Artificial Intelligence

“There are lies, damned lies and statistics” B.Disraeli

Why such distaste for a field that, according to Webster’s Merriam-dictionary, is simply “a branch of mathematics dealing with the collection, analysis, interpretation and presentation of masses of numerical data. ”1 Why is the field of statistics in such a negative light by so many people?

There are four main reasons

• It’s a complex field. Even the basic concepts are not easily accessible and are very difficult to explain.
• Even the best-intentioned experts can misapply the tools at their disposal.
• The third reason behind all this hatred is that those with an agenda can easily create statistics to lie about when communicating with us.
• The final reason is that statistics can often seem cold and distant, making them very difficult for the public to grasp.

Descriptive statistics are intended to summarize the main characteristics of a data set. However, incorrect or inappropriate use can lead to misleading conclusions. A typical example is the use of the mean to summarize a distribution, without taking into account variability or skewness. Another common error is to present percentages without explaining the total number of people, which can be misleading as to the true extent of a phenomenon. It is therefore crucial to understand the assumptions and limitations of each descriptive measure in order to use it correctly.

Let’s take the example of analyzing salaries within a company. If we simply look at average salaries, we might conclude that the company is paying its employees well. However, if management salaries are very high compared to the rest of the employees, the average would be biased upwards. It would be more relevant to use the median, which gives the salary in the middle, or to look at the complete salary distribution for a more accurate view.

Statistical inference aims to draw conclusions about a population from a sample of that population. However, this process is subject to error. Sampling errors and Type I and II errors are common. In addition, errors can be exacerbated by confusion between correlation and causation. A solid understanding of the principles of statistical inference is essential to avoid these pitfalls.

Let’s imagine a public health study seeking to establish a link between a particular dietary habit (such as eating organic) and better overall health. If the study finds a positive correlation, it doesn’t necessarily mean that eating organic causes better health. There could be confounding factors, such as income level or lifestyle, that influence both eating habits and health status. Here, we can fall into the trap of confusing correlation with causation.

Sampling is a crucial stage in any data collection process. Yet many errors can occur at this stage. The sample may not be representative of the target population, due to selection bias or non-response. What’s more, the sample size may be insufficient to detect an effect. Careful sample planning is therefore essential to obtain reliable results.

Consider a customer satisfaction survey conducted by an e-commerce company. If the company only solicits opinions from customers who have made a recent purchase, it runs the risk of obtaining a distorted picture of overall customer satisfaction. Indeed, dissatisfied customers may have stopped making purchases and therefore not be included in the sample. This is an example of selection bias.

A common mistake in data analysis is to ignore the impact of sample size on results. A large sample size can make a very small effect significant, while too small a sample size may not have sufficient power to detect an existing effect. Furthermore, statistical significance does not necessarily mean practical significance. So it’s important to consider sample size when interpreting results.

Suppose you’re conducting a study to assess the effect of a drug on lowering blood pressure. If you have a very large sample of patients, you may see a statistically significant drop in blood pressure. However, this drop may be very small, say 0.1 mm Hg, a clinically insignificant value despite its statistical significance. This is an example where sample size can make a small effect significant. On the other hand, if the sample is too small, a real effect may be missed. It is therefore important to consider clinical or practical significance in addition to statistical significance.

Digging deeper into this issue, Ben Jones (see author who inspired this article) managed to find figures on kidney cancer rates as well as demographics for every US county, and he created an interactive dashboard (figure below) to visually illustrate the fact that Kahneman, Wainer and Zwerlink are doing quite clearly in words.

Notice a few elements in the dashboard. On the choropleth map (filled in), the darkest orange counties (high rates relative to the overall U.S. rate) and the darkest blue counties (low rates relative to the overall U.S. rate) are often side by side.

Also, note how in the scatterplot below the map, the marks form a funnel shape, with less populated counties (on the left) more likely to deviate from the reference line (the overall U.S. rate), and more populated counties like Chicago, L.A. and New York are more likely to be close to the overall reference line.

