School-age outcomes among IVF and ICSI-conceived children: a causal inference analysis using linked population-wide data.

A causal inference framework was employed to answer the research question. This approach attempts to simulate a randomised trial by [ 1 ] requiring an a priori statistical analysis protocol that specifies both a hypothetical target randomised control trial and its proposed emulation using observational data; [ 2 ] addressing a causal question reflecting the effect of an intervention at a specific clinical decision point on a prespecified outcome; and [ 3 ] using inverse probability weighting via a propensity score (PS) model to balance covariate distributions between exposed and control populations, with the aim of producing exchangeable comparison groups and eliminating selection bias [ 20 , 21 ]. By conducting the study from a causal perspective and adhering to the conditions of a hypothetical target trial, causal conclusions can be made about the association between fertilisation by ICSI and childhood developmental vulnerability, compared with standard IVF. The first step of the target trial emulation was to describe the ideal hypothetical intervention trial required to answer the research question. Each component of this hypothetical trial was then assessed against the data available in our retrospective cohort (Additional file 1 : Target trial emulation). This process helped determine how closely we could emulate the results of the ideal target trial using the observational data available for analysis, and thus minimise as many sources of bias as possible. Limitations to trial emulation were acknowledged and strategies for overcoming them were outlined in our a priori statistical analysis plan (SAP) (Additional file 2 : Statistical analysis plan [ 3 , 21 – 32 ]). The source population included Victorian singleton live births between 2005 and 2013 conceived via assisted reproductive techniques (ART). The three largest ART units operating in Victoria at the time provided data on all cycles that resulted in a live birth during the study period. Linked maternal and child data pairs were obtained using Victorian Perinatal Data Collection data and birth records from the Victorian Births, Deaths and Marriage registry. Data linkage was performed by the Centre for Victorian Data Linkage (CVDL), a third-party government-funded data linkage unit [ 33 ]. Post-linkage false matches and duplicates were removed. A modified Delphi survey [ 34 ] was performed to ensure robustness of our exposure definition and analysis plan, including which covariates to include in our analysis model (Additional file 3 : Delphi survey results). This process was necessary since practices vary globally, especially relating to what constitutes a moderate or an absolute indication for ICSI [ 35 , 36 ]. Fertility specialists, embryology scientists and experts in perinatology, epidemiology and education were surveyed (22 of 30 respondents completed three survey rounds). Our study sought to examine outcomes for children conceived via ICSI in the absence of any absolute indication for ICSI. Delphi consensus established the absolute indications for ICSI (standard IVF would not be possible), and thus which births to exclude from our analysis. These included surgically extracted sperm (e.g. percutaneous epididymal sperm aspiration, testicular sperm aspiration, microscopic testicular retrieval of sperm), severe male factor infertility (e.g. OAT oligoasthenoteratozoospermia), oocyte thaw cycles, pregestational testing for monogenetic disorders or chromosomal translocations and total motile sperm count less than 2 million on the day of egg collection. In keeping with causal methodology, these cases with Delphi consensus absolute indications for ICSI were excluded since they could not be feasibly be assigned to the non-treatment (standard IVF) group in a target trial. Exposure was thus defined as conception via in vitro fertilisation (IVF) using intracytoplasmic sperm injection (ICSI) (non-absolute indications) compared with IVF conception using standard insemination. Method of oocyte insemination for the embryo transferred was clearly documented within the IVF database. Singleton pregnancies resulting from double embryo transfer where two different methods were used for insemination, and consequently fertilisation method of the implanted embryo was unable to be known were excluded. Based on Delphi consensus, the covariates considered to be potential determinates of the exposure were year of oocyte insemination, maternal age (at birth of the child), socioeconomic status, parity, language background other than English (LBOTE), both maternal and paternal highest obtained level of education, as well as laboratory related indices such as number of eggs on day of egg collection, presence of mild-moderate male factor infertility and specific categories of female factor subfertility (e.g. tubal factor, ovulation disorder). Determinants of the outcome included maternal age (at birth of the child), socioeconomic status, parity, language background other than English (LBOTE), both maternal and paternal highest obtained level of education, census year, child’s age in years at assessment and sex of the child. Language background other than English is recorded by the AEDC to identify children from non-English speaking families, which may impact the assessment of language-based aspects of childhood development. Gestational age at birth, mode of delivery and birthweight were considered mediators on the causal pathway and were therefore not adjusted for in this analysis. Our post-Delphi statistical analysis plan and directed acyclic graph were agreed upon and signed off by all authors prior to the commencement of data analyses (Additional file 2 : Statistical analysis plan). Childhood developmental outcome was assessed using a standardised, national assessment, The Australian Early