As participation in epidemiologic studies wanes, there is increasing concern about the potential influence of nonparticipation selection bias impacting study findings.1–4 There have been increasing calls for the collection and evaluation of validation data in epidemiologic studies and quantitative approaches to bias analysis in both peer-review and regulatory settings,5,6 including examination of potential selection bias in case–control studies of weak associations.7 In the INTERPHONE study, a 13-country case–control study, which examined the association between cellular telephone use and risk of tumors of the head and neck in adults,8,9 participation was low, particularly among controls, and there was a higher prevalence of ever-regular cellular telephone use, and an earlier start date of use, among interviewed participants compared with nonparticipants responding to a brief nonrespondent questionnaire.10,11 It was estimated that selective nonparticipation could lead to an underestimation in ever cellular telephone use odds ratios (ORs) of approximately 10%, further emphasizing the importance of bias prevention and control in such studies.12
This article seeks to examine the potential impact of nonparticipation selection bias in the MOBI-Kids study. MOBI-Kids is a 14-country case–control study designed to investigate the potential effects of exposure to radiofrequency and extremely low-frequency electromagnetic fields from mobile telecommunications devices on brain tumor risk in children and young adults.13 This work describes differences in cellular telephone use and personal characteristics among interviewed MOBI-Kids study participants and refusers responding to a brief nonrespondent questionnaire. It also assesses the potential impact of nonparticipation selection bias on study findings by deriving selection bias factors under different hypothetical scenarios reflecting cellular telephone use in other nonparticipants without a nonrespondent questionnaire. Inverse probability of selection weights were estimated based on relevant predictors of study participation for use in analysis of cellular telephone associations in MOBI-Kids.
Details of the MOBI-Kids study are described elsewhere.13 Briefly, MOBI-Kids is a prospective case–control study conducted in 14 countries (Australia, Austria, Canada, France, Germany, Greece, India, Israel, Italy, Japan, Korea, New Zealand, Spain, and the Netherlands) from May 2010 through March 2016. Eligible cases were recruited from neurosurgery, radiology, and oncology units from tertiary centers, were between 10 and 24 years of age residing in the study regions, and were diagnosed with a first primary eligible brain tumor (benign or malignant brain tumor except those located at the base of the skull). We verified completeness of case ascertainment through cancer registry and/or hospital discharge records where available. We excluded case participants with language difficulties (i.e., not speaking the local language) or a known genetic syndrome related to brain tumors (i.e., neurofibromatosis). We recruited case participants rapidly following diagnosis with a maximum of 12 months between the case interview and the reference date (first image with suspicion of a space-occupying lesion). If an identified case was too ill to participate or had died, a proxy respondent (usually the parent) was sought.
Two hospital-based controls were recruited for each case among patients undergoing an appendectomy for suspected diagnosis of appendicitis (in an attempt to improve participation rates compared with population controls) in general and pediatric surgery departments from both smaller and tertiary hospitals with catchment areas reflecting those of the neurosurgery departments. Controls were matched by age (within 1 year of age for cases <17 years and within 2 years of age for cases ≥17 years to allow for closer matching among younger participants where patterns in cellular telephone use change more rapidly with age), sex, date of surgery/interview date (within 3 months), and region of residence. These criteria were relaxed somewhat where there were difficulties in control recruitment. We did not seek a proxy for controls as the number who died or were too ill was small.
Hospital staff recruited participants, generally presenting MOBI-Kids as a study of environmental risk factors including communication technologies in an attempt to minimize possible selection bias by cellular telephone use. Ethics approval was obtained from all appropriate national or local research ethics boards, and written informed consent was obtained prior to the study interview.
A total of 899 (72%) out of 1257 eligible cases and 1910 (54%) out of 3539 eligible controls agreed to participate and completed the study interview (Table 1). The main reasons for nonparticipation among both cases and controls were refusal (14% and 27%, respectively) and inability to trace (12% and 18%, respectively). There were also a small number of participants who did not participate due to medical refusal (n = 10), had died or were too ill and there was no proxy (n = 15), or another reason (n = 63) (i.e., unknown reason, data loss).
