About the Authors:
Adeniyi Francis Fagbamigbe
Roles Conceptualization, Data curation, Formal analysis, Methodology, Visualization, Writing – original draft, Writing – review & editing
* E-mail: [email protected], [email protected]
Affiliations Department of Epidemiology and Medical Statistics, College of Medicine, University of Ibadan, Ibadan, Nigeria, Health Data Science Group, Division of Population and Behavioural Sciences, School of Medicine, University of St Andrews, St Andrews, United Kingdom, Populations, Evidence and Technologies Group, Division of Health Sciences, University of Warwick, Coventry, United Kingdom
ORCID logo https://orcid.org/0000-0001-9184-8258
Emma Norrman
Roles Data curation, Supervision, Writing – review & editing
Affiliations Department of Obstetrics and Gynecology, Institute of Clinical Sciences, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden, Sahlgrenska University Hospital/Östra, Gothenburg, Sweden
Christina Bergh
Roles Data curation, Investigation, Writing – review & editing
Affiliations Department of Obstetrics and Gynecology, Institute of Clinical Sciences, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden, Sahlgrenska University Hospital/Östra, Gothenburg, Sweden
Ulla-Britt Wennerholm
Roles Investigation, Supervision, Writing – review & editing
Affiliations Department of Obstetrics and Gynecology, Institute of Clinical Sciences, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden, Sahlgrenska University Hospital/Östra, Gothenburg, Sweden
ORCID logo https://orcid.org/0000-0003-2475-2226
Max Petzold
Roles Conceptualization, Methodology, Supervision, Writing – review & editing
Affiliation: School of Public Health and Community Medicine, Institute of Medicine, University of Gothenburg, Gothenburg, Sweden
Introduction
An exploratory analysis of age at onset of type-1 diabetes among children born in Sweden between 1985 and 2015 and conceived either spontaneously or by assisted reproductive technology (ART) showed that (i) skewed discrete age distribution because most ART-conceived children are younger (ART is increasing and most children were born in the more recent years), (ii) a probable peak in the risk of type-1 diabetes at 10–14 years of age, (iii) there is an increased risk of developing diabetes in more recent years [1]. The ART was either the standard in-vitro-fertilization (IVF) or intracytoplasmic sperm injection (ICSI) with fresh or frozen-thawed embryo transfer. This study used the Cox proportional hazard (CPH) model to evaluate if the mode of conception (ART vs spontaneous) was associated with an increased risk of type-1 diabetes [1]. However, it is not known whether this method captured the peculiarities of the data. There is a need to investigate if the differences were caused by skewness in age, having more ART-conceived children during the later years when diabetes is more prominent.
There have been contradictions in the reported risk of conception mode-type-1 diabetes in the literature [1–3]. Norrman et al. reported an insignificant risk of Type-1 diabetes by maternal fertility (adjusted hazard ratio (aHR) = 1.07; 95% CI: 0.93–1.23) but significant differences in risk were found between frozen embryo transfer and type-1 diabetes (aHR 1.52; 95% CI: 1.08–2.14 and 1.41; 95% CI: 1.05–1.89) for frozen versus fresh and frozen versus spontaneous conception (SC), respectively in a subgroup analyses of a Swedish data [1]. Using Danish data, Hargreave et al. found that the risk of type-1 diabetes was not significant with conception mode (aHR = 1.01; 95% CI: 0.90–1.13) after adjusting for other covariates [3]. Similarly, Kettner et al. reported no association between fertility treatment and childhood type-1 diabetes mellitus but found that ovulation induction or intrauterine insemination with follicle-stimulating hormone was associated with a higher risk of type-1 diabetes mellitus (aHR = 3.22; 95% CI: 1.20 to 8.64) [2]. The authors found no association with diabetes type-1 diabetes in other types of fertility treatment indications [2]. The data analysed by Hargreave et al. and Kettner et al. were both Danish but differed in sizes with over 1, 550,519 and 565,116 births respectively whereas there were 3,138,540 births in the study by Norrman et al. [1–3]. The study setting, sample sizes and methods of analysis might have caused these contradictions. The reason for these contradictory findings could be that the analyses used in the prior literature depended on the appropriateness of proportional hazards models in the settings. This paper is therefore designed to revisit and assess the effect of ART and its subgroups, on the development of type-1 diabetes. To achieve this, we assessed the fit of a range of different survival analysis regression models in assessing the effect of ART in having type-1 diabetes among a birth cohort in Sweden.
The data used in Norrman et al. presented right-censored data on the onset of type-1 diabetes which makes survival analysis a natural choice of analysis [1]. Survival analysis, also called analysis of time-to-event, is a statistical procedure for a follow-up study in which some subjects may not have experienced the event of interest due to loss to follow up, withdrawal of subjects from the study or the study coming to an end [4, 5]. Besides the estimation of the survival function of such time-to-event, greater interest is the identification of prognostic factors that affects the timings [6–9]. Several methodologies, known as survival regression methods, have been developed to achieve this goal. While some of the models are traditional survival regression models, some are modern [10, 11]. The traditional survival regression models consist of the non-parametric, semi-parametric and parametric models. We describe the assumptions, advantages and disadvantages of these models in the next paragraph.
The Kaplan-Meier estimator is the nonparametric maximum likelihood estimate (NPMLE) of the survivor function when the model is a non-parametric survival model. It does not allow the investigation of the effects of covariates [5]. Hence the effects of risk factors on the time-to-event cannot be estimated. Nonetheless, the Kaplan-Meier estimator allows the exploration of the effect of covariates graphically and by comparison of distinct estimates of the survival function but this gets cumbersome when several covariates are to be investigated together. The CPH model is a semi-parametric regression model and it is the most commonly used survival analysis regression models [12–16]. This model has a unique advantage of not making any assumptions about the shape of the underlying hazard function and can directly use the hazard ratios to make a reliable estimation of the treatment effects [14] but rather assumes that the hazard ratios of the covariates in the model are constant over time [16, 17]. Although the CPH model is very useful and often accurate for a large sample size study, the major criticism against the CPH models is that this assumption is often violated, especially in a very long follow-up study. The CPH model has another disadvantage in that the survival function hence the cumulative hazard function, are continuous with respect to Lebesgue measure [18]. The estimators are, however, step functions rather than being smooth [11, 19].