Development Census (AEDC) [ 37 – 42 ] conducted every 3 years across Australia. All schools, including special schools, are eligible to participate. Children are assessed in their first year of primary school and all children in Victoria start school in February in the year they turn five (by end of March) or six. The AEDC is a teacher-reported measure that assesses broad childhood functional development at school entry (age 4–6) across five domains: physical health and wellbeing, social competence, emotional maturity, language and cognitive skills (school-based) and communication skills and general knowledge. Our primary outcome used a validated global measure—developmental vulnerability—defined as scoring less than the 10th percentile in two or more of the five developmental domains—“DV2” [ 43 ]. Secondary outcomes included developmental vulnerability (less than 10th percentile) in each of the five individual domains. Within the AEDC, teachers are also able to record any diagnosis that may affect development, including vision or hearing impairment, autism spectrum, attention deficit hyperactive disorder or other physical and cognitive conditions. Children with these conditions are noted as “special needs” if a diagnosis has been made prior to school entry or “emerging needs” if a diagnosis if suspected or under assessment. The presence of a special or emerging need was also examined as an outcome in a secondary analysis. Descriptive statistics are presented for each cohort by exposure status, according to type of data. We performed a detailed examination of the patterns of missing data and its frequency by exposure status. For the AEDC outcome data, the major mechanism for missing outcome was considered “informative” or “missing not at random” (MNAR), where a child was flagged as having “special needs”, the domain score of the metric is arbitrarily deemed invalid by the AEDC, and was not provided to us in the dataset, resulting in missing outcome data. Estimation bias related to missing outcomes due to “special needs” status was managed conservatively by deterministically imputing all missing outcomes as “developmentally vulnerable” in DV2 and the individual domains. Rarely, the reason for missing outcome data was considered “non-informative” that is the mechanism was considered “missing at random” (MAR) conditional on measured covariates. For example, where a teacher was unable to sufficiently complete response tool due to time constraints. If the frequency of this outcome MAR missingness was both of small magnitude and did not differ by exposure status so its exclusion is very unlikely to result in biased inference. Consequently, there were 13 children that were excluded from analysis. We considered that covariate data were likely to be MAR and imputation was performed using fully conditional specification that included birth mother to provide standard errors adjusted for maternal clustering [ 31 , 32 , 44 ]. Provided the imputation model is correctly specified and data are missing at random, multiple imputation methods can provide least biased estimates even with high proportion of covariate missingness [ 45 ]. All imputation models included exposure; both overall and domain outcomes; potential analysis model covariates including interaction terms; and auxiliary variables (see Additional file 4 : Missing data diagnostics—for details). Standard diagnostics were performed (Additional file 4 ). To obtain analysis model standard errors that account for the variability induced by imputation, we used the two-stage method detailed by von Hippel and Bartlett [ 35 ]. This involved creating 1000 bootstrapped samples from the original dataset; performing two imputations on each of these 1000 samples and then running the analysis model on each of these 2000 datasets, allowing us to obtain pooled point estimates and associated standard errors using one-way ANOVA. The target estimand for the primary and secondary outcomes is the difference in the potential outcome means between exposure groups for AEDC defined vulnerability at school entry. These estimands are presented as the ATE (average treatment effect) relative risk (RR) and risk difference (RD) point estimates and 95% confidence intervals (CIs), based upon the potential outcomes framework for causal inference [ 46 ]. Each estimator required three sequential stages (i) bootstrap and impute datasets (vide supra), (ii) run an augmented doubly robust inverse-probability-weighted regression adjustment (AIPW model [ 47 , 48 ] on each of the 2000 datasets and (iii) combine estimates using one way ANOVA to compute the ATE RD and RR estimates. Initial modelling included a selection model with 16 terms including two interaction terms (child age at Feb 1st in test year with exposure and gender) and one quadratic term (child age at Feb 1st in test year) and a regression adjustment model with 12 terms. Prespecified sensitivity modelling using Targeted Maximum Likelihood Estimation (TMLE) that accounts for both interaction and non-linear relationships between exposure and covariates was also undertaken for both the cohort excluding and the cohort including donor egg, donor sperm and IVF laboratory (see Additional file 5 : Sensitivity analysis). Standard regression diagnostics were used to assess the adequacy of covariate balance (Additional file 6 : Covariate balance). All statistical analyses were performed using Stata MP version 18.0, including the t -effects and multiple imputation suites of commands. An exploratory secondary analysis was performed to determine the association between male subfertility and childhood developmental outcome. This was an analysis that did not attempt to generate an effect estimate with a causal interpretation, comparing cases of ICSI with male factor subfertility to cases of ICSI without male factor indication/subfertility; all cases of standard insemination were excluded.

