Performance Data Primer: Assumptions

The assumptions we make

Sport science is built on assumptions.

Every measurement we collect, every metric we calculate, every model we build, and every recommendation we make requires us to simplify reality in some way. We assume that a sensor adequately represents the construct we care about. We assume that our processing decisions preserve the meaningful parts of the signal. We assume that our statistical approach appropriately represents the athlete or population. Eventually, we assume that the information we have generated is useful enough to inform a decision.

None of this is inherently problematic.

The problem begins when we forget that those assumptions exist.

As technology has become increasingly embedded within sport, collecting and analyzing data has become easier than ever. Wearable systems generate thousands of observations from a single training session. Force plates produce dozens of variables from a jump. Athlete management systems can automatically calculate rolling averages, z-scores, ratios, and risk flags. Machine-learning models can search through enormous datasets for relationships that would be impossible to identify manually.

But increased analytical capability does not necessarily mean increased certainty.

In many cases, the opposite may be true.

The further data travels from what was directly measured, the more assumptions it tends to accumulate along the way.

Understanding those assumptions does not mean abandoning the technology, metric, or statistical approach. It means understanding what the information can reasonably tell us…and what it cannot.

Measurement is already an assumption

Before we ever open a spreadsheet, we have already made assumptions.

Consider something as seemingly straightforward as an athlete's maximum velocity.

If we use a local positioning system or GNSS/GPS device, position is measured repeatedly over time and velocity is derived from changes in position. If we use an inertial measurement unit, acceleration measured at the sensor may be integrated and combined with other information to estimate velocity.

Both systems may ultimately provide a variable labeled max velocity.

That does not mean they arrived there in the same way.

Sensor location, sampling frequency, filtering, signal noise, integration error, satellite geometry, positional accuracy, algorithms, and proprietary processing can all influence the final value.

The number on the dashboard may say 9.2 m/s, but that apparent precision can disguise a much messier measurement process underneath it.

The important question is not simply:

What number did we get?

It is:

How did we get that number?

This distinction becomes increasingly important as we move further away from directly measured quantities. Some metrics are relatively close to the underlying signal. Others are derivatives of derivatives, composites of multiple variables, or outputs from proprietary algorithms we may only partially understand.

That does not make them useless.

But it should change our confidence in what they represent.

A metric having a name does not guarantee that it perfectly represents the construct implied by that name.

"Player load" is not load itself. "Fatigue" derived from a countermovement jump is not fatigue itself. "Readiness" is not something a wearable directly observes. These are representations of larger constructs.

Useful representations, potentially.

But representations nonetheless.

Our foundational understanding of injury is full of assumptions

Assumptions are not limited to technology.

They can exist in something as foundational as deciding what counts as exposure in injury epidemiology.

We commonly report injury incidence as injuries per 1,000 athlete-exposures or per 1,000 hours of exposure. These approaches are useful because they allow comparisons across populations with different amounts of participation.

But consider what happens when we use hours.

Embedded within that denominator is an assumption that an hour is a meaningful unit of comparable exposure.

An hour of low-intensity technical practice and an hour of high-intensity competition both contribute an hour.

An hour for an athlete returning from injury and an hour for a fully healthy athlete both contribute an hour.

An hour containing repeated maximal accelerations, decelerations, jumps, and contacts may be treated similarly to an hour with substantially lower physical demands.

We know intuitively that those hours are not equivalent.

Yet the denominator requires us to treat them as comparable units in order to calculate the rate.

That does not invalidate injury incidence.

It tells us what the number represents.

It is a population-level simplification that helps us describe injury occurrence relative to exposure. It is not a perfect representation of the biological risk contained within every minute of participation.

And importantly, this assumption is introduced at the point of data collection. Long before we select a statistical model or calculate a risk estimate, we have already decided what constitutes exposure and how it will be quantified.

Those decisions determine what questions the dataset can eventually answer.

Some of the most consequential assumptions in an analysis are made before the analysis ever begins.

That distinction becomes especially important when we begin moving from epidemiology toward individual decision-making.

When observation becomes causation

There is another assumption embedded in the way we study injury that is easy to overlook.

We often assume that the biomechanics we observe at the moment of injury are the biomechanics that caused the injury.

Video analysis has given us an incredible window into injury mechanisms. We can slow an event down frame by frame and identify trunk position, hip motion, knee flexion, foot placement, or the sequence of events leading into injury.

Those observations are valuable.

But observing a movement during an injury does not necessarily mean that movement caused the injury.

Consider the false step during acceleration. An athlete may initially step backward before propelling themselves forward. If we only observed false steps during injuries, it would be tempting to identify that movement as problematic. Yet athletes perform false steps repeatedly during sport without consequence.