The parametric models assume that the survival and hazard functions follow a specific distribution in which case the parameters can be estimated. The parametric survival models, such as the exponential and Weibull models in addition to the log-logistic and generalized gamma models are often used to obtain smooth hazard rates and cumulative hazard functions of the risk factors and to extrapolate the survival functions and cumulative functions [5, 20]. A major disadvantage of the parametric models lies in their insufficient flexibility to ensure adequate representation of the data been fitted, and the fact that the cumulative hazard function or survival function may be biased [19]. These inflexibilities often lead to biased estimates for the cumulative hazard and/or survival functions. Moreover, parametric models with complex underlying hazard fail to capture the true (covariate) effects [21, 22].
The Flexible parametric survival regression (FPSR) models were developed by Royston and Parmar in the early 2000s [23] as a more flexible alternative. The FPSR attempted to eliminate the shortcoming of the CPH and the parametric models by building on the strengths of these models. The flexible models are an extension of several parametric survival models [23] and were formulated to relax the assumption of linearity of log time [24]. These flexible models represent the log of the baseline cumulative hazard function with a restricted cubic spline function of the logarithm function of time [4, 21, 24]. The literature is replete with several advantages of the FPSR models [4, 5, 14, 21, 22, 25, 26].
This study is a method paper with an application. The analyses were carried out to assess the effect of ART on the development of type-1 diabetes and compare the performance of the CPH, parametric and the FPSR models in the identification and quantification of the hazards of prognostic factors of the timing of onset of type-1 diabetes among ART (including fresh and frozen embryo transfer) and spontaneously conceived children in Sweden between 1985 and 2015. We ascertained the effect of the differences in age distribution between ART and spontaneously conceived children on the risk of having type-1 diabetes. In this study, we applied and compared the fit and performances of (i) Cox proportional hazard (CPH) models, (ii) The proportional hazard (PH) (Weibull PH, exponential PH and Gompertz PH) and the Accelerated Failure Time (AFT) (Weibull AFT, exponential AFT, log-logistic, lognormal, and the generalized gamma models) parametric survival models and (iii) Royston-Parmar (RP) Flexible parametric survival regression (FPSR) models with a different number of knots. The assumption of the PH can be relaxed in the PH models. The FPSR models have the advantage that the PH assumption can be relaxed rather easily in comparison to the other survival models.
Methods
The data
The data used for this study were collected in Sweden and represent all singleton children (n = 3,138,540) born between 1985 and 2015, excluding singletons conceived after oocyte donation. The data was extracted from the national health data registers hosted by the Swedish National Board of Health and Welfare, linked with several national quality registers and information from Statistics Sweden (SCB). The SCB maintains a constantly updated and quality-checked register from the Swedish National Board of Health and Welfare, and the Swedish National Data Service which ensures the completeness of the data. The national quality registers include registers on morbidity and mortality of children born in Sweden. The SCB did not state any change in the inclusion criteria of morbidities and mortalities over time.
Further details on the data have been reported [1]. Ethical permission was given from the Regional Ethical Committee ([email protected]) at the University of Gothenburg (Dnr 214–12, T422-12, T516-15, T233-16, T300-17, T1144-17, T121-18). The Swedish National Board of Health and Welfare obtained informed written consent from patients to have data/samples from their medical records used in research. We followed the prescribed ethical regulations on confidentiality when the database was accessed in October 2019. The Swedish National Board of Health and Welfare, and Statistics Sweden (SCB) placed ethical restrictions on sharing the data publicly as the data contain potentially identifying or sensitive patient information. The restriction was concurred by the Regional Ethical Committee at the University of Gothenburg, Sweden.
Statistical models
Cox proportional hazard models.
The CPH model was developed based on the assumption that hazards are multiplicatively proportional to baseline hazards [27] but without any assumption about the distribution of the hazards.
(1)
From Eq (1), coefficients β1, β2,………, βk could be estimated but a direct estimate of the baseline hazard (h0(t)) or its distribution cannot be estimated. The model nonetheless provides an avenue to estimate the baseline cumulative hazard (H0(t)) and baseline survival (S0(t)) which can be used to estimate the h0(t) [19].
The CPH uses maximum partial likelihood methods to estimate coefficients (βi). Numerically, let xi be the row vector of covariates for the time interval (t0i; ti) for the ith observation in a dataset with N subjects (i = 1, 2, 3,………., N). The coefficient (βi) of the covariates (Xi) can be estimated by maximizing the partial log-likelihood function(2)where j indexes the ordered failure times t(j), j = 1, 2, ……., P; Pj is the set of Pj observations that fail at t(j); dj is the number of failures at t(j); and Rj is the set of observations k that are at risk at time t(j) (that is, all k such that t(0k) < t(j) ≤ t(k)).
Parametric survival models.
As reported in the literature [5, 20], the accelerated failure-time (AFT) models and the multiplicative or proportional hazards (PH) model are the most-used parametric models for adjusting survivor functions for the effects of covariates. The PH models include the Weibull, Gompertz, and exponential while the exponential, Weibull, lognormal, log-logistic, and generalized gamma are the commonest AFT models. In the AFT model, log(t) (the natural logarithm of the survival time) is expressed as a linear function of the covariates which yields a linear model(3)where xj is a vector of the covariates, β is a vector of the regression coefficients, and zj is the error function whose distributional form determines the form of the regression model. For instance, a normal density leads to the lognormal regression model just as the logistic density yields the log-logistic regression. The AFT models applied to change the time scale by a factor of . The time is accelerated if the factor is <1 and decelerated (degraded) if >1. In the PH model, the concomitant covariates have a multiplicative effect on the hazard function and contain the covariates which have a multiplicative effect on the hazard function(4)for some h0(t) and for g(Xj), a non-negative covariates function. In some cases, g(Xj) is expressed as . The parametric PH model becomes a CPH model if the function h0(t) is not specified [5, 20].
Flexible parametric survival regression model.
These models use restricted cubic splines to model a transformation of the survival function and can be modelled on different scales, including the hazard scale, the odds scale, and the probit scale in flexible parametric survival analyses. We focussed on the hazard scale of the models to ensure that the estimates from the FPSR and CPH models are comparable. This approach has an advantage in that the corresponding function is more stable and the process of capturing the shape of the function is easier [28]. The Weibull model is one of the commonest parametric models for survival data and can be criticized due to its lack of flexibility in the shape of the hazard function. The log cumulative hazard function of Weibull distribution is written as(5)
In Eq (5), ln H0(t) is the baseline log cumulative hazard at time point t, λ is the scale parameter, γ is the shape parameter, xi is the covariate and βi is the coefficient of the covariate [28]. The log cumulative hazard function of the FPSR models are based on a log cumulative hazard scale with the use of restricted cubic spline function of the log time to transform the function in Eq (5) is(6)
In Eq (6), x = ln(t) and s(x) is a restricted cubic spline function with k knots k1, k2,…………….kk and parameters γ0, γ1,…………….γk−1 is expressed as
Where z1 = x = ln(t) and zj(j≥2) are derivable functions that are determined by the numbers and the positions of the knots is computed as
Where ϕj = (kk−kj)/(kk−k1) [4, 19, 23].
