Statistically speaking – a Plain English glossary and visualisations of terms used in statistical methodology research

Inter-rater Reliability (IRR) is a measure to help people know if they agree or not by checking if they make the same choices or decisions.

Bayesian refers to a specific way of doing statistics, by learning and making our predictions better using new information. Take predicting the weather as an example. We make a first guess about how likely it is to be sunny based on historical data. As new information becomes available, such as weather patterns, the Bayesian approach updates our guess using both the historical and the new information.

Bootstrapping is a way of making guesses about things for which we don’t have all the answers. A simplified example of the method is as follows: imagine you have a tub of chocolates, and you only know a few of the flavours in the tub.

Bootstrapping is like taking those few chocolates you do know, shaking them up, and from that, making guesses about what all the chocolates in the tub might be.

Borrowing strength is a way to improve your understanding of something by using information from others. Suppose you wish to understand how well a school is doing.

Instead of only looking at the test results from a few students at that school, you can ‘borrow strength’ by also looking at the overall performance of schools across the town.

Burn-in is a process similar to when you play a board game for the first time, and you need to play it a little bit before you fully understand how it works.

In statistics, it means you have to use some data at the beginning to make sure your calculations are right, just like warming up before you start playing a game smoothly.

When we check calibration we are seeing if a tool measures things well. For example, we may check that a thermometer or a scale is working correctly so we can trust the numbers it shows us.

Checking the calibration of a prognostic model is a way of seeing if our model is working well. We can do this by seeing if the predictions it makes match what actually happened.

For example, a prognostic model has good calibration if it correctly predicts that a high risk person is high risk and bad calibration if it predicts a high risk person is low risk (or vice versa).

Suppose you notice people that have lower blood pressure exercise more regularly.

Causal inference is a way of trying to understand if exercise is the direct reason for the lowered blood pressure, and not something else such as better nutrition. It helps us figure out if something actually makes something else happen or if it’s just by chance.

‘Censoring’ is a statistical word to refer to a situation where we don’t know if and when things happen.

Imagine tracking the time until light bulbs stop working. If a bulb is still shining at the end of the study, it is called ‘censored’ data because we don’t know when it actually stopped working.

Competing risks is like when you have different things that could happen, but only one can happen at a time. Imagine you are out for a walk, and you see two paths. You have to decide which one to take because you can’t walk down both paths at the same time.

In statistics, competing risks help us understand which thing is most likely to happen when there are many possible things that could happen.

Component network meta-analysis helps researchers figure out which medicines work best by looking at lots of different studies. It is a tool that helps find the best combination of medicines to make the most helpful treatment.

A “composite outcome” in statistics looks at different pieces of important information all at once to understand something better. Instead of just checking one thing, we look at many things all together.

It’s like when you bake cookies and you want to know if 1) they taste nice, 2) they are soft, and 3) they have chocolate chips.

A “continuous outcome” in statistics means that we’re looking at things that can have lots of different values. Imagine you are measuring the heights of sunflowers, instead of saying that a sunflower’s height is exactly 172cm, we could find out that its height is 172.7cm or 172.72cm. So, we use numbers to show how much or how big something is.

In statistics, convergence is when we gather more information which makes the results get closer to the right answer.

Coverage (in a simulation study) – A simulation study gives a range of values that it thinks the answer could be in. The “coverage” refers to the number of times that the actual answer is within that range. This shows how much we can trust the statistical model.

Cross-validation is like testing how well you can solve a puzzle by trying different pieces in different spots. Instead of just doing it once and hoping you’re right, you try fitting the pieces in different ways to see which one works best.

In statistics, cross-validation is where we split our data into parts, using some parts to build the model and other parts to test how well it works. This helps us check our model is working properly, even on new data it hasn’t seen before.

Deviance is a way of understanding how well our model fits the data. It’s a measure of how far our predictions are from the actual data. If our estimates are very close to the actual data, then we have a small deviance, and vice versa. It’s like seeing how far from the bullseye our throws are in a game of darts.

A diagnostic model can be used to work out how likely it is that a person has a medical condition or illness, by using information such as their age or symptoms.

– In statistics, the term “discrimination” has a different meaning than it does in everyday language. It is a way of checking how well a prognostic model is working.

We might have a prognostic model that predicts if someone’s health will get worse over the next year.

If the model can separate groups of people into those who will stay healthy and those whose health will get worse, we can say that it is able to discriminate. However, if it cannot separate these groups it cannot discriminate.

An electronic health record, or EHR for short, is a patient’s record held in digital form. It consists of all the different information the NHS keeps about a patient’s physical and mental health. This can include letters, assessments, medication given, and any treatments received.

