There was clear heterogeneity in non-chromosomal congenital anomaly rates between register regions as shown in more detail previously . Only a small part of this was explained by differences in maternal age and deprivation. Some of the remaining regional variation may reflect real differences in other risk factors, for example nutritional factors for spina bifida, but some variation is likely to be due to ascertainment differences . For chromosomal anomalies as a whole, we found some evidence of regional variation, but not for Down Syndrome specifically after maternal age had been taken into account. Regional and hospital variation in other chromosomal anomalies may be due to differences in prenatal screening methods, which can pick up anomalies that would not otherwise be evident until later life. However, there was no regional variation in the well diagnosed Edwards and Patau syndrome specifically.
This paper shows clear evidence for heterogeneity between hospital catchment areas also for non-chromosomal anomalies in general (90% range of underlying relative risks 0.84–1.17) and some subtypes (hydrocephaly, congenital heart disease, cleft lip). These may well have been caused at least in part by differences in ascertainment. If this is true, it would suggest that adjustment for hospital catchment area may be necessary to control for ascertainment bias in studies of congenital anomalies in relation to putative risk factors. It is likely that some ascertainment differences between hospitals would be present in most registers. This was shown also in a geographical analysis in New York State That study however looked at individual birth hospital of delivery, which could also reflect selection of high risk pregnancies by some hospitals, whereas our method of comparing catchment areas distinguished ascertainment (and other area) differences from such selection effects. The types of differences between hospitals reporting to the same register can include hospital practices with regard to prenatal and neonatal screening, the thoroughness with which newborns are examined for more minor anomalies and how these are recorded, referral practices, linkage with specialist services such as paediatric cardiology, how efficiently and accurately congenital anomaly diagnoses can be retrieved from hospital databases, and the quality of collaboration between the hospital and the register.
Below hospital catchment area level, there was very little evidence for heterogeneity or clustering. The finding that syndromes and two non-chromosomal subtypes (neural tube defects and specifically spina bifida) had nominally significant heterogeneity across EDs is compatible with chance findings from multiple testing. Further surveillance should establish whether this finding holds. The borderline significant gastroschisis finding is of interest given the increasing trends in prevalence over the last few decades reported elsewhere and rather strong regional differences in prevalence [17–22].
Our results should be interpreted in the light of the limitations in power to detect clusters. In particular for rarer congenital anomalies, geographical heterogeneity (clustering) would have to have been substantial for our tests to be able to detect it. For example, we estimated a 5th–95th percentile range of underlying limb reduction (n = 348) rates across EDs, relative to the overall rate, of 0.19–2.32 (a 12-fold range), but the significance level was only p = 0.17. Nevertheless, there need not be many EDs with more than one case to cause large and statistically significant heterogeneity. For example, exploratory analyses showed that for heterogeneity to be very strongly significant (p = 0.002) for a rare anomaly (121 cases), it was sufficient to have just two EDs with two anomalies and one with three anomalies. Indeed, our strongest finding of variation related to syndromes, where some EDs had more than one case from the same family. Much smaller variation is detectable between registers, and somewhat so between hospital catchment areas. For example the 5th–95th RR range estimated between registers for limb reduction was just 0.76–1.27, but this approached significance at p = 0.097.
The scan statistic complemented the other tests of geographic variation by picking up clustering at geographical levels other than ED and ward, and by not respecting ward boundaries in examining areas larger than ED size. The two nominally significant "clusters" found by the scan statistic should be interpreted cautiously in view of the large number of congenital anomaly types tested, and the dominance of the more unusual cluster by one family. That three of the four cases are to the same mother does not preclude a common environmental explanation (in particular an exposure specific to this home), but it could also be familial susceptibility to the environmental factor, or alternatively familial inheritance unrelated to any local environmental exposure. Overall therefore we found a remarkable sensitivity of all clustering methods to family clusters, which are particularly difficult to interpret in relation to possible environmental factors. We performed careful cleaning of the data for duplicate registrations, but it can be easily seen from our consideration of familial cases that duplicate registrations would cause detectable clustering, an important warning for further research.
The absence of evidence for general heterogeneity below the hospital catchment area cannot be taken as evidence against a modest degree of more specific heterogeneity, for example between areas of different deprivation level , or specific clustering close to sources of environmental exposures. Data from these registers contributed towards the finding of clustering around hazardous waste land-fill sites , and reanalysis with these data have confirmed this for non-chromosomal anomalies although with reduced odds ratios . Further, the absence of evidence for generalised local clustering reassures that the excess found close to waste landfill sites is indeed as unusual as conventional statistical methods suggest.
We used statistical methods of intermediate complexity. Could more sophisticated spatial analysis reveal more of interest? It had been our intention to explore spatial patterns further using Bayesian and other "multi-level" hierarchical models, including spatial adjacency models (investigating whether adjacent areas have similar rates). However, exploratory analyses using even the simpler of such models to investigate spatial patterning in variation below hospital catchment area (random effects between wards or EDs) suggested that estimates of heterogeneity were often sensitive to specific model specification, in particular "prior" distribution for variance. This seems likely to be a further consequence of the sparseness of occurrence of anomalies in EDs and wards – it is too much to expect these data to give information on subtle spatial patterns in rates at the local level. The problem seems likely to be less for analyses of all anomalies combined than for specific anomaly groups, and less for variation between wards than variation between EDs, but we sought an approach applicable routinely in all analyses. We did not pursue use of these models at higher geographic levels (investigate whether adjacent hospital catchments have similar rates), because the likelihood of variation between hospital catchments being artefacts of case ascertainment made interpretation less interesting.
We sought here to describe geographical variation. Another use of models incorporating geographical variation is to allow for it when addressing specific risk factors in "ecologic correlation studies". In general ignoring such variation, if present, exaggerates certainty in such analyses, and in some situations can bias the estimated effects. For this use the negative binomial model – which can only include random variation at one level of the hierarchy in any one analysis and does not incorporate other modes of spatial autocorrelation (e.g. similarity in adjacent areas) – is more limited. Models which do incorporate spatial structure more fully reduce the risk of underestimating uncertainty, though where data is sparse conclusions may depend strongly on prior information rather the empirical observation of spatial structure.
Our results give encouragement that well-cleaned register data, limited to a predefined set of well diagnosed and reported congenital anomalies, can be adequate to investigate putative specific environmental causes of anomalies for which exposure varies geographically below the hospital catchment area level. Major data errors that would bedevil such studies would have been likely to have shown up in this study as localised clustering. Such studies would generally need estimates of exposure to fine geographical resolution and should, where possible, control for "hospital catchment" effects.