The desired complexity of the FSPR models is specified as the number and positions of the knots (connection points) in log time of the spline’s cubic polynomial segments [19]. The k knots, a maximum of 9 knots, has k+1 degrees of freedom (df). The placement of knots is at the centiles (computed as 100/df) of event times. For 4 Knots, the df is 5 and the knots will be located at centiles 20, 40, 60 and 80. The internal knots are bounded by the “boundary knots” which are placed at the minimum and maximum of the distribution of uncensored survival times. The FSPR model reduces to the Weibull model if the number of knots is 0, with γ0 and γ1 as the estimates of its scale and shape parameter respectively. Royston et al. suggested either 1 or 2 knots for smaller (<10,000) datasets and 4 or 5 knots for larger (> = 10,000) datasets [19]. The FSPR model was implemented in Stata using “stpm2” command. The stpm2 can be used with single- or multiple-record or failure survival data on the log cumulative hazard, the log cumulative odds, the probit scales, or on a scale based on theta-value using the Aranda-Ordaz family of link functions [9, 24]. We computed the incidence rate as the probability of having diabetes per person-year and the attributable fraction to assess the excess risk attributable to been conceived with ART compared with spontaneous conception. The survival analysis was performed using Stata 16.0 (Stata-Corp LP, College Station, TX, USA). Microsoft Excel in Office 365 and R Software were used for data visualization.
Dependent variable.
The outcome is a paired survival time and indicator of whether the participant had experienced type-1 diabetes or not. For those that had type-1 diabetes, the survival time was their age as of its onset, while the time (age) to type-1 diabetes was censored on the date of the data collection, emigration and death for those with no type-1 diabetes.
Independent variable.
Our choice of covariates was informed by Norman et al. [1]. The main explanatory variable in this study is the mode of the conception of the children: spontaneous or ART conception. The types of ART considered in this study are IVF and ICSI with fresh and frozen-thawed embryo transfer. The other independent variables are sex of the child, maternal age, parity, year of birth (birth cohort), mother’s and father’s country of birth, mothers’ smoking status, mothers’ and fathers’ diabetic status, and maternal and paternal education at the birth of the child.
Model selection criteria.
The log-likelihood, Akaike information criteria (AIC) [29] and the Bayesian information criteria (BIC) [30] were used to assess the fit of the models. The lower these quantities the better the fitness of the models. The AIC and the BIC are usually computed and compared separately among different models to determine the best fitting model. In all cases, the lower the AIC and BIC, the better the model [14]. Literature suggests that AIC will choose a more complex model irrespective of sample size while BIC is more likely to choose a simpler model [14]. AIC is often preferable in situations when a false negative finding would be considered to be more misleading than a false positive, and BIC is better in situations where a false positive is as misleading as, or more misleading than, a false negative [14]. We used these criteria to assess which of the models been compared was the optimal model.
Results
Among the 3,138,540 children included in the analysis, 98.5% (3,090,602) were conceived spontaneously while 1.5% (47,938) were conceived through ART. Of the 47,938 ART-conceived children, 36,727 were through fresh embryo transfer and 11,211 were from frozen-thawed embryo transfer. About 51.4% of the children were males, nearly half (47.8%) were from mothers aged 30–39 years, 81.7% were from Swedish mothers, from mothers (0.4%) and fathers (0.6%) who were diabetic. In total, 18,118 (0.58%) became type-1 diabetic during the follow-up period comprising 50,936,586 person-years at risk. Among those conceived spontaneously, 0.58% (n = 17,916) became diabetic compared with 0.42% (n = 202) among the ART-conceived (Table 1). The median (Interquartile Range (IQR)) age at onset of type-1 diabetes among spontaneously conceived children was 10 (6–14) years, 8 (5–12) for ART, 9 (5–12) years for fresh embryo transfer and 6 (4–10) years for frozen-thawed embryo transfer. Children conceived by ART had a higher incidence rate of type-1 diabetes compared with those conceived spontaneously. The overall incidence rate of those conceived through ART and spontaneously were 43 and 36 per 100 000 person-years at risk respectively. The highest incidence of type-1 diabetes was found among children whose fathers (320/100,000) and mothers (244/100,000) were diabetic compared with those whose parents were not diabetic (34/100,000 and 35/100,000). The distribution of the incidence of type-1 diabetes by selected children characteristics is presented in Fig 1.
[Figure omitted. See PDF.]
Fig 1. Distribution of incidence of type-1 diabetes by selected children characteristics.
https://doi.org/10.1371/journal.pone.0253389.g001
[Figure omitted. See PDF.]
Table 1. Distribution of children’ characteristics and incidence rate of type-1 diabetes.
https://doi.org/10.1371/journal.pone.0253389.t001
Test of equality of incidence rates of type-1 diabetes
The overall difference in unadjusted incidence rate between the children conceived spontaneously and by ART was 0.000791 (95% CI: 0.000019–0.000139, incidence rate ratio was 1.222589 (95% CI: 1.058–1.40448), attributable fraction exposed was 0.18206 (95% CI: 0.05568–0.28799). The two tail hypothesis test that the incidence rates for each of the groups are not different was significant (p = 0.0058) (Not shown in the Tables).
Distribution of the hazard and survival functions under different distributions
On the choice of the survival regression analysis to be adopted, we first explored the CPH model. As shown in Fig 2 (left panel), the assumption of proportionality was not violated. In the same way, a test of non-proportionality of Kaplan Meier survival curves of the observed and predicted ART estimates at various ages of the children was not significant (Fig 2: right panel).
[Figure omitted. See PDF.]
Fig 2. Proportionality of survival curves of onset of type-1 diabetes among ART and spontaneously conceived children.
https://doi.org/10.1371/journal.pone.0253389.g002
In Fig 3, we present the survival and hazard functions of developing type-1 diabetes by modes of conception using the Kaplan Meier, smoothed, CPH and the FSPR models. In all, survivorship was higher among the spontaneously conceived children compared with those conceived through ART. In the same way, the hazard of type-1 diabetes was higher among those conceived through ART compared with those that were spontaneously conceived.
[Figure omitted. See PDF.]
