Intended for healthcare professionals

Letters

Maternal age and risk of type 1 diabetes in children

BMJ 2001; 322 doi: https://doi.org/10.1136/bmj.322.7300.1489/a (Published 16 June 2001) Cite this as: BMJ 2001;322:1489

Flawed analysis invalidates conclusions

  1. Graham Byrnes, biostatistician (Graham.Byrnes{at}mh.org.au)
  1. Clinical Epidemiology and Health Service Evaluation Unit, Royal Melbourne Hospital, Parkville, Victoria 3050, Australia
  2. Department of Epidemiology and Public Health, Queen's University Belfast, Belfast BT12 6BJ
  3. Department of Clinical Science, Umeå University, S-90185 Umeå, Sweden
  4. Department of Paediatrics, Pécs University, H-7623 Pécs, József A. u. 7, Hungary On behalf of the EURODIAB Substudy 2 Study Group.
  5. Department of Paediatrics, University of Auckland, Auckland, New Zealand
  6. Section of Epidemiology, National Institute of Public Health, N-0403 Oslo, Norway
  7. Aker Diabetes Research Centre, Aker University Hospital, Oslo, Norway
  8. Diabetes and Metabolism, Division of Medicine, University of Bristol, Southmead Hospital, Bristol BS10 5NB

    EDITOR—The conclusion drawn by Bingley et al, of dependence on both birth order and maternal age, is a tempting explanation of previous ambiguous results.1 I believe, however, that their study is seriously flawed in the analysis of data, to an extent that the results are completely meaningless.

    There are three apparent problems.

    Firstly, the cohort studied consists of children with a sibling diagnosed with type 1 diabetes. Hence every only child in the study must be a case. Roughly half of the children from two child families will be cases, and so on. This immediately produces an (inverse) association between the number of siblings and the diagnosis of type 1 diabetes. To see this clearly, suppose that cases occur purely at random. To make the numbers easy, suppose the incidence is 1%. Now select 1000 only child cases and 1000 cases with a single sibling. Almost half of the 2010 expected cases will be only children, but none of the non-cases.

    Moreover, birth order relates to family size: an only child cannot be a second born. So of our hypothetical 3000 children with 2010 cases, 1505 cases will be expected to be first born, compared with only 495 of 990 non-cases; a risk-ratio of 1.5. The dependence on maternal age then follows immediately if we allow that women who have their first child later tend to have fewer children. The results Bingley et al obtain are therefore consistent with the null hypothesis of no association between maternal age or birth order with type 1 diabetes.

    Secondly, Cox regression is essentially a form of nested case-control study and relies on the assumption that the events are independent. This does not seem a reasonable assumption when the study contains many sets of siblings, so the confidence intervals calculated are probably too small.

    Thirdly, it is not clear how the result would generalise to a population of women who have not had a diabetic child. The first of the above points is at first sight the most serious, although reanalysis could overcome it: only children should be excluded and a single control selected from each family to eliminate the family size effect. It is hard to see, however, how you would ever circumvent the third argument.

    References

    1. 1.

    Relative risks by maternal age are biased

    1. C C Patterson, senior lecturer in medical statistics (c.patterson{at}qub.ac.uk),
    2. G Dahlquist, professor of paediatrics,
    3. G Soltész, professor of paediatrics
    1. Clinical Epidemiology and Health Service Evaluation Unit, Royal Melbourne Hospital, Parkville, Victoria 3050, Australia
    2. Department of Epidemiology and Public Health, Queen's University Belfast, Belfast BT12 6BJ
    3. Department of Clinical Science, Umeå University, S-90185 Umeå, Sweden
    4. Department of Paediatrics, Pécs University, H-7623 Pécs, József A. u. 7, Hungary On behalf of the EURODIAB Substudy 2 Study Group.
    5. Department of Paediatrics, University of Auckland, Auckland, New Zealand
    6. Section of Epidemiology, National Institute of Public Health, N-0403 Oslo, Norway
    7. Aker Diabetes Research Centre, Aker University Hospital, Oslo, Norway
    8. Diabetes and Metabolism, Division of Medicine, University of Bristol, Southmead Hospital, Bristol BS10 5NB

      EDITOR—Statisticians familiar with survival analysis techniques may have spotted methodological problems in the paper by Bingley et al,1 but other readers might have been suspicious about the very high risks of diabetes among the offspring of families in the study depicted in figure 1, especially in the oldest maternal age groups, where fewer than 40% of offspring would be predicted to be free of diabetes by 20 years of age.

      Bingley et al ascertained families through affected children (probands) and then applied survival analysis techniques to all the offspring in these families, including probands, with age at diagnosis of affected children as the end point and follow up in unaffected children censored at their age on the date of last contact. This strategy will result in ascertainment bias, which will lead to overestimation of the risk of diabetes in these families. The mothers in this study will grow older, and some will have more children, most of whom will not develop diabetes. These children, together with ageing younger children in the family, will contribute to improving the rates of diabetes free survival in the analysis depicted in figure 1, particularly in the oldest maternal age categories. By ignoring these children Bingley et al introduce biases into the relative risks for maternal age categories shown in their table. Their plot of estimated risks in figure 3 obtained by extrapolating these relative risks to birth cohorts in England and Wales is also flawed.

