Study design

Epilymph is a case-control study of incident lymphoma conducted between 1998 and 2004 in the Czech-Republic, France, Germany, Ireland, Italy, and Spain. Cases were recruited as soon as possible after the diagnosis to minimize the possibility of any survival bias. Controls were frequency matched by 5-year age groups, sex and country of residence and randomly selected from population registries in Italy and Germany (population-based recruitment) and among patients admitted to the same hospital as cases in the remaining countries (hospital-based recruitment). The proportion of hospital-controls who had the same diagnosis for hospital admission was kept lower than 10%. Participation consisted of an in-person interview on the subject`s socioeconomic status (SES), lifestyle factors and medical history, including medical radiological history. Interviews were conducted by trained personnel using a structured questionnaire. The detailed methods have been published elsewhere [24, 25].

Exposure estimation

Regarding medical radiological history, the interviewer presented a list of common radio-diagnostic examinations: conventional X-ray to the thorax, abdomen (with and without contrast), kidney, bones and face; dental X-ray; gammagraphy (not collected in Germany), and Computerized Tomography (CT) scans. Participants were asked to report the total number of procedures they underwent and the age at first and last procedure for each type of examination. The precise body part was not asked for bone X-rays, CT-scans or gammagraphy. An additional question about any additional examinations was asked, with the answer recorded as free text. The list of questions asked can be found in S1 File.

We excluded examinations taken during the two years prior to the date of diagnosis/reference date to allow for an appropriate latency period between exposure and diagnosis as well as to exclude any procedure possibly related to the diagnosis of the cases. As only age at first and last procedures were available, we assumed a uniform distribution of examinations between the two reported ages, excluding those that fell within the 2-year exclusion period. A sensitivity analysis using a 5 year exclusion period was also performed. We excluded CT-scans reported before the year 1980 (44 CT-scans reported by 42 subjects) since CT-scans were not routinely performed until the late 1970´s-early 1980’s [26, 27].

Since there were no data on the specific dosimetry protocols followed for each procedure and body part, we based our organ dose estimations on typical protocols used in different decades, using published estimates of organ doses. This approach has been used in other similar studies [12, 13, 28]. For conventional X-rays, Melo [29] provided a dose value for each 10-year period from 1930 to 2010. For CT-scans, Kim et al. [30] provided a dose value by sex for two different time periods (pre- and post-2001) and by year of age (0–22 years). We subsequently calculated a mean dose value for the following age groups: 0–5; 6–10; 11–15; 16–20; and >20 years. Since no body part specification was available for CT-scans or bone X-rays, we attributed to each body part a weighted average of doses to different anatomical regions, according to the frequency of examinations of these regions in the general population [3, 31, 32]. A look-up table was derived to assign doses by type of procedure, body part, and time period (see S1 and S2 Tables).

For the cumulative dose calculation, we excluded 169 gammagraphy procedures (reported by 107 subjects) because of lack of information on the procedure in the study questionnaire, and no information was collected in Germany. We also excluded 9,125 dental X-ray procedures (reported by 2,296 subjects) since the fraction of BM in the mandible is close to 0 [33]. Indeed, the dose to the bone marrow from one intraoral dental X-ray is estimated to range from 10⁻³ to 10⁻⁵ mGy, depending on the time period [34]. An additional 52 examinations reported in free text by 51 subjects were also excluded as they were too poorly defined to classify.

Cumulative lifetime BM dose was calculated summing doses from all procedures up to the exclusion period. As only age at first and last procedures was available, we attributed: a) to the first and last procedures, the doses in the look-up table corresponding to the time periods and type of procedures. Since there was no data on the precise time period for the remaining “n-2” examinations, we used three different scenarios for attributing dose values, as follows: a) the mean of the dose over the entire time period between first and last exposure for that type of procedure (main analysis); b) the dose corresponding to the time period of the first procedure; c) the dose corresponding to the time period of the last procedure. Therefore, for each participant and examination we estimated three possible dose values. Sensitivity analyses using the “b” and “c” scenarios were conducted to compare results under the alternative exposure assumptions. Total cumulative BM dose was then estimated for each scenario by summing across all examination types (S2 File). Cumulative BM dose was categorized into 4 categories, with cut points chosen a priori: 1, 5 and 15 mGy (which approximately correspond to the 25^th, 75^th and 95^th percentiles of the dose distribution in the exposed controls).