The total original cohort of ART conceived children included 16,054 singleton births in Victoria between 2005 and 2013. Among this cohort, 4700 were linked to AEDC outcome data. As the AEDC is conducted triennially, the biggest determinate of successful linkage was year of birth (Fig.  1 and Table  1 ). To remove the potential of confounding by indication and ensure consistency with the causal framework, we excluded cases with documented severe male factor infertility ( n  = 630), donated gametes ( n  = 379), egg thaw cycles ( n  = 11) and missing exposure status ( n  = 9). This left a final study cohort of 3656 children including 1489 IVF-conceived and 2167 ICSI-conceived children (Fig.  1 ). Fig. 1 Participant flow chart. Graphical representation of study population illustrating datasets used in data linkage Table 1 Descriptive statistics—baseline demographics of primary analysis population Whole cohort IVF with standard oocyte insemination Intracytoplasmic sperm injection N 3656 1489 2167 Child baseline data Sex (% female) 48.7 44.9 51.3  (number) % missing (0) 0.0 (0) 0.0 (0) 0.0 Year of birth ( n )  2005 23 6 17  2006 869 370 499  2007 188 83 105  2008 37 16 21  2009 925 408 517  2010 170 84 86  2011 27 12 15  2012 1173 431 742  2013 244 79 165  (number) % missing (0) 0.0 (0) 0.0 (0) 0.0 Age of assessment (median) [IQR] 5.50 [5.19 to 5.73] 5.52 [5.20 to 5.74] 5.49 [5.18 to 5.72]  (number) % missing (1) 0.03 (1) 0.07 (0) 0.0 Language background other than English (%) 13.8 12.8 14.4  (number) % missing (0) 0.0 (0) 0.0 (0) 0.0 Aboriginal/Torres Strait Islander (%) 0.4 0.4 0.4  (number) % missing (0) 0.0 (0) 0.0 (0) 0.0 Birthweight in grams (mean (SD)) 3349 (± 571) 3344 (± 570) 3353 (± 573)  (number) % missing (0) 0.0 (0) 0.0 (0) 0.0 % SGA (< 10th centile for gestation) 0.8 0.7 0.8  (number) % missing (0) 0.0 (0) 0.0 (0) 0.0 Gestational at delivery (median) [IQR] 39.1 [38.1 to 40.1] 39.0 [38.0 to 40.1] 39.1 [38.3 to 40.1]  (number) % missing (7) 0.19 (5) 0.34 (2) 0.09 Method of delivery (% of each category)  Normal vaginal birth 33.7 33.0 34.1  Instrumental vaginal birth 20.2 20.3 20.2  Caesarean section 46.0 46.5 45.7  (number) % missing (2) 0.05 (1) 0.07 (1) 0.05 Maternal baseline data Maternal age (median) [IQR] 35.6 [32.8 to 38.4] 35.7 [33.0 to 38.4] 35.4 [32.7 to 38.4]  (number) % missing (0) 0.0 (0) 0.0 (0) 0.0 Maternal high school education level (% for each category)  Year 9 or below 0.5 0.5 0.6  Year 10 2.7 2.3 2.9  Year 11 4.0 4.0 4.0  Year 12 and above 61.1 59.6 62.0  (number) % missing (1160) 31.7 (500) 33.6 (660) 30.5 Maternal post-school education level (% for each category)  No post-school education 7.6 6.7 8.3  Certificate (including trade) 10.2 9.7 10.6  Advanced diploma 9.9 8.6 10.8  Bachelor degree or above 39.3 40.0 38.7  (number) % missing (1207) 33.0 (521) 35.0 (686) 31.7 Parity (% for each category)  0 62.7 62.9 62.7  1 31.0 31.2 31.0  2 5.0 4.8 5.2  3 +  1.3 1.1 1.1  (number) % missing (0) 0.0 (0) 0.0 (0) 0.0 SEIFA (Socio-Economic Index for Areas) quintile  1 (most disadvantaged) 5.7 4.9 6.3  2 10.7 10.2 11.0  3 18.6 18.1 18.9  4 29.3 31.4 27.9  5 (least disadvantaged) 35.5 35.2 35.8  (number) % missing (6) 0.16 (2) 0.13 (4) 0.18 Second parent high school education level (% for each category)  Year 9 or below 1.2 0.9 1.3  Year 10 4.7 3.5 5.6  Year 11 5.9 6.2 5.7  Year 12 and above 52.2 51.3 52.8  (number) % missing (1316) 36.0 (566) 38.0 (750) 34.6 Post-school education (% for each category)  No post-school education 7.8 6.9 8.4  Certificate (including trade) 14.6 14.5 14.6  Advanced diploma 8.6 7.7 9.3  Bachelor degree or above 31.6 31.2 31.8  (number) % missing (1369) 37.4 (591) 39.7 (778) 35.9 Laboratory based variables Year of fertilisation  2004–2005 16.8 17.4 16.4  2006–2007 12.9 13.6 12.4  2007–2008 30.1 33.6 27.8  2009–2010 39.1 34.7 42.2  (number) % missing (38) 1.04 (12) 0.81 (26) 1.20 Number of eggs at egg collection  0–1 4.1 4.4 3.9  3–5 7.9 8.0 7.8  6–10 16.7 16.6 16.7  10–20 33.0 31.1 34.2  > 20 11.2 11.8 10.8  (number) % missing (995) 27.2 (419) 28.1 (576) 26.6 Male factor  Mild-moderate 26.0 4.6 40.7  Severe Excluded Female factors a  Tubal 9.2 14.0 6.0  Polycystic ovary syndrome 17.1 16.9 17.2  Endometriosis 6.8 7.8 6.1  Other b 6.8 10.6 7.3  Specified as “unexplained subfertility” 19.4 19.2 19.5  Poor prognosis female c 5.4 5.6 5.4  Specified as “no female factor” 17.4 13.8 19.8  (number) % missing (656) 17.9 (249) 16.7 (407) 18.8 Severe male factor (see Additional file 7 for detailed demographics of