Funding a new drug: A Devil’s Advocate, back-of-the-envelope approach


The Guardian headline reads: “New heart treatment is biggest breakthrough since statins” and the article goes on to claim that “cancer deaths were also halved”.

Sounds impressive. So how should we decide whether to fund this drug?

The New England Journal of Medicine published the findings of a randomised controlled trial into this new treatment just two days ago.

The study was well constructed, with an impressive sounding 10,061 participants. All participants had previously had a heart attack and had high levels of one marker for inflammation.

The study counted whether participants had another heart attack, stroke, or died during the follow-up period of approximately 4 years.

During that time 16% of people taking the placebo suffered one of the primary outcomes (535/3344) compared with 14% (320/2284) in the best-performing treatment group (the group that took a medium dose).

Basically, if I was a random member of the study population taking the placebo, I would have had a 16% chance of having another heart attack, stroke or dying in that 4 year period. Crudely, this is a risk of about 4% per annum.

If I’d been taking the study drug, I’d have had a 3.5% risk per annum.

So what does a reduction of 0.5% mean? It means (very roughly) that to prevent one additional heart attack, stroke, or death, we need to treat 200 people for one year.

Now, I have no idea what price this drug will be sold for, but new drugs of this kind often command prices of $10,000 per year, or more.

That’s $2 million dollars.

And there was no reduction in ‘all cause mortality’ in the treatment group.

That means that the ‘halving of cancer deaths’ was balanced out by increases in deaths from other causes, such as serious infections.

That’s $2 million dollars without saving a life.

So, what else can we buy for $2 million? What is the opportunity cost of funding this drug?

Well $2 million is a lot of dietician appointments, a lot of personal trainers, a lot of quit smoking programmes, a lot of health insurance, a lot of income protection insurance, a lot of cardiologists.

Basically that’s a lot of prevention and resilience against future problems that could benefit all 200 of those patients, not just the one who would go on to have the extra heart attack.

At Adapt Research we provide objective health research analysis and plain writing. Send us your question here.

How to survive the next big pandemic


The New Zealand Ministry of Health has recently published its 2017 pandemic action plan. Flicking through it I noted that it tends to cite other governmental publications rather than academic sources.

Of course, many of the publications cited may well cite the academic literature themselves, but I decided to take a quick independent look at what has been published since the 2009 H1N1 pandemic.

So this afternoon I searched PubMed for ‘pandemic, public health, virus’ and limited the results to the last five years, and review articles only. This turned up 354 results. I read the titles and selected 45 for abstract review. Please note, this is a quick look, so I have not read any full-texts.

The findings of these recent reviews can be collated under the following seven headings:

Travel Restrictions

  • Travel restrictions delay pandemics if implemented within 6 weeks, but only reduce case numbers by 3%.

Although note a modelling study by Adapt Research Ltd, which suggests good cost-benefit for border closure in island nations.


  • Vaccines are effective but there are cognitive facilitators and barriers to vaccination
  • Some cross-protection occurs between strains, so any vaccination might be better than no vaccination
  • There are attempts around the world to develop a universal vaccine effective against all pandemic influenza
  • The Influenza Risk Assessment Tool (IRAT) can be used to prioritize vaccine development for those strains with most pandemic potential

A problem with vaccines is the clear difficulty in changing knowledge, attitudes and behaviours related to influenza and influenza vaccination, particularly on the scope and scale needed to greatly improve uptake.

Modeling the spread

  • Modelling is potentially useful in real time, but its effectiveness still needs evaluating in real pandemic situations
  • During the 2009 influenza pandemic modeling work struck problems with data availability, dissemination, heterogeneity, and unclear assumptions
  • Modelling might be able to be used to identify higher risk populations on whom to target interventions
  • Modeling can provide a quantitative estimate of the impact of various interventions

The challenges that modelling faced in 2009 were: (i) expectations of modelling were not clearly defined; (ii) appropriate real-time data were not readily available; (iii) modelling results were not generated, shared, or disseminated in time; (iv) decision-makers could not always decipher the structure and assumptions of the models; (v) modelling studies varied in intervention representations and reported results; and (vi) modelling studies did not always present the results or outcomes that are useful to decision-makers.