One final observation: if you hover over a county with a small population in the interactive online version, you’ll notice that the average

number of cases per year is extremely low, sometimes 4 cases or less. A small deviation – even just 1 or 2 cases – in a subsequent year will pull a county from the bottom of the list to the top, or vice versa.

In the next article, we’ll explore the 5th type of error we may encounter when using data to illuminate the world around us: Analytical aberrations.

This article is heavily inspired by the book “Avoiding Data pitfalls – How to steer clear of common blunders when working with Data and presenting Analysis and visualization” written by Ben Jones, Founder and CEO of Data Litercy, WILEY edition. We recommend this excellent read!

You can find all the topics covered in this series here: https://www.businesslab.mu/blog/artificial-intelligence/data-7-pitfalls-to-avoid-the-introduction/

We’ve all expressed disbelief at the relevance of mathematics to our daily lives. What purpose could this dense, complex subject possibly serve? Well, in a world where data is everywhere and infuses every strategic decision made by organizations, mathematics is vitally important (editor’s note: it always has been!).

In our data analysis projects, mathematical errors can occur as soon as a calculated field is created to generate additional information from our initial dataset. This type of error can be found, for example, when :

• We perform aggregations (sum, mean, median, minimum, maximum, count, separate count etc.) at different levels of detail
• We make divisions to produce ratios or percentages
• We work with different units

These are obviously just a few of the types of operation where errors can occur. But in our experience, these are the main causes of the problems we encounter.

And, in each of these cases, it doesn’t take a genius engineer or scientist to correct them. It just takes a little care and a lot of rigor!

In this article, we won’t dwell too much on this common mistake. In fact, there are a large number of articles and anecdotes which illustrate this type of problem perfectly and in detail (which we also discussed in the previous article).

The most famous and costly example is the crash of the Mars Orbiter probe. If you’d like to find out more, please click here: Mars Climate Orbiter – Wikipedia

You may argue that none of us is part of NASA and has to land a probe on a distant planet, so we’re not concerned. Well, you may well come across this type of error when handling time data (hours, days, seconds, minutes, years), financial data (different currencies), or managing stocks (units, kilos, pallets, bars etc.).

We aggregate data when we group records that have an attribute in common. There are all sorts of such groupings that we deal with in our world as soon as we can establish hierarchical links; time (day, week, month, year), geography (cities, region, country), organizations (employees, teams, companies) and so on.

Aggregations are a powerful tool for apprehending the world, but beware, they involve several risk factors:

• Aggregations summarize a situation and do not present detailed information. Anyone who has taken part in a datavisualization training course with our teams is familiar with Anscombe’s quarter:

The statistical summary is a typical example of what aggregates can hide. In this example, the four data sets have exactly the same sums, means and standard deviations on both coordinates (X,Y). When we plot each of the points on curves, it’s easy to see that the 4 stories are significantly different.

As soon as data is aggregated, we try to summarize a situation. We must always remember that this summary masks the details and context that explain it. So be careful when, in a discussion, your interlocutors only talk about average values, sums or medians, without going into the details of what may have led to that particular scenario.

• Aggregations can also mask missing values and be misleading. Indeed, depending on the way we represent information, the fact that data is missing may not be clearly visible at first glance.

Take, for example, a dataset in which we observe the number of bird strikes on aircraft for an airline.

Our objective is to determine the month(s) of the year with the most incidents. This gives :

July appears to be the month with the highest number of impacts counted. However, if we look at the details by year, we realize that the aggregation chosen to answer our question did not allow us to determine that the seizures for the year 2017 stopped during this famous month of July:

The answer to our question was therefore August, if we exclude the data for the year for which we didn’t have all the records.

• Totals and aggregations :

This is the last example of the problems linked to aggregations that we’re going to discover in this article. This is one of the author’s “favorite” mistakes. Some might even call it a specialty!