this population), egg thaw cycles, donor egg and sperm cases excluded from this population baseline demographics a As recorded in IVF database, primary diagnosis only b Smaller categories combined, e.g. genetic disorders, fibroids, single sex, non-PCOS ovulation defects c Advanced age, low ovarian reserve, previous failed IVF Participant flow chart. Graphical representation of study population illustrating datasets used in data linkage Descriptive statistics—baseline demographics of primary analysis population Severe male factor (see Additional file 7 for detailed demographics of this population), egg thaw cycles, donor egg and sperm cases excluded from this population baseline demographics a As recorded in IVF database, primary diagnosis only b Smaller categories combined, e.g. genetic disorders, fibroids, single sex, non-PCOS ovulation defects c Advanced age, low ovarian reserve, previous failed IVF The baseline characteristics of the IVF-conceived and ICSI-conceived populations were similar. The frequency of infant sex at birth differed between exposure groups, with 45% female after IVF conception and 51% female after ICSI. There were minimal differences across other demographic variables between the two groups with similar maternal age, education, socioeconomic status indicator and parity, as well as second parent level of education (Table  1 ). The laboratory-based characteristics were also similar between the two exposure groups, excluding the frequency of documented male factor subfertility. The frequency of mild-moderate male factor subfertility was 4.6% in the IVF group compared with 40.7% in the ICSI group. Female factor subfertility was grouped into board categories; tubal factor subfertility was more common in the IVF than ICSI group (14.0% vs 6.0%) (Table  1 ). Unadjusted rates of developmental vulnerability were similar between exposure cohorts for the primary outcome (DV2; IVF with standard—10.6%, versus ICSI—9.8%) and secondary outcomes (individual domains) (Table  2 ). Table 2 Unadjusted results and adjusted casual analysis results—IVF with standard insemination versus IVF with ICSI Unadjusted analysis Adjusted causal analysis Observed proportions (%) Unadjusted effect estimates Estimated proportions (%) Treatment effect estimates IVF with standard IVF with ICSI Relative risk (95% CI) Risk difference (%) (95% CI) IVF with standard IVF with ICSI ATE relative risk (95% CI) ATE risk difference (%) (95% CI) Primary outcome Developmental vulnerability (DV2) (< 10th in ≥ 2 domains) 10.6 9.8 0.93 (0.75 to 1.11)  − 0.79 (− 2.90 to 1.32) 10.79 10.12 0.90 (0.68 to 1.21)  − 1.11 (− 4.23 to 2.01) Secondary outcomes Individual domains  Physical health and wellbeing 10.9 9.9 0.91 (0.74 to 1.09)  − 0.92 (− 3.04 to 1.20) 12.11 10.65 0.81 (0.61 to 1.09)  − 2.37 (− 5.75 to 1.02)  Social competence 10.9 9.9 0.91 (0.74 to 1.09)  − 0.90 (− 3.06 to 1.26) 11.05 10.21 0.91 (0.68 to 1.22)  − 1.02 (− 4.20 to 2.17)  Emotional maturity 11.3 10.1 0.89 (0.72 to 1.06)  − 1.26 (− 3.30 to 0.77) 11.37 10.61 0.90 (0.70 to 1.56)  − 1.21 (− 4.04 to 1.64)  Language and cognitive skills (school-based) 8.6 6.9 0.81 (0.62 to 0.99)  − 1.57 (− 3.45 to 0.31) 8.70 6.90 0.79 (0.58 to 1.08)  − 1.82 (− 4.35 to 0.70)  Communication skills and general knowledge 8.7 7.5 0.86 (0.67 to 1.05)  − 1.17 (− 3.01 to 0.68) 8.81 7.53 0.86 (0.64 to 1.18)  − 1.19 (− 3.76 to 1.38) Special and emerging needs 17.7 15.0 0.85 0.72 to 0.97  − 2.17 (− 5.92 to − 0.34) 18.0 15.5 0.86 (0.71 to 1.05)  − 2.48 (− 5.75 to 0.79) IVF in vitro fertilisation, IVF with standard IVF with standard oocyte insemination, ICSI intracytoplasmic sperm injection, 95% CI 95% confidence interval AIPW adjustment model (linear in covariate terms and no interaction terms): Treatment assignment—year of oocyte insemination, maternal age (at birth of the child), socioeconomic status, parity, language background other than English (LBOTE), maternal school education and maternal highest obtained level of education, number eggs on day of egg collection, presence of mild-moderate male factor infertility and categories of female factor subfertility Outcome adjustment model—maternal age (at birth of the child), socioeconomic status, parity, language background other than English (LBOTE), maternal school education and maternal highest obtained level of education, AEDC census year, child’s age in years at assessment and sex of the child Unadjusted results and adjusted casual analysis results—IVF with standard insemination versus IVF with