Bias due to non-consent in assisted reproductive treatment cohort studies: consent for disclosure to non-contact research in the Human Fertilisation and Embryology Authority register

Data availability in cohort studies is often subject to participants consenting to using their data for research purposes. As such, available cohorts are potentially made up of limited and selective groups of individuals, making participants increasingly less representative of the target populations. This is particularly concerning if prevalences and associations are inherently heterogeneous across different groups of individuals ( Hernán, 2017 ; Huitfeldt et al. , 2019 ); hence, results may not be generalizable to the target population. The main emphasis and discussion here will be on the simple selection bias, which can also be referred to as collider restriction bias ( Hernán et al. , 2004 ; Hammerton and Munafò, 2021 ; Nilsson et al. , 2021 ; Lu et al. , 2024 ), where the decision to consent is conditional on a set of factors related to the outcome(s) of interest (see Supplementary Fig. S1 ), leading to bias in the associations between ART outcomes and patient or treatment characteristics. In reproductive epidemiology, ART cohorts are a rich data source for studying ART effects and have been useful in studying growth and long-term health outcomes for ART offspring using follow-up or linkage studies ( Williams et al. , 2013 ; Hann et al. , 2018 ; Turner et al. , 2020 ; Magnus et al. , 2021 ; Terho et al. , 2021 ; Kondowe et al. , 2023 ; Sutcliffe et al. , 2023 ). Some of these cohorts may have specific patient consent procedures, which may result in patients being excluded from study samples. Although the exclusions due to non-consent may be minimal, a significant percentage of the study population can be excluded, resulting in potentially unrepresentative study samples prone to bias. Still, studies have been published using different ART cohorts, yet the consent issues, if they exist, have not been explored. In this study, we used the UK ART dataset from the Human Fertilisation and Embryology Authority (HFEA) as an example to explore the possible impact of consent on population characteristics and selected outcomes. The HFEA holds data on treatments and birth outcomes for all ART treatments conducted in licenced clinics in the UK since 1991. Since the HFE (Disclosure of Information for Research Purpose) Regulations 2010 came into effect in October 2009, the HFEA introduced a concept of specific prospective patient consent for the disclosure of their identifying information, termed ‘consent for disclosure’ (CD), which is requested from the patient at the start of an ART treatment cycle. This is applied in the context of clinical treatment and research, including both contact and non-contact research. At the same time, the regulations gave ‘presumed consent’ for patient-identifying information for all ART treatment cycles from 1991 to 2009 to be made available for research requiring bespoke data unless patients explicitly did not give CD if they returned for another treatment cycle after 2009 ( HFEA, 2010 ). As such, patients with cycles before October 2009 who returned for treatment after October 2009 and did not give CD have their pre-2009 cycles retrospectively recorded in the register as non-CD. This is also applied to patients starting their first treatment cycle post-2009, initially giving CD but failing to do so in a subsequent cycle. These pre-2009 exclusions also include a number of patients where CD cannot be unambiguously ascertained due to uncertainty in patient linkage. From October 2009 onwards, only the data for cycles in which patients gave CD for non-contact research can be made available for any research projects requiring linkage or detailed de-identified data with a potential risk of patient re-identification. Excluding the non-CD sample implicitly conditions on CD, leading to concerns that the available data may not be representative of all ART treatment cycles conducted after September 2009 (hereafter referred to as post-2009 cohort), hence prone to selection bias. For HFEA Register studies conducted prior to 2009, the loss due to lack of consent was small, but for later linkage studies ( Raja et al. , 2023 ) and any requests for identifying information, a substantial subset of patients are excluded. This study used an anonymized version of the HFEA ART dataset, including both CD and non-CD treatment cycles, to explore the impact of consent. We aimed to investigate differences between CD and non-CD groups, highlighting associations between giving CD and population characteristics. We also examined the relationship between CD and live birth (LB), child low birthweight (LBW), and macrosomia (common outcomes of interest in ART vs naturally conceived studies) outcomes. It has been demonstrated in recent statistics literature that weighting methods can potentially restore sample representativeness under specific assumptions ( Lesko et al. , 2017 ; Bonander et al. , 2019 ; Westreich et al. , 2019 ; Nilsson et al. , 2021 ), using auxiliary data to develop appropriate models. We obtained a limited set of auxiliary data to explore the degree to which such methods could be utilized to correct prevalence estimates when only CD data are available.