Trained interviewers generally interviewed participants in person. The main study questionnaire captured detailed data on demographics, history of communication technology use, medical history, and other factors. We only considered participants reporting ever having made or received on average at least one call per week for at least 3 months to be ever-regular users; they completed the detailed cellular telephone history questionnaire, which included questions regarding when they started using their phone (year and month, or if they could not recall a range, season, or age was provided), whether they still use the phone (i.e., current regular user), and if they had stopped, a stop year and month (or a range, season, or age). Participants were also asked to provide the average length of time spent making and receiving calls in the last 3 months during which they were using their current phone (or last phone if a former user) among other factors. Parents or guardians helped to complete the questionnaire for younger participants.
Interview refusers, including proxies, were asked to complete a brief nonrespondent questionnaire, where the ethics committee allowed, to examine whether they differed from interviewed participants according to cellular telephone use or other personal characteristics. These questionnaires were generally completed with a trained interviewer either in person or by phone directly following participant refusal, though in some cases they were self-completed at home and returned by mail (eAppendix 1; https://links.lww.com/EDE/B420). The nonrespondent questionnaire captured whether they were ever a regular user (ever having made or received on average at least one call per week for at least 3 months) (yes/no), start year of use (or range), if they were still using a phone regularly (yes/no), and the average length of time spent making and receiving calls in the last 3 months during which they were using their phone (for both current and former users) as well as sex, age, and maternal education level (high school or less, medium level technical/professional school, university/post-graduate, other). For interviewed nonrespondent questionnaire respondents, we also captured person interviewed (index subject, parent, other), place of interview (hospital, home, other), and responsiveness of person interviewed (not at all, fairly, or very cooperative, responsive, and interested). Three study countries (Austria, Greece, and India) did not collect nonrespondent questionnaires. In five countries (Australia, Canada, Korea, New Zealand, and the Netherlands), only a limited number of nonrespondent questionnaires were collected (n < 10) and were not examined here. Results are based on the remaining six study countries (France, Germany, Israel, Italy, Japan, and Spain) representing the majority of the interviewed MOBI-Kids study population (76% of cases and 79% of controls).
Cellular Telephone Use
We compared the percentage of ever-regular cellular telephone use and time since start of use of 5+ years (as an indicator of longer-term use) between interviewed participants and refusers responding to the nonrespondent questionnaire among both cases and controls. Multivariable logistic regression models were used to examine whether associations of interviewed status (interviewed participant versus nonrespondent questionnaire) and either ever use or time since start of use (5+ years) varied by case/control status, age group, or country.
Selection Probabilities and Bias Factors
We estimated selection probabilities for cases and controls according to ever-regular cellular telephone use and time since start of use (5+ years) in order to derive selection bias factors with which to estimate the magnitude of the potential impact of nonparticipation selection bias on cellular telephone associations in MOBI-Kids according to the methodology used in INTERPHONE.11 Eligible study participants were classified into four categories: (1) interviewed participants (i.e., completed the full study interview); (2) nonparticipants who refused to participate but completed the nonrespondent questionnaire; (3) nonparticipants who refused to participate and did not complete the nonrespondent questionnaire; and (4) nonparticipants for other reasons (i.e., unable to trace, medical refusal, et cetera) (eAppendix 2 and 3; https://links.lww.com/EDE/B420).
Because data on cellular telephone use were only available for the first two categories of participants, namely (1) interviewed participants and (2) refusers with a nonrespondent questionnaire, cellular telephone use was imputed for both (3) refusers without a nonrespondent questionnaire and (4) nonparticipants for other reasons according to the following hypothetical scenarios:
- (1) using a weighted mean of responses among interviewed participants and refusers with a nonrespondent questionnaire;
- (2) applying nonrespondent questionnaire response frequencies to refusers without a nonrespondent questionnaire, then for other nonparticipants, using a weighted mean of responses among interviewed participants and refusers;
- (3) applying nonrespondent questionnaire response frequencies to both refusers without a nonrespondent questionnaire and other nonparticipants;
- (4) assuming refusers without a nonrespondent questionnaire and other nonparticipants had more extreme differences in use from interviewed participants than nonrespondent questionnaire respondents, including scenarios of both lower and (5) greater prevalence of use; and
- (6) finally, a reference scenario was also considered assuming all refusers and other nonparticipants had the same pattern of use as interviewed participants.