ICSI IVF in vitro fertilisation, IVF with standard IVF with standard oocyte insemination, ICSI intracytoplasmic sperm injection, 95% CI 95% confidence interval AIPW adjustment model (linear in covariate terms and no interaction terms): Treatment assignment—year of oocyte insemination, maternal age (at birth of the child), socioeconomic status, parity, language background other than English (LBOTE), maternal school education and maternal highest obtained level of education, number eggs on day of egg collection, presence of mild-moderate male factor infertility and categories of female factor subfertility Outcome adjustment model—maternal age (at birth of the child), socioeconomic status, parity, language background other than English (LBOTE), maternal school education and maternal highest obtained level of education, AEDC census year, child’s age in years at assessment and sex of the child Our findings support the null hypothesis of no causal effect of mode of oocyte fertilisation on developmental vulnerability at school entry (less than 10th percentile in more than 2 of 5 domains of the AEDC), with 11.06% of ICSI-conceived children predicted to be developmentally vulnerable in two or more AEDC domains compared with 9.95% of IVF-conceived children. The ATE RD was − 1.11% (95% CI: − 4.23 to 2.01%), and the ATE RR was 0.90 (95% CI: 0.68 to 1.21) consistent with no causal effect at the population level of method of oocyte insemination for children who were conceived by ICSI compared with children born after standard IVF (Table  2 ). We examined each of the five AEDC domains individually as secondary outcomes, in addition to special and emerging needs status. The unadjusted and causal model results for each individual domain are reported in Table  2 . There were no differences between ICSI and IVF-conceived children in adjusted risk difference for any of the individual AEDC domains nor special and emerging needs status. There were two cases with missing exposure status and seven cases of double embryo transfer with different fertilisation methods where we were unable to tell which embryo implanted. These nine cases (0.25% of the cohort) were excluded. Outcome data were missing for 256 (5.5%) of the AEDC-linked cohort. The majority (95.0%) of these missing cases were children with special needs ( n  = 243, 5.2% of overall cohort) and these were coded as developmentally vulnerable. The 13 remaining cases with missing outcome data were considered as MAR as there was no evidence of association with exposure status (χ 2 -test p  = 0.58) and they were removed from analysis. Most covariates had minimal or no missing data (< 1.0%). Maternal education level was missing for 31.7% and maternal post-school education was missing for 33.0%. The number of eggs collected at ovum pick up (OPU) was missing in 27.2% of cases. Missing data for documented female factor fertility was 17.9%. Exposure and outcome distributions for these cases are presented in Table  1 and Additional file 4 . The imputed values for missing covariate data, in particular the four covariates with greater than 15% missingness (number of oocytes, type of female factor and maternal levels of school and post-school education), produced plausible values and data distributions were consistent with a MAR assumption (Additional file 4 ). We found that an AIPW model that included both donor egg, sperm and IVF laboratory were computationally unstable due to perfect prediction errors; consequently the primary AIPW estimator models were run on a cohort without both (Table  2 ) and then containing one or the other (Additional file 5 : Sensitivity analysis). Further, the addition of the two interaction terms and/or the non-linear term to each of these models also resulted in perfect prediction errors. The final primary DV2 outcome model was linear with ten covariates in the selection model and nine in the outcome model (see Table  2 legend). The influence of interactions and non-linear terms on the estimates was assessed using alternate TMLE estimators, including a cohort containing donor egg, donor sperm and IVF laboratory. The results did not meaningfully differ from the findings of the primary analysis (Additional file 5 : Sensitivity analysis). An exploratory analysis was performed to determine the association between ICSI indication and childhood outcome and assess the marginal effect of [ 1 ] ICSI with documented male factor compared with ICSI performed without documented male factor and [ 2 ] ICSI in