Warning: This post is a bit technical and very long. It may be more useful as a reference, should you happen upon a medical research study and want to understand some of the terminology better.
There is a vast body of medical research papers and reports available on-line. The abstract is almost always free and the paper itself is frequently free. There are two sources, PubMed (http://www.ncbi.nlm.nih.gov/pubmed) and Google Scholar (http://scholar.google.com/). They both appear to have effectively all the research available on-line, but Google Scholar has one big advantage: If the full paper can be found outside a pay-wall, Google Scholar will show it. On the other hand, PubMed has a more elaborate search engine.
Certain words appear repeatedly. Here is the translation from Medicalese to English:
Cohort – Group of people
Etiology – Cause or reason
Efficacy – Effectiveness
Modality – method
Outcome or Endpoint – what the research is looking for
Control for this or that – keep the comparison apples to apples.
Confounding variables – something that confuses the outcome and must be accounted for.
Most clinical research is trying to determine if some variable, use of a drug, for instance, alters some outcome, like number of heart attacks, or all-cause mortality.
Always bear in mind that there is some reason the research was published. The researcher may be an academic at a university or government research lab, who received research grants and is expected to publish. The paper may be published by a drug company research team, in which case the objective is to show off the wondrous new drug. Little happens without funding of some sort, which puts an inordinate amount of unseen influence in the hands of those directing the funding. The era of the 19th century ‘gentleman scientist’ who had a lab in his basement and researched whatever interested him has long passed.
We are specifically interested in ‘clinical’ research, meaning research done on people. Most report’s research can be sorted into one of six categories:
Cohort studies – this studies the effect of some variable on a group of patients. For example Dr. Nichols could (and does) use his patients as a cohort. These patients are involved in his Quantitative Medicine program and typically get quarterly blood draws. Dr. Nichols could pose this question: Do those that practice periodic fasting have better glucose numbers? He could then gather the glucose numbers from the fasters and non-fasters, wait a year, gather the numbers again and compare. Since he is going forward with the study, it is a prospective cohort study. However, if Dr. Nichols wanted to save time, he could dig out records from the past and sort them into fasters and non-fasters (provided he had gathered that information) and compare the glucose numbers. This would then be called a retrospective cohort study.
Cross sectional study – cross sectional studies are a one shot event typically. A group of people could be asked to fill out a medical history and current eating and drinking habits. This can answer questions like “How much less lung cancer do people have that smoked 10 years and then quit, versus people that have smoked that long and are still smoking?”
Case-control studies – Here the outcomes are gathered, and the histories examined for possible causes. Dr. Nichols could divide his cohort into a high triglyceride group (the “Cases”) and a normal triglyceride group (the “Controls”), and explore what prior lifestyle choices had been made.
Randomized blinded trials – this is the ‘clinical trial’. Two groups are matched in age, health, etc., and half are given a drug and half either a placebo, or another drug. Who gets what is kept secret (blind) from both the patient and the people administering the trial in order to avoid various biases.
Epidemiological Studies – You can see the ‘epidemic’ root in this, an epidemiology is just a large database of medical records. For example, the SEER is a huge government maintained database of people who have been diagnosed with cancer. It also contains other sorts of information about these people, age, other diseases, habits, drugs they took, etc. This database is then said to be an epidemiology database. If you download it and do data analysis on it, perhaps to see if people who smoked got prostate cancer more frequently, you would be doing an epidemiological study.
This is a study of studies. Meta-analyses are typically used to determine long term consequences of a drug use – statins for instance. The meta-analysis would perhaps look at a dozen other studies and attempt to determine some outcome, all-cause mortality, for instance. The total number of people involved in the meta-study is summed from all the studies they looked at and is often in the 100’s of thousands.
Since these are studies of effects on people, and because people differ in complex and unpredictable ways, the results will commonly spread out and will need to be subjected to some statistical treatment. The statistical treatment can be arcane and is remarkably subject to mischief. We will delve into the reported statistics in a future post, but here are a couple of the most common statistical outcomes.
