Research Questions

Research involves pursuing answers to questions.  How can I reset the clocks on my kitchen appliances?  A Google search usually provides a ready answer to this question.  One would not think of publishing an article on having answered this question, nor would any media outlet encourage such a publication.

Do I have any bakers’ yeast?  This leads to researching the cupboards and perhaps the refrigerator.  This research leads to a yes or no conclusion.  One would not think of publishing an article on having found the yeast.  Despite the success of this research, there is no generalizable finding to be reported.

Who killed Roger Ackroyd?  Agatha Christie reported on the research to answer this question.  Hercule Poirot sleuths his way to determining James Sheppard to be the killer.  Poirot confronts Sheppard who commits suicide.  The research question was answered but, as intriguing as the story may be, little is added our knowledge of murder.

Scientific research involves seeking answers to real, rather than fictional, questions to which there are no “off the shelf” answers.  Scientific research also involves addressing questions whose answers are of broad importance.  No one knows what I want for lunch, although no one but me or the waiter at the pub really knows the answer.

So, scientific research addresses important questions where are no existing sources of answers.  Research sponsors can tell us what questions are important to them, but it may be that answers are available, but not readily apparent.  Thus, the first consideration is whether existing answers can be found.

In recent years, most people would start by searching Google.  I just searched on “Do pickles affect icebergs?” It resulted in 28 million “hits.”  At 10 minutes per hit, it would take me 25 years to reviews this corpus.  Fortunately, there are much more powerful and efficient ways to approach this need.

I recently pursued the question, “What are the patient states and transition probabilities in substance abuse?”  Using the Curis Meditor research portal (, hosted on the Northern Light SinglePoint platform, I searched 40 million research articles to identify 250 highly relevant articles, mostly medical research articles.

I then used Northern Light’s machine learning capabilities to “read” these articles and provide relevant insights.  From these insights, and fully reading several articles, I formulated an evidence-based state model of substance abuse.  This process required 6 hours, not the months required if I had solely relied on Google.

This state model of substance abuse provided an evidence-based hypothesis.  We might have used this model to predict the future course of substance abuse.  However, we were interested in whether new social interventions might help diminish opioid abuse.  In other words, we wanted to  explore “what if?” rather than “what is?”  How do you address a research address questions regarding what might happen versus what will happen if the future replicates the past?

We could assume that estimates of current states are good predictions of future states.  If the future interventions of interest are intended to affect the independent variables of the statistical model of current outcomes, it might be reasonable to assume that this the model of “what is?” can reasonably predict the future.  But, if these interventions have not been previously employed, this assumption is questionable. 

More typical are situations where models of the past cannot provide acceptable predictions of the future.  Then, we cannot construct our model by totally relying on past and current data.  Other than relying on psychics or other prognostications, we need to start with first principles to formulate a mathematical model of the phenomena underlying the prospect of predicting possible answers to our research question.

What are first principles?  There are a rich set of principles for dynamic motion, e.g., F=MA and E=MC2, electrical current, e.g., E=RI, and fluid mechanics, e.g., Q=VA.  There are also principles that cut across physical domains such as continuity (e.g., of mass and energy) and conservation of momentum. 

What about behavioral and social phenomena?  There are models of human behaviors related to driving cars and piloting airplanes, troubleshooting failures, coordinating multiple tasks and other human-system oriented tasks.  The ability to formulate such models is abetted by the requirements for humans to conform to the constraints of the systems they are operating and maintaining.

Social phenomena of interest for our research on opioid abuse concern how people relate to each other via their social networks.  In particular, it has been found that current addicts can be coached into recovery by former addicts – termed peer recovery coaches.  Thus, the model used to address this research question had to include a representation of each person’s social network and the status of the members of the network, i.e., susceptible, addicted, in recovery, etc.

We used this model to address a rather unfortunate real-world experiment as the coronavirus caused the isolation of people from their social networks.  Our model predicted that this would lead to increasing instance of overdoses and deaths. This is what happened in Washington, DC and elsewhere.  The predictions of our model were in the ballpark of the actual measurements.

Does this validate our model?  This question involves several subtleties.  At one level, did we answer the research question?  Yes, we did.  Was the answer “correct”?  We certainly did not accurately predict the exact number of overdoses and deaths.  So, perhaps the answer is “no.”  Were our predictions in the range of possibilities that could have happened?  The answer is arguably “yes.”

Consider our abilities to predict the paths and intensities of hurricanes.  There are multiple American and European models for making these predictions.  Weather forecasters rely on all of them to inform them on the range of possible futures.  The key is to understand what might happen by computing possible futures, knowing that a point prediction of what will happen is inevitably wrong.  I explore this notion in great depth in Computing Possible Futures (Oxford, 2019).

Leave a Reply