the setting of severe male factor compared with ICSI with mild or mild-moderate male factor. The baseline population characteristics between exposure groups were similar (Additional file 7 : Male factor cohort). Both unadjusted (RR 1.05, 95% CI: 0.81 to 1.28) and adjusted (aRR 0.99, 95% CI: 0.78 to 1.26) results found little difference between the comparative groups in outcomes of DV2, the five individual AEDC domains or special and emerging needs (Table  3 ). Table 3 Unadjusted results and adjusted analysis—ICSI with no male factor versus ICSI with male factor Unadjusted analysis Adjusted analysis Observed proportions (%) Unadjusted effect estimates Estimated proportions (%) Marginal effect estimates ICSI with no male factor ICSI with male factor Relative risk (95% CI) Risk difference (%) (95% CI) ICSI with no male factor ICSI with male factor Adjusted relative risk (95% CI) Adjusted risk difference (%) (95% CI) Primary outcome Developmental vulnerability (DV2) (< 10th in ≥ 2 domains) 9.6 10.0 1.05 (0.81 to 1.28) 0.4 (− 1.8 to 2.7) 9.8 9.7 0.99 (0.78 to 1.26)  − 0.1 (− 2.4 to 2.3) Secondary outcomes Individual domains  Physical health and wellbeing 9.6 10.1 1.05 (0.81 to 1.28) 0.4 (− 1.8 to 2.7) 9.6 10.2 1.07 (0.84 to 1.35) 0.6 (− 1.7 to 3.0)  Social competence 9.7 9.7 1.00 (0.77 to 1.23) 0.0 (− 2.2 to 2.2) 9.9 9.5 0.96 (0.75 to 1.21)  − 0.4 (− 2.7 to 1.9)  Emotional maturity 9.9 9.2 0.93 (0.71 to 1.14)  − 0.7 (− 2.9 to 1.5) 9.9 9.1 0.93 (0.74 to 1.17)  − 0.7 (− 2.9 to 1.5)  Language and cognitive skills (school-based) 6.1 7.8 1.28 (0.92 to 1.64) 1.7 (− 0.2 to 3.6) 6.3 7.7 1.21 (0.91 to 1.62) 1.3 (− 0.7 to 3.4)  Communication skills and general knowledge 6.9 8.0 1.15 (0.84 to 1.46) 1.1 (− 0.9 to 3.0) 10.1 9.0 0.89 (0.70 to 1.13)  − 1.1 (− 3.4 to 1.1) Special and emerging needs 15.8 14.9 0.94 (0.78 to 1.11)  − 0.9 (− 3.6 to 1.8) 16.0 14.8 0.93 (0.77 to 1.12)  − 1.2 (− 4.1 to 1.7) ICSI intracytoplasmic sperm injection, 95% CI 95% confidence interval Adjustment model: outcome regression adjustment model—maternal age (at birth of the child), socioeconomic status, parity, language background other than English (LBOTE), both maternal and paternal highest obtained level of education, child’s age in years at assessment and sex of the child Unadjusted results and adjusted analysis—ICSI with no male factor versus ICSI with male factor ICSI intracytoplasmic sperm injection, 95% CI 95% confidence interval Adjustment model: outcome regression adjustment model—maternal age (at birth of the child), socioeconomic status, parity, language background other than English (LBOTE), both maternal and paternal highest obtained level of education, child’s age in years at assessment and sex of the child

Intracytoplasmic sperm injection (ICSI) is an essential technique in the in vitro fertilisation (IVF) armamentarium. It involves the direct injection of a single spermatozoon into the cytoplasm of an oocyte [ 1 ]. This is in contrast to standard insemination where an egg and sperm are simply co-incubated to facilitate fertilisation—a cellular process that is comparable to fertilisation in vivo [ 2 ]. The introduction of ICSI to clinical practice in 1992 aimed to improve fertilisation rates in the setting of severe male factor infertility—and indeed it did [ 3 ]. The high fertilisation rate associated with ICSI has since led to its increasing use in cases of non-male factor infertility [ 4 , 5 ]. By 2010, the frequency of use of ICSI in IVF cycles reached 100% in some regions globally [ 6 ]. Data from the USA have shown ICSI utilisation increased from 15.4% in 1996 to 66.9% in 2012 for non-male factor indications [ 7 ]. As of 2018, reports demonstrated that ICSI rates vary from 53 to 90% across different American states with the national average close to 80% [ 8 ]. In Australia, ICSI use peaked in 2016 where it was the fertilisation method of choice in 69% of cycles [ 4 ]. Despite this widespread use, there is still no proven benefit on the use of ICSI outside of severe male factor infertility and concerns have evolved about the safety of its widespread use [ 5 , 9 ]. The actual technique itself—with arbitrary selection of a spermatozoon for injection and bypassing of the zona pellucida—has led to concerns about the transmission of abnormal genetic and epigenetic traits and subsequent implications for the offspring [ 10 , 11 ]. Multiple studies investigating the risk of congenital malformations associated with ICSI have produced conflicting evidence [ 5 , 12 , 13 ]. A