MPH Master pieces

Sudden infant death

I haven't seen much written about sudden infant death recently but here is a good paper from Bristol. A nice simple case-control design and a clear message.

There was a huge drop in postneonatal death from about 1985 - see this graph. Further reductions, but less dramatic, followed the 1991 'Back to sleep' campaign. Case control studies had shown a problem with babies being placed on their tummies. Cases of sudden infant death were far more likely than controls to have been laid to sleep prone. Hence the 'Back to sleep' campaign. 

You may wonder why anyone would put a baby to sleep on its tummy. The answer is that in the 1960s had shown that very premature babies do badly when placed on their backs. This finding, from intensive care units, was unwsiely - indeed fatally - generalised to all babies as general advice to parents. I guess we'd need a historian to document fully quite how this happened. But for now the moral is clear: when you interpret research, be very careful about the situation to which the research does, and does not, apply.

Autism

Autism in on the increase. Or is it? The diagnostic criteria are pretty vague, and they keep changing. And to ignite this explosive mix, a diagnosis of autism now secures a hefty range of services and benefits.

This paper from California is quite tricky to follow. The basic idea is simple, which is to count how many individuals had a diagnosis of autism added to an existing diagnosis of mental retardation. This is called 'accretion' of an extra diagnosis. Autism and mental retardation (MR) are related because two key features of autism - poor social skills and little speech - would of course also be features of learning disability. Hence in young children they are difficult to tell apart. In California, most people receiving services for autism or MR are tested each year; so the database used has individual information plus individual test scores for the period 1992 to 2005.

The complication arises because in earlier years an IQ test would be done and MR diagnosed, with the extra diagnosis of autism following later; in later years children were more likely given a diagnosis of autism and then not tested for IQ. So a statistical model is needed to estimate how much accretion would have happened if people weren't in later years jumping straight to a diagnosis of autism without any IQ testing.

Another paper, from Australia , is simpler - no modelling involved! The key graph shows prevalence each year from 1983 to 1999 and marks on the graph which years saw the major changes in diagnostic criteria. Note that this paper focusses on autism spectrum disorder - a wider category than the Californian study.

Elsewhere in the same issue of IJE is a paper from Nottingham looking at trends in schizophrenia and other psychotic disorders, based on evidence from three big studies which happen to have been done there around 1978, 1993 and 1997. The incidence of schizophrenia seems to have gone down but other psychoses - especially drug related - have gone up. Overall the proportion of of people using illicit drugs hasn't changed much in this time (up from 22% to 33% in young men from 1994 to 1998) so the authors speculate that illict drugs are stronger than they used to be, or used more heavily.

Surgical mortality

We have known for over 50 years that surgery is safer in some hospitals than others. But we are still plumbing the reasons. Here is a worthy addition to the literature.

A very large database was used, covering 84 730 patients with a wide range of clinical information available to allow for risk adjustment. As expected there was variation in death rates after surgery, from 3.5% in the best quintile to 6.9% in the worst. But interestingly the complication rate was about the same across the quintiles. So the difference is that in the good hospitals, when you get a complication they deal with it. The largest differences were for stroke, deep wound infection and septic shock.

Flu masks

Ontario got hit by SARS so you can assume people there are interested in face masks.

This trial compared the simple surgical masks with the fitted N95. In theory the surgical mask might let a cough or sneeze in through the side, whereas the fitted mask should not. In practice however (always do the trial) there was no difference.

The trial used laboratory confirmed influenza as the outcome. Nurses were randomised to use either the surgical mask or the N95 when caring for patients with acute respiratory symptoms. In both groups about 23% of the nurses caught flu. As the authors say this implies that droplet transmission is not a major route of infection. (It's also a frightening high infection rate once you start to think about a pandemic.)

The trial is called a 'non-inferiority' trial. Don't worry too much about this - it's still a randomised trial with an outcome in two group. The 'non-inferiority' bit affects how the study power is calculated since what you are interested in is the lower confidence interval: we need to be confident that the intervention is no worse (or 'not more than 9% worse') than the control group.

Halfway through the trial H1N1 flu reached Ontario and the health ministry ordered everyone onto the N95. Sometimes politics trumps evidence.

Avoiding CT scans

The temptation to order a CT scan is great indeed when faced with a child, a history of head injury, and frightened parents (with or without their lawyers in attendance). But CT scan carries a heavy radiation dose, and radiation induces cancer. For every 1000 to 5000 paediatric CT of the brain, you get one extra brain cancer, and every year about 150 000 children in the USA have a CT for head injury. All of this info comes from a paper which describes the development and validation of a rule intended to allow a doctors safely and accurately to say that a head injury does not warrant a CT.

This is a model of how to develop and validate a classification: hige database, separate derivation and validation samples etc.

The rule had a negative predictive value of 100% and sensitivity of 100% in children under 2 years old (1176/1176 with no clinically important injury and 25/25 with an injury correctly classified). The equivalent figures for a different rules in patients over two were 3798/3800 and 61/63. Of course even in under 2s the performance is not necessarily perfect, with lower confidence intervals on the 100%s being 99.7% and 86% respectively. So there is room for absolutists to argue that you still need a CT to be absolutely certain.

Hopefully this definitive paper will stop lawyers suing when doctors don't take a CT. (But knowing how lawyers behave in the US, we probably need a lawyer to sue a doctor who did order a CT when the child goes on to develop a brain cancer.)

Veggies

I try to stick to papers with original, primary data because summaries / editorials / even meta-analyses are so often, so very often, misleading. But here is a readable editorial by the great Jim Mann about being a vegetarian.