Association: When they report association, it means that two items are observed together. It does not tell us if one cause the other or vice versa or something else entirely. If two actions or outcomes always occur together, the correlation wound be 1. If one never effect the other, it would be 0. If one they ‘sort of’ track, they would be said to associate or correlate. When one number goes up when the other goes down, it is called a negative correlation.
Relative Risk – often abbreviated RR, and often called Odds Ratio or Hazard Ratio. (These are in fact different, but, for now, that is nit picking). For instance, if over a five year period, 10 out of 1000 smokers got lung cancer, and 3 out of 1000 non-smokers got lung cancer, then the Relative Risk of getting lung cancer, smokers to non smokers, would be 10/3, or 3.33. This is an example of a pretty strong result. Often, the relative risk will be something like 1.10, meaning that in one group perhaps 11 out of 1000 got the disease, and in the other same sized group, 10 did. 1.10 isn’t a particularly strong result.
Here are actual abstracts from two research papers illustrating the various types of studies, along with a few annotations [in italics]. These were ‘bombshell’ results, but you would never guess it from the tone.
Here is a Cohort Study example.
The influence of glucose-lowering therapies on cancer risk in type 2 diabetes
J. Currie & C. D. Poole & E. A. M. Gale
Received: 19 May 2009 / Accepted: 18 June 2009 / Published online: 2 July 2009 # Springer-Verlag 2009
Aims/hypothesis The risk of developing a range of solid tumours is increased in type 2 diabetes, and may be influenced by glucose-lowering therapies. We examined the risk of development of solid tumours in relation to treatment with oral agents, human insulin and insulin analogues. <-Objective: They will see if taking extra insulin causes cancer
Methods This was a retrospective cohort study of people treated in UK general practices. Those included in the analysis developed diabetes >40 years of age, and started treatment with oral agents or insulin after 2000. A total of 62,809 patients were divided into four groups according to whether they received monotherapy [the only drug they were taking] with metformin or sulfonylurea, combined therapy (metformin plus sulfonyl- urea), or insulin. Insulin users were grouped according to treatment with insulin glargine, long-acting human insulin, biphasic analogue and human biphasic insulin. The out- come measures were progression to any solid tumour, or cancer of the breast, colon, pancreas or prostate. Confounding factors were accounted for using Cox proportional hazards models.
Results Metformin [the normal drug given to diabetics] monotherapy carried the lowest risk of cancer. In comparison, the adjusted HR [Hazard Ratio] was 1.08 [1.08 means people who took Metformin only had an 8% greater incidence of cancer. This would be considered a slight increase.] (95% CI 0.96–1.21) [CI is Confidence Interval – this means 95% of the results were in this range] for metformin plus sulfonylurea, 1.36 (95% CI 1.19–1.54) for sulfonylurea monotherapy, and 1.42 [this means cancer was 42% more common in the group taking addition insulin – the main result!] (95% CI 1.27–1.60) for insulin-based regimens. Adding metformin to insulin reduced progression to cancer (HR 0.54, 95% CI 0.43–0.66). The risk for those on basal human insulin alone vs insulin glargine alone was 1.24 (95% CI 0.90–1.70). Compared with metformin, insulin therapy increased the risk of colorectal (HR 1.69, 95% CI 1.23–2.33) or pancreatic cancer (HR 4.63, 95% CI 2.64–8.10), but did not influence the risk of breast or prostate cancer. Sulfonylureas were associated with a similar pattern of risk as insulin. Conclusions/interpretation Those on insulin or insulin secretagogues were more likely to develop solid cancers than those on metformin, and combination with metformin abolished most of this excess risk. Metformin use was associated with lower risk of cancer of the colon or pancreas, but did not affect the risk of breast or prostate cancer. Use of insulin analogues was not associated with increased cancer risk as compared with human insulin.
The ‘real’ conclusion here is that taking additional insulin is a very bad idea. Note how understated this is in the abstract.
Here is an example of a Meta-Analysis
Drug Treatment of Hyperlipidemia in Women
Judith M. E. Walsh, MD, MPH Michael Pignone, MD, MPH
JAMA. 2004;291:2243-2252 [Journal of the American Medical Association – one of the most prestigious ones.]