  • Precautionary behaviours are less frequent than expected or intended given the threat during a pandemic
  • Difference in behavior between populations within countries is marked (this suggests targeting interventions might be done better, and a one-size-fits-all response may not be appropriate)
  • Misconceptions about risk are common and vaccination uptake is low
  • Risk communication needs to be tailored to the perceptions/behavior being seen in real time, monitoring social media might help
  • Hand washing has modest efficacy, and dental hygiene may be useful, but other interventions have not been fully assessed
  • Healthcare workers’ willingness to work in a pandemic is variable (this will need to be accounted for in any planning/workforce assumptions)
  • Effectiveness of school closures was unclear in a Japanese review

It seems that a comprehensive, longitudinal study is needed to clarify the effects of school closure and other public distancing measures during pandemics.


  • Emergency response planners should consider leveraging social media to track population beliefs and behavious in real time, and consider individually tailored engagement and communication
  • There are a number of potential predictors of behavioral compliance with preventive recommendations, these might help focus interventions


  • Cost-effective are: hospital quarantine, vaccination, antiviral stockpile usage
  • Not cost-effective are: school closures, antiviral treatments, social distancing (at $45,000 willingness to pay per QALY)
  • These interventions are potentially more cost-effective the more severe the pandemic
  • However, cost-effectiveness modeling in the local context is needed.


  • Pandemic plans need ethics frameworks that can be used in unique infectious disease pandemic situations, yet most pandemic plans copy and paste ethical approaches from previous influenza plans.

To summarize: How ought we prepare for the inevitable next pandemic?

Before the pandemic hits:

  • Seasonal influenza vaccine
  • Personal protective gear stockpiles
  • Strategic drug stockpiles
  • Risk communication strategy in place
  • Plan for modelling and data needs
  • Regional cooperation plan
  • Plan to research during pandemic (to inform future plans)

During the pandemic

  • Real time PCR for diagnosis (recommended by the CDC)
  • Case surveillance
  • Surge capacity ensured
  • Antiviral drug delivery
  • Risk communication implemented
  • Adherence to strict sanitary and hygienic measures
  • Regional collaboration and cooperation
  • Focus on high-risk groups
  • Data collection and research to inform future response

The above information is consistent with research published in the last five years, and ought to be considered for further research or evaluation, or inclusion in any local, national or international pandemic plan.

Artificial Intelligence: Fake news, digital evolution and free will


The moot at a recent Oxford Union debate reads, “This house believes that fake news is a serious threat to democracy and truth.” the fact is, it’s far worse than that.

Artificial intelligence (AI) is poised to catastrophically transform the information ecosystem and in the process destroy all semblance of truth, fact, knowledge and our ability to act freely and autonomously.

This is because AI will win the fake news game, and it will evolve and adapt to perfectly exploit the psychological weaknesses of human beings. This is foreseeable, and unleashing such intelligences to fight human information wars will be negligent at the least and malicious at worst.

Reasons why AI will destroy the notion of truth

Fake news can already be engineered to discredit journalists and cause real life political demonstrations over issues that do not exist. A Trend Micro report claims it costs $200,000 at present to cause such events.

Autonomous agents now account for 45% of social media posts in some countries. Masses of fake content can give the impression of popularity and cause conformist behavioural effects.

Fake news resembling, at first glance, legitimate sources is now widespread. This can drive belief and behavior through prestige-based psychological effects.

Social media databases harbor vast quantities of psychological and preference information about billions of humans. This data can be exploited to psychologically profile every user on Earth.