Mann makes the interesting comment that 

'Although vegetarians, notably young women, do have lower iron stores than those who eat meat, there is little evidence that meat avoidance itself is associated with malnutrition.'

Let's unpack a couple of phrases in this - 'lower iron stores' and 'malnutrition'. Firstly, severe iron deficiency leads to anaemia, and one of the features of severe anaemia is extreme lassitude - you feel tired all the time. But note the two 'severe's in there. It is false logic to argue that if severe anaemia causes tiredness, then mild anaemia causes mild tiredness. This false logic leads to huge numbers of women being given iron supplements. They arrive at the surgery saying they feel tired all the time, a blood test shows mild iron-deficiency anaemia, a prescription follows. But lots of people (9% of a general population sample) have unexplained fatigue, so mild anaemia plus tiredness is at least as likely to be coincidence as cause and effect. 

One of the hallmarks of good public health practice is a sense of proportion - whenever someone says that x% of people with a disease (or worse, some biochemical result) have tiredness / headache / depression / wheeze, you should immediately ask yourself - 'What proportion of the general population have this symptom?'

Indeed this raises the whole question of what we mean by 'anaemia'. If defined as 'outside the normal range' for a blood test, what is 'normal'? Values found in 95% of the population - in which case we have by definition declared 5% as ill, or at least as 'not normal'. Or is anaemia a value 'low enough to be associated with ill health'? Which takes us back to where we started.

--------------------------------------------------------------------------------------------------------------------------

While we're on diet, here's a good story from Edwin Moore ("Lemmings don't leap' - 180 myths, misconceptions and urban legends exploded").

"in 1870 a Dr von Wolf put a decimal point in the wrong place and it became a 'scientific fact' (until the error was spotted in 1937) that spinach contained ten times more iron than it actually does."

Unfortunately Moore gives no citations so this story may itself be an urban myth.

Dangerous drugs

 Here is a good example of using routine statistics.

Co-proxamol is a pill containing paracetamol and an opiate. It is both dangerous and unnecessary: dangerous because the opiate stops your breathing in the mildest of overdoses; unnecessary because there are plenty of other mild analgesics.

Between 2005 and 2007 co-proxamol was withdrawn from use in the UK. This paper documents the results. Crucially, patients did not turn to other drugs to commit suicide; there was no increase in death from drug overdose. (This paper can't tell us whether depressed people turned to other methods such as hanging or jumping, but we know that the national trend for suicide is downwards.)

As with a number of papers I've blogged recently, the statistical methods are tricky but the graph speaks clearly.

Vaccine safety

Every now and then there's a vaccine safety scare, and the results are disastrous for public health. Pertussis and 'brain damage'; MMR and autism; hepatitis B vaccine and multiple sclerosis. Did the last one surprise you? Vaccine scares are surprisingly country specific. Only in the UK did we worry about pertussis and brain damage, or about MMR; only in France were there popular concerns about hep B and multiple sclerosis (see this paper for some background and follow up).

This paper uses the nation-wide system of vaccine adverse event reporting to look at HPV vaccine in the USA.

Two things strike me. Firstly, it gives some perspective on other vaccine scares. When you do anything to millions of people, someone deaths are going to happen in the subsequent few weeks. In this case, 32 girls died after HPV vaccination. There hasn't (as far as I know) been an outcry, and rightly so: but what must the parents have thought, especially for the the small number labelled as 'unknown cause'? Secondly, the reporting system is designed to be very inclusive, so any report is accepted, even one found during this research to be at fifth hand.

The VAERS was also used to document the association between the flu vaccine used in 1976 and Guillain Barre syndrome.

Counting intravenous drug users

Here's a classic of its type. Counting the number of Iv drug users in an area is, for some obvious reasons, difficult.

Capture-recapture methods come from ecology. If you want to count butterflies in a wood, capture ten and mark them with a spot. Come back a week later and capture another ten. How many of this recapture sample are already marked? Two - i.e. 20%? Then your first effort must have captured 20% of the entire population; there are 50 butterflies in the wood altogether.

The Bristol team use a variant of this. Rather than capture- recapture at different points in time, they look for overlap (patients already 'marked,' as it were) in different data sources. For this to work properly the sources must be independent - getting arrested should make no difference to whether or nor you are attending a treatment agency. Then they add a sophistication I don't fully understand - the 'covariate' method.

All in all they found lower death rates than expected (though still appallingly high); a lot of hepatitis B and a lot of undiganosed hepatitis C. The result of all this?

"As a result Bristol PCT increased investment in needle and syringe provision, HCV diagnostic testing, and renewed its efforts to reduce injecting risk and BBV transmission locally. Without the local information showing the scale of the risk the commitment and a sense of urgency by providers and policymakers may not have emerged as quickly. In contrast, the public health message from the mortality study was that the high coverage of O[piate] S[ubstitution] T[herapy] was improving the health of the population, and led to a local commitment to maintain current levels of provision and increase the uptake of OST."

Confirms the opinion expressed 20 years ago in an obscure (but brilliant!) doctoral thesis that the purpose of local epidemiology is not to generate new knowledge but to prompt action; and hence the requirement is for info that is local and immediate. It doesn't have to be accurate to the last decimal point. (But it does have to be rigorous enough to be honest.)

1000 families

The International Journal of Epidemiology has been running an interesting series summarising various cohort studies around the world. This month it's the turn of one of the best: the Newcastle 1000 families study. Comment is superfluous - read and enjoy!