The HFEA provided data for all 819 512 (568 919 post-2009 and 250 593 pre-2009) non-donor IVF cycles conducted between 2004 and 2018. LBW and macrosomia outcome analyses were based on 155 986 IVF cycles that resulted in a singleton LB (post-2009: n = 115 334 and pre-2009: n = 40 652) ( Fig. 1 ). CD rates were lowest (16%) during the October–December 2009 inception of CD, rising to 64% in 2018, the last year available. Retrospective CD exclusions were highest (27%) immediately preceding CD introduction in 2009 (January to September 2009) and in the immediately preceding years ( Fig. 2 ). CD rates in the post-2009 and pre-2009 (retrospective non-CD) cohorts. The vertical red line represents the point at which CD was introduced. CD, consent for disclosure. Table 1 presents patient characteristics and ART treatment information for the CD and non-CD groups in the pre- and post-2009 cohorts for the whole sample and LB subset. Patient and treatment characteristics for the full dataset (all IVF cycles cohort) and LB subset (live births cohort) for the post-2009 and pre-2009 cohorts. Post-2009 (October 2009–2018), pre-2009 (January 2004–September 2009). 3.4% and 6.1% of cycles had missing maternal age in the All IVF cycles cohort non-CD post-2009 and non-CD pre-2009, respectively. LB, live birth; CD, consent for disclosure; eSET, elective single embryo transfer. Logistic regression results for both unadjusted and adjusted (including available pre-treatment factors) models presented in Table 2 show that maternal age and ethnicity were associated with giving CD and retrospective CD in the post- and pre-2009 cohorts, respectively. In the post-2009 cohort, fewer older and ethnic minority patients gave CD. The results were reversed in the pre-2009 cohort, where older patients and cycles conducted in the earlier years were likelier to remain in the assumed retrospective CD cohort. At the same time, the differences between ethnicities were less apparent since most (86%) of the ethnicity data were unavailable during this period. Previous LBs, pregnancies, and IVF treatments were also associated with CD in both the post- and pre-2009 cohorts ( Table 2 ). These associations persisted in sensitivity analysis models, including embryo transfer type and stage. Since 2009, CD rates have improved over time for all categories of ethnicity and maternal age ( Supplementary Fig. S2 ). Factors associated with CD in the post-2009 and pre-2009 cohorts. Unadjusted estimates show an association between each factor and CD without including other variables in the model. Adjusted estimates generated from models with all variables presented in the table plus year of treatment and causes of infertility (tubal, endometriosis, male factor, ovulatory, unknown). CD, consent for disclosure; OR, odds ratio; Ref, reference group. Supplementary aggregated data were used to explore CD rates by funding/clinic type (NHS/private/both NHS and private patients). During the study period, CD rates were consistently higher in NHS-funded cycles ( Supplementary Fig. S3 ) and lower for all funding types in women over 40 years old ( Supplementary Fig. S4 ). Unadjusted LBRs were slightly higher for the CD group in the post-2009 cohort, while the differences disappeared when the models were adjusted for the known and available confounders ( Fig. 3 ). The adjustment did not fully correct the differences in the pre-2009 cohort and the striking differences in 2008 and 2009 ( Fig. 3 ). In the retrospective pre-2009 cohort, LBRs (adjusted and unadjusted) were comparable for the CD subset and full population (CD plus non-CD) between 2004 and 2007. Although the non-CD patients had lower LBRs pre-2008, there were too few to materially bias the LBR estimates. Live birth rates in CD and non-CD groups in the post-2009 and pre-2009 cohorts alongside those of the full population (CD plus non-CD). The adjusted plot is generated from LBR models adjusted for year of treatment, mother’s age, ethnicity, infertility causes, number of previous IVF treatments, pregnancies and live births, embryo transfer type and stage, ICSI, ovarian stimulation, and elective single embryo transfer. The vertical red line is the point at which CD was introduced. CD, consent for disclosure; LBR, live birth rate. Compared to the full population, LBW rates were consistently but slightly higher in the non-CD group and lower in the CD group, and there were minimal differences in the prevalence of macrosomia during the post-2009 (2009–2018) period ( Fig. 4 ). Unadjusted low birthweight and macrosomia trends for the full population, CD, and non-CD groups. Low birthweight (birthweight  4000 g). CD, consent for disclosure. Although the differences between CD and non-CD groups were minimal, regression results show that there were slightly fewer LBW babies born to patients who gave CD (compared to those who did not give CD) in the post-2009 group in both unadjusted (OR: 0.86, 95% CI [0.83, 0.90]) and adjusted (OR: 0.92, 95% CI [0.86, 0.97]) models ( Table 3 ); the adjusted regression model reduced but did not fully remove the bias. These differences in LBW were not observed in the pre-2009 group. There were no significant differences in the proportion of macrosomic babies between CD and non-CD groups in the pre- and post-2009 cohorts ( Table 3 ). Association between CD and birthweight outcomes (LBW and macrosomia). Unadjusted estimates for the association between birthweight outcomes and CD. Model adjusted for the variables: child sex, gestation (≥36 weeks/<36 weeks), embryo transfer type and stage, ovarian stimulation, ICSI, year of treatment, mother’s age, infertility causes, number of previous ART treatments, and live births. Model further adjusted for ethnicity. CD, consent for disclosure; LBW, low birthweight; OR, odds ratio. The results demonstrated that conditional on covariates included in the weighting model, the distribution of variables (especially patient factors) in the post-September 2009 CD sample can be weighted to be comparable, though not identical, to that of the full ART population. Weighting the CD subsample did reduce the bias in LBR estimates in the post-2009 cohort but not in the pre-2009 cohort ( Supplementary Table S2 ). The estimated BW outcomes were less well corrected, where the CD sample was already very close to the full population, and the bias was dominated by stochastic variation.