We derived selection bias factors in terms of selection ORs applying either the same scenario to both cases and controls or different scenarios. For each scenario, we calculated selection ORs as follows: (S1aS0b)/(S0aS1b), where S1a represents the selection probability of participating in the full study interview for cases who were ever-regular cellular telephone users (or had a time since start of use of 5+ years) and S1b for controls; S0a for cases who were never regular cellular telephone users (or had a time since start of use of <5 years); and S0b for controls.11,14 Selection bias factors of 1.0 represent no selection bias, whereas those <1.0 represent the magnitude of the underestimation in the observed OR due to selective nonparticipation and those >1.0 the magnitude of the overestimation.
Inverse Probability of Selection Weights
Inverse probability of selection weights were estimated using data from the six included study countries here for application to all interviewed MOBI-Kids participants in an attempt to account for nonparticipation selection bias in analysis of cellular telephone associations in the full study population in a multivariable manner. We estimated the weights using mixed-effects logistic regression models with random country intercepts. First, age-adjusted associations of interviewed status (versus nonrespondent questionnaire) with case or control status, sex, maternal education, ever-regular cellular telephone use, current cellular telephone use, time since start of use (years), and average length of calls in the last 3 months during which they were using their phone (minute/week) were examined. An initial multivariable model was then fit including predictors significant at the P < 0.20 level. In the final model, predictors significant at the P < 0.05 level were retained, and interactions between all included variables were examined on the multiplicative scale. The functional form of age (continuous) was assessed using generalized additive models and was not found to deviate from linearity (not shown).
We then extended the population considered in the final multivariable model to include refusers without a nonrespondent questionnaire and other nonparticipants in order to derive inverse probability of selection weights based on the full eligible study population (i.e., interviewed status versus all nonparticipants). We used different hypothetical scenarios to impute missing data on cellular telephone use by age group as only data on case/control status and age were available for refusers without a nonrespondent questionnaire and other nonparticipants here. We also assessed the significance of including a random country effect for each included model variable. We then derived the inverse probability of selection weights from the inverse of the fitted values of the regression models, using the individuals’ probability of selection from an intercept-only model as the numerator to stabilize the weights, and applied them to participants in all 14 study countries. Statistical analysis was conducted using Stata version 12.1 and R version 3.3.15,16
Among the 127 case refusers in the six included study countries, 61% (n = 77) completed the nonrespondent questionnaire (eFigure 1; https://links.lww.com/EDE/B420). Among 788 control refusers, 63% (n = 498) completed the nonrespondent questionnaire. Nonrespondent questionnaires were completed only by interview refusers and not by other types of nonparticipants (i.e., unable to trace, medical refusal, et cetera). We compared nonrespondent questionnaires completed by the 77 case and 498 control refusers with responses from 683 interviewed cases and 1501 controls (suspected appendicitis patients) in the six countries. The distribution of responses by respondent type is presented in eAppendix 4; https://links.lww.com/EDE/B420. The total number of nonrespondent questionnaires ranged from 37 in Japan to 158 in Spain. There were more males among nonrespondent questionnaire cases than interviewed cases. Nonrespondent questionnaire controls tended to be older than interviewed controls and had a lower maternal level of education.
Cellular Telephone Use
The prevalence of ever-regular use was generally similar among interviewed and nonrespondent questionnaire controls ranging from 92% to 98% in the 15- to 19- and 20- to 24-year-old age groups (Table 2). In the youngest 10- to 14-year-old age group, the prevalence of ever-regular use was somewhat higher among interviewed (68%) versus nonrespondent questionnaire (62%) controls. The prevalence of ever-regular use was also similar among interviewed and nonrespondent questionnaire cases ranging from 97% to 100% in the 20- to 24-year-old age group, whereas it was somewhat higher among interviewed versus nonrespondent questionnaire cases in both the 15- to 19-year-old (95% vs. 89%) and 10- to 14-year-old (63% vs. 48%) age groups.
Interviewed controls were more likely to have a time since start of use of 5+ years than nonrespondent questionnaire controls in both the 15- to 19-year-old (64% vs. 47%) and 20- to 24-year-old (94% vs. 88%) age groups (Table 2). A similar finding was also observed among cases (66% vs. 37% 15- to 19-year-old age group and 94% vs. 82% 20- to 24-year-old age group). Fewer cases or controls in the 10- to 14-year-old age group had a time since start of use of 5+ years. Associations of interviewed status and either ever use or time since start of use (5+ years) were not found to vary by case/control status, age group, or country (P > 0.05).
