At “publication time” for this blog, I experienced the sinking realization that I had failed to achieve the purpose I set out to accomplish, leaving me with three choices:
- To postpone our vacation visit to our son,
- To publish and hope no one reads it, or
- To just describe the purpose and warn against reading the rest of the blog (which I chose).
The purpose of this blog was to assure the reader that improving employees’ health and wellbeing is Job One at Switch Healthcare and we are better positioned to accomplish that than anyone else in the health industry.
But as far as the following: “Reader Discretion Advised” – the blog turned into a treatise on statistics.
Switch Healthcare is built on the premise that:
- Better decisions can dramatically increase the value of healthcare,
- Decision-making can be improved, and
- Good decisions are dependent on the effective analysis of reliable, effectively peer-reviewed studies.
Several Conversations have already reviewed elements of good decision-making.
This Conversation focuses on the question, “How can we improve the accuracy of our interpretation of research data?” In this quest we will explore three concepts:
- Correlation is not causation,
- Randomized controlled trials are not always the gold standard, and
- Epidemiological data can add significant value to decisions.
Correlation is not causation.
This may be the most popular maxim in statistics, reminding us to be cautious when interpreting research that demonstrates the correlation of factors.
An example of a mistaken correlation with significant ramifications occurred with the use of hormone replacement therapy (HRT) for menopausal symptoms. From the early 1980’s and into the 1990’s, a frequently touted “added benefit” of HRT was that it also reduced the likelihood of heart disease.
However, as more definitive studies were done, the correlation turned out to be misleading. It was the patient’s financial status, not HRT, which caused the apparent relationship between HRT and decreased heart disease. When adjusted for the wealth of subjects in the study, HRT was found to correlate with an increase in the likelihood of heart disease in postmenopausal women. (For a clinically balanced discussion, read the Mayo Clinic web site HRT advice.)1
The flip side of the correlation story can also be a problem. A devastating example was the decades-long tobacco industry contention that correlation does not prove causation vis-à-vis smoking and lung cancer.
Although there had been correlation studies on smoking and lung cancer published in the 1920’s and 1930’s, some commentators say2 that not until the British Doctor’s Study3 was published in 1956 was the causation between smoking and lung cancer adequately established, although even after that study, tobacco lobbyists still disagreed.
The key point is that correlation studies alone are not adequate to make valid conclusions about cause and effect. However, weighing the evidence of correlation studies with other confirming data extends one’s ability to make prudent health decisions.
Randomized controlled trials are not always the gold standard.
A randomized control trial (RCT) is a methodology designed to reduce bias in determining the cause and effect relationship of an intervention. The RCT has become the most influential study for FDA approval decisions because of the effectiveness of an RCT design to reduce bias. Does this make a RCT the gold standard in making decisions? Sometimes it does; sometimes it doesn’t.
There can be major challenges in taking the findings from an RCT and applying them to a particular patient’s situation. I will describe three:
- The studied population does not represent a particular patient’s demographics. Effective RCT designs seek to minimize confounding factors as much as possible to increase the certainty that the results are caused by the intervention. For this reason, people over 65, minorities, and those with multiple medical conditions are often excluded from clinical trials. Example: Our knowledge of statins in heart disease comes primarily from randomized trials that only had a small percentage of women participants, and yet as is well documented, there are key differences in the details of heart disease between the sexes. Extrapolating statin data from men to women without a history of heart disease may have resulted in millions of women being treated with statins with no benefit and potentially significant side effects.4
- Large studies can make small differences “statistically significant”, but in the context of a particular patient, these findings might not be that helpful. Example: The median life expectancy of Stage 4 non-small cell lung cancer is still only eight months despite aggressive therapy. The differences in lifespan are often related more to the characteristics of the particular cancer than to the treatment administered. The high variability makes treatment decisions quite challenging for patients (and their physicians). Statistically significant differences are often measured in terms of weeks, not months of increased survival.