It comes into play when it’s necessary to count the distinct individuals in a given population. Let’s say we’re looking at our customer base and want to know how many unique individuals are in it.

Counting the distinct ids for the whole company gives us a count of our unique customers:

But if we look at each product line and display a sum without paying attention :

We found 7 more customers!

This happens simply because there are customers in the customer base of the company studied who take both services AND licenses, and who end up being counted twice in the total!

This is a problem with simple solutions in all modern datavisualization and BI software, but it tends to hide itself in a series of calculations and aggregations, causing sometimes surprising discrepancies at the end of the chain.

We’ll illustrate this point with an example taken from one of the dashboards we made for one of our customers. With all our expertise, we also sometimes jump headlong into this type of error:

And yes, we’re talking about an occupancy rate that’s “slightly” over 100%!

How is this possible? A simple oversight!

The sum of the divisions is not equal to the division of the sums…

In this case, we had a data set similar to the one below:

Is the occupancy rate equal to :

The sum of the individual occupancy rates? FALSE!

This gives us a total of 30% + 71% + 100% + 50% + 92% + 70%, i.e. 414%.

And that’s exactly the error we made on an even larger data set…

Or the ratio of total passengers to total available capacity? 125/146 = 86%. That’s more accurate!

Note: the average of individual occupancy rates would also be wrong.

In short, whenever a ratio is manipulated, it’s a question of dividing the total of the numerator and denominator values to avoid this type of problem.

This is just one example of a ratio error. Honorable mentions can be given to the treatment of NULL values in a calculation, or to the comparison of ratios that are not calculated with the same denominators.

In the next article, we’ll explore the 4th type of obstacle we may encounter when using data to shed light on the world around us:

Statistical slippage. (Spoilers: “There are lies, damned lies and statistics” B.Disraeli)

This article was strongly inspired by the book “Avoiding Data pitfalls – How to steer clear of common blunders when working with Data and presenting Analysis and visualization” written by Ben Jones, Founder and CEO of Data Litercy, WILEY edition. We recommend this excellent read!

You can find all the topics covered in this series here: https://www.businesslab.mu/blog/artificial-intelligence/data-7-pitfalls-to-avoid-ep-2-7-technical-errors-how-is-data-created/

Having defined a few key data-related concepts, we can now delve into the technical issues that can lead to errors. This article deals with the problems associated with the process of obtaining the data that will subsequently be used. It’s about building the foundations of our analyses.

And it goes without saying that we don’t want to build a house of cards on sand!

To stay with the construction metaphor, if problems of this nature exist, they will be hidden and barely visible in the final building. Particular care must therefore be taken during the data collection, processing and cleaning stages. It’s not for nothing that it’s estimated that 80% of the time spent on a data science project is spent on this type of task.

To avoid falling into this trap, and to limit the load required to carry out these potentially tedious operations, we need to accept three fundamental principles:

• Virtually all datasets are not clean and need to be cleaned and formatted.
• Each transition (formatting, join, link, etc.) during the preparation stages is a potential source of new error
• It is possible to learn techniques to avoid the creation of errors arising from the first two principles.

Accepting these principles does not remove the obligation to go through this preliminary work before any analysis, but the good news is that knowing how to identify these risks, and learning as we go along, helps to limit the scope of this second obstacle.

Data is dirty. I’d even go so far as to say that all data is dirty (see first principle above), with problems of formatting, data entry, inconsistent units, NULL values and so on.

Some well-known examples of this trap

Take the crash of NASA’s Mars Climate Orbiter in 1999, for example. A $125 million error caused by a dual unit system: imperial and metric units. This led to an erroneous calculation that affected the power sent to the probe’s thrusters and its destruction.

Fortunately, not all errors of this nature will cost us so much money! But they do have a significant impact on the results and ROI of the analyses we carry out.