Selection Probabilities and Bias Factors
Scenarios of cellular telephone use among refusers and other nonparticipants without a nonrespondent questionnaire for both cases and controls are presented in eAppendix 2 and 3; https://links.lww.com/EDE/B420. The ratios of selection probabilities according to both ever use and time since start of use (5+ years) were small but somewhat greater among controls than cases for each of the different scenarios examined, due to differences in study participation rates.
Table 3 presents selection ORs estimated for various combinations of scenarios of cellular telephone use among cases and controls. For ever-regular use, selection ORs ranged from 0.96 to 0.97, applying nonrespondent questionnaire results to nonrespondent questionnaire respondents only (Aa), all refusers (Bb), or all nonparticipants (Cc) using the same scenario for both cases and controls. Selection ORs ranged from 0.92 (Ca) to 1.00 (Ac) applying different combinations of these scenarios to cases and controls. The largest selection bias factors were observed with assumptions of larger differences in use among nonparticipants without a nonrespondent questionnaire, ranging from 0.73 under a scenario of 10% less use among refusers and other nonparticipating controls without a nonrespondent questionnaire and 10% greater use among cases (De) to 1.34 under the opposite scenario (Ed).
For time since start of use of 5+ years, selection ORs ranged from 0.92 to 0.94 under scenarios Aa, Bb, or Cc, applying the same scenario to both cases and controls, and 0.86 (Ca) to 1.00 (Ac), applying different combinations of these scenarios. The largest selection bias factors, of 0.76 and 1.16, were obtained with assumptions of larger differences in use among refusers and other nonparticipants without a nonrespondent questionnaire (scenarios De and Ed, as well as Rd, respectively).
Inverse Probability of Selection Weights
Age-adjusted associations of interviewed status (versus nonrespondent questionnaire) with various individual demographic and cellular telephone use characteristics are presented in Table 4. Variables retained in the final multivariable model included age, case or control status, and, due to collinearity, a combined time since start of use (years) and average weekly length of calls variable (minute/week), collapsing the 5–9 and 10+ years time since start of use categories. There was no evidence for interactions between any included model variables (P > 0.05), though an interaction term between case/control status and the combined cellular telephone use variable was forced into the model to allow for differential participation by cellular telephone use history. Last, a random country effect for case/control status was also included in the final model (P < 0.05) (eAppendix 5; https://links.lww.com/EDE/B420).
On extending models to include all refusers and other nonparticipants without a nonrespondent questionnaire, ORs were largely similar across the different model scenarios, imputing missing data on cellular telephone use history by age group. The best model fit, according to the Akaike Information Criterion, was for scenario C, imputing missing cellular telephone use history data according to the distribution among nonrespondent questionnaire respondents. Mean stabilized inverse probability of selection weights estimated for interviewed participants in all 14 MOBI-Kids study countries were 1.09 for controls and 0.80 for cases and ranged from 0.60 to 1.71 in scenario A to 0.58 to 2.31 in scenario C (eAppendix 6; https://links.lww.com/EDE/B420).
Inverse probability of selection weights were also estimated in an alternative model including only time since start of use (years) rather than the combined time since start of use (years)/average weekly length of calls (minute/week) variable because data on average length of calls represent use in the previous 3 months, which may be affected by the developing tumor in cases, with a similar distribution of inverse probability of sampling weights obtained (eAppendix 7 and 8; https://links.lww.com/EDE/B420).
Table 5 presents an example of the weighted distribution of ever use and time since start of use of 5+ years by age group in the six included countries here. There were only small changes in the distribution of cellular telephone use observed.
This article sought to describe differences in cellular telephone use and personal characteristics among interviewed participants and refusers responding to a nonrespondent questionnaire as well as the potential impact of nonparticipation selection bias in MOBI-Kids. The prevalence of ever-regular use was generally similar among controls and cases completing either the full study interview or the nonrespondent questionnaire in the 15- to 19- and 20- to 24-year-old age groups, whereas in the 10- to 14-year-old age group, it was somewhat higher among controls and cases completing the full study interview (68% vs. 62% for controls and 63% vs. 48% for cases). Interviewed controls and cases, particularly in the 15- to 19-year-old age group, were more likely to have a time since start of use of 5+ years (64% vs. 47% for controls and 66% vs. 37% for cases).