The HFEA makes a highly anonymized dataset publicly available at https://www.hfea.gov.uk/about-us/data-research/ ]. To avoid the risk of re-identification, this dataset is severely limited, providing only a subset of the variables in the register (e.g. no treatment centre identification or details) and with important variables banded. An augmented version of this public dataset containing all CD and non-CD ART treatment cycles conducted from 2004 to 2018 with an extra field flagging which cycles were CD or non-CD was obtained from the HFEA for this study. The 2018 date represents the latest available data at the time of the study. We also obtained an aggregated version of this dataset to be able to describe the proportion of CD cycles by funding source or clinic type. In the UK, public funding via the National Health Service (NHS) is available for a minority of IVF treatment cycles, with strict eligibility criteria applied and with geographical constraints depending on the area of the UK. The majority of cycles are thus funded by patients (termed fee-paying or private). IVF clinics can (rarely) host only NHS-funded cycles or (much more commonly) only private cycles, with most clinics hosting both NHS and fee-paying/private cycles. Before 2009, data on funding for ART treatments were not recorded on the HFEA register. After 2009, we categorized centres as ‘private’ or ‘NHS’ funded if ≥90% of cycles recorded were labelled as private/predominantly private or public/primarily public clinics, respectively; otherwise, the clinics were classed as mixed ‘private and NHS’ funded. Ethical approval was obtained through the UK NHS Health Research Authority [20/PR/0505] and HFEA Register Research Panel [HFEA_2019_017]. Data on 819 512 autologous (own-gamete) ART cycles started between 2004 and 2018 were obtained from the HFEA ( Fig. 1 ). The primary analysis utilized the 819 512 cycles to explore factors associated with CD and live birth rates (LBRs). The cohort was split into two, before and after the introduction of CD: pre- and post-September 2009. We created LB subsets from the same periods to explore the birthweight outcomes for 155 986 singletons. For the birthweight analysis, we included only singletons listed as one LB with valid birthweight, gestation, and sex entries ( Fig. 1 ). Exclusions for the birthweight dataset of treatment cycles before and after the introduction of CD. Post-2009 (October 2009–2018) and pre-2009 (January 2004–September 2009). Retrospective non-CD is applied in the pre-2009 cohort only for returning patients (after 2009) who did not give CD. CD, consent for disclosure. The primary analysis utilized univariable and multivariable logistic regression models to explore the association between CD and available parental factors in order to estimate the influence of pre-treatment patient characteristics on CD rates over time, with and without adjusting for the confounding with other factors. The factors considered were: maternal age groups (18–34, 35–37, 38–39, 40–42, 43–44, 45–50 years old), ethnicity (White, Asian, Black, Mixed, or Other), previous IVF treatments (0, 1, 2, 3, 4, or 5+), pregnancies (0, 1, 2, or 3+), and LBs (0, 1, or 2+), and causes of infertility (tubal, endometriosis, male factor, ovulatory, unknown). In a sensitivity analysis, we explored the association of CD with embryo transfer type and stage, which (although possibly influenced by clinic policy) are events that may have some dependency on early treatment outcomes and happen after CD has already been given (or not). We also explored CD trends over time and for the different maternal age and ethnicity categories. In a secondary analysis, we explored the relationship between CD and LB (defined as one or more LBs following the treatment cycle) and birthweight (LBW (birthweight < 2500 g) and macrosomia (birthweight ≥4000 g)) as exemplar outcomes of interest which were available in this restricted dataset (noting that only banded birthweight was available). We fitted unadjusted and adjusted models to assess the magnitude of the bias and the degree to which this could be reduced by adjusting for the readily available confounders in the dataset. To explore the association between giving CD and LB or singleton birthweight (LBW and macrosomia) outcomes, binary logistic regression models were used to explore unadjusted (only including CD indicator in the model) and adjusted (for available predefined confounding covariates) models. Adjusted models included maternal age groups, ethnicity (including not-stated as a separate category), previous IVF treatments, pregnancies, and LBs, causes of infertility (binary indicators for the categories listed above), embryo transfer type (fresh or frozen) and stage (cleavage/blastocyst/missing), ovarian stimulation, ICSI, and elective single embryo transfer (eSET). Child sex and gestation (≥36 weeks/<36 weeks) were additionally included in the LBW and macrosomia models. We also compared LB rate (number of LBs per yearly treatment cycles started) trends in the CD and non-CD subsets and the full ART population (CD plus non-CD). The analyses were conducted separately for the pre- and