Fig 3. Comparison of hazard and survival functions under different survival regression models.
https://doi.org/10.1371/journal.pone.0253389.g003
In Fig 4, we compared the hazard and survival functions of the flexible models at different degrees of freedom (knots/splines). The functions appear more realistic and flexible at higher degrees of freedom than the Weibull distribution. The greatest flexibility was noticeable at df = 6. The lower panel of Fig 4 shows the predicted incidence of diabetes using the flexible model.
[Figure omitted. See PDF.]
Fig 4. Comparing the hazard and survival functions of the flexible models at different knots/splines.
https://doi.org/10.1371/journal.pone.0253389.g004
Comparison of the models
The statistics of all the model selection criteria considered in this study (the loglikelihood, the AIC and the BIC) are shown in Table 2. The AIC and BIC of the FPSR models, the statistics obtained from the parametric models including the Weibull AFT, exponential AFT, log-logistic, lognormal, and the generalized gamma models were only slightly higher than that of the FPSR models. Among the FPSR models evaluated at different degrees of freedom, the largest loglikelihood, the lowest AIC and the lowest BIC were at 6 degrees of freedom (df = 6). Although the model fits are similar for all the models, the FPSR has the lowest values for each of the criteria. This is further supported by the significance of restricted cubic splines (rcs) of each covariate otherwise called the slope of the hazard curve within each of the knots generated at 6 degrees of freedom (Table 2).
[Figure omitted. See PDF.]
Table 2. Comparison of the loglikelihood, AIC and BIC of the models considered.
https://doi.org/10.1371/journal.pone.0253389.t002
Table 3 shows the coefficients of the covariates included in the adjusted models and their respective standard errors (s.e). In all, the estimates from the CPH model, the parametric PH models and the FPSR model at different degrees of freedom were similar and generally lower than the estimates from the parametric AFT models.
[Figure omitted. See PDF.]
Table 3. Coefficients of the covariates from the fitted models.
https://doi.org/10.1371/journal.pone.0253389.t003
Modelling the risk factors of type-1 diabetes
The estimates from each of the model were similar. We have used the estimates from the FPSR model at 6 degrees of freedom to interpret the adjusted determinants of timing of the onset of type-1 diabetes among the children. The age of the children was used as a timescale to correct for the problem associated with lesser exposure time among the late entries. Although insignificant, the adjusted hazard of type-1 diabetes was 7% higher (adjusted Hazard Ratio (aHR) = 1.070; 95% Confidence Interval (CI): 0.929–1.232, p = 0.346)) among children conceived through ART than those conceived spontaneously while controlling for other variables. A sub-analysis with emphasis on whether the transferred embryo was fresh or frozen-thawed transferred while controlling for other covariates showed that the hazard of type-1 diabetes was 37% (aHR = 1.368; 95% CI: 1.013–1.847, p = 0.041) higher among frozen-thawed transfer compared with those conceived spontaneously. Inversely, the hazard of type-1 diabetes was higher among frozen-thawed embryo transfer than fresh embryo transfer (aHR = 1.407; 95% CI: 1.007–1.965, p = 0.046; not shown in the Tables). While controlling for other variables, the hazard of type-1 diabetes was about 17% (aHR = 1.171; 95% CI: 1.137–1.206, p<0.001) higher among males than the females. The adjusted hazard of type-1 diabetes was significantly associated with the birth cohort of the children. The younger the children the higher the hazard of type-1 diabetes. For instance, those born in 1991–1995 had 18%, 1996–2000 (33%), 2001–2005 (38%), 2006–2010 (38%) and 2011–2015 (36%) higher hazard of type-1 diabetes than those in 1985–1990 birth cohort.
The hazard of type-1 diabetes was higher among children whose mothers do not smoke (aHR = 1.296; 95% CI: 1.240–1.354, p<0.001) than those whose mother smokes. The adjusted hazard of type-1 diabetes was 60% higher among children of both Swedish and other Nordic mothers than those of other European mothers while the hazard was about 76% higher among the children of both Swedish and other Nordic fathers than those of other European fathers. The hazard of type-1 diabetes increased with a lower level of paternal and maternal education. The highest hazard of type-1 diabetes was found among children whose parents were diabetic. Children of diabetic mothers and fathers were about 6 times (aHR = 6.419; 95% CI: 5.852–7.041, p<0.001) and 9 times (aHR = 8.808; 95% CI: 8.221–9.437, p<0.001) respectively more likely to be diabetic than those whose parents were not diabetic as shown in Table 4. However, maternal age and the parity of the index child were not significantly associated with the risk of type-1 diabetes. The visualization of the adjusted hazard ratios is presented in Fig 5.
[Figure omitted. See PDF.]
Fig 5. The hazard ratios of the prognostic factors of type-1 diabetes among the studied Swedish cohort.
https://doi.org/10.1371/journal.pone.0253389.g005
[Figure omitted. See PDF.]
Table 4. Adjusted prognostic factors of type-1 diabetes among the studied Swedish cohort from CPH and FPSR model with 6 degrees of freedom.
https://doi.org/10.1371/journal.pone.0253389.t004
Discussions
This study was designed to compare the performance of different survival models in examining the risk of type-1 diabetes viz-a-viz mode of conception and assess if there are differentials in the risk of type-1 diabetes between children from fresh or frozen-thawed embryo transfers. In all, the models considered in this study are comparable as far as modelling the effect of conception modes on the risk of diabetes is concerned. While there were no significant differences in the risk of diabetes between children conceived spontaneously and by ART, the risk differed significantly among children from fresh and frozen-thawed embryo transfers. The measures of goodness of fit of models considered in this study were relatively lower in the flexible models. However, rather than likelihoods, the CPH model uses a different procedure and returns partial likelihood, which is also lower than the likelihood of the other models. Similarly, the AIC and the BIC which are measures of information lost by each of the models were more favourable to the flexible models. Similar approaches have been used to identify better models in the literature [31, 32]. Besides having the lowest information lost, the flexible model fitted the data most but this did not lead to large differences between the estimates from the models. At higher degrees of freedom, the flexible model fitted the data most and also showed that the occurrence of type-1 diabetes is not proportional to the baseline hazard but that the risk changed severally as the children grew older. We have applied the proportional hazard and other parametric, as well as the flexible parametric models to the data. A major strength of the flexible parametric model is its ability to account for non-proportional hazards, hence individual non-PH models were not explored in this study. However, the estimates from the flexible parametric and CPH models are similar when the proportional hazard assumptions were not violated.