Context Several clinical trials have evaluated the effects of lipid-lowering medications [statins] on coronary heart disease (CHD). Many of the trials have not included enough women to allow sex-specific analyses or have not reported results in women separately.
Objectives To assess and synthesize the evidence regarding drug treatment of hyperlipidemia [high cholesterol] for the prevention of CHD events in women and to conduct a meta- analysis of the effect of drug treatment on mortality. [<- They are going to ‘average’ the results from several studies others have done.]
DataSources Wesearched MEDLINE, the Cochrane Database, and the Database of Abstracts of Reviews of Effectiveness for articles published from 1966 through December 2003. We reviewed reference lists of articles and consulted content experts.
Study Selection and Data Extraction Studies of outpatients that had a treatment duration of at least 1 year, assessed the impact of lipid lowering on clinical out- comes, and reported results by sex were included. Outcomes evaluated were total mortality [research that ignores overall mortality is suspect], CHD mortality, nonfatal myocardial infarction, revascularization, and total CHD events. Summary estimates of the relative risks (RRs) with therapy were calculated using a random-effects model for patients with and without a previous history of cardiovascular disease.
Data Synthesis Thirteen studies were included. Six trials included a total of 11435 women without cardiovascular disease and assessed the effects of lipid-lowering medications. Lipid lowering did not reduce total mortality (RR, 0.95; 95% confidence interval [CI], 0.62-1.46) [Actually there was a 5% reduction, but this isn’t considered ‘statistically significant’, meaning such a small reduction could easily be due to chance.], CHD mortality (RR, 1.07; 95% CI, 0.47-2.40) [deaths from heart attacks is slightly higher for low risk women taking statins – a surprise.], nonfatal myocardial infarction (RR, 0.61; 95% CI, 0.22-1.68), revascularization (RR, 0.87; 95% CI, 0.33-2.31), or CHD events (RR, 0.87; 95% CI, 0.69-1.09). However, some analyses were limited by too few CHD events in the available trials. Eight trials included 8272 women with cardiovascular disease and assessed the effects of lipid-lowering medications. Lipid lowering did not reduce total mortality in women with cardiovascular disease (RR, 1.00; 95% CI, 0.77-1.29) [No benefit (Relative Risk=1.0) on total mortality for women known to have heart disease – a real bombshell] . However, lipid lowering reduced CHD mortality (RR, 0.74; 95% CI, 0.55-1.00), nonfatal myocardial infarction (RR, 0.71; 95% CI, 0.58-0.87), revascularization (RR, 0.70; 95% CI, 0.55-0.89), and total CHD events (RR, 0.80; CI, 0.71-0.91). [No effect on total mortality, but reduced heart deaths. Therefore, statins increased some other sorts of deaths.]
Conclusions For women without cardiovascular disease, lipid lowering does not affect total or CHD mortality. Lipid lowering may reduce CHD events, but current evidence is insufficient to determine this conclusively. For women with known cardiovascular disease, treatment of hyperlipidemia is effective in reducing CHD events, CHD mortality, nonfatal myocardial infarction, and revascularization, but it does not affect total mortality.
The real conclusion: No woman should take statins. Regardless of heart disease risk, statins don’t improve overall mortality for women. For women at low risk, it doesn’t even prevent heart attacks.
Once you have worked your way through a few of these, it will get a lot easier to interpret them. There will be a bit more about the use and abuse of statistics in the next UNDERSTANDING MEDICAL RESEARCH post.
Here is the list of UNDERSTANDING MEDICAL RESEARCH posts (past, present, and future):
PART 4 – Medical Research Problems – Confirmation Bias.
PART 5 – Medical Research Problems – Commercial Subversion
PART 6 – Medical research Problems – Political subversion
PART 7 – The Cart drives the horse – what causes what?
PART 8 – Playing games with the included items, eg. smoking plus red meat
PART 9 – Using Fear, Uncertainty, and Doubt