An estimand is like a target or goal in statistics. It’s what we’re trying to figure out or measure from our data.  For example, if we’re studying the effect of a new plant food for sunflowers, the estimand might be the difference in height between the sunflowers that had the new food and those that didn’t.

External validation in statistics is like double-checking your answers. It helps us make sure our calculations and predictions are right by testing them with different data.

A flexible parametric model is a tool that can be used in survival analysis to help us understand how likely it is that an event will happen in the future. Imagine you have lots of points on a graph, like dots on a piece of paper.

A flexible parametric model is like a bendable wire that can take on many shapes to help us draw a line through those points to see how they relate to each other.

Frequentist refers to a specific way of doing statistics, by counting and measuring things many times to understand how likely something is to happen. Instead of guessing, maths is used to look at the actual numbers and data to understand how often things happen and what’s most likely to be real.

A “generalized linear model” is a tool that helps us understand and explain how things work together. Imagine you have lots of points on a graph, like dots on a piece of paper.

Generalised linear models help us draw a line through those points to see how they relate to each other by using preset options that can fit different shapes and curves to the data.

Granular data is information that has been divided into smaller more detailed parts. For example, knowing a person’s address instead of just which city they live in. It helps us understand things better by looking closer at the small details in the data.

In statistics, a “hazard” refers to the rate of something happening. It helps us understand how likely it is for a specific event to happen over time. Hazards are used in survival analysis to study things like how long it takes for a broken bone to heal.

A hazard ratio is a tool that helps us compare the rate of something happening, like getting sick, between two groups. Imagine a race between two teams.

The hazard ratio relates to how much faster one team is compared to the other in reaching the finish line. It helps us study how different things, like treatments or activities, affect the speed of reaching that finish line.

Individual patient data is the information we have about each person in a study. Researchers can use this data to learn more about how people are different and what helps them stay healthy.

Prognostic models use information about a person to figure out the risk of something happening to them in the future. Internal validation means testing if our prognostic model works well in our data to make sure it makes good predictions and could be useful.

Interpolation is like drawing a line between two dots on a graph to find the missing dot in the middle. It helps us guess what the unknown number might be when we know some of the numbers nearby.

Imagine you have a ruler with marks every inch, but you need to know the length of something that’s somewhere between the 2-inch mark and the 3-inch mark. Interpolation helps you guess its length as 2.5 inches.

“Jackknife” is a method where we leave one piece of data out at a time and see how much our answer changes. Imagine you are interested in the weight of a bag of marbles. Instead of weighing them all at once, you could leave one marble out and weigh the rest. Then repeat this leaving out a different marble and see how much the weight changes.

“Likelihood” in statistics means how probable or possible something is. It helps us understand the chance of different things happening. If something has a high likelihood, it’s very likely to happen, but if it has a low likelihood, it’s not very likely at all.

Linear regression is like drawing a straight line on a graph to show how two things are connected. For example, if we are interested in seeing if there is a relationship between the height of a plant and how much sunlight it gets, linear regression would allow us to do this.

Logistic regression helps us make predictions when there are only two possible results, like “yes” or “no”. For example, a logistic regression model could use information like how many clouds there are and the temperature to tell us the chance that it will rain.

Calculating the “mean difference” helps us find out how similar or different things are on average. For example, if we are comparing the heights of two groups of plants, we find the average, or the ‘mean’, of each group and then compare them. If one group is taller on average, the mean difference tells us by how much.

Markov Chain Monte Carlo, or MCMC for short, is a method that helps researchers solve problems and make predictions by taking little steps randomly to explore the data. It’s like moving step-by-step through a maze to find the answers you’re looking for.

Calculating the “mean difference” helps us find out how similar or different things are on average. For example, if we are comparing the heights of two groups of plants, we find the average, or the ‘mean’, of each group and then compare them. If one group is taller on average, the mean difference tells us by how much.

A statistical model is a tool that helps us understand numbers and information better. It can be used to find patterns in data, helping us see how things are connected and allowing us to make predictions.

For example, we might be interested in seeing if there is a relationship between the height of a plant and how much sunlight it gets, a statistical model can help us find the answer in our data.

Model fit is about checking if the model we’ve created matches well with the data we have. If it is a good fit, it means our model does a good job of explaining the data we collected.

Multiple imputation is a way to handle data that has some missing pieces. It makes several predictions for what the missing parts might be and combines the results, to give a better picture of the whole data.

It’s like making multiple guesses for questions that weren’t answered in a neighbourhood survey to get a more complete picture of the neighbourhood’s opinions.

Multiplicity happens when we test lots of different things using our one set of data. For example, we might want to know if a new medicine for the flu helps people feel better.