The estimates from the adjusted hazard ratios from the CPH model was close to those obtained from the parametric PH and the flexible models. Lambert et al. had stated that rather than just a comparison of the AIC or BIC, it is a strength to draw conclusions from the estimates that do not change largely when different reasonable models are chosen [32]. Therefore, we can conclude that although the FPSR provided a relatively better model fit, the estimates from all the models considered in the current study are similar. With such, any of the models could be used. Among the FPSR models evaluated at different degrees of freedom, the loglikelihood, as well as the AIC and the BIC unanimously identified the 5-knot (6 degrees of freedom) flexible model as the best model. Our finding agreed with the recommendation of Royston et al., that flexible models with 5 or 6 degrees freedom will be more ideal for a sample size greater than 10,000 [19].
Nonetheless, the parameter estimates from the flexible and the CPH model are very close, despite the wide differences in all the measures of model fitness adopted in this study. These similarities underscore the strength of the CPH model when a “large” sample is followed up for a sufficiently “long” time [14]. There are possibilities that the proportional assumption could be violated, especially in a long follow-up study of this nature. Bower et al affirmed the need to account for non-proportionality, if present, to get an insight into the natural history of the disease and biological process, and to make accurate predictions [14]. Rutherford et al. emphasized that the flexible models provide a more accurate estimate of complex hazard functions as well as unbiased estimates of hazard ratios than the parametric models [22]. The flexible model has a unique advantage of providing a direct estimate of relative and absolute effects besides the quantification of differences between survival and hazard functions using a time scale of interest [33–35].
However, literature has suggested that AFT models are better alternatives to the CPH in the analysis of time to event data, especially when there is no permanent effect in the context of the follow-up period [36–39]. The AFT models assume that the effect of a covariate is either to accelerate or decelerate the life course of an outcome, say diabetes type-1, by some constant. Whereas, the CPH model assumes that the hazard functions for any two patients with baseline x vectors x1 and x2 are constrained to be proportional, upon which the estimation method is based [37–39]. Additionally, Lambert et al remarked that, unlike a CPH model, regression parameter estimates from AFT models are robust to neglected covariates and less affected by the chosen probability distribution [36].
Although the hazard of type-1 diabetes was higher among children conceived through ART than those conceived spontaneously in the bivariable analysis, we found no significant difference in the hazard of type-1 diabetes among the two groups of children in the adjusted models. This finding is corroborated in the literature by recent Danish studies [2, 3]. However, a further sub-group analysis controlling for other variables showed that the children conceived from frozen-thawed transfer had a significantly higher hazard of type-1 diabetes than those from fresh embryo transfer and children from spontaneous conception. This is in agreement with the previous study on the same cohort by Norrman et al. who reported a higher risk of type-1 diabetes among children from frozen-thawed embryo transfer than those from fresh embryo transfer (aHR: 1.52; 95% CI: 1.08–2.14) and those conceived spontaneously (aHR: 1.41; 95% CI: 1.05–1.89) [1]. This implies that the type of ART adopted is associated with the risk of type-1 diabetes.
The current study established a higher risk of type-1 diabetes in the offspring of diabetic parents. Children of diabetic mothers were about 540% at higher risk of type-1 diabetes compared with children whose mothers are not diabetic. A similar pattern but a higher likelihood of risk (780%) was found among children whose fathers were diabetic. This finding aligns with the findings of previous studies [40, 41]. Diabetic parents pose a higher risk of type-1 diabetes to their children independent of the conception method. This is plausible as a systematic review has reported a 51% significant increase in diabetes mellitus among parents who conceived by ART compared with those who conceived spontaneously [42].
The sex of the children was significantly associated with the age at the onset of type-1 diabetes. The males were at higher risk of type-1 diabetes than females. Further research is necessary to understand the biological configuration of males that put them at higher risk of type-1 diabetes. Our study is however corroborated by earlier researches which showed that about three-fifths of children who developed type-1 diabetes were boys [43].
The median age at onset of type-1 diabetes among diabetic children was 10 years (Interquartile range: 6–14 years). There appears to be a generational shift in the age at the onset of type-1 diabetes among the children. The increase in the risk of type-1 diabetes in the studied population was linear with the lowest risk among the children born between 1985 and 1990 and highest (about 36% on average) among the birth cohorts of 2001–2005, 2006–2010, and 2011–2015. Our finding is at variance with the results of the Danish study which found no association between risk of type-1 diabetes and birth cohorts [3]. Although not reported in the results, our stratified birth cohort analysis showed that covariate effects remain constant over different cohorts.
We found a higher preponderance of type-1 diabetes among children whose mothers do not smoke compared with those whose mothers smoke. This finding is at variance with findings of a comparative study on the associations of parental smoking during pregnancy and childhood-onset of type-1 diabetes in the Norwegian Mother and Child Cohort Study and the Danish National Birth Cohort by Magnues et al. [44]. Magnus et al. reported that in both cohorts, maternal smoking beyond gestational week 12 was inversely associated with type-1 diabetes” and that “in the Norwegian register-based cohort, children of mothers who still smoked at the end of pregnancy had a lower risk of type-1 diabetes” [44]. The authors noted that their findings were in contrast to teratogens operating very early during development, but similar to what has been reported earlier for effects of maternal smoking on birth weight [45]. There could be pathways to the modification of the effect of smoking on diabetes. One such is the effect of low birth weight. Reduced risk of type-1 diabetes in children with low birth weight has been reported and low birth weight is more prevalent among smoking mothers [46].
The country of birth of the parents significantly predicted the risk of type-1 diabetes with higher risks among children of Swedish and other Nordic mothers and fathers compared with children whose either parent is born outside the Nordic countries. Existing literature had suggested that people from certain countries are more likely to be diabetic than others [1, 47–51]. Studies have shown a higher preponderance of type-1 diabetes in the Scandinavian countries [1, 49, 50]. Although the reasons for these differences are yet to be well articulated in the literature.
Educational attainment of parents was significantly associated with the risk of type-1 diabetes among the children. The risk of type-1 diabetes was higher among children of those who had a lower level of paternal and maternal education. Educational attainment could be associated with the amount of information at the disposal of parents which may or may not predispose a child to type-1 diabetes. Maternal age and the parity of the index child were not significant predictors of type-1 diabetes.
Strength and limitations of the study
Our study has a major strength of a detailed analysis and comparison of the fit and performance of a different range of regression models including the semi-parametric proportional hazard, parametric proportional hazard and accelerated failure time models as well as the flexible model at different degrees of freedom to arrive at our conclusions. Also, we have used a large dataset that consists of over three million births spanning over 30 years of follow-up and nearly 19,000 cases from high validity registries. Hence, the estimated risks of type-1 diabetes were made with high statistical precision. Furthermore, our study is based on Swedish nationwide population-based registries, we, therefore, used data with almost complete coverage with no sampling or recall bias, leading to high generalizability, no loss to follow up, and almost complete ascertainment of type-1 diabetes cases. Besides, we adjusted for several potential confounders including the sex of the children, parental type-1 diabetes status, and country of birth of the parents. One of the limitations of our study was the limited time of exposure for those born after 2000. Some data were missing hence we may not have accounted for some confounders. The control group consisted of all non-ART singletons. A better control group might have been children to subfertile couples conceiving spontaneously. However, such a control group is almost impossible to identify.