Human minds are hackable:

Human psychology is flawed and open to exploitation and manipulation as classic experiments in power and authority, conformity, bias and ideology demonstrate.

AI will learn to harvest this social information, and draw associations and temporal connections between information and behaviour. A recent systematic review details progress to date in using social media to predict the future.

Rival human factions will deploy such AIs to create, distribute, target, and deploy fake content individualized to the susceptibilities of individual human beings on a massive scale. The limits on human productivity will not apply to content generated by AI. The output will be unimaginably vast.

These AIs will be programmed as swarms of intelligences able to evolve and adapt to the defenses used by the AIs of rival human factions. Fake news spam identifiers will struggle to keep up in this evolutionary arms race.

Almost all Internet traffic and content will become AI generated.

Humans will be misled into beliefs and courses of action many steps ahead of being aware they are being manipulated, if even aware at all.

Human beings will fade into the background awash in a polluted information ecosystem, unable to discern fact from fiction or reality from revision.

We will lose all ability to act on information and evidence and thereby lose all freedom and autonomy.

This is the real threat of AI.

Adapt Research promotes the importance of clean information and the notions of risk management, and evidence-informed policy.

To discuss collaborating to understand the issues of the social and institutional threat from AI, fill out our contact form here.

Artificial Intelligence: the great unknown


Artificial intelligence has arrived, is here to stay, and is likely to transform our work-lives, personal lives and social structures. Exactly how no one is entirely sure.

The potential of AI was very apparent from discussions at the IBM Watson Summit in Auckland on August 16, 2017, and the New Zealand AI Forum ‘Connect’ event that followed.

With the development of data analysis that uses natural conversation as commands rather than code, expert practitioners in various disciplines who are not trained in programming will be able to navigate complex data structures to gain evidence-based insight without the need for analysts. Neural networks can be programmed without coding by using IBM’s Darviz tool.

For more on the future of analysis and AI see 13 year old Tanmay Bakshi’s YouTube channel with over 100 instructional videos.

The attendees at IBM’s event were at pains to point out that AI will not replace humans but will augment what humans can do. However, I wonder how truck drivers feel about autonomous vehicles?

Later in the day at the AI Forum event, New Zealand lawyer Bruce McClintock used historic case law to demonstrate how the issues of foreseeability and negligence are well covered by existing law. But how will we negotiate the issues around human autonomy and freedom that AI is likely to impinge upon. These are societal and moral rather than legal issues.

It is clear that much more thought and research is needed into the social, psychological, ethical, and legal aspects of AI and it’s rapid introduction into our lives.

At Adapt Research, we are very interested in this space, and in collaboration with one of our clients we’ve submitted an opinion piece for publication on these issues (details to come). We will update this blog with further commentary as it emerges.

Click here to contact us if you would like a copy of our report once it is published.

For information on AI issues, see for example:

Strategic Implications:

Societal Response:

Further research needed:

Read back is critical for healthcare communication and safe teamwork

doc photo

Safe and effective healthcare is frustrated by failures in communication. We know that double checking drug names and doses and using checklists are huge boons to patient safety. Effective communication is important too.

Repeating back important information (read back) enhances the effectiveness of communication across many industries.

However, formal communication protocols are uncommon in healthcare teams.

In our study we quantified the effect of read back on the transfer of information between members of a healthcare team during a simulated clinical crisis.

To do this we gave post-anaesthesia care unit nurses and anaesthetic assistants clinically relevant items of information at the start of simulations. A clinical crisis was prompted so that participants called an anaesthetist, who had no prior knowledge of the patient.

We analysed video recordings of the simulations and found that anaesthetists who read back the information were eight times times more likely to know the information at the end of the scenario compared to times when they didn’t respond.

Anaesthetists who gave any response at all were still three times more likely to know the information compared with no verbal response.

This means that in a critical healthcare situation, if information is not read back, there is a good chance that communication has failed.

Training healthcare teams to use read-back techniques should increase information transfer between team members with the potential for improved patient safety.