- The study is not designed to cover related significant health issues. Example: Multiple research studies have shown that the Aromatase Inhibitors (AI), a class of drugs that block estrogen production, are beneficial in treating post menopausal women with breast cancer. In the late 1990’s, AI began to replace Tamoxifen® as the drug of choice. A 2008 meta-analysis5 (a study that combines the “best” available studies) demonstrated that the adverse effect profiles strongly favored the AI drugs. It was not until AI’s were prescribed for over twenty years that those taking Tamoxifen were found to have a significantly lower rate of coronary artery disease.6
I had a very wise mentor during my medicine residency who stressed that if a new drug was clearly better than an old drug, use it. However, even if the new drug was a little bit better, consider the old drug first because unknown side effects will be discovered down the road.7
If these limitations are not taken into account, RCT results can lead to overconfidence in the general applicability of the RCT.
The conclusion on randomized controlled trials is that they are the “gold standard” in overcoming biases, and I believe these expensive trials should be strongly supported. However, considering a randomized trial as the “gold standard” without examining its specific limitations often leads to suboptimal decisions.
Epidemiological data can add significant value to decisions.
An example is the challenges in understanding the interaction between the rise in US obesity rates since 1990 and the amount of persistent organic pollutants (most often pesticides and industrial chemicals – POP) exposure in individuals.
There is a clear and well-known association between an increase in weight and an increase in Type 2 diabetes.
There is less well-known but also important findings that show that increasing POP (persistent organic pollutants) exposure may also be key in the development of diabetes.9 Because of the complexity of associations, epidemiology data alone cannot sort out such complex interactions.
- Another major limitation of epidemiology is in evaluating the effectiveness of programs to improve population health. Again, the number of potentially conflicting events reduces validity.
- Perhaps the most important limitation of epidemiological data is that it is routinely ignored in the healthcare setting. The medical field has become so focused on prescribing medications and doing procedures that the epidemiological data that is a driving force in population management is largely ignored.
Switch Healthcare is designed to better assess data validity, weighing the risks and benefits from multiple sources to enhance the effectiveness of health decisions.
For those of you who did not follow the initial warning to the reader but still slogged through this tome, I apologize for the length, difficulty, and the medical lingo of this blog.
However, rest assured that our approach to these challenges will improve health, lower healthcare costs, and be much appreciated by your employees and their beneficiaries.
Breakthrough To Better,
1Hormone Replacement Therapy poses potential heart risk
2History of the prevalence of lung cancer surgeries
3British Doctors Study
4Statins affect women and men differently
5Meta-analysis favors aromatase inhibitors
6Aromatase inhibitors found less effective
7Prevalence of unpublished adverse events
8Limitations of epidemiology
9Persistant organic pollutants impact linked to diabetes
Switch Conversations is a bi-weekly blog exclusively for our potential employer partners.
Edition 1 – Solving a Well-Entrenched Problem
Edition 2 – A Case of Dr. Jekyll and Mr. Hyde
Edition 3 – Best marketing tagline of all time?
Edition 4 – Post-Truth Killed a President
Edition 5 – What’s an employer to do?
Edition 6 – Profiting From the Opioid Epidemic
Edition 7 – The Keys to Unlocking Better Decisions
Edition 8 – When Difficult Things Need to be Done Well
Edition 9 – Fixing Healthcare
Edition 10 – Beware of a Singing Cow
Edition 11 – Wise Reflections
Edition 12 – Warning: Reader Discretion Advised
Edition 13 – Can AI save healthcare? (Part 1)
Edition 14 – Can AI save healthcare? (Part 2)
Edition 15 – Can AI save healthcare? (Part 3)
Edition 16 – Embracing Reality to Improve Healthcare
Edition 17 – Everything I Needed To Know…
Edition 18 – The Eighth Circle of Hell
Edition 19 – So… What’s Our Solution?
Edition 20 – Protecting Integrity as a Core Strategy
Edition 21 – An Unadorned Legacy
Edition 22 – Time to Grow Up
Edition 23 – Against All Odds
Edition 24 – When Everyone Has Stopped Listening
Edition 25 – Focusing on What’s Important
Edition 26 – Don’t Give Up Your Shot
Edition 27 – Join the Goodhood
Edition 28 – Fixing Healthcare (Recycled)
Edition 29 – Taming the Healthcare Beast
Edition 30 – Leadership
Edition 31 – Better Health Requires Good Sense
Edition 32 – Little Decisions With Big Consequences
Edition 33 – Transformational Courage
Edition 34 – Transformational Courage – Part 2
Guest Post – Happy Thanksgiving! By Jeff Novick, RD
Edition 35 – Transformational Courage – Part 3