Title, Date of Publication & Journal‘

Screening Prevalence of Fetal Alcohol Spectrum Disorders in a Region of the United Kingdom – a population-based birth-cohort study.  Preventative Medicine; McQuire C et al, November 2018.

Study’s main claims – and are they supported by the data

The paper does not support the claim that the UK has higher than average prevalence of FASD compared to other countries.

The wording in the paper and press statement does reflect the uncertainty, where the words “could have” are used to outline the uncertainty in the conclusions based on the high amounts of missing data.

In her quote in the press release, Dr McQuire uses the word “significant” in terms of children with symptoms and the health risk and this appears to be speculation, as there is no quantification of these in the paper and the data were from a period before the latest changes in guidance.

Strengths/Limitations

Strengths:

Recognition of the impact of missing data on the outcome measure.

Use of the MICE technique; although note the assumptions of missingness (see below)

Limitations:

Use of Missing at Random (MAR) to impute missing values – this assumes that the presence/absence of data are not related to the values of other data that are not missing.  This assumption appears likely to not hold true which is acknowledged by the authors. This is because,  based on the results of the complete case small sample and baseline characteristics summary in Appendix 5, it shows differences in educational profile, maternal age, and paternal social class when comparing the complete case subgroup (no missing data) with the multiple imputed data (the one that gives 17%). The results are dependent on the adjustment applied when imputing the data and whilst better than not attempting any adjustment, it has not been validated against other sources of data, therefore while some attempt has been made to partially account for this, these factors (like educational profile, maternal age, paternal social class etc.) may still influence the results.

It is unclear as to whether the uncertainty involved in the imputation of missing data has been accounted for in the overall uncertainty of the results. The greater the percentage of the dataset where the values are imputed from missing data, the more uncertainty there is around the resulting analysis. I would have expected more uncertainty for multiple imputation in both lower and upper bound; if I am correct then the lower bound may well have been more similar to the upper bound of the complete case and the upper bound may well be much lower than 17% i.e. nearer 11%

Glossary

Multiple Imputation using Chained Equations (MICE) is a complex but recognised technique for imputing values where data are missing for multiple parameters.

Data are missing at random (MAR) if any systematic differences between the missing and observed values can be fully explained by differences in the observed data. For example, if missing blood pressure measurements are lower than the observed measures but only because younger people are less likely to have their blood pressure monitored then data are said to be MAR given age.

Any specific expertise relevant to studied paper (beyond statistical)?

None.