular this is the case for studies of thyroid cancer by Tsuda et al (Tsuda et al. 2016) and by Ohira et al (Ohira et al. 2020) (see Table 5); the study of Tsuda et al (Tsuda et al. 2016) has been much criticised on other grounds (Wakeford et al. 2016).

Studies of computed tomography (CT) in children
Large datasets  In these CT studies there is some assessed potential for bias associated with reverse causation (Boice 2015; Walsh et al. 2014), that is that the CT scan might have been taken because of early symptoms from pre-existing (latent) disease and was therefore not a cause of the disease (Schubauer-Berigan et al. 2020; Walsh et al. 2014). A priori it is unlikely that reverse causation would have much role to play for leukaemia, as most cases are acute. Confounding by indication, in other words the possibility that high-risk conditions lead to increase in prevalence of CT imaging, is also a concern in these studies Walsh et al. 2014). Confounding by indication is quite distinct from reverse causation, although the two terms are often used interchangeably (Kummeling and Thijs 2008 (Nikkilä et al. 2018) studies, excluding patients with possibly predisposing syndromes (PS) did not much affect the trends with dose. A number of different methods were employed: in the UK study pathology reports from the cancer registries and radiologists' notes were used to determine whether the cases and non-cases had any of a large number of PS (Berrington de Gonzalez et al. 2016); in the French study patient data from the hospital discharge data was used to diagnose patients with these PS, using a slightly smaller list (Journy et al. 2015); in the Dutch study only patients with tuberous sclerosis (one of the PS ascertained in the British but not in the French study) were ascertained via linkage with two hospitals treating most of the patients; and in the Finnish study (Nikkilä et al. 2018) only patients with Down syndrome were ascertained (one of the PS ascertained in the British and French studies). The theoretical study of Meulepas et al. examined plausible scenarios for confounding by indication, via inclusion of individuals with cancer susceptibility syndromes, and suggested that confounding by indication associated with inclusion of such persons was not expected to be substantial (Meulepas et al. 2016). Another theoretical study suggested that reverse causation is unlikely to result in bias away from the null for brain cancer (Little et al. 2021c in press). However, the study of Mathews et al (Mathews et al. 2013) suggests that care is required in the design and conduct of CT scan studies if confounding by indication is to be avoided.

Doses and dose rates in the studies considered and for in vivo radiobiological data
Doses from natural background radiation (Table 1, Supplementary Table S6), the medical in utero studies (Table 2) and the studies of fallout (Table 5) are exceptionally low (generally < 0.03 Gy), but in all studies considered here the dose rates are generally low or moderate -the highest dose rates were generally in the CT scan studies and in a study of persons receiving multiple fluoroscopic exposures as part of the monitoring of tuberculosis treatment (Little and Boice 1999), approach 0.1 Gy per hour; only in the study of Swedish children treated for skin haemangioma (Lundell et al. 1999) did dose rates greatly exceed 0.1 Gy per hour (Table 2). Natural background radiation exposures are ubiquitous and adequately sized cohorts can be assembled, though the studies need to be at a national or super-national scale in order to achieve reasonable statistical power (Little et al. 2010).
A particularly interesting finding is that relative risks associated with lung cancer were significantly lower in the moderate and high dose-rate studies than in the low dose-rate studies, also lower in the moderate and high dose studies than in the low dose studies (Table 7, Supplementary Table S5). This is not what would be expected conventionally, running somewhat counter to the general observation, based largely on radiobiological data, that cancer risks following low dose and low dose rate radiation would be below those at high dose rate (United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 1993; 2008). These results took no account of smoking data in the relevant datasets (Cahoon et al. 2017b;Ronckers et al. 2010), which could conceivably confound; nevertheless, analyses concentrating on lifelong smokers, or in which the baseline rates in the LSS were adjusted for smoking status and numbers of cigarettes per day smoked yielded ERR that were fairly close to the unadjusted ERR (Table 4). However, recent reanalysis of some large animal datasets did not yield very strong evidence for the ameliorating effects of low dose-rate or low dose exposure on cancer risk (Tran and Little 2017), although evidence of such dose rate effects is stronger when the less relevant endpoint of life shortening is used (Haley et al. 2015). This evidence relating to possible effects of dose rate is fairly weak, since we are comparing risks in moderate and high dose rate studies with those at low dose rate among very different study populations, with different periods of follow-up; nevertheless, what we have done is in the spirit of similar exercises that have been conducted in the epidemiological literature that attempt to assess dose rate effects (Hoel 2018;Jacob et al. 2009;Kocher et al. 2018;Little et al. 2021d;Shore et al. 2017;Walsh et al. 2021).