Conclusion
Our estimates from the different regression models including the semi-parametric proportional hazard, parametric proportional hazard models and the flexible model at different degrees of freedom were similar. This is an indication that the models performed similarly and any of them can be used to model the data. For the Swedish birth data used in the current study, we did not find any significant differences among the performances of the regression methods. There may be a need for a similar comparative analysis of Danish data used by Hargreave et al. and Kettner et al. [2, 3] so as to ascertain that the contradictions in their findings were not due to the methods of analysis. As shown in this study, for a very large sample followed up for a long time, the CPH model provided a comparative estimate with the FPSR model. Any of the models, especially the Cox semi-parametric model, parametric proportional hazard models and the flexible model at higher degrees of freedom are appropriate in modelling the risk of type-1 diabetes among ART and spontaneously-conceived children.
The results from our large population-based nationwide data were generally reliable, indicating that mode of conception could affect the risk of type-1 diabetes among the children especially when they are from frozen-thawed embryo transfer. Although we found no significant difference in the risk of type-1 diabetes among ART and spontaneously conceived children generally, our sub-analysis showed that children from frozen-thawed embryo transfer had a higher risk than other children. This is of some concern in view of the increasing number of frozen cycles all around the world. Estimates suggest that some 390,000 babies are now born each year globally through assisted fertility techniques including IVF and ICSI while the total number of ART children have exceeded eight million by 2019 [52, 53].
It’s hard to generalise these findings to other settings. Although the considered models performed similarly in this specific setting, there’s no guarantee that this will be the case with other data sources or diseases or other settings.
Acknowledgments
The authors appreciate the logistic supports provided by the Consortium for Advanced Research and Training in Africa (CARTA) to AFF to visit the University of Gothenburg as part of his fellowship at the University of Warwick. The data analysis was done during the visit.
Citation: Fagbamigbe AF, Norrman E, Bergh C, Wennerholm U-B, Petzold M (2021) Comparison of the performances of survival analysis regression models for analysis of conception modes and risk of type-1 diabetes among 1985–2015 Swedish birth cohort. PLoS ONE 16(6): e0253389. https://doi.org/10.1371/journal.pone.0253389
1. Norrman E, Petzold M, Clausen TD, Henningsen A-K, Opdahl S, Pinborg A, et al. Type 1 Diabetes in Children Born After Assisted Reproductive Technology: A Register-Based National Cohort Study. Hum Reprod. 2020;35:221–32. pmid:31976535
2. Kettner LO, Matthiesen NB, Ramlau-Hansen CH, Kesmodel US, Bay B, Henriksen TB. Fertility treatment and childhood type 1 diabetes mellitus: a nationwide cohort study of 565,116 live births. Fertil Steril. 2016;106:1751–6. pmid:27773424
3. Hargreave M, Kjaer SK, Jørgensen ME, Jensen A. Type 1 diabetes risk in children born to women with fertility problems: a cohort study in 1.5 million Danish children. Acta Obstet Gynecol Scand. 2016;95:1441–6. pmid:27681848
4. Orsini N. Review of Flexible Parametric Survival Analysis Using Stata: Beyond the Cox Model by Patrick Royston and Paul C. Lambert. 2013.
5. Du X, Li M, Zhu P, Wang J, Hou L, Li J, et al. Comparison of the flexible parametric survival model and Cox model in estimating Markov transition probabilities using real-world data. PLoS One. 2018;13:1–13. pmid:30133454
6. Fagbamigbe AF, Akintayo AO, Oshodi O, Makinde FT, Babalola M, Damilola EA, et al. Survival Analysis and Prognostic Factors of Time to First Domestic Violence after Marriage among Married Women in Africa. Public Health. 2020;181:122–34. pmid:32007782
7. Fagbamigbe AF, Karlsson K, Derks J, Petzold M. Performance evaluation of survival regression models in analysing Swedish dental implant complication data with frailty. PLoS One. 2021;16:1–16,e0245111. pmid:33411801
8. Fagbamigbe AF, Afolabi RF, Alade KY, Adebowale AS, Yusuf BO. Unobserved Heterogeneity in the Determinants of Under-five Mortality in Nigeria: Frailty Modeling in Survival Analysis. African J Appl Stat. 2019;6:565–83.
9. Lambert P. STPM2: Stata module to estimate flexible parametric survival models. Stat Softw Components. 2010. https://ideas.repec.org/c/boc/bocode/s457128.html. Accessed 14 Feb 2021.
10. Crowther MJ. Multilevel mixed-effects parametric survival analysis: Estimation, simulation, and application. Stata J Promot Commun Stat Stata. 2019;19:931–49.
11. Nelson CP, Lambert PC, Squire IB, Jones DR. Flexible parametric models for relative suvival, with application in coronary heart disease. In: Statistics in Medicine. 2007. p. 5486–98. pmid:17893893
12. Fagbamigbe AF, Adebowale AS, Morhason-Bello IO. Survival analysis of time to uptake of modern contraceptives among sexually active women of reproductive age in Nigeria. BMJ Open. 2015;5:1–8. pmid:26671948
13. Otwombe KN, Petzold M, Martinson N, Chirwa T. A review of the study designs and statistical methods used in the determination of predictors of all-cause mortality in HIV-infected cohorts: 2002–2011. PLoS ONE. 2014;9. pmid:24498313
14. Bower H, Crowther MJ, Rutherford MJ, Andersson TML, Clements M, Liu XR, et al. Capturing simple and complex time-dependent effects using flexible parametric survival models: A simulation study. Commun Stat Simul Comput. 2019;:1–17.
15. Hinchliffe SR, Lambert PC. Flexible parametric modelling of cause-specific hazards to estimate cumulative incidence functions. BMC Med Res Methodol. 2013;13:13. pmid:23384310
16. Jatoi I, Anderson WF, Jeong J-H, Redmond CK. Breast cancer adjuvant therapy: time to consider its time-dependent effects. J Clin Oncol. 2011;29:2301–4. pmid:21555693
17. Hernán MA. The Hazards of Hazard Ratios. Epidemiology. 2010;21:13–5. pmid:20010207
18. Schechter E. Positive Measure and Integration,. In: Handbook of Analysis and Its Foundations. First. London, UK: Academic Press, Elsevier; 1997. p. 547–72. http://staffwww.dcs.shef.ac.uk/people/R.Chisholm/docs/EricSchechterHandbookofAnalysisandItsFoundations1996.pdf.