We may do three tests using the same data to see if it helps with different symptoms like a fever, cough and sore throat.

A multi-state model helps us understand how things change over time. For example, they can be used to study how people move between being healthy, getting sick, and recovering. It helps us understand complex processes that have multiple possible outcomes.

Network meta-analysis is a method to compare more than two treatments for a particular disease by combining data from different studies. It looks at how all the treatments relate to each other and helps researchers understand which treatment might be the best overall.

Parametric refers to an approach where we assume data follows a specific pattern, like a curve or a straight line. This assumption helps us use mathematical equations to describe and analyse the data more easily.

A pooled estimate is when you combine estimates from different groups or studies to get one more accurate overall estimate. It can help us make better decisions or conclusions by combining all the information together.

A posterior distribution combines a prior distribution with the data that has been collected to give the best result possible. It helps us to understand the probability of different results after looking at all the evidence.

Precision is an estimate of how consistently you have measured something. If you get very similar results each time you measure something, you have high precision.

A prior distribution is a guess about what could be true based on what you already know. A vague prior distribution means you are not very sure about the truth and so your guess covers a lot of possibilities. An informative prior distribution means you have a good idea of the truth based on previous knowledge or research and your guess covers a smaller set of possibilities.

A prognostic model is a tool that uses information about a person, like their age, blood test results and how they feel, to figure out what might happen to them in the future. If the model predicts that a person is at a high risk of their health getting worse, they may be given a treatment to try to lower the risk of this happening.

Regression is a tool that helps us find patterns in data. Imagine you have a lot of points on a graph, showing the height and number of leaves on a plant. Regression helps us draw a line through the points on the graph to see relationships, like how taller plants have more leaves.

Residuals are the differences between what we see and a statistical model’s predictions. For example, if we predict that a cake will bake in 40 minutes because it weighs 500 grams, but it takes 45 minutes, then the residual is 5 minutes. Residuals help us understand how well our model fits the data by showing us where our predictions are off.

The risk difference is a way of comparing two groups to see, for example, which group is more likely to get sick. If group A has 100 people and 6 get sick (6%), and group B has 100 people and only 4 get sick (4%), the risk difference is 2%. It helps us understand how much more or less likely it is that something will happen to one group compared to another.

A risk ratio helps us understand how likely something is to happen by comparing chances between two groups. If the risk ratio is 1, it means both groups have the same chance. If it’s more than 1, it means one group has a higher chance, and if it’s less than 1, it means the other group has a higher chance.

Shrinkage is a method that can be used when we have limited data and it’s affected the model we’ve created. It does this by adjusting parts of the model so that they are more trustworthy.

This helps to reduce the influence of extreme data points, leading to better model predictions.

Imagine getting a computer to work out what equipment would work best to dig a hole instead of trying them all yourself. A simulation study is similar – it uses a computer program to test how well statistical tools work under different conditions, rather than doing a real-world experiment. This helps researchers select the best tool to analyse their data.

The standardised mean difference is a method that helps us see if there are differences between groups when their results have been measured differently. It’s like using the same measuring stick to check if one group of plants is taller or shorter than another, even if they were measured in different ways (e.g. centimetres vs inches).

Statistical methodology are tools and techniques we can use to make sense of lots of data and answer questions like whether a new medicine works well. It’s like assembling a puzzle to see the whole picture where statistical methods help us put the pieces together in the right way.

Research in this area focuses on finding the best tools for different data.

A “surrogate outcome” in statistics is when we use something easier to see or measure to understand something that is harder to check directly.

For example, to check that a plant is healthy you can’t check their roots but instead, you could look at its leaves. If the leaves are green and strong, you can assume that the plant is healthy.

Survival analysis is a tool that helps us figure out how long something lasts. By studying the times when events happen, it’s used to understand things like how long people live for or how long it takes for something to happen, like a machine to break down.

Time-to-event outcomes are a way of measuring how long it takes for something important to happen. For example, if a machine stops working today, the manufacturer may want to know how long it will be before another one breaks.

By keeping track of these events, we can predict when similar events might happen in the future.

A Type I error is when you say something is true when it’s not. For example, in a medical test, a type I error would happen if the test showed that a person has a disease when they actually don’t. This type of error is also known as a “false positive”.

A Type II error is when you say something is false when it is actually true. For example, in a medical test, a type II error would happen if the test showed that a person doesn’t have a disease when they actually do. This type of error is also known as a “false negative”.

Validation is like checking if puzzle pieces fit together perfectly so that we can see the whole picture clearly. In statistics, it means making sure that the models we fit and the results we get are correct and can be trusted.