Outcome definition

For all cases, diagnosis was histologically confirmed and classified according to the World Health Organization (WHO) classification of neoplastic diseases revision 3 [35]. For the main analysis we considered all lymphomas together. Additional analyses were conducted by subtype: NHL, HL and CLL. In addition an analysis with all lymphomas excluding CLL was performed for consistency with previous radiation epidemiology papers [10].

Definition of covariates

Data on personal and family medical history were collected as part of the study interview by asking if the subject or one of his/her family members had ever suffered from any of a number of medical conditions and, if so, the age at onset. The conditions were grouped as follows: a) family history of lymphoma; b) any autoimmune disease (arthritis, lupus, celiac disease and rheumatic fever); c) any atopic disease (asthma, eczema, allergy, urticaria and sinusitis); d) any infectious disease (mononucleosis, tuberculosis, hepatitis, brucellosis and typhoid fever); e) osteoarthritis; f) cardiovascular disease (diabetes, hypertension). A measure of global health status in childhood was also captured by asking the participants about their level of health during childhood using five questions (having been more sick/absent from school comparing with schoolmates; taking more medicines/antibiotics comparing with siblings; having been a very healthy child) [25]. Two measures of Socio-Economic Status (SES) were also captured: a) years of school attendance (categorized using cut points of: 10, 13 years); and b) socio-economic status based on the Standard International Occupational Prestige Scale (SIOPS) [36] for the longest held occupation (categorized using quartiles of the SIOPS score distribution among controls). The correlation between the two SES measures was low (Pearson r = 0.39).

Statistical analysis

We examined associations between cumulative BM dose and sex, age, SES and medical history among controls (both overall and by type of control) in multivariate linear regression models. Only medical history variables were found to be associated with cumulative BM dose and were retained as possible confounders and included in the fully adjusted risk model.

We estimated the association (odds ratios (OR) and 95% confidence intervals (CI)) between lymphoma risk and cumulative medical IR dose using multivariate unconditional logistic regression, adjusting for the matching variables (sex, age, and country), SES (categories of education and of SIOPS) and medical condition (listed above). Linearity of the dose-response relation was evaluated using a Generalized Additive Model (GAM) by entering dose as a continuous variable.

Sensitivity analyses were conducted by a) restricting the analysis to only hospital or only population-based controls; b) excluding subjects reporting any history of osteoarthritis, atopic, autoimmune or infectious disease; c) restricting analyses to participants with at least one high-dose examination (kidney, bone X-ray or CT-scan); d) excluding subjects reporting gammagraphy or any additional IR procedures not included in the dose calculation; e) using a 5 year exclusion period between dose and disease; f) using dose estimates calculated under different assumptions related to the examination time period (above); g) excluding cases interviewed over one year after diagnosis to minimize a potential survival bias; h) adjusting for smoking as a potential confounder; and i) testing for homogeneity of risk between sexes. All tests were likelihood ratio based, two-sided with a significance level of 0.05. The analysis was performed using R version 3.3 [37] and STATA 14.0 [38].

Ethics approval

Ethics approval was obtained in 1997 from the Ethics Committee of Bellvitge hospital (coordination centre) under the project name “Eurolymph II”. Additional approval (extension of the project, reanalysis of existing data) was obtained from the International Agency for Research on Cancer (IARC) ethics review board (January 2010, project number 09–34). Ethics approval was also obtained from the relevant ethics committees of the participating countries. The list of participating institutions/hospitals can be found in S3 File. The study protocol followed was in accordance with the ethical standards of the responsible committees on human experimentation and with the Helsinki Declaration. Written consent was obtained from all participants whose data were used in the present publication.

Selection bias

Selection bias arises if participation is related to the exposure of interest. Hospital controls may have been suffering from a chronic disease (and were possibly more likely to have undergone medical IR): this could explain the higher levels of exposure among controls and the reduced ORs for lymphoma risk (S5 File). However, analyses by type of controls showed similar results and we saw no substantial difference in disease prevalence between population and hospital controls. Population controls also had a higher BM dose when compared to controls recruited in the hospital. Participation rates were lower among controls compared to cases, particularly among population controls. It is possible that the controls who participated represent a population receiving a different type of medical attention (i.e. controls who were more worried about their health or with poorer general health may be more likely to participate in a medical research study) than the population from which they were drawn. Indeed, when excluding subject reporting any chronic disease, we didn’t found evidence of a decrease trend.

