Setting and eligibility
Consecutive newborns of an urban teaching hospital in London, England, and those referred by family physicians were eligible for this prospective cohort study, conducted from November 2010 to January 2013. This hospital provides ante and postnatal services for the local population and acts as a tertiary referral unit for pregnant women and neonates. The local population is mixed representing all socioeconomic groups in a multi ethnic community. Newborns were ineligible if they had a syndrome or neurological disorder, which by definition excludes DDH as a possible diagnosis. The Institutional Review Board approved the study. Caregivers provided written informed consent.
Predictor selection and measurement
A dedicated research team examined each newborn and community referral and ascertained all variables prospectively. The median newborn age at this assessment was 1 day (interquartile range, 0-1 days). We assessed newborns for the presence of widely-accepted DDH risk factors, based on systematic reviews,[1, 5, 6] clinical guidelines,  and a consensus document of experts from 34 countries . These included a family history of DDH in a first degree relative (parental self-report); breech lie in the third trimester with cephalic presentation or breech presentation at birth; and history of oligohydramnios (ultrasound-based diagnosis at 18th to 20th week of gestation with amniotic fluid index ≤5). We also recorded sex, birth weight; whether it was a first born child; a twin pregnancy; a vaginal or a caesarean delivery.
All newborns underwent an examination of their hips and we ascertained the presence of standardized diagnostic criteria for DDH  – Ortolani or Barlow sign; asymmetry in hip abduction ≥20°; leg length discrepancy (Galeazzi sign). We recorded the presence of a torticollis. We included all foot deformity and grouped them (based on the assessment of a physiotherapist) into those with marked postural deformities (e.g. metatarsus adductus; calcaneovalgus) that we followed on at least one occasion with a physiotherapist to ensure improvement; and those with clubfeet that we immediately referred to a surgeon). Neonatology residents, overseen by attending neonatologists and dedicated research nurses, carried out these assessments. Newborns with abnormal or equivocal clinical examination of the hip underwent a concurrent examination by an attending neonatologist or attending pediatric orthopaedics surgeon to confirm eligibility. During the course of the study 68 infants were referred by family physicians for a possible diagnosis of DDH due to risk factors and we included them all. Central project staff ensured quality through training sessions and regular on-site data quality audits. We ensured recruitment of consecutive patients by daily comparing the number of newborns with the number of case report forms completed (birth occurring on weekends were compared on Mondays). The same was done weekly for referrals made by family physicians. We randomly performed double examination of 230 hips (115 infants) to determine the inter-rater reliability in the hip examination between the team of residents and a pediatric orthopaedics surgeon: hips were examined reliably (inter-rater κ=0.90).
One outcome was assessed for all patients – ultrasound-based diagnosis of DDH at age 6-8 weeks (median, 8 weeks). We classified sonograms with use of standardized diagnostic criteria based on international consensus  which define an α-angle  <55° as DDH. This outcome is robust as any such finding warrants monitoring or treatment . In rare instances where hips were scanned before 6 weeks because of suspected dislocation, we defined α <43° as DDH .
Two senior attending pediatric radiologists and 3 sonographers specifically trained in infant hip ultrasonography performed (and reported) all ultrasound tests in dedicated clinics according to a standardized protocol  (GE Medical Systems, Chalfont St. Giles, United Kingdom). We (AR, PH) evaluated all scans that had been reported as α <60°, blinded to predictors and original report; we performed our own measurement of the α-angle, which we used for this study. We did the same for a random sample of scans that had been reported as α >60°; we found that our measurements were consistent with the initial report in all cases and thus assumed robustness of the initial report. Inter-rater reliability studies for the α-angles in the original scan reports and that reported by the study team showed an intra-class correlation coefficient of 0.88 (95% confidence interval, 0.86-0.90).
Of 13,208 livebirths and 68 community referrals, 2,276 (17%) newborns were eligible (Figure 1). Of those, 80 (3%) caregivers declined participation. Of 2,191 consented newborns, 1,953 (89%) completed the ultrasound and were included in the analysis; 238 (11%) infants who did not were excluded (Table 1).