      We conducted a large case-control study including over 1000 cases of childhood type 1 diabetes and over 2000 controls drawn from eight European centres in which we collected data on the family history of type 1 diabetes in the siblings of both cases and controls.2 To explore the possible extent of these biases, we reanalysed the data from the cases and their siblings using the same strategy as Bingley et al, with follow up ending at the time of our contact with the families. The table shows that the results are similar to those of Bingley et al.

      When, however, we analysed our data using appropriate statistical methods comparing the distribution of maternal age in cases with the corresponding distribution in controls with the Mantel-Haenszel odds ratio pooled over our eight centres, we found a much weaker relation with maternal age (table). These results are more in keeping with published data and were little altered by adjustment for potential confounding variables.

      Comparison of association between maternal age and type I childhood diabetes using strategy of Bingley et al1 and Mantel-Haenszel odds ratio in EURODIAB study2

      View this table:

      References

      1. 1.
      2. 2.

      Other secular trends may explain associations with diabetes risk

      1. Alistair J Gunn, paediatrician (aj.gunn{at}auckland.ac.nz),
      2. Wayne S Cutfield, paediatrician,
      3. Paul L Hofman, paediatrician,
      4. Craig Jeferries, clinical fellow
      1. Clinical Epidemiology and Health Service Evaluation Unit, Royal Melbourne Hospital, Parkville, Victoria 3050, Australia
      2. Department of Epidemiology and Public Health, Queen's University Belfast, Belfast BT12 6BJ
      3. Department of Clinical Science, Umeå University, S-90185 Umeå, Sweden
      4. Department of Paediatrics, Pécs University, H-7623 Pécs, József A. u. 7, Hungary On behalf of the EURODIAB Substudy 2 Study Group.
      5. Department of Paediatrics, University of Auckland, Auckland, New Zealand
      6. Section of Epidemiology, National Institute of Public Health, N-0403 Oslo, Norway
      7. Aker Diabetes Research Centre, Aker University Hospital, Oslo, Norway
      8. Diabetes and Metabolism, Division of Medicine, University of Bristol, Southmead Hospital, Bristol BS10 5NB

        EDITOR—The report by Bingley et al of an association between age of the mother and birth order, and risk of type 1 diabetes mellitus in children is of interest, but the findings do not seem to be consistent.1 If the age of the mother is associated with a real increase in risk, then it is surprising that subsequent children could possibly have a reduced risk given that maternal age must have increased. The hypothesis given to explain the apparent relation, of more mature immune responses in older women priming the fetal system in some way, is biologically improbable given the functional immaturity of the fetal immune system.

        Since the cohort spans a significant interval, it is not clear that Bingley et al have adjusted for secular trends that might have been associated with greater risk. The most important of these factors is the trend towards childbirth in older women over the study period. The greatest risk of type 1 diabetes mellitus was reported in the offspring of a very small group of mothers giving birth after 45 years of age.1

        Several other factors in older women might have influenced the reported observation, including assisted reproduction, ethnicity, maternal illness, and obesity that were not presented.

        Other potentially confounding factors not examined include the changing incidences of type 1 diabetes mellitus and the increasing rate of obesity in childhood, which tends to lower the age of onset of children at risk of type 1 diabetes, because of insulin resistance. 2 3

        Type 1 pre-diabetes antibodies were apparently not measured to validate the diagnosis of type 1 diabetes mellitus. Clinical diagnosis based on apparent requirement for insulin treatment or presentation with ketoacidosis is not specific at a time of increasing type 2 diabetes mellitus in children.

        References

        1. 1.
        2. 2.
        3. 3.

        Association may disappear after adjusting for year of birth

        1. Lars Christian Stene, research fellow (lars.christian.stene{at}folkehelsa.no),
        2. Geir Joner, paediatrician
        1. Clinical Epidemiology and Health Service Evaluation Unit, Royal Melbourne Hospital, Parkville, Victoria 3050, Australia
        2. Department of Epidemiology and Public Health, Queen's University Belfast, Belfast BT12 6BJ
        3. Department of Clinical Science, Umeå University, S-90185 Umeå, Sweden
        4. Department of Paediatrics, Pécs University, H-7623 Pécs, József A. u. 7, Hungary On behalf of the EURODIAB Substudy 2 Study Group.
        5. Department of Paediatrics, University of Auckland, Auckland, New Zealand
        6. Section of Epidemiology, National Institute of Public Health, N-0403 Oslo, Norway
        7. Aker Diabetes Research Centre, Aker University Hospital, Oslo, Norway
        8. Diabetes and Metabolism, Division of Medicine, University of Bristol, Southmead Hospital, Bristol BS10 5NB

          EDITOR—Bingley et al reported an association between increased age of the mother at delivery and increased risk of type 1 diabetes in children.1 The simultaneous increase in diabetes and maternal age in the population during the period of recruiting families may have led to a spurious association confounded by unknown factors that have caused the rise in the incidence of diabetes and are simply correlated with increased maternal age in the population over time.