Selection bias factors ranged from 0.96 to 0.97 for ever-regular use and 0.92 to 0.94 for time since start of use of 5+ years applying nonrespondent questionnaire results to either nonrespondent questionnaire respondents only (Aa), all refusers (Bb), or all nonparticipants (Cc) for both cases and controls indicating a potential 3%–4% and 6%–8% downward bias in cellular telephone ORs, respectively. Although there were somewhat stronger selection ORs with a time since start of 5+ years, possibly suggesting greater differential nonparticipation among longer-term users, supplemental analysis examining selection bias factors according to increasing time since start of use categories, from 1+ through 10+ years, for these same scenarios (Aa, Bb, and Cc) generally revealed similar findings, though with somewhat stronger bias factors with either 4+ or 5+ years since start of use observed (eAppendix 9; https://links.lww.com/EDE/B420) paralleling greater differences in use between interviewed and nonrespondent questionnaire participants (P2/P1 ratio), particularly in the younger 10- to 14- and 15- to 19-year-old age groups. There were fewer users with longer times since start of use, largely in the 20- to 24-year-old age group, and somewhat smaller differences in use between interviewed and nonparticipating subjects resulting in more limited differences in selection probabilities by use category. Findings in INTERPHONE suggested a more recent time of start of use among nonparticipating cases and controls.11
Selection bias factors also varied and were larger when applying different scenarios of use among refusers and other nonparticipants without a nonrespondent questionnaire to cases and controls and were sometimes >1.0 including in scenarios assuming bias among nonparticipating cases only but not controls or in more extreme scenarios of use among nonparticipating cases. However, differences in use among interviewed and nonrespondent questionnaire participants were generally similar by case and control status here.
Potential nonparticipation selection bias was examined in other studies of cellular telephone use. A greater prevalence of regular cellular telephone use among participating than nonparticipating controls was reported in a study of uveal melanoma risk.17 In the CEFALO study of adolescent brain tumors, there were higher participation rates in both cases (83%) and controls (71%) and only minor differences in participation among cellular telephone users and nonusers.18
Strengths of this study include the collection of data on personal and cellular telephone use characteristics from study refusers with which to examine potential nonparticipation selection bias and to estimate inverse probability of selection weights for use in analysis of cellular telephone associations in the MOBI-Kids study. Limitations include small numbers of nonrespondent questionnaire respondents (n = 77 cases and 498 controls). Although a substantial effort was made to maximize nonrespondent questionnaire completion rates, the small number of these questionnaires, particularly among cases, limits our assessment of potential nonparticipation selection bias here, including in the estimation of inverse probability of selection weights. Although nonrespondent questionnaires were completed by a majority of refusers in included countries here (61% and 63% of case and control refusers, respectively), they may differ from other refusers and nonparticipants, including those in the other eight MOBI-Kids study countries, in unpredictable ways.19 Reasons for nonparticipation and willingness to complete the nonrespondent questionnaires are also likely different for case and control participants, with cases possibly more likely to refuse due to reasons surrounding the seriousness of a brain tumor diagnosis, whereas controls may simply be less interested or distracted. As such, selection ORs were examined in different scenarios of hypothetical use among other nonparticipants without a nonrespondent questionnaire. There was also little evidence for differential nonparticipation in inverse probability of selection weight modeling with interaction terms between case or control status and cellular telephone use weak and nonsignificant. There were only small changes in the distribution of personal and cellular telephone use characteristics following application of inverse probability of selection weights to interviewed cases and controls, indicating that their likely impact on overall study findings is likely to be modest.
There were limited data collected from nonrespondent questionnaire respondents; as such it is possible that there may be other unmeasured factors that may influence study participation not captured here.20 Data on average length of calls represent use in the previous 3 months and may not reflect longer-term use patterns. There were also no data on changes in cellular telephone use over time or on stop year for former users. Included cases here (both interviewed and nonrespondent questionnaire responders) were somewhat more likely to be former users (6%) than controls (3%). There were also no data on age or cellular telephone use history of the parent(s).
There are also differences in interview characteristics which may affect the quality and accuracy of responses. Responses may differ as to whether the participant completed the lengthy full study or brief nonrespondent questionnaire interview. Interviewed participants were more likely to be interviewed in person either in the hospital (58% interviewed cases, 53% interviewed controls) or at home (33% interviewed cases, 32% interviewed controls), while nonrespondent questionnaire respondents were more likely to complete the questionnaire in another location (i.e., telephone interview, self-complete the questionnaire at home) (66% nonrespondent questionnaire cases, 76% nonrespondent questionnaire controls). Nonrespondent questionnaire respondents were also more likely to be considered uninterested or not at all cooperative or responsive (10% of nonrespondent questionnaire cases and 17% of nonrespondent questionnaire controls) than interviewed participants (2% interviewed cases and controls) by the study interviewer. However, results were similar when excluding such participants from analysis (not shown). There were also no data on brain tumor histology available for nonparticipating cases, as these data were only captured during the full study data collection process.