number of studies have also examined longer term developmental outcomes for children conceived via ICSI, with some providing reassurance about childhood outcomes [ 10 , 14 – 19 ]. However, past studies have been limited by their sample size, use of historical cohorts, limited accounting for confounding factors and misleading comparator cohorts (using spontaneously conceived children where neither subfertility nor fertility treatments are involved). Given the rising use of ICSI for non-male factor infertility and the ongoing concerns about its potential harms, we aimed to examine the long-term developmental outcomes for children conceived via ICSI for non-male factor infertility compared with those in which standard IVF was used. Using a state-wide birth cohort and a causal inference framework, our study was designed to emulate a real-world target trial to investigate the causal effect of ICSI for non-male factor infertility compared with standard IVF on the school-age developmental outcomes for children.

Using population level data and statistical methods with a formal framework for causal inference, our findings suggest that ICSI for non-absolute indications does not confer an altered risk of childhood developmental vulnerability in two or more domains of the AEDC, compared with IVF with standard insemination. While there is growing evidence of a modest increase in congenital abnormalities associated with ICSI (e.g. hypospadias [ 13 ]), it appears that the mechanism for this difference does not extend to broader neurodevelopmental concerns. Given the limitations of previous research and the rising use of ICSI globally, our study has generated much needed evidence. The strengths of our study lie in the use of state-wide birth and IVF data, deployment of record linkage between birth and longer-term childhood data and our rigorous application of formal methods for causal inference. Many previous studies in this field were published over two decades ago, when ICSI was used almost exclusively for severe male factor subfertility or previous failed fertilisation with standard IVF, and the practice and technique of ICSI had not yet been refined to its current standard [ 15 , 17 , 18 ]. Additionally, past prospective observational cohort studies looking at detailed health and neurodevelopmental assessments were likely too underpowered (e.g. n  = 76 to 201 ICSI-conceived cases) to detect small yet clinically relevant effects [ 15 , 16 , 30 , 49 ]; have focused on early perinatal outcomes such as congenital abnormalities; have relied upon simplistic, coarse or low-prevalence binary outcome measures such as cerebral palsy or autism diagnosis [ 13 , 29 , 50 – 53 ]; or have, where differences were detected, suggested that these were largely explained by parental factor like maternal age, parity, socioeconomic status, as well as parental education and parents’ cognitive ability [ 54 – 56 ]. Uniquely, our study was able to delineate between indications for ICSI—we were able to assess outcomes for ICSI cases were no severe male factor was evident. Previous studies have clustered all children conceived by ICSI together, without separating ICSI for severe male factor infertility from ICSI for non-essential indications. This is a crucial difference because ICSI is widely considered essential for fertilisation in cases of severe male factor infertility but for other more elective indications, and thus modifiable as an exposure. The ability to exclude cases of severe male factor infertility is an integral element of our target trial emulation. These cases could never be randomised to standard IVF in an intervention trial and thus should not contribute data in a randomised trial context. By designing our study cohort in the context of a potential clinical trial, our study has arguably provided a stronger conclusion that an association analysis of observational data would not. We argue that existing evidence in this field is currently limited by the frequent use of the spontaneously conceived children as the comparison group. A 2008 large systematic review by Middelburg et al. evaluated the impact of IVF and ICSI conception on infant neurodevelopmental outcomes [ 14 ]. Reassuringly, they found no increased risk of adverse development, in children conceived by standard IVF or ICSI compared with children conceived without assistance. We were able to further echo this reassurance with a recent analysis making the same comparison [ 21 ]. The spontaneously conceived children