Confounding

The aetiology of lymphoma is complex, and putative risk factors may not be consistently associated across all lymphoma subtypes [22]. We adjusted all our analysis for family history, which has been found to be the most consistent risk factor across all NHL subtypes [22]. Chronic diseases may also be relevant confounders as they are associated with both diagnostic IR exposure and lymphoma risk (S5 File). A negative association between lymphoma and atopic, autoimmune and osteoarthritis has been reported [22, 23] and this could explain our inverse ORs. The association between medical conditions and lymphoma risk has been evaluated in the Epilymph study [25, 40], with reduced OR observed for allergies, osteoarthritis, diabetes, and those reporting to have been “more sick than siblings during childhood”.

Subject who reported any atopic, autoimmune, or infectious disease as well as osteoarthritis had a higher average estimated BM dose, as did participants with poorer health during childhood. This higher dose is expected as such diseases may require radio-diagnostic evaluation.

Although we were unable to explore the potential impact of chronic disease in depth here, as the medical reason/indication for the diagnostic procedure was not available, we adjusted analyses for atopic, autoimmune, osteoarthritis, infectious, and childhood diseases. This did not change the risk estimates, neither did excluding subjects who reported any chronic disease (atopy, autoimmune, infectious disease or osteoarthritis). However, as information on medical history was self-reported and may be subject to recall bias, adjustment may not be adequate.

We adjusted for socioeconomic status, as it has been reported that the exposure to medical ionizing radiation may be related to the socioeconomic status [41]. Again, adjustment did not change results, however we may not fully capture the socioeconomic status of the subject, as we measured it with proxies (occupational SIOPS score, education level).

Uncertainty in the dose estimation

For the individual dose estimation, we had limited data on lifetime radiological history, leading to uncertainties in dose estimates. First, we lacked information on the exact time period and age at each procedure. In general, doses have decreased in recent time periods, especially in younger patients as a result of increased awareness and optimization [29, 42]. To address this, we developed three exposure scenarios with different assumptions regarding the timing of examinations; sensitivity analyses using the two alternative scenarios showed that these assumptions did not influence our risk estimates.

X-ray procedures were defined in broad terms in the questionnaire. The advantage of using such broad, generic terms is to make the questions shorter and simpler. However, when reconstructing dose, it is a major source of uncertainty since doses could vary substantially depending on the anatomical region examined. Reported BM doses from the most common CT-scan type performed in adults can vary between 6.9 mGy (chest CT-scan), 3 mGy (abdomen CT-scan) and 1.3 mGy (head CT-scan) [30]. As no information on body part was specified for CT-scans, we used a weighted average considering the distribution of body part scanned in different age groups (for adults we attributed 3.32 mGy to each CT-scan). Thus, in adults we have potentially overestimated the dose for 40% of the procedures (head CT-scans) and underestimated it for 30% (chest CT-scan). However, the percentage of participants who reported a CT scan is low (18%) and the body-part uncertainty is unlikely to be differential between cases and controls.

Furthermore, our cumulative BM dose estimation does not include all types of medical radiological procedures, only the most common. Misclassification due to the omission of other radiological procedures cannot be excluded. To further address this issue, we excluded subjects who reported a previous tumor because they may have received radiotherapy. In a sensitivity analysis, we also excluded subjects reporting gammagraphy and additional procedures (reported as free text) with no change in ORs. However, differential misclassification due to omission of uncommon radiological procedures, though unlikely to affect a large proportion of the population, cannot be excluded if controls (in particular hospital controls) were more likely to have undergone procedures not considered here, due to specific underlying chronic medical conditions.

Finally, the choice of BM as the target organ may also lead to uncertainty as it may be inadequate for some lymphoma subtypes: all lymphoid cells originate in the BM, however, their differentiation occurs only partially in the BM and, in adults, forms of lymphomas that arise from peripheral cells are more frequent [43]. However, in radiation epidemiology, BM is traditionally accepted as the target organ for lymphomas and leukemia and using it allows comparisons with the published literature. Although some studies have used alternative target organ doses, for example mean lymphocyte dose in a study of pediatric CT-scans [14], we were limited by the available information in the study. [14].