So, at DATANALYSIS, we’re currently running several projects specifically on data quality in the context of DATA Marketing, and we’re dealing with two types of subject:

• Data validation, which aims to improve data quality through data processing, by :

-Standardizing fields (phone number, email, etc.): +262 692 00 11 22 / 00262692001122 / 06-92-00-11-22 correspond to the same line, and we can automate much of this work thanks to appropriate processing;

– Filling in empty fields using other data in the table. For example, we can deduce the country of residence from telephone numbers, zip codes, cities, etc.

• Deduplication, by :

-Using adapted rules to identify potentially identical lines. Two records with the same e-mail address, telephone number or company ID;

-Using distance calculation algorithms to define similar values in terms of spelling, pronunciation, common characters, etc.

From these examples and our own experience, we can see that this type of error mainly stems from data entry, collection or “scrapping” processes, whether implemented automatically or by humans. So, in addition to the solutions that can be implemented in data preparation processes, improving these preliminary steps will also greatly improve the quality of the data to be processed, and this requires education, training and the definition of rules and standards that are clearly known and shared (data governance is never far away).

Finally, we should also ask ourselves when we can consider this stage to be sufficiently clean. After all, we can always do more and better, but the costs involved can often outweigh the expected returns.

In the IT world, there’s an image that sums up this type of problem:

Often, the mistake lies between the screen and the seat!

And yes, even the best data scientists, data analysts or data engineers can make mistakes in the data cleansing, transformation and preparation stages.

Frequently, we manipulate several files from different sources and different applications, which multiplies the risks associated with dirty data issues and the risks when manipulating the files themselves:

• Different levels of granularity
• Joins on fields whose values are not exactly identical (e.g. ST-DENIS vs SAINT DENIS).
• Different file perimeters

And this problem can also be made more complex depending on the tools used in our analyses:

• In Tableau, for example, we can perform data joins, relations or links to link several datasets together. Each type of operation has its own rules and constraints.
• In Qlik, you need to understand how the associative engine works and the associated modeling rules, which differ from those of a traditional BI model.

In this case, it’s often a question of technical constraints linked to the very business of data preparation, and taking the time to understand the risks and processes in place will save a great deal of time in delivering reliable, high-performance data analysis.

In the next article, we’ll explore the 3rd type of obstacle we may encounter when using data to shed light on the world around us: Mathematical errors.

This article was strongly inspired by the book “Avoiding Data pitfalls – How to steer clear of common blunders when working with Data and presenting Analysis and visualization” written by Ben Jones, Founder and CEO of Data Litercy, WILEY edition. We recommend this excellent read!

You can find all the topics covered in this series here : https://www.businesslab.mu/blog/artificial-intelligence/data-7-pitfalls-to-avoid-ep-1-7-epistemological-errors-how-do-we-think-about-data/

Let’s start by defining what epistemology is.

Epistemology (from the ancient Greek ἐπιστήμη / epistémê “true knowledge, science” and λόγος / lógos “discourse”) is a field of philosophy that can refer to two fields of study: the critical study of science and of scientific knowledge (or scientific work).
In other words, it’s about how we construct our knowledge.

In the world of data, this is a central and critical topic. We are familiar with the process of transforming data into information, knowledge and wisdom:

Here, the problem lies in the way we consider our starting point: data! Indeed, the use of data and its transformation in the following stages are the result of conscious and controlled processes and procedures:

==>I clean up my data, process it in an ETL / ELT, store it, visualize it, communicate my results and share them, and so on. This mastery gives us control over the quality of each step. However, we tend to embark on this work of transforming our primary resource while overlooking a crucial point, the source of our first obstacle:

DATA IS NOT AN EXACT REPRESENTATION OF THE REAL WORLD!

Indeed, it’s all too easy to work with data by thinking of data as reality itself, and not as data collected about reality. This nuance is essential:

It’s not crime, but reported crime
It’s not the diameter of a mechanical part, but the measured diameter of that part.
It’s not public sentiment on a subject, but the declared feeling of those who responded to a survey.