19. Royston P, Lambert PC. Flexible Parametric Survival Analysis Using Stata: Beyond the Cox Model. 1st edition. Texas: Stata Press; 2011.
20. Williams C, Lewsey JD, Mackay DF, Briggs AH. Estimation of Survival Probabilities for Use in Cost-Effectiveness Analyses. Med Decis Mak. 2016;37:0272989X1667061. pmid:27698003
21. Ng R, Kornas K, Sutradhar R, Wodchis WP, Rosella LC. The current application of the Royston-Parmar model for prognostic modeling in health research: a scoping review. Diagnostic Progn Res. 2018;2:1–15. pmid:31093554
22. Rutherford MJ, Crowther MJ, Lambert PC. The use of restricted cubic splines to approximate complex hazard functions in the analysis of time-to-event data: a simulation study. J Stat Comput Simul. 2015;85:777–93.
23. Royston P, Parmar MKB. Flexible parametric proportional-hazards and proportional-odds models for censored survival data, with application to prognostic modelling and estimation of treatment effects. Stat Med. 2002;21:2175–2197. https://pubmed.ncbi.nlm.nih.gov/12210632/. pmid:12210632
24. Lambert PC, Royston P. Further development of flexible parametric models for survival analysis. Stata J. 2009;9:265–290. https://journals.sagepub.com/doi/pdf/10.1177/1536867X0900900206.
25. Miladinovic B, Kumar A, Mhaskar R, Kim S, Schonwetter R, Djulbegovic B. A Flexible Alternative to the Cox Proportional Hazards Model for Assessing the Prognostic Accuracy of Hospice Patient Survival. PLoS One. 2012;7:1–8. pmid:23082220
26. Kearns B, Stevenson MD, Triantafyllopoulos K, Manca A. Generalized Linear Models for Flexible Parametric Modeling of the Hazard Function. Med Decis Mak. 2019;39:867–78. pmid:31556792
27. Cox D. Regression Models and Life Tables (with Discussion). J R Stat Soc Ser B. 1972;34:187–220.
28. Syriopoulou E, Mozumder SI, Rutherford MJ, Lambert PC. Robustness of individual and marginal model-based estimates: A sensitivity analysis of flexible parametric models. Cancer Epidemiol. 2019;58:17–24. pmid:30439603
29. Akaike H. Information theory as an extension of the maximum likelihood principle. In: Petrov BN, Csaki F, editors. In Second International Symposium on Information Theory. Ed. Budapest: Akademiai Kiado; 1973. p. 267–281.
30. Schwarz G. Estimating the Dimension of a Model. Ann Stat. 1978;6:461–4.
31. Brewer MJ, Butler A, Cooksley SL. The relative performance of AIC, AICC and BIC in the presence of unobserved heterogeneity. Methods Ecol Evol. 2016;7:679–92.
32. Lambert PC, Dickman PW, Rutherford MJ. Comparison of different approaches to estimating age standardized net survival. BMC Med Res Methodol. 2015;15:64. pmid:26275405
33. Kandala N-B, Connock M, Pulikottil-Jacob R, Sutcliffe P, Crowther MJ, Grove A, et al. Setting benchmark revision rates for total hip replacement: Analysis of registry evidence. BMJ. 2015;(mar09 1): pmid:25752749
34. King NB, Harper S, Young ME. Use of relative and absolute effect measures in reporting health inequalities: Structured review. BMJ. 2012;sep03 1:e5774. pmid:22945952
35. Faruk A. The comparison of proportional hazards and accelerated failure time models in analyzing the first birth interval survival data. In: J. Phys, Conf. Ser. 974 012008. 2018. p. 1–10.
36. Lambert P, Collett D, Kimber A, Johnson R. Parametric accelerated failure time models with random effects and an application to kidney transplant survival. Stat Med. 2004;23:3177–92. pmid:15449337
37. Pourhoseingholi MA, Hajizadeh E, Dehkordi BM, Safaee A, Abadi A, Zali MR. Comparing cox regression and parametric models for survival of patients with gastric Carcinoma. Asian Pacific J Cancer Prev. 2007;8:412–6. https://pubmed.ncbi.nlm.nih.gov/18159979/. pmid:18159979
38. Teshnizi SH1, Ayatollahi SMT. The Cox-Snell Residual Plots for Assessing Goodness of Fit Parametric and Cox Models. Asian Pac J Cancer Prev,. 2017;18:981–5. pmid:28545196
39. Kargarian-Marvasti S, Rimaz S, Abolghasemi J, Heydari I. Comparing of Cox model and parametric models in analysis of effective factors on event time of neuropathy in patients with type 2 diabetes. J Res Med Sci |. 2017. pmid:29184573
40. Chernausek SD, Arslanian S, Caprio S, Copeland KC, Kelsey MM, Koontz MB, et al. Relationship Between Parental Diabetes and Presentation of Metabolic and Glycemic Function in Youth With Type 2 Diabetes: Baseline Findings From the TODAY Trial. Diabetes Care. 2016;39:110–7. pmid:26577415
41. Law JR, Stafford JM, D’Agostino RB, Badaru A, Crume TL, Dabelea D, et al. Association of parental history of diabetes with cardiovascular disease risk factors in children with type 2 diabetes. J Diabetes Complications. 2015;29:534–9. pmid:25784087
42. Mohammadi M, Khedmati Morasae E, Maroufizadeh S, Almasi-Hashiani A, Navid B, Amini P, et al. Assisted reproductive technology and the risk of gestational diabetes mellitus: a systematic review and meta-analysis. Middle East Fertility Society Journal. 2020;25:6.
