The core expertise that Statisticians offer to the world is drawing conclusions from small samples. Therefore, knowing how to design surveys, estimate the right sample size, decide on the right way to ask the question or measure a property are all essential skills for any statistical thinker. The skills you need to be competent in Sampling & Surveys are best captured by my Survey Wheel.
My Survey Wheel has 6 Arcs.
- Objectives – What it is that you need to know and from who or what?
- Design – How to sample & measure your target population.
- Fieldwork – How to collate, clean & code our data.
- Results – What are the best charts & tables to summarise our data and what to do with missing data?
- Insight – What are the key drivers or segments in our data?
- Decisions – What actions can be taken and is the client capable of doing them?
The reason I call this a Survey Wheel rather than a survey line is that my 25+ years experience of advising clients on sampling has taught me the starting point of any sampling project can be any one of these arcs. As a result, the other 5 arcs have to adapt to the requirements of the starting arc in order to prevent the wheel from breaking.
To learn more about my Survey Wheel, why not download this 7-page PDF document “The 6 Arcs of the Survey Wheel“? Should you engage me as a consultant to work on your survey or sampling project, you will find me following what is described in this document.
Alternatively, listen to me explain my Survey Wheel in this 50 minute webinar produced in collaboration with Captive Health in 2015. In this webinar, I focus on many of the common mistakes that people make in designing surveys and samples.
Below is a list of various materials that you can use to learn more about Surveys & Sampling. I have organised these by the 6 Arcs of my Survey Wheel. By the way, there is a reason for the colour coding of the headings. Red headings are the most statistically advanced arcs where serious statistical thinking is needed. Green arcs are more straightforward computational tasks that can be automated in some cases. Blue headings require softer skills in consulting, facilitation and elicitation but statisticians still have much to offer in these arcs.
There are 2 parts to the Objectives arc:
- What do I need to know? (Goals)
- Is this linked to a decision that is pending?
- Who do I need to ask? (Target Population)
- How many sub populations are needed?
For an example of objective setting, take a look at the 5 questions I pose at the start of this post “Who reads fake news?“.
This arc has two very distinct spokes, Sampling & Measuring.
The Sampling spoke seeks to answer these two questions:
- How many people should be measured?
- What are the objectives?
- How should they be selected?
- What are your sampling cells?
- How random is your selection?
- What biases may be inherent in your sample selection?
- Do you need to weight your results?
- Do you have good quality external data to allow you to weight?
For the Measuring spoke, there are 3 questions to answer:
- Do you need to use a questionnaire to take measurements?
- Are your measurements repeatable, reproducible & unbiased (whether taken by questionnaire or instrument)
- Have you tested your measurements prior to the full survey?
I explore some of the issues of sample design in this blog post “Is all-white alright?” which explores how large a sample is needed to decide if all white workforce is statistically plausible or could be an indicator of discrimination.
Often, a statistician needs to link the sample size to the Results arc, especially when there is a requirement to report an estimate to within a specified margin of error. There are many methods available and in my post “Life on Mars” I show how simulation can be used to measure the reliability of gender pay gap statistics and how this is correlated with sample size.
Other times, the sample size will be linked to the Decisions arc. The COVID19 pandemic is a perfect case study on how to make decisions under huge uncertainty and consequences so I wrote this post on what sample size would be needed to make a decision to lift all restrictions and you may be surprised at the answer!
The issue of question design came to the fore during the run-up to the 2016 EU Referendum in Britain. There were persistent differences between polls undertaken by phone and those undertaken online. Matt Singh of Number Cruncher Politics carried out an experiment looking at the impact of including or excluding a Don’t Know option in the question. The results were fascinating and whilst I thought at the time, Matt had drawn the wrong conclusions from his data, it was an excellent piece of work. My conclusion at the time was that Leave voters were more certain on their vote than Remain voters.
A very good introduction to the issue of survey & question design can be found in one of lesson plans that the Royal Statistical Society provides to schools. The Marketing Statistician example was developed by me based on data from a dating site client I worked with years ago. The lesson takes you through the issues of survey modes and biases, weighting of respondents and ways of designing questions given an objective.
The key questions to answer are:
- How will you collate the responses? Web, text, face to face, other?
- How will you ensure data collation is free from error?
- How will you code the responses? (if needed)
If you want to create a web survey yourself, then I have found Google Forms to be quite easy to use. I used this to create a 10-question questionnaire to help people work out the odds of Donald Trump being re-elected as President of the USA in 2020.
The key questions to answer are:
- How will you handle non-response?
- How will you summarise the responses?
Non response or missing data is a serious statistical issue which can affect any sample. The key question is whether those not responding are different in any way from those who did respond.
Summarising data is a standard statistical skill and requires an understanding of basic statistical concepts. You can find a list of relevant posts in section A of this link “Basic Statistical Concepts“.
The key spokes of this arc are:
- What are the key drivers of the key questions?
- Do these drivers differ by segments within the target population?
The 1st spoke, Key Driver Modelling, requires expertise in Statistical Modelling. This allows you to explore and explain the relationship between a set of key questions and a set of potential driver questions. If you would like to learn more, then why not take a look at my training course “Understanding your world with Statistical Modelling“?
The 2nd spoke, Segmentation, requires expertise in Multivariate Analysis. You can find out a list of relevant posts in this link “Multivariate Analysis” but one post that is particularly relevant is one I wrote at the beginning of 2017 which looks at the similarities between Brexit & Trump.
The key spokes of this arc are:
- What are the constraints on any action that the organisation can take?
- What are the capabilities of the organisation to take action?
This arc covers a lot of potential topics but you may find my post on “How to close your gender pay gap with DMAIC” of interest.
For more information about my other training courses in statistics, please visit my Statistical Training homepage.