Let’s go into the details of this obstacle with a few examples:

Let’s take a look at this dashboard showing all the meteorite impacts on Earth between -2500 and 2012. Can you identify what’s strange here?

Source: Meteorite fall on Earth visualized | Tableau Public

Meteorites seem to have carefully avoided certain parts of the planet – a large part of South America, Africa, Russia, Greenland, etc. And if we focus on the graph showing the number of meteorites per year, these have tended to fall more in the last 50 years (and almost not over the whole period covering -2055 to 1975). And if we focus on the graph showing the number of meteorites per year, these have tended to fall more in the last 50 years (and hardly at all over the whole period from -2055 to 1975).

Is this really the case? Or rather flaws in the way the data was collected?

• We have recently begun to systematically collect this information and rely on archaeology to try and determine the impacts of the past. As erosion and time have taken their toll, the traces of the vast majority of impacts have disappeared and can no longer be counted (and no, meteorites didn’t start raining in 1975).
• For a meteorite impact to be included in a database, it has to be recorded. And to do that, you need an observation, and therefore an observer, who knows who to report it to. These two biases have a major impact on data collection, and help to explain the large areas of the Earth that seem to have been spared by the meteorite fall.

Sometimes, the cause of this discrepancy between data and reality can be explained by a defect in the collection equipment. Unfortunately, anything manufactured by a human being in this world is liable to fail. This applies to sensors and measuring instruments, of course.

What happened on April 28 and 29, 2014 on this bridge? There seems to have been a huge spike in bicycle traffic across the Fremont Bridge, but only in one direction (blue curve).

Source : 7 datapitfalls – Ben Jones

Time series of the number of bicycles crossing the Fremont Bridge

You’d think it was a beautiful summer’s day and everyone was on the bridge at the same time? That it was a one-way bike race? That everyone who crossed the bridge on the outward journey had a flat tire on the return journey?

More prosaically, it turns out that the blue counter had a fault on those particular days and was no longer counting bridge crossings correctly. A simple change of battery and sensor solved the problem.

Now, ask yourself how many times you’ve been misled by data from a faulty sensor or measurement without being aware of it?

And yes, our own human biases have a major effect on the values we record when gathering information. We tend, for example, to round off measurement results:

Source : 7 datapitfalls – Ben Jones

If we go by his data, diaper changes take place more regularly every 10 minutes (0, 10, 20, 30, 40, 50) and sometimes over certain quarters of an hour (15, 45). Wouldn’t that be incredible?

It is an incredible story. In fact, we need to look at how the data was collected. As human beings, we have this tendency to round up information when we record it, especially when we look at a watch or clock: why not indicate 1:05 when it’s 1:04? Or even simpler, 1:00, because it’s even simpler?

The final example we’d like to highlight here is the so-called “Black Swan” effect. If we think that the data we have at our disposal is an accurate representation of the world around us, and that we can extract from it assertions to be set in stone; then we are fundamentally mistaken about what data is (see above).

The best use of data is to learn what isn’t true from a preconceived idea, and to guide us in the questions we need to ask ourselves to learn more?

But back to our black swan:

Before the discovery of Australia, every swan sighting ever made could confirm to Europeans that all swans were white – wrongly! In 1697, the sighting of a black swan completely challenged this common preconception.

And the link with the data? In the same way that we tend to believe that a repeated observation is a general truth – wrongly so – we can be led to infer that what we see in the data we manipulate can be applied generally to the world around us and to any era. This is a fundamental error in the appreciation of data.

All it takes is a little mental gymnastics and a little curiosity:

• Clearly understand how measurements are defined
• Understand and represent the data collection process
• Identify possible limitations and measurement errors in the data used
• Identify changes in measurement methods and tools over time
• Understand the motivations of data collectors

In the next article, we’ll explore the 2nd type of obstacle we may encounter when using data to illuminate the world around us: Technical Mistakes.

