Paul Madley-Dowd

1/ NEW R PACKAGE! For estimating the impact of potential interventions on multiple mediators in countering exposure effects (led by @cttc101.bsky.social)

- Paper👉 tinyurl.com/ye26jsps
- Package👉 tinyurl.com/yuh4kens

Thread shows published examples of how the method can be used! #EpiSky #CausalSky

🚨 Funded PhD opportunity 🚨 Work with large-scale electronic health record data from #OpenSAFELY to optimise vaccine effectiveness estimation for respiratory viruses.

Apply here 👉 www.findaphd.com/phds/project...

Moving beyond risk ratios in sibling analysis: estimating clinically useful measures from family-based analysis Objective: Findings from family-based analyses, such as sibling comparisons, are often reported using only odds ratios or hazard ratios. We demonstrate how this can be improved upon by applying the ma...

For those of you using sibling analysis, perhaps you'll find this useful 🥳: www.medrxiv.org/content/10.1...

Just adding "The meta-analyst decides that the accumulated evidence is in fact a pileup." as an additional favourite

Congratulations Viktor!!

Safer medications for pregnant women and children Viktor H. Ahlqvist, postdoc at the Institute of Environmental Medicine (IMM), receives 3,000,000 SEK in a Postdoctoral Grant from the Swedish Society for Medical Research (SSMF) for the project “Advan...

Fantastic news! This is entirely attributable to all my colleagues—everyone from mentors to students—who have made this possible🎉🎉

news.ki.se/safer-medica...

It’s my understanding that with the parametric g-formula you use the outcome model to predict the outcome for each subject, independently of whether they are censored. And you take its mean considering ALL N subjects. If the pot. outcome has some NA, I’d still sum and divide by N.
Stupid question:

I'm even less convinced by these certificates now

bsky.app/profile/bkle...

New publication led by @proflouisemarston.bsky.social using multiple imputation to target a hypothetical estimand in a pandemic restriction-free world for a trial in
schizophrenia - demonstrating the potential impact of the pandemic on the trial results

There's already been a very interesting preprint commentary on our paper by Maya Mathur and Ilya Shpitser which I highly recommend people take a look at :

osf.io/preprints/os...

Conclusions:
- Use auxiliary variables that are completely observed, or have smaller amounts of missing data
- Explore the missing data mechanisms of incomplete auxiliary variables
- Aim to use auxiliary variables that are independent of their own missingness.

"Bias was larger when the auxiliary had a stronger correlation with the outcome...In terms of absolute bias in the MI estimate, this equates to around ... 17% of the true effect size." We would tend to treat such an auxiliary as preferable, but we need to show caution when it has missing data in it

Subsection from Figure 2 of the paper. The image shows a plot with relative bias on the Y axis, and the proportion of missing data in the auxiliary variable on the Z axis. Four coloured lines are on the plot representing different correlations between the outcome and the auxiliary variable. This plot, plot H, displays results for an example where the missingness mechanism for the outcome leads to an unbiased estimate of an exposure outcome

The most striking finding to me was that when there was no bias in CRA (and we are using MI to reduce SEs only), including an auxiliary variable with an open path to its own missing data can introduce substantial quantities of bias.

We looked at different missing data mechanisms for both an outcome and an auxiliary variable.

Where the outcome missingness mechanism led to a biased complete records analysis, increasing proportions of missing data reduced the ability of auxiliary variables to remove bias.

But what happens when those auxiliary variables have missing data in them?

We didn't know what consequence including incomplete auxiliary variables has on bias of exposure-outcome estimates made using regression models - so we did some simulating.

Analyses using multiple imputation need to consider missing data in auxiliary variables Abstract. Auxiliary variables are used in multiple imputation (MI) to reduce bias and increase efficiency. These variables may often themselves be incomple

Final version published so time to talk about it: doi.org/10.1093/aje/...

When using multiple imputation to account for missing data we often use auxiliary variables (variables included in the imputation model but not the analysis model) to 1) reduce bias and 2) improve statistical efficiency.

Work by @ahlqvistviktor.bsky.social @draipsych.bsky.social @karolinskainst.bsky.social @aarhusuni.bsky.social
@neilmdavies.bsky.social @danielberglind.bsky.social ... and many more... @natureportfolio.bsky.social

The phrase 200% higher looks a lot more alarming than twice the risk of 0.9%.

This is nothing new in the area of risk communication - but today it has annoyed me

I think there is an issue with people communicating relative risk, they often use language relating to risk difference (i.e. X% higher risk). They’re not wrong but I feel that people should be using phrases like twice the risk instead of 200% higher risk.

- At 12 years of age, the children of unexposed mothers had a ... 0.9% risk of intellectual disability ...
- With intellectual disability ... polytherapy was associated with a risk of 1.8%.
- Taken together, the risk of intellectual disability was ... 200% higher with polytherapy

Antiseizure medications in pregnancy tied to child neurodevelopment risks Researchers investigate how anti-seizure medication use during pregnancy may increase the risk of neuropsychological conditions in children.

One thing that has frustrated me with the reporting on this is the communication of risk.

In an overall well written article (www.news-medical.net/news/2024112...) about our work we have the following (see next):

In our latest work we investigated the effect of antiseizure medication prescribing/dispensation in pregnancy on offspring neurodevelopmental outcomes using over 3 million pregnancies from the UK and Sweden. Article out now in Nature Comms: nature.com/articles/s41...

We will be covering multiple imputation methods to address
1) where data are missing not at random
2) imputation for multilevel models
3) imputation for survival models, and
4) imputation for propensity score analysis

We are running our short course on Advanced Multiple Imputation Methods to Deal with Missing data this December (5th and 6th).
link for further info: tinyurl.com/ybu982ru