Where did I leave that data?
8 Jul, 09 | by Bob Phillips
Data can be lost or go missing for lots of different reasons, and it’s quite important to know why as it might make you fundamentally muck-up the results of your study of you deal with it badly.
The most obvious reason for data to get lost is by bad luck, for example a freak accident like power failure in the lab meaning a blood test can’t be analysed. In this setting, the data are missing for no reason but random chance, and are described as “missing completely at random” (MCAR). more…
The separation of ‘risk’ factors and ‘prognostic’ factors at first seems the sort of obsessive fine detail that gives epidemiologists and statisticians a bad name. Sadly, the difference is actually worth understanding for any clinician that’s going to try to cut through an observational study and understand what it might be truthfully telling us. (This isn’t the true of the difference between a Peto odds ratio meta-analysis and a DerSimion & Laird random effects meta-analysis. That is a pointlessly academic difference.) Fortunately, the difference between risk and prognostic factors is straight forward. ‘Risk’ factors are those which as associated with causing a condition (like smoking for lung cancer, being premature for chronic lung disease, or soft light and wine for falling in love). 
Sometimes we are in situations where we think that something causes problems, and we can’t do a trial randomising one group to get something which we think causes problems! How do we then go about finding out - how to we avoid the problems of ‘confounding’ - and what is that anyway? For example, think about necrotising enterocolitis. Which babies develop NEC? 
If you were offered a choice of medication to treat an ailment you were suffering from, and you’d asked about how effective they were (and there’s a huge chunk of the population that wouldn’t, and would be happy to just do as they are told), then what information would you like? 
A blog post of questions that are calling out to be answered.
In examining a diagnostic test, we make the assumption that the characteristics of the test - its sensitivity and specificity (or likelihood ratios, the way I prefer to think) - will stay constant across different populations, although the positive and negative predictive values will change * . This is sort of true, and sort of false.
It’s a great sport of journalists and commentators to look back at predictions of the future from decades past, and see just how badly they have gone astray. We do this as clinicians too, but with a sense of guilt … looking back to an unexpected relapse of a low-risk tumour, or a fulminant hepatitis that presented with mild nausea, and ask ‘Why didn’t we predict that?”. 
The days of a meta-analysis being the simple adding up of lots of studies, pretending that they are all just tiny pieces of the One Big Trial that was performed before the world was made are on their way out. Newer ways of using synthesised evidence - like meta-regression and individual patient data analysis - are coming up quickly.