          Since siblings are usually born within a few years, the year of birth has been controlled to some extent, but not completely. Adjusting the analyses for year of birth should control such potential confounding. Since Bingley et al did not state that adjustment for birth year was done, the observed association may be at least partly the result of such confounding.

          Furthermore, it is possible that genetic susceptibility as indicated by having a sibling with type 1 diabetes increases the relative risk conferred by a high maternal age. If this is the case, the data presented by Bingley et al would not be generalisable to the general population.2

          The claim that matching for genetic susceptibility is an advantage compared with studies using population controls1 applies only if genetic susceptibility is a confounder and not an effect modifier. Genetic susceptibility to type 1 diabetes in the children must be associated with the mother's age at delivery in order to be a confounder in studies using population controls; something that is hard to envisage.

          We did a cohort analysis of all live births in Norway between 1984 and 1998, including 1105 children who developed type 1 diabetes, by linking the medical birth registry of Norway with the national childhood diabetes registry. 3 4 Our study should have sufficient power to detect even weak associations. We found no significant association between maternal age and the incidence of type 1 diabetes, neither crude nor after adjustment for birth order and year of birth. In Norway, mean maternal age has increased in recent years, but the incidence of diabetes has not. However, the criteria for registration of cases dictate that during any period of registration, those who are born later have a younger age at onset. We observed a crude association between maternal age and age at onset among cases similar to those found by Bingley et al, but the association disappeared after adjusting for year of birth. We suspect that this would be the case for the data of Bingley et al, too.

          References

          1. 1.
          2. 2.
          3. 3.
          4. 4.

          Authors' reply

          1. Polly J Bingley, reader,
          2. Edwin A M Gale, professor of diabetic medicine
          1. Clinical Epidemiology and Health Service Evaluation Unit, Royal Melbourne Hospital, Parkville, Victoria 3050, Australia
          2. Department of Epidemiology and Public Health, Queen's University Belfast, Belfast BT12 6BJ
          3. Department of Clinical Science, Umeå University, S-90185 Umeå, Sweden
          4. Department of Paediatrics, Pécs University, H-7623 Pécs, József A. u. 7, Hungary On behalf of the EURODIAB Substudy 2 Study Group.
          5. Department of Paediatrics, University of Auckland, Auckland, New Zealand
          6. Section of Epidemiology, National Institute of Public Health, N-0403 Oslo, Norway
          7. Aker Diabetes Research Centre, Aker University Hospital, Oslo, Norway
          8. Diabetes and Metabolism, Division of Medicine, University of Bristol, Southmead Hospital, Bristol BS10 5NB

            EDITOR—We accept Byrne's trenchant comments about birth order and retract this conclusion. The effects of maternal age, paternal age, and sex were, however, still present when we reanalysed the data exactly as he suggested. In this reduced sample of 2314 children, Cox proportional hazard regression showed that the risk of type 1 diabetes increased by 12.6% (95% confidence interval 6.0% to 9.6%) for each five year band in maternal age at delivery, and by 8.5% (3.0% to 14%) for each five year increase in paternal age at delivery. As in the published analysis, standard errors have been adjusted for clustering within families.

            Readers need not be suspicious about the high risk of diabetes in this cohort since probands were included in the analysis. Potential bias resulting from shorter follow up of later born siblings was minimal since, in contrast to the EURODIAB study, this is a long term prospective family study. The median age of randomly selected control siblings in the reanalysis was 16, while probands were diagnosed at a median age of 10.7 years.

            These correspondents question whether our observations can be generalised to the population as a whole, but they may have misunderstood the assumptions underlying this approach. Our study is population based,1 and the families therefore represent the subset of children in the region who are genetically susceptible to type 1 diabetes. Relative risks within genetically susceptible families were based on comparison of cases and sibling controls. To relate these to the incidence of diabetes in the region we assumed genetic susceptibility and other factors determining progression to diabetes were independent of maternal age and that the proportion of children with genetic susceptibility remained constant. Simple algebra then shows the incidence of diabetes is proportional to the product of the relative risk for each maternal age group and the proportion of children in that group, thus permitting the potential impact of changes in maternal age distribution to be assessed.

            Previous studies using population controls show a similar effect of older maternal age, as do Patterson et al, although the failure to reproduce this in the Norwegian sample is clearly important. We acknowledge the risk of a spurious association when both maternal age at delivery and the incidence of diabetes are rising. We were unable to adjust for date of birth as suggested because this was not a birth cohort study and many of the children were born outside the ascertainment period of 1985-99. The distributions of dates of birth were, however, similar in probands and sibling controls (P=0.6). Misclassification of diabetes was not a problem since 96% of 491 probands tested had antibodies to glutamate decarboxylase or IA-2, or both.2

            Future studies will need to take account of the important considerations raised by these correspondents.

            References

            1. 1.
            2. 2.
            View Abstract