The movement itself is not enough.

The same problem exists when we analyze cutting injuries.

We may identify a particular trunk position, knee angle, foot placement, or cutting strategy during an ACL injury and conclude that the mechanics explain what happened. Yet ask the athlete about the movement and you may hear something remarkably simple:

"I've done that move hundreds of times before."

And they probably have.

They may have planted their foot in a similar position, reached a similar knee angle, or executed a similar cut hundreds—or thousands—of times without injury.

So why did the tissue fail this time?

That is a much harder question.

The biomechanics observed at the moment of injury may certainly contribute to the loads experienced by the tissue. But they exist within a much larger system: tissue capacity, fatigue, previous exposure, perception, decision-making, timing, opponent behavior, contact, speed, accumulated loading, and countless other factors that may differ from one repetition to the next.

The mechanics present when an injury occurs are not automatically the mechanics that caused the injury.

This is an important distinction between mechanism and causation.

Video can help us describe how the injury occurred. It may help identify the positions and forces that were present when tissue capacity was exceeded. But moving from that observation to the claim that a particular movement pattern caused the injury requires another inferential step—and another set of assumptions.

The absence of injury during hundreds of similar movements should matter to how confidently we make that claim.

In some ways, the repetitions where nothing happened may be just as informative as the one where something did.

If we only study the injury event, we risk treating the movement associated with failure as inherently dangerous while ignoring the enormous number of times the same or similar strategy was successfully tolerated.

If a movement occurs hundreds of times without consequence and once with injury, the movement alone is unlikely to tell the entire story.

Again, the answer is not to dismiss biomechanical analysis.

It is to be more precise about what it tells us.

Biomechanics can help us understand the conditions present when an injury occurred. It can identify potential loading strategies, describe mechanisms, and generate hypotheses about contributors to injury.

What it cannot do by itself is establish causation.

And recognizing that limitation should influence how confidently we move from observing an injury mechanism to trying to "correct" the movement we saw.

Processing adds another layer

The assumptions continue after data collection.

Raw data rarely arrives ready for decision-making. Signals are filtered. Trials are excluded. Variables are normalized. Missing values are handled. Outliers are removed—or retained. Time windows are selected. Baselines are created.

Every one of these decisions changes the information we ultimately see.

Filtering provides an obvious example. Too little filtering may leave noise that contaminates the signal. Too much filtering may remove meaningful information. Neither choice is universally correct. The appropriate decision depends on the signal, measurement system, movement, sampling frequency, and question being asked.

Even seemingly simple decisions such as averaging can hide important assumptions.

If we average three jumps, we assume the average provides the representation we care about. If we take the best jump, we assume maximal performance is more important. If we use the most recent trial, we make a different assumption again.

The same three jumps can therefore generate three different answers.

This is why data processing should not be viewed as a neutral step between collection and analysis.

Processing is part of the analysis.

There is no completely untouched version of data once we begin turning a signal into information. Every processing decision reflects a belief about what matters.

The goal is not to eliminate these decisions. We cannot.

The goal is to make them intentionally.

Statistics come with assumptions too!

Once the data are processed, we often introduce another layer of simplification through statistics.

Take z-scores.

Z-scores are extremely useful in sport science because they allow variables with different units and scales to be placed into a common framework. Jump height, peak force, eccentric impulse, and reactive strength can suddenly be visualized together.

But the simplicity of that visualization can make it easy to forget what sits underneath it.

Traditional interpretation of z-scores around standard deviations is most natural when the distribution behaves reasonably like a normal distribution. Sport performance data do not always cooperate.

Some metrics are skewed. Others have hard boundaries. Some contain meaningful outliers. Small athlete samples can make the mean and standard deviation unstable. An individual athlete's longitudinal distribution may look very different from the distribution across the team.

Yet we may still convert everything into the same standardized scale.

Again, that does not mean we should stop using z-scores.

It means we should understand what we are asking them to do.

A z-score can be an excellent descriptive tool without pretending that every decimal point carries the same inferential meaning.

This same principle applies to rolling averages, acute-to-chronic ratios, smallest worthwhile changes, regression models, clustering approaches, and nearly every other analytical technique commonly used in sport science.

The question should not be:

Is this method perfect?

No method is.

A better question is:

Are its assumptions reasonable enough for the decision I am trying to make?

Prediction magnifies the problem with assumptions

The stakes become even higher when we move from describing what happened to predicting what will happen.

Injury prediction provides perhaps the clearest example.

A common modeling approach might classify injury as a binary outcome:

0 = no injury
1 = injury

That structure allows us to use methods such as logistic regression or, when time-to-event is considered, survival approaches such as Cox proportional hazards models.