post-September 2009 cohorts to capture differences in the derivation of CD (explicitly consenting in the post-2009 cohort and retrospective assignment in the pre-2009 cohort). Inverse probability of sampling weights (IPSW) were estimated using a logistic regression model for CD in the augmented dataset for the entire HFEA UK ART population. The estimated weights were then applied to the CD sample to assess if the distribution of population characteristics in the weighted CD sample and unweighted full ART population were comparable. The estimated IPSW models included the pre-treatment variables: maternal age, ethnicity, year of treatment, previous pregnancies, previous LBs, previous IVF treatments, and infertility causes, and further details are given in the Supplementary Table S1 . All statistical analyses were performed in Stata (version 17; Stata Corporation, College Station, TX, USA). 95% confidence intervals for the estimates are presented.

We have used the HFEA dataset to demonstrate that ART datasets where consent is necessary for inclusion in research studies can deliver significantly biased estimates of treatment effects and possibly erroneous conclusions. Such biases are likely in other datasets where explicit or implicit consent is required to use the data. Our work clearly shows that naïve analyses of HFEA data that do not take the above biases into account may potentially yield misleading conclusions. The HFEA, researchers, and journals publishing such research must be made aware of this. Careful consideration of the potential biases introduced by the required exclusion of non-consenting individuals is essential in analysing such data, along with appropriate statistical adjustment if possible. Standard adjustment approaches would involve conditioning on predictors of missingness while paying careful attention to model specification and avoiding complicating the analyses. Alternatively, if auxiliary data are available, weighting methods may be used to reweight the available CD sample to represent the target population in the distribution of important population characteristics. At least for the post-2009 treatments, the biases due to retrospective application of CD in the 2007–2009 period may not be correctable.

The study explored differences in population characteristics and outcomes between consenting and non-consenting groups in a national ART register dataset. The results highlight that in the context of CD of identifying information for non-contact research (CD), the CD sample in the HFEA dataset differs from the non-CD sample (and thus the full UK ART population) in the distribution of important population characteristics and study outcomes. This is the first study exploring the characteristics of UK ART patients who provide CD and thereby make their data available for research, as required for those studies wishing to use personal data from the UK HFEA national register. Overall, 55% of the ART cycles have given CD since its introduction in October 2009; rates of CD initially were very low (∼20%) but have improved steadily with time. We interpret this as a result of improved patient and clinic awareness of the value of CD to outcome and follow-up studies of maternal and ART offspring health and also probably linked to the enforcement of related regulatory requirements. There is an argument that moving to an opt-out system (similar to the NHS) may improve consent rates and increase the utility of the HFEA register for population-level research. In this system, the patients who withdraw CD may choose to change their choice and give CD at a later stage (and vice versa), even if they do not return for another treatment cycle. However, the possibility of improving CD rates under this system cannot be guaranteed. In many cohort studies, only consented data are available; thus, it is impossible to address the questions in this study. The use of the HFEA’s anonymized dataset, which includes all ART treatment cycles (CD and non-CD), made exploring differences between consenting and non-consenting cycles possible. The introduction of CD in 2009 means that bespoke data projects requiring more detailed data and linkage studies to other databases lose about 45% of the total dataset after 2009 ( Raja et al. , 2023 ). Using the remaining data would be appropriate if the CD cycles do not differ from the non-CD cycles on important variables. However, our results show that patient characteristics and health outcomes are associated with whether a patient gives CD (e.g. fewer older women and women from ethnic minorities give CD). It is known that ART cycles in younger mothers and mothers of white ethnicity are linked to improved LBRs and better neonatal outcomes ( Jackson et al. , 2015 ; Lean et al. , 2017 ; HFEA, 2021 ; Ghidei et al. , 2022 ; Henderson et al. , 2023 ), which are ultimately linked to later child outcomes. Based on this literature and our present study, it is likely that utilizing only the CD cohort may bias the results towards favourable outcomes, as more patients with a known poor prognosis are likely to be excluded from the CD group. Additional analyses revealed that CD rates are lower in