This article is heavily inspired by the book “Avoiding Data pitfalls – How to steer clear of common blunders when working with Data and presenting Analysis and visualization” written by Ben Jones, Founder and CEO of Data Litercy, WILEY edition. We recommend this excellent read!

You can find all the topics covered in this series here : https://www.businesslab.mu/blog/artificial-intelligence/data-7-pitfalls-to-avoid-the-introduction/

DATA! DATA ! DATA everywhere!

These days, data is everywhere, featuring prominently in all new projects and corporate strategies. It’s the key to performance in these uncertain times. At Business Lab consulting, we’re the first to be convinced that it’s a powerful tool that accelerates performance…when it’s well used, well understood and well mastered!

In this new series of articles, we’re going to talk about the big bad wolf; the devil that hides in the detail (or sometimes reveals itself in broad daylight) and discuss with you the 7 main types of pitfalls posed by data and its use. As far as possible, we’ll try to illustrate them with an example from our own experience, because as experts we’ve had the good fortune to come up against each of them in our missions…

Note: these are the pitfalls discussed in Ben Jones’ book, “7 data pitfalls”, which we highly recommend!

Enough suspense, let’s now unveil the 7 families of DATA deadly sins that we’ll be exploring in greater detail over the next 7 weeks:

We often use data with the wrong frame of mind, or with erroneous preconceptions. So, if we go into an analysis project thinking that the data is a perfect representation of reality; if we draw definitive conclusions based on predictions without questioning them; or if we look in the available information for anything that might confirm an opinion already made; then we can create critical errors in the very foundations of these projects.

Technical and technological issues are often a major source of error in the world of data. Once you’ve identified the information you need, there’s a whole series of obstacles to overcome. Are my sensors working? Do my processes not generate duplicates? Is my data clean or up to date? Complex issues in our projects! After all, isn’t it said that a data analyst spends most of his time and energy preparing and cleaning his data?

So now you know what your math lessons from school, college and high school are all about! There’s something for every level and taste! If you’ve never combined data at different levels of detail, or made mistakes when calculating ratios, or forgotten that you shouldn’t mix carrots and bananas, we’d love to hear from you!

As the saying goes, “There are lies, damned lies and statistics”. This is the most complex trap to get to grips with, because it takes a lot of skill to fully understand what’s at stake. However, in a world where machine learning, datamining and AI are king, it’s a family of errors that’s only becoming more common!

Do the measures of central tendency or variation we use lead us astray? Are the samples we work with representative of the population we want to study? Are our comparison tools valid and statistically significant?

So now you know what your math lessons from school, college and high school are all about! There’s something for every level and taste! If you’ve never combined data at different levels of detail, or made mistakes when calculating ratios, or forgotten that you shouldn’t mix carrots and bananas, we’d love to hear from you!

Golden rule: we’re all analysts (whether we have that title or not).

As soon as we use data to make decisions, then we are analysts, and therefore prone to making decisions based on aberrant analyses. For example, are you familiar with vanity metrics? Or have you ever made extrapolations that don’t make sense in the light of the data used?

Unlike statistical errors or analytical aberrations, graphical blunders are well known and easily identifiable. Why? Because they can be seen (often from a distance). Have we chosen the right type of chart for our analysis? Is the effect I want to show clearly visible?

What’s the difference with graphic blunders?

Here we’re talking about the overall design of the final product and the interactions we’ve defined within it to ensure that the audience we’re trying to convince has the most ergonomic and aesthetically pleasing experience possible! Does the choice of colors we’ve made confuse or simplify the analysis? Have we used our creativity to make our dashboards pleasing to the eye, and have we used aesthetics to bring impact to the analysis we’re making? Is the final product easy to use and ergonomic, or are the interactions complex and time-consuming?

Are you ready to follow us through the twists and turns of everything that can go wrong with your data analysis projects, so that you don’t fall into these traps?

See you next week!