Exposure information bias

This study is based on self-reported data and it was not possible to check medical records to validate responses. The informed consent signed by the participants allowed the researchers to access medical records only in the hospital where they were recruited, and hence we are unable to validate all of the procedures reported by the patient. In addition, considering the age range of the participants, most of the procedures (in particular those performed in childhood) are likely to have been recorded in paper charts, not in electronic medical records, thus making them very difficult to access. A medical records review that cannot assess the complete lifetime diagnostic exposure of the subjects would be difficult to interpret [44].

As we based our analyses on self-reported medical radiological history, recall error (random and systematic, differential or not between cases and controls) is likely. Previous studies have attempted to estimate the impact of recall bias when collecting medical radiological history [44–46]. Dreger et al. (2015) found evidence of underreporting in older populations and of over-reporting of CT-scans possibly due to confusion with Magnetic Resonance Imaging (MRI), however comparison was only for recent procedures (not covering the entire lifetime of the study subjects), thus the measurement error in our study may be greater. Berrington et al. [44] reported substantial disagreement between medical record based and self-reported medical radiological history. Such disagreement was shown to be independent of sex, age at interview, time since exposure and calendar year at exposure. Disagreement was also non-differential between cases and controls. However, the accuracy of reporting tended to decrease with increasing number of procedures reported in medical records. The study also compared risk estimates of thyroid cancer from medical IR using information from medical records with risk estimates based on self-reported information; no substantial difference in risk estimates was observed. Similarly, Pogoda et al. [45] did not report important differences when comparing acute myeloid leukemia risk from medical diagnostic IR exposure based on medical records or self-reported.

Even when non-differential between cases and controls, recall error can affect risk estimates, though the direction of the effect is unpredictable [47]. The observed reduced ORs in our analyses may be partially explained by non-differential misclassification, if the misclassification occurs between the higher and the lower dose categories [48]. Individuals with higher dose may have been classified as non-exposed simply because they failed to report a single high dose examination (CT-scan) or if they underwent other high dose examinations that weren´t considered in the present study. An individual with no radiation exposure could also be classified as highly exposed if diagnostic procedures that do not involve IR (e.g. magnetic resonance imaging, ultrasound) were confused with X-ray procedures (CT-scan, other X-rays). Participants with chronic diseases may also have greater misclassification as they may be a sub-group that receives more “uncommon” radiological examinations (categorized as unexposed, but truly high exposed), or non-IR diagnostic examination (magnetic resonance imaging) that may be confused with CT-scans. In this sense, the analyses excluding subjects who reported the medical conditions considered should be less affected by potential non-differential misclassification. Such analysis did not shown a decrease trend.

Our findings could also be explained by a tendency of controls to over-report or a tendency among cases to under-report the number of procedures, however, such differential bias has not previously been reported in similar studies [44, 45].

Reverse causation bias

In our analysis, we could not specifically identify examinations taken for the lymphoma diagnosis, and including them may lead to reverse causation bias. To avoid this, we excluded all examinations taken in the two years prior to diagnosis/reference date. The median number of excluded examinations in cases is in line with the number of examinations required for a lymphoma diagnosis (S3 Table), which is normally one CT-scan and one thoracic X-ray [49]. In addition, if reverse causation bias had an impact on cases, we would have expected a positive bias in ORs. Further, since analysis stratified by type of controls lead to similar results, it is unlikely that our risk estimates were affected by hospital controls possibly receiving medical X-rays during hospitalization.

Statistical power

To our knowledge, this is the largest case-control study evaluating medical IR exposure and subsequent lymphoma risk. However, the doses estimated were very low and hence the statistical power was limited to detect the small increase in risk that might be expected, based on current radiation risk estimates, from such dose levels. In addition, the literature review approach to exposure assignment could have introduced Berkson error [50], as we attributed exposure based on average dose values by time period, likely to further reduce the statistical power of the study. Analyses by sub-type of lymphoma were based on lower numbers and hence had even lower power. In our analysis, however, the ORs for NHL, CLL, and excluding CLL all showed similar and consistent reduced risks.

Our results should be interpreted in the context of the literature of lymphoma risk after exposure to medical diagnostic IR. The low study power and low exposure level makes it difficult for our study to reach adequate conclusions.