private clinics than in NHS-funded clinics. This may reflect on clinic practices regarding CD, attitudes towards research, and differences in patient populations beyond those recorded in the register. With retrospective removal of cycles (explained above) in place, researchers need to be aware that both the pre-2009 and post-2009 cohorts are not static but change with time. Therefore, the cohort utilized in the first UK register linkage study ( Williams et al. , 2013 ) differs from that used in subsequent studies, e.g. Hann et al. (2018) . These retrospective non-CD rates are notably higher in the immediate years preceding the introduction of CD. On average, the retrospective non-CD proportion is 10% for cycles performed between January 2004 and September 2009. These would most likely be patients returning after 2009 because of an earlier failed treatment cycle or seeking a second IVF child, both of which will disproportionately be those with private funding, given the limitations on the number of publicly funded cycles, and will, in turn, bias estimates of success rates (albeit in different directions). These results highlight differences in the distribution of population characteristics between consenting and non-consenting patients, resulting in a consented subset that is not representative of the total ART population, leading to biased estimates in our exemplar analyses. We note that these characteristics are assessed here at the treatment cycle level rather than truly representing patient-level per se . This reflects the usual use of ART data, where analyses are conducted at cycle level with adjustment as appropriate for intra-patient correlations. The results also demonstrate differences in LB, LBW, and macrosomia rates between the CD subgroup and the full population. This scenario is not specific to the HFEA dataset but is possible in other ART cohorts. Here, the missing/non-consenting sample data cannot be assumed to be missing completely at random as the selection is (in part) shown to be associated with some important characteristics of the participants that may be related to outcomes of interest. Estimates of outcomes at the population level and naive comparisons between ART and naturally conceived children are thus potentially significantly biased, while cross-sectional comparisons within the ART samples are less likely to be substantially biased (see also Supplementary Fig. S1 ). Using externally derived weights based on patient characteristics did substantially reduce the bias in LBR estimates in the post-2009 cohort. However, it was not effective for this outcome pre-2009. This is not unexpected as CD in this cohort was only ascertained if patients returned for later cycles; such patients will predominantly be those whose earlier treatment had failed and will not be well identified from just the pre-treatment characteristics. BW estimates did not differ much between CD and non-CD patients, and so weighting was less useful for these outcomes. These weights will be particularly useful for following up on adverse health outcomes (using linkage studies) identified in the 1991–2009 cohort and studying the more recent ART generation (post-2009 cohort), as rapid changes in ART technology may impact these risks. It is imperative to note that the estimated weights were limited due to the unavailability of some important factors to include in the weighting models. For the weights to be truly effective, the HFEA should look to estimate and provide a more comprehensive set of weights based on more detailed sets of covariates which are not available or accessible in the anonymized dataset used in this paper. We could not explore important unmeasured confounders possibly associated with the decision to give CD or not, such as treatment centre, socio-economic status, and funding (private vs NHS), as these were not available in the primary dataset or recorded on the HFEA register. In addition, we used anonymized data with banded/categorized age, gestation, and birthweight, limiting our estimates’ precision. Data were unavailable beyond 2018, so we cannot explore how consent has changed in the most recent period. However, this period was significantly affected by the COVID-19 pandemic and may well be atypical. Even though our analysis was limited to available factors in the dataset, the results are still informative and show notable differences between CD and non-CD groups, although care is required in the interpretation as there are non-trivial correlations between the factors. Our results remain valid since the main aim is to understand the variables’ association with CD, which might then inform the readers’ understanding of the propensity score model. The current non-CD flagged cycles include some cycles where, due to issues linking multiple treatments for individuals, the HFEA cannot be certain if a patient gave CD, and it is possible that the numbers excluded may be reduced in future revisions of the register. The exemplar analyses are intended as illustrative rather than definitive, with limited covariates and no consideration of potential correlations between repeat cycles in the same patients.