We did not consider here other sources of bias, including recall bias.18,21,22 Recall and motivation may differ between interviewed participants and nonrespondent questionnaire respondents.23 A related validation study, MOBI-Expo, comparing questionnaire responses to recorded data on use based on software-modified smartphones reported that participants tended to underestimate number of calls but overestimate call duration (minute/week).24 A similar pattern was also observed in INTERPHONE.22 Another validation study using operator records for number and length of calls for consenting participants was also conducted as part of MOBI-Kids and is currently being analyzed.13 Further work to account for both recall and selection biases, including in probabilistic multiple-bias modeling, may be useful.25,26
Last, this study examines only one possible source of selection bias, namely nonparticipation selection bias. There may also be some sort of bias due to the largely hospital-based nature of the study, both in cases due nonparticipating hospitals and difficulties in accessing eligible participants, and in the use of suspected appendicitis controls possibly limiting their representativeness.13 Though we used appendicitis patients in an attempt to improve control participation rates, control participation remained low here with 54% of eligible controls completing the study interview, similar to that of INTERPHONE.11 Control participation, however, was greater among hospital controls in Germany, the only country that also recruited population controls.13 Although we assume that appendicitis patients are representative of the general population and that appendicitis is unrelated to both socioeconomic status and use of mobile telecommunications devices, selection bias factors should be interpreted with caution should by chance this be found to differ.
In conclusion, although results were limited by small numbers of nonrespondent questionnaire respondents, selection bias factors for both ever-regular use and time since start of use of 5+ years generally indicated a small underestimation in cellular telephone ORs. We estimated inverse probability of selection weights in an attempt to account for nonparticipation selection bias in analysis of MOBI-Kids.
The authors would like to thank Dr Xavier Basagaña for helpful comments throughout the course of this work. We would like to thank all of the research assistants and interviewers in the study centers for their enormous efforts to ensure that the study was carried out with care and with due consideration for the participants. All the hospital services (neurosurgery, neurology, oncology, pediatric oncology, general surgery, et cetera) are gratefully acknowledged for their cooperation and collaboration in recruiting participants. Finally, we would like to thank all participants and their relatives who took part in this study. We appreciate it very much. France: we would like to thank Dr Martine Hours for her valuable advice and assistance in the implementation of the study in France. We are grateful to Dr Luc Bauchet for supporting the implementation of the study and for helping us in the solicitation of neurosurgeons. Germany: we are very grateful to Vanessa Kiessling, Jenny Schlichtiger, and Iven-Alex Heim for their dedication to the conduct of the study in Germany. Also, we would like to thank Susanne Brilmayer and Silke Thomas for their contribution to the planning and implementation of MOBI-Kids in Germany. Italy: we thank all cooperating professionals who made this study possible, Antonio Argentino, Anna Maria Badiali, Emma Borghetti, Valentina Cacciarini, Laura Davico, Laura Fiorini, Francesco Marinelli, Sara Piro, and Caterina Salce. Japan: we appreciate generous support and advice offered by the Japanese Epidemiological Committee. Spain: we would like to thank Carmen Caban for setting up the complex electronic database. We appreciate the enormous time and effort of the regional coordinators (Marta Cervantes, Eva Ferreras, Irene Gavidia, Rebeca Sánchez, Angeles Sierra, and Angela Zumel).
1. Karvanen J, Tolonen H, Härkänen T, Jousilahti P, Kuulasmaa K. Selection bias
was reduced by recontacting nonparticipants. J Clin Epidemiol. 2016;76:209217.
2. Lim S, Immerwahr S, Lee S, Harris TG. Estimating nonresponse bias in a telephone-based health surveillance survey in New York City. Am J Epidemiol. 2013;178:13371341.
3. Mazloum M, Bailey HD, Heiden T, Armstrong BK, de Klerk N, Milne E. Participation in population-based case-control studies: does the observed decline vary by socio-economic status? Paediatr Perinat Epidemiol. 2012;26:276279.