Catheter Ablation is cost-saving if we choose the right patients with Atrial Fibrillation


Catheter ablation (CA) for atrial fibrillation (AF) is a procedure with high up-front costs but is superior to pharmacologic treatments for reducing symptoms1 and hospital presentations2. In patients with mild symptoms or few hospitalisations the cost of CA may not be justified.

However, for patients with severe symptoms and/or frequent hospital admissions CA could be preferred when downstream health system costs and quality of life are taken into account. Several international cost effectiveness analyses have been published on CA for AF3, but few have stratified the target patient group by hospitalisations avoided, or by heterogeneity of quality of life gained.

Adapt Research developed a macroeconomic model to define a patient population for whom CA is economically rational.

We compared scenarios where CA is offered to different sub-groups of patients with AF. International literature and local New Zealand health system data informed heterogeneity of procedure success by type of AF, time since procedure, and age of patient. Disability weight and number of hospital presentations were varied. Costs of CA, downstream outpatient care, and subsequent hospitalisations were estimated from New Zealand health datasets and international literature. CA and pharmacologic management were compared to obtain incremental cost-effectiveness ratios (ICERs). Scenarios were modeled over five years and no difference in the rate of mortality or stroke was assumed between CA and drug treatment.

It turns out that the ICER for CA compared to pharmacologic management ranged from cost-saving to NZD$169,308 (USD$112,680).

Variables tending to increase the ICER were: lower cost of drug treatment, increased cost of CA, offering CA to older patients, and to those with non-paroxysmal AF.

Variables tending to decrease the ICER were lower procedure cost, increased disability weight assigned to AF, and increased number of hospitalisations avoided.

The ICER under present provision in New Zealand is estimated to be NZD$55,994 (USD$37,249). Targeting only those patients with the most severe symptoms reduces the ICER to NZD$35,750 (USD$23,782).

CA is cost-saving for patients having more than one hospitalisation per year for AF.

If QALYs and absentee days are monetized using GDP, then CA for a wide range of patients is cost-saving from a societal perspective. Time to recoup costs ranged from zero to 17 years.

So it seems that the cost-effectiveness of CA for AF is highly dependent on the patient population to whom CA is offered. This is important given heterogeneity of the target population. Using severity of AF scales4, which have been validated against quality of life metrics and number of hospital presentations, could help identify an appropriate target patient group.

Click here to request a copy of our full technical report.

References: 1 Shi, LZ. et al. 2015. Exp Ther Med, 10(2):816-22. 2 Bulkova, V. et al. 2014. J Am Heart Assoc, 3(4) e000881. 3 Neyt, M. et al. 2013. BMC Cardiovasc Disord, 13(78). 4 Ha, AC. et al. 2013. J Interv Card Electrophysiol, 36(2):177-84.

Should we close borders in a pandemic?


There will almost certainly be future pandemic diseases that pose a grave threat to human lives. Pandemic influenza, novel emerging infectious agents and possible synthetic bioweapons all pose serious risks. It seems biologically plausible that a new infectious agent might have the transmission characteristics of influenza and the death rate of Ebola.

In our modeling study we explored the costs and benefits of complete border closure to protect the island nation of New Zealand during a global pandemic.

Our cost-benefit analysis took a societal perspective and included case-study specific epidemiological data from past influenza pandemics. Country-specific healthcare cost data, valuation of life, and lost tourism revenue as well as a complete end to trade.

Even in the face of a complete end to tourism, exports and imports, a net benefit was estimated for scenarios where the mortality rate was very high at 2.75% of the country’s population dying. In this situation the net benefit was NZ$54 billion. Even for lower mortality rates there was a period of closure between 12 and 26 weeks at which the net benefit switched from favorable to unfavorable.

This “proof-of-concept” modeling work suggests that in some extreme pandemic scenarios it may make sense for New Zealand to close its borders.