and their parents are a vastly different population to that of the IVF-conceived. Correct comparison of vastly different exposed and unexposed population groups (even in studies where data for a rich set of covariates is available for adjustment [ 54 ]) requires complex methodology to appropriately account for these differences, rather than traditional regression methods [ 21 , 57 ]. However, comparison with spontaneous conception represents a different research question and fails to tease out the specific impact of ICSI over and above that of IVF as a whole and subfertility as a condition. To achieve this comparison, we have implemented a target trial emulation, which by explicitly defining inclusion criteria reduces observational cohort size, but if well-emulated leads to a closer estimation of causal effect. In assessing baseline characteristics, our study found notable differences in the sex proportions of babies born after IVF with standard insemination versus ICSI. This phenomenon, with a higher incidence of female babies conceived via ICSI, has been described previously in the literature [ 58 , 59 ]. As sex is a known determinant of performance in developmental assessments, sex at birth was included as a covariate to ensure this difference between exposure groups was accounted for. There are data-based limitations inherent in all studies. In our study, as in others, documentation of male factor infertility was highly variable within and between the data from different IVF laboratories. Severe male factor infertility was documented as [ 1 ] surgical sperm extraction, [ 2 ] free text data which included entries such as “oligoasthenoteratozoospermia” or “OAT”, “CBAVD” and “azoospermia” and [ 3 ] clear abnormalities on the laboratory analysis parameters on day of egg collection. The conclusions drawn from this study depend on the reliability and accuracy of data on which they are based. We recognise that even the most sophisticated analysis techniques cannot compensate for inaccuracies or inconsistencies in the data itself. Due to the nature of record-linkage studies and the AEDC as a metric, our study may be limited by selection bias. A small fraction of children who did not attend mainstream school due to severe disability or home schooling were not captured. In Australia, the majority of children with a disability attend mainstream school, with only approximately 1% of the population attending specialist schools [ 60 ] and 0.5% home-schooled [ 61 ]. All Victorian schools including specialist schools are invited to participate in the AEDC but can elect not to. However, our study was not designed to assess severe disability or developmental delay, but rather an overall measure of global development and school readiness—within our study cohort of 2468 ICSI cases, no increased risk of developmental vulnerability was identified. Finally, as with all observational studies it is possible that unmeasured confounders may have led to bias in estimating the treatment effects. To the best of our knowledge, the important covariates have been identified, measured and accounted for.

Within a causal framework, the developmental outcomes at school entry for children conceived by ICSI, without severe male factor infertility, are equivalent to those conceived via standard IVF. Furthermore, conception via ICSI in the setting of male factor subfertility is not associated with an increased risk of childhood developmental vulnerability compared with conception via ICSI in the absence of male factor subfertility. Our findings suggest that empirical use of ICSI does not affect early childhood developmental outcomes. These findings provide further important reassurance for current and prospective parents and clinicians alike.

Additional file 1: Target trial emulation. Additional file 2: Statistical analysis plan. Additional file 3: Delphi…Results_04Sep2023 (short title: Delphi survey results). Additional file 4: Missing data diagnostics. Additional file 5: Sensitivity analyses. Additional file 6: Covariate balance after IPW (short title: covariate balance). Additional file 7: Male factor cohort. Additional file 1: Target trial emulation. Additional file 2: Statistical analysis plan. Additional file 3: Delphi…Results_04Sep2023 (short title: Delphi survey results). Additional file 4: Missing data diagnostics. Additional file 5: Sensitivity analyses. Additional file 6: Covariate balance after IPW (short title: covariate balance). Additional file 7: Male factor cohort.