The sample was the Twins Early Development Study (TEDS), a study of twins born in the UK between 1994 – 1996 12. The TEDS sample has been shown to be reasonably representative of the UK population 1213. All twins and parents involved in TEDS provide informed consent for each stage of the study. Ethical approval for the Twins Early Development Study has been provided by the King's College London ethics committee, reference number: 05/Q0706/228. Response rates based on active families were 65%, 63%, and 65%, respectively at the three ages. For the purposes of the current study, we excluded from the analyses families in which at least one member of the twin pair had a specific medical syndrome or was an extreme outlier for perinatal problems such as extreme low birth weight. Data was included from monozygotic (MZ) male and female twins; dizygotic (DZ) male and female twins, and opposite-sex DZ twins. Across all genotypes the sample was 42% MZ (MZ female: 23%; MZ male: 19%), 32% DZ (DZ female: 17%; DZ male: 15%), and 26% opposite-sex DZ.

The mean ages of the twins when questionnaires were returned were 4.05 (s.d. = .16), 7.05 (s.d. = .25), and 9.93 (s.d. = .87) respectively. Zygosity was assessed through a parent questionnaire of physical similarity, which has been shown to be over 95% accurate when compared to DNA testing 14. For cases where zygosity was unclear from this questionnaire, DNA testing was conducted.

Height and weight data were collected by postal questionnaire from the parents when their children were 4, 7 and 10 years of age. Correlations with measured heights and weights in a sub-sample of these children at age 11 were 0.83 and 0.90. The sample included 5000 children at each age. BMI was calculated from the height and weight data (BMI = weight (kg)/(height (m)²)) and converted to BMI z-scores. BMI z-scores take into consideration the child's age and sex. They were based on 1990 UK growth reference curves 15 and were calculated using the program lmsGrowth (available from http://homepage.mac.com/tjcole). Children were categorized as normal weight, overweight or obese based on International Obesity Task Force (IOTF) criteria 16; which are based on age-specific and sex-specific cut-off points linked to growth curves, and correspond to BMI criteria for normal weight, overweight and obesity in adults 16. Although there are disagreements about the correct cut-off for childhood obesity [see for example 171819] we opted for the IOTF criteria because it was important that the cut-offs used in this study were internationally meaningful, in addition, we are not using the cut-offs for clinical purposes. Finally, the focus of our paper is not on the categorical analyses, but on the more powerful continuous analyses.

The percentages of the sample in each category (normal weight, overweight and obese) were 83%, 12% and 5% at 4 years; 88%, 9% and 3% at 7 years; and 86%, 11% and 3% at 10 years.

The mean raw BMI scores (i.e. unstandardized for age and sex) were 15.83 (sd = 1.96) at 4 years; 15.73 (sd = 1.99) at 7 years; and 17.24 (sd = 2.93) at 10 years. The raw BMIs of the 3 groups (normal weight, overweight and obese) at each age were: 4 years: 15.19 (sd = 1.29); 18.06 (sd = 0.53); and 20.95 (sd = 1.81); 7 years: 15.22 (sd = 1.29); 18.92 (sd = 0.76); and 22.53 (sd = 1.96); and 10 years: 16.36 (sd = 1.68); 21.60 (sd = 1.41); and 27.41 (sd = 5.07).

Candidate SNPs were selected based on the 2005 update of the human obesity gene map 8. Nine SNPs were selected on the following criteria: i) the association must have been shown in a sample with at least 500 participants, ii) the measured phenotype had to be BMI, waist circumference or obesity, and iii) the SNP had to have a minor allele frequency greater than 0.1 because statistical power to detect associations is too low for alleles with lower frequencies. The SNPs that were selected are shown in Table 1, which also includes information about the genes in which they reside, the risk allele (that is, which of the two alleles is associated with higher BMI), and the phenotype of interest. Genotyping of these nine SNPs was performed by Kbiosciences, UK 20 using the competitive allele specific PCR (KASPar) method. The error rate of this method is low (< 0.5% in our analyses of blind duplicates).

In addition, SNP rs7566605 was genotyped in an attempt to replicate a recent report of an association between rs7566605 and obesity in four out of five populations 21, although it subsequently failed to replicate in several studies 22. Genotyping for rs7566605 was performed in-house using Taqman on a subsample of 3000 children. The overall genotyping call rate was 93%, and we only called the genotypes if the TaqMan quality value exceeded 95. The allele frequencies for all 10 candidate SNPs are included in Table 1; all the frequencies in our sample are similar to those for Europeans in HapMap.

Results were analyzed using the QTDT total association model 23, which takes into account the twin pair structure by comparing variance between and within pairs. We did not have parental genotypes, so were unable to include IBD status in the model. The total association model was chosen after we had verified that there was no significant population stratification. In the QTDT analyses we specifically modeled a polygenic variance component, as well as a non-shared environment and twin-specific environmental variance components. Variance components were tested for significance before being included in the model. The heritability of BMI estimated by the QTDT analyses was 41% at 4 years, 60% at 7 years and 75% at 10 years.

Our sample of 5000 (2500 twin pairs) provides 80% power to detect an effect size (r²) of 0.2% (p = .01, one-tailed, as calculated using 11). We corrected for multiple testing using the False Discovery Rate 24. This multiple testing correction was applied across all ages (i.e. 30 tests) rather than at each age separately, because the measures of BMI at each age are not independent. We analyzed the entire sample of 5000 in order to maximize power to detect associations of small effect size. In addition, because the original reports compared obese and overweight cases versus controls, we also divided the sample into groups (obese, overweight, and normal) and compared genotypic frequency differences between the groups using chi-square analysis. We compared the obese group to the normal weight group and also the obese and overweight groups (combined) were compared to the normal-weight group. These analyses were done on one randomly selected member of each twin pair.

A specific limitation of the present study is the use of parental reports of children's heights and weights, although previous research has shown that they are reasonably reliable 25. This was confirmed in our sample on a sub-sample at 11 years: Home measurements of height and weight, taken by trained researchers, correlated .83 and .90, respectively, with parental report. A further limitation is that we were only able to study BMI. It would have been beneficial to include other measures of adiposity, but this was not feasible in this large sample. In addition BMI was measured only at ages 4, 7 and 10 years. It is possible that we could have got different results if we had more frequent measurements of BMI during childhood, so that we could more accurately estimate when in development the association with the candidate gene emerged. However, as we found no significant associations, this limitation is unlikely to impact on the interpretation of these results.

Finally, this analysis does not consider genetic influences that impact on change during development. This could have been done using growth curve analyses, similar to Podolsky et al. 26. However, growth curve analyses are best performed with at least four measurements, so we were unable to include such analyses in this study. All of the results should be interpreted in light of these limitations.