4. Tolonen H, Ahonen S, Jentoft S, Kuulasmaa K, Heldal J; European Health Examination Pilot Project. Differences in participation rates and lessons learned about recruitment of participants–the European Health Examination Survey Pilot Project. Scand J Public Health. 2015;43:212219.
5. Fox MP, Lash TL. On the need for quantitative bias analysis in the peer-review process. Am J Epidemiol. 2017;185:865868.
6. Lash TL, Fox MP, Cooney D, Lu Y, Forshee RA. Quantitative bias analysis in regulatory settings. Am J Public Health. 2016;106:12271230.
7. Mezei G, Kheifets L. Selection bias
and its implications for case-control studies: a case study of magnetic field exposure and childhood leukaemia. Int J Epidemiol. 2006;35:397406.
8. Cardis E, Richardson L, Deltour I, et al. The INTERPHONE study: design, epidemiological methods, and description of the study population. Eur J Epidemiol. 2007;22:647664.
9. INTERPHONE Study Group. Brain tumor risk in relation to mobile telephone use: results of the INTERPHONE international case–control study. Int J Epidemiol. 2010;39:675694.
10. Vrijheid M, Deltour I, Krewski D, Sanchez M, Cardis E. The effects of recall errors and of selection bias
in epidemiologic studies of mobile phone use and cancer risk. J Expo Sci Environ Epidemiol. 2006;16:371384.
11. Vrijheid M, Richardson L, Armstrong BK, et al. Quantifying the impact of selection bias
caused by nonparticipation in a case-control study of mobile phone use. Ann Epidemiol. 2009;19:3341.
12. Saracci R, Samet J. Commentary: Call me on my mobile phone.or better not?–a look at the INTERPHONE study results. Int J Epidemiol. 2010;39:695698.
13. Sadetzki S, Langer CE, Bruchim R, et al. The MOBI-Kids study protocol: challenges in assessing childhood and adolescent exposure to electromagnetic fields from wireless telecommunication technologies and possible association with brain tumor risk. Front Public Health. 2014;2:124.
14. Greenland S. Rothman KJ, Greenland S. Basic methods for sensitivity analysis and external adjustment. In: Modern Epidemiology. 1998:2nd ed. Philadelphia, PA: Lippincott, Williams & Wilkins; 343358.
15. StataCorp. Stata Statistical Software: Release 12.1. 2011.College Station, TX: StataCorp LP.
16. R Core Team. R: A language and Environment for Statistical Computing
. 2016.Vienna, Austria: R Foundation for Statistical Computing.
17. Stang A, Schmidt-Pokrzywniak A, Lash TL, et al. Mobile phone use and risk of uveal melanoma: results of the risk factors for uveal melanoma case-control study. J Natl Cancer Inst. 2009;101:120123.
18. Aydin D, Feychting M, Schüz J, et al. Impact of random and systematic recall errors and selection bias
in case–control studies on mobile phone use and brain tumors in adolescents (CEFALO study). Bioelectromagnetics. 2011;32:396407.
19. Lash TL, Fox MP, MacLehose RF, Maldonado G, McCandless LC, Greenland S. Good practices for quantitative bias analysis. Int J Epidemiol. 2014;43:19691985.
20. Cole SR, Hernán MA. Constructing inverse probability weights for marginal structural models. Am J Epidemiol. 2008;168:656664.
21. Kleinbaum DG, Morgenstern H, Kupper LL. Selection bias
in epidemiologic studies. Am J Epidemiol. 1981;113:452463.
22. Vrijheid M, Armstrong BK, Bédard D, et al. Recall bias in the assessment of exposure to mobile phones. J Expo Sci Environ Epidemiol. 2009;19:369381.
23. Goedhart G, Vrijheid M, Wiart J, et al. Using software-modified smartphones to validate self-reported mobile phone use in young people: a pilot study. Bioelectromagnetics. 2015;36:538543.
24. Goedhart G, van Wel L, Langer CE, et al. Recall of mobile phone usage and laterality in young people: the multinational Mobi-Expo study. Environ Res. 2018;165:150157.
25. Greenland S. Multiple-bias modelling for analysis of observational data. J R Stat Soc Series A. 2005;168:267306.
26. Momoli F, Siemiatycki J, McBride ML, et al. Probabilistic multiple-bias modeling applied to the Canadian data from the interphone study of mobile phone use and risk of glioma, meningioma, acoustic neuroma, and parotid gland tumors. Am J Epidemiol. 2017;186:885893.