We summarized all variables as means and standard deviations or frequencies and proportions, respectively. We treated the occurrence of DDH as a binary outcome (based on α <55° at age >42 days, or α <50 at age <42 days). We considered for inclusion in the risk prediction model only candidate predictors with a prevalence of >2% in order to avoid imprecise estimates of regression coefficients . We combined less frequently occurring predictors if clinically plausible (e.g. we created the variable abnormal hip examination, encompassing any hips showing any of the following: positive Barlow, Ortolani, or Galeazzi signs, or asymmetry in abduction ≥20° - hips exhibiting multiple of these criteria were counted once only). We also combined all foot deformities warranting followup with a physiotherapist or with a surgeon into one variable, ‘foot deformity’. The variables oligohydramnios and torticollis were omitted from analysis due to their low prevalence. This resulted in 9 candidate predictors to be analyzed (Appendix 1).
Variables associated with DDH at p<0.10 in simple logistic regression were taken to multivariate analysis. They were: female sex, family history of DDH, first born, birth weight, foot deformity warranting followup, and abnormal hip examination. Because ‘breech presentation/breech in last trimester” is a widely accepted risk factor, we explored this variable further despite it not meeting this threshold. We tested if the mode of delivery (vaginal or caesarean) would affect its association with DDH. No such effect was seen. Hence, we omitted ‘breech presentation/breech in last trimester’ from further analyses. We dichotomised the variable birth weight (because it would be easier to use for clinicians) based on an accepted threshold of 4000g [4, 11].
For derivation of a risk prediction model, we included all 6 candidate predictors in a multivariable logistic regression model and used backward elimination and retained predictors at the 5% significance level. Four of 6 predictors retained in the model: female sex; family history; birth weight >4000g; and abnormal hip examination. We also developed a model without the variable ‘abnormal hip examination’ as this may be of interest for certain groups of clinicians.
Birthweight was missing for 61 cases and first born was missing for 35 cases. We used multiple imputation (assuming data were missing at random) to account for this missing data when fitting the final multivariable models. The imputation models included all risk factors considered in the univariate analysis (which we assume includes all predictors of missingness). We generated 20 data sets and ran logistic regression, using the whole data set and implementing a bootstrap (200 samples) for each imputed data set to correct for overfitting. Results from analyses using 20 imputed datasets were compared with those only including infants without missing data and no differences were found. We derived the final models by fitting a logistic regression model for all significant predictors. Estimates were combined using Rubin rules .
Outcome data was missing for 238 infants. Summaries of risk factor data showed these cases were not different to cases with available ultrasound information. Furthermore, we ran a sensitivity analysis assuming all these 238 cases did not have DDH and found that results of univariate analyses were similar to those for 1,953 cases; hence we ran multivariable models for cases with available outcome data only.
We assessed the performance of the model in terms of the C statistic (a value of 1.0 is perfect) and calibration. The C statistic represents the probability that for any randomly selected pair of newborns with and without DDH, the newborn who had DDH had a higher predicted risk. Calibration refers to how similar the model-estimated likelihood of DDH was to the observed likelihood of DDH (Hosmer–Lemeshow goodness-of-fit test – a well-fitting model will result in p >0.05).
We validated the model to correct measures of predictive performance for optimism by bootstrapping 200 samples of the derivation data. We repeated the model development process in each bootstrap sample as outlined above in order to produce a model, we applied the model to the same bootstrap sample to quantify apparent performance, and we applied the model to the original dataset to test model performance (C statistic and calibration) and optimism (difference in test performance and apparent performance). We then estimated the overall optimism across all models.
Sample size — in planning the study, we anticipated to evaluate 6-7 independent covariates with a high-enough (>2%) prevalence in our sample. Our study revealed 77 cases of DDH, which allows for 8 covariates to be examined in multivariable analysis .
The local research ethics board approved this study (research ethics committee reference: 14/LO/0420).