Other reviews of the literature
A group of NCI collaborators conducted a systematic review of 26 recently published epidemiological studies with mean doses<0.1 Gy (range 0.0001-0.082 Gy). These comprised eight environmental, four medical, and 14 occupational studies Daniels et al. 2020;Gilbert et al. 2020;Hauptmann et al. 2020;Linet et al. 2020;Schubauer-Berigan et al. 2020). The review included siX studies of cancer after childhood exposure, all included here. The review considered a critical appraisal of dosimetry methods, confounding and selection bias, outcome assessment problems and the effects of dose measurement errors. Meta-analysis was conducted of leukaemia risk after childhood exposure (Hauptmann et al. 2020). Both for solid cancers and for leukaemia, the majority of the studies reported positive ERRs per unit dose. Several limitations were identified, but only a few positive studies were potentially biased away from the null. These studies therefore directly supported excess cancer risks from low-dose ionising radiation Daniels et al. 2020;Gilbert et al. 2020;Hauptmann et al. 2020;Linet et al. 2020;Schubauer-Berigan et al. 2020), and the magnitude of cancer risks was statistically compatible with those in the LSS (Hauptmann et al. 2020).
It is interesting to note how little overlap there is between the datasets considered in the NCI review (Hauptmann et al. 2020)  The NCI study was limited to those datasets (excluding the Mayak workers and the Kerala natural background radiation studies) in which the mean dose was under 0.1 Gy (Hauptmann et al. 2020), unlike the present review, which limited coverage based on a combination of the maximum cumulative dose and maximum dose rate, so that studies which had dose rate > 0.1 Gy/hour and maximum dose > 1 Gy were excluded. The NCRP review was likewise limited to "a comprehensive review of recent (within ∼ 10 y) relevant epidemiologic studies with quantitative dose-response analyses" (National Council on Radiation Protection and Measurements (NCRP) 2018). Nevertheless, as inspection of Tables 1-5 demonstrates there are certain other studies that might have been included in previous reviews. For example, the lung cancer mortality part of the scoliosis study of Ronckers et al (Ronckers et al. 2010) and the German CT study of Krille et al (Krille et al. 2015) were both apparently eligible for inclusion in the NCI review, but were both excluded because of the abstract-based screening employed there. Likewise the mortality part of the scoliosis study of Ronckers et al (Ronckers et al. 2010) was apparently not considered for inclusion in the NCRP review.
This highlights the difficulty of reviews based on automatic searches of databases such as PubMed, but may also reflect exclusions made in reviewing the searched articles using established criteria that were not made clear in the publications. Bearing on this it should be noted that a very large number of the papers we reviewed were not found in our PubMed search, but were found from assessments of other literature (see Supplementary Methods).

Biological data pertaining to the plausibility of low dose cancer risk
In this section we consider biological data that may support linearity of dose response for cancer, and why departures from linearity can be expected, as well as dose rate effects. The discussion here is mostly quite general, and is applicable to exposures at any age. Although curvature of dose response is not addressed in the low dose/low dose rate data considered in this paper, it is certainly observed at slightly higher levels of dose (United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 1993; 2008), and the presence of curvature is the reason why it is important that attention be paid to low dose or low dose rate effects -otherwise one would just consider the full dose range.
Cancer is thought to result from mutagenic damage to a single cell, specifically to its nuclear DNA, which in principle could be caused by a single radiation track, and this argues against the existence of a threshold of dose below which cancer risk is not elevated, as discussed elsewhere (Little et al. 2009). A more recent evaluation of the biological mechanisms relevant for low dose radiation cancer risk inference concluded that 'There remains good justification for the use of a nonthreshold model for risk inference for radiation protection purposes, given the present robust knowledge on the role of mutation and chromosomal aberrations in carcinogenesis' and, in relation to the potential targets in addition to nuclear DNA, 'The potential contributions of phenomena such as transmissible genomic instability, bystander phenomena, induction of abscopal effects and adaptive response remain unclear.' (United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR) 2021a . The variation in efficacy of repair that undoubtedly occurs will affect the magnitude of unrepaired and misrepaired damage and, whereas unrepaired damage is likely to result in cell death, non-lethal misrepaired damage by definition results in mutation. Some information is available on the age-dependence of induction of DNA and chromosomal damage, and while not many studies are available, any differences between the young and adults are not great Oestreicher et al. 2018).

Conclusions
Here we have considered the overall question of the relationship between low-level exposure to low LET radiation in childhood and the consequent risk of cancer. Attention was mainly focused on leukaemia, the most common and best studied of the childhood cancers, but we have also presented evidence of excess risk of brain/CNS and thyroid cancer and thyroid nodule risk. The data presented here, particularly that in Tables 1 and 2, indicate that there is now little reasonable doubt that the childhood leukaemia risk extends into the low dose range conventionally considered to be doses<0.1 Gy of low LET radiation. We would suggest that the evidence for elevated leukaemia risk now extends down to 0.05 Gy, and indeed for acute lymphoblastic leukaemia, as shown in a recent pooling study of low dose studies, excess risk extends down to around 0.02 Gy (Little et al. 2018b). These studies ( Studies should concentrate on all these endpoints (leukaemia, thyroid cancer and brain/CNS tumours), because in these the excess risk will be most clearly established, and at lower levels of dose, although it is likely that different patterns of low-dose response may emerge. Mechanistic understanding of the development of cancer at a cellular and molecular level makes such a possibility biologically plausible. Further integration of information, particularly quantitative information, on the biological mechanisms of radiation carcinogenesis and leukaemogenesis making use of the adverse outcome pathway framework (National Council on Radiation Protection and Measurements (NCRP) 2020) with the further refinement of epidemiological investigations such as those reviewed here will serve to characterise radiation-related cancer risks more precisely.

Supplementary Material
Refer to Web version on PubMed Central for supplementary material.  Table 2 Risks from childhood and in utero exposure in studies of medical diagnostic radiation exposure.  Stockholm 1920-1959and Sahlgrenska University Hospital Göteborg 1930-1965and followed up 1958-1993 Refitted to data in paper <    (Gilman et al. 1988) by year for 1959, using dose estimate of 6.1 mGy per obstetric radiograph of Mole ( Mole 1990) for that year All cancer mortality  Table 3 Risks resulting from in utero man-made environmental and maternal occupational exposure.  Table 4 Risks from postnatal and in utero exposure in the Japanese atomic bomb survivors.   Table 6 Restricted maximum likelihood (REML) and DerSimonian and Laird 1-step random effects model fits to various subsets (as given by Supplementary  Table 7 Maximum likelihood fits to assess significance of improvement in fit of various explanatory variables (to data as specified by Supplementary Table S2). lung cancer data restricted to < 1 Gy.