43. Rawshani A, Rawshani A, Franzén S, Eliasson B, Svensson AM, Miftaraj M, et al. Mortality and cardiovascular disease in type 1 and type 2 diabetes. N Engl J Med. 2017;376:1407–18. pmid:28402770
44. Magnus MC, Tapia G, Olsen SF, Granstrom C, Mårild K, Ueland PM, et al. Parental Smoking and Risk of Childhood-onset Type 1 Diabetes. Epidemiology. 2018;29:848–56. pmid:30074542
45. Wickstrom R. Effects of Nicotine During Pregnancy: Human and Experimental Evidence. Curr Neuropharmacol. 2007;5:213–22. pmid:19305804
46. Cardwell CR, Stene LC, Joner G, Davis EA, Cinek O, Rosenbauer J, et al. Birthweight and the risk of childhood-onset type 1 diabetes: A meta-analysis of observational studies using individual patient data. Diabetologia. 2010;53:641–51. pmid:20063147
47. Spurr S, Bally J, Bullin C, Allan D, McNair E. The prevalence of undiagnosed Prediabetes/type 2 diabetes, prehypertension/hypertension and obesity among ethnic groups of adolescents in Western Canada. BMC Pediatr. 2020;20:31. pmid:31973728
48. Whincup PH, Nightingale CM, Owen CG, Rudnicka AR, Gibb I, McKay CM, et al. Early Emergence of Ethnic Differences in Type 2 Diabetes Precursors in the UK: The Child Heart and Health Study in England (CHASE Study). PLoS Med. 2010;7:e1000263. pmid:20421924
49. Fadl HE, Simmons D. Trends in diabetes in pregnancy in Sweden 1998–2012. BMJ Open Diabetes Res Care. 2016;4:e000221. pmid:27547412
50. Wandell P, Carlsson A, Steiner K. Prevalence of Diabetes Among Immigrants in the Nordic Countries. Curr Diabetes Rev. 2010;6:126–33. pmid:20201798
51. Banerjee AT, Shah BR. Differences in prevalence of diabetes among immigrants to Canada from South Asian countries. Diabet Med. 2018;35:937–43. pmid:29663510
52. Adamson GD, de Mouzon J, Chambers GM, Zegers-Hochschild F, Mansour R, Ishihara O, et al. International Committee for Monitoring Assisted Reproductive Technology: world report on assisted reproductive technology, 2011. Fertil Steril. 2018;110:1067–80. pmid:30396551
53. De Geyter C, Calhaz-Jorge C, Kupka MS, Wyns C, Mocanu E, Motrenko T, et al. ART in Europe, 2015: results generated from European registries by ESHRE. Hum Reprod Open. 2020;2020. pmid:32123753
You have requested "on-the-fly" machine translation of selected content from our databases. This functionality is provided solely for your convenience and is in no way intended to replace human translation. Show full disclaimer
Neither ProQuest nor its licensors make any representations or warranties with respect to the translations. The translations are automatically generated "AS IS" and "AS AVAILABLE" and are not retained in our systems. PROQUEST AND ITS LICENSORS SPECIFICALLY DISCLAIM ANY AND ALL EXPRESS OR IMPLIED WARRANTIES, INCLUDING WITHOUT LIMITATION, ANY WARRANTIES FOR AVAILABILITY, ACCURACY, TIMELINESS, COMPLETENESS, NON-INFRINGMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Your use of the translations is subject to all use restrictions contained in your Electronic Products License Agreement and by using the translation functionality you agree to forgo any and all claims against ProQuest or its licensors for your use of the translation functionality and any output derived there from. Hide full disclaimer
© 2021 Fagbamigbe et al. This is an open access article distributed under the terms of the Creative Commons Attribution License: http://creativecommons.org/licenses/by/4.0/ (the “License”), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Abstract
The goal is to examine the risk of conception mode-type-1 diabetes using different survival analysis modelling approaches and examine if there are differentials in the risk of type-1 diabetes between children from fresh and frozen-thawed embryo transfers. We aimed to compare the performances and fitness of different survival analysis regression models with the Cox proportional hazard (CPH) model used in an earlier study. The effect of conception modes and other prognostic factors on type-1 diabetes among children conceived either spontaneously or by assisted reproductive technology (ART) and its sub-groups was modelled in the earlier study. We used the information on all singleton children from the Swedish Medical Birth Register hosted by the Swedish National Board of Health and Welfare, 1985 to 2015. The main explanatory variable was the mode of conception. We applied the CPH, parametric and flexible parametric survival regression (FPSR) models to the data at 5% significance level. Loglikelihood, Akaike and Bayesian information criteria were used to assess model fit. Among the 3,138,540 singletons, 47,938 (1.5%) were conceived through ART (11,211 frozen-thawed transfer and 36,727 fresh embryo transfer). In total, 18,118 (0.58%) of the children had type-1 diabetes, higher among (0.58%) those conceived spontaneously than the ART-conceived (0.42%). The median (Interquartile range (IQR)) age at onset of type-1 diabetes among spontaneously conceived children was 10 (14–6) years, 8(5–12) for ART, 6 (4–10) years for frozen-thawed embryo transfer and 9 (5–12) years for fresh embryo transfer. The estimates from the CPH, FPSR and parametric PH models are similar. There was no significant difference in the risk of type-1 diabetes among ART- and spontaneously conceived children; FPSR: (adjusted Hazard Ratio (aHR) = 1.070; 95% Confidence Interval (CI):0.929–1.232, p = 0.346) vs CPH: (aHR = 1.068; 95%CI: 0.927–1.230, p = 0.361). A sub-analysis showed that the adjusted hazard of type-1 diabetes was 37% (aHR = 1.368; 95%CI: 1.013–1.847, p = 0.041) higher among children from frozen-thawed embryo transfer than among children from spontaneous conception. The hazard of type-1 diabetes was higher among children whose mothers do not smoke (aHR = 1.296; 95%CI:1.240–1.354, p<0.001) and of diabetic mothers (aHR = 6.419; 95%CI:5.852–7.041, p<0.001) and fathers (aHR = 8.808; 95%CI:8.221–9.437, p<0.001). The estimates from the CPH, parametric models and the FPSR model were close. This is an indication that the models performed similarly and any of them can be used to model the data. We couldn’t establish that ART increases the risk of type-1 diabetes except when it is subdivided into its two subtypes. There is evidence of a greater risk of type-1 diabetes when conception is through frozen-thawed transfer.
You have requested "on-the-fly" machine translation of selected content from our databases. This functionality is provided solely for your convenience and is in no way intended to replace human translation. Show full disclaimer
Neither ProQuest nor its licensors make any representations or warranties with respect to the translations. The translations are automatically generated "AS IS" and "AS AVAILABLE" and are not retained in our systems. PROQUEST AND ITS LICENSORS SPECIFICALLY DISCLAIM ANY AND ALL EXPRESS OR IMPLIED WARRANTIES, INCLUDING WITHOUT LIMITATION, ANY WARRANTIES FOR AVAILABILITY, ACCURACY, TIMELINESS, COMPLETENESS, NON-INFRINGMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Your use of the translations is subject to all use restrictions contained in your Electronic Products License Agreement and by using the translation functionality you agree to forgo any and all claims against ProQuest or its licensors for your use of the translation functionality and any output derived there from. Hide full disclaimer