There are six broad questions which can be answered in data analysis according to an article called “What is the question?” By Jeffery T. Leek, Roger D. Peng. These questions help to frame our thinking of data science problems. Here, I propose that these questions also provide a unified framework for relating statistics to data science.
The six questions according to Jeffery Leek and Roger Peng are
A descriptive question seeks to summarize a characteristic from a dataset. You do not interpret the results. For example: number of fresh fruits and vegetables served in a day.
An exploratory question is a question where you try to find a pattern of a relationship between variables i.e. you aim to generate a hypothesis. At this stage, you do not test the hypothesis. You are merely generating a hypothesis. More generally, you could say that you are proposing a hypothesis which could hold in a new sample from the population.
An inferential question restates the proposed hypothesis in the form of a question that would be answered by analyzing the data. You are validating the hypothesis i.e. does the observed pattern hold beyond the data at hand. Most statistical problems are inferential problems.
A predictive question would be one where you predict the outcome for a specific instance.
A causal question: Unlike the predictive and inferential questions, causal questions relate to averages in a population i.e. how changing the average in one measurement would affect another. Causal questions apply to data in randomized trials and statistical experiments where you try to understand the cause behind an effect being observed by designing a controlled experiment and changing one factor at a time.
A mechanistic question asks what is the mechanism behind an observation i.e. how a change of one measurement always and exclusively leads to a deterministic behaviour in another. Mechanistic questions apply typically to engineering situations.
Implications for Data Science
The six questions framework raise awareness about which question is being asked and aim to reduce the confusion in discussions and media. However, they also help to provide a single framework to co-relate statistics problems to data science.
Re the six questions:
- Descriptive and exploratory techniques are often considered together
- Predictive and Inferential questions can also be combined
So we could consider four questions:
Why this framework matters?
That’s because it provides questions which may not have been encountered before. Here are three examples
- Typically, the starting point of data science is ‘big data’. Most data scientists are not used to dealing with small data. When you do not have a large amount of data you need to consider a more statistics driven approach see the difference between statistics and data science big data and inferen…
- The significance of causal questions i.e. Corelation vs Causation which I discuss in Correlation does not equal causation but how exactly do you determi…
- And finally – the mechanistic question which is elaborated more in Why do some traditional engineers not trust data science
The six questions provide rigor and simplicity of analysis. I also find that these questions provide a comprehensive set of questions that link statistics to data science. They help you to think beyond the norm i.e. beyond problems that you encounter to consider all possible problems.
What is the question? By Jeffery T. Leek, Roger D. Peng