Mathematically, this can be useful.

Biologically, the situation is considerably more complicated.

An athlete does not move cleanly from "healthy" to "injured" because a spreadsheet changed from zero to one.

Tissue capacity may fluctuate. Symptoms may develop gradually. Exposure differs between athletes. Previous injury changes context. Training demands vary. Competition creates different demands than practice. Contact and non-contact mechanisms are different. Some athletes continue participating despite symptoms that another athlete might report as an injury.

The binary outcome is therefore not reality.

It is a representation of reality that allows us to perform a particular analysis.

That distinction matters.

The model can still be informative. But its output should not suddenly be treated as biological truth simply because sophisticated statistics produced it.

The complexity of the model does not remove the assumptions underneath it. Sometimes it simply makes them harder to see.

Can we navigate the prediction problem?

One approach that may help us navigate some of these problems is Bayesian analysis.

Rather than asking us to interpret an estimate as though it exists in isolation, Bayesian approaches allow us to begin with what we already believe, or what previous evidence suggests, and update that belief as new information becomes available.

Conceptually, this fits sport remarkably well.

We rarely encounter an athlete with no prior information.

We know their injury history. We have previous testing. We understand their position, training history, recent exposure, and where they are in the competitive calendar. As new information arrives, our understanding of the athlete changes.

In that sense, Bayesian thinking may better resemble how practitioners actually make decisions.

If an athlete has repeatedly demonstrated a certain level of performance and suddenly produces one abnormal test, we probably do not immediately abandon everything we previously believed about them. We update our confidence based on the new observation.

As additional evidence accumulates, our belief may continue to shift.

That is an appealing alternative to approaches that can encourage us to think in more absolute terms.

But Bayes does not solve the assumption problem.

It moves some of those assumptions somewhere we can see them.

The most obvious example is the prior.

A Bayesian model requires us to specify what we believe about a parameter before incorporating the current data. That prior might come from previous research, historical team data, previous observations of the athlete, expert knowledge, or a deliberately weakly informative distribution.

That can be a major strength.

A reasonable prior allows existing knowledge to contribute to the analysis rather than pretending every new dataset begins from zero. This can be particularly valuable in elite sport, where sample sizes are often small, and we may have substantial historical information about an athlete or population.

But priors can also create problems.

What happens if our previous evidence is poor? What if our historical population does not resemble the athlete in front of us? What if the prior reflects our own biases more than meaningful information?

A strong prior built on the wrong assumptions can pull our interpretation in the wrong direction. A very weak prior may avoid that problem but provide little advantage when the available data are sparse.

Bayesian approaches do not eliminate assumptions. They force us to decide which assumptions we are willing to bring with us before seeing the next piece of evidence.

There are assumptions elsewhere in the model as well. We still have to decide which variables matter, how they relate to the outcome, what probability distributions appropriately represent them, and whether the observations entering the model are trustworthy in the first place.

Bad measurement does not become good measurement because we put it into a Bayesian model.

And uncertainty does not disappear simply because we express it probabilistically.

That may actually be one of the most useful features of the approach.

Instead of forcing the question into "Is this athlete at risk or not?" we can begin thinking in terms of how strongly the available evidence should change what we currently believe.

That is a much more honest representation of many problems in sport.

The goal of prediction may not be to perfectly identify what happens next. It may be to appropriately update how confident we are about what could happen next.

Bayesian approaches can help us do that.

But, like every other method discussed here, their usefulness depends on understanding the assumptions underneath them…and being willing to adjust our confidence accordingly.

Awareness, not perfection

It would be easy to read all of this and conclude that sport science data are fundamentally flawed.

That is not the point.

Every scientific discipline simplifies reality.

We need measurement models. We need statistics. We need classifications. We need thresholds. We need ways to reduce enormous amounts of information into something humans can actually interpret.

The goal is not to eliminate assumptions.

The goal is to know where they are.

If I know my measurement system becomes less accurate at certain velocities, I can account for that when interpreting the data.

If I know that my exposure metric treats very different sporting experiences as equivalent units, I can avoid pretending an incidence rate tells me more than it actually does.

If I know that observing a movement at the time of injury does not establish that the movement caused the injury, I can be more careful about the interventions I build from that observation.

If I know a variable is highly skewed, I can reconsider whether mean and standard deviation-based profiling is the best representation.

If I know my injury model reduces an extraordinarily complex process into a binary outcome, I can be more cautious about how strongly I interpret its predictions.

This is where good sport science becomes less about finding the "best" technology or statistical method and more about selecting the most appropriate approach for the question being asked.

The best method is rarely the one with the fewest assumptions. It is the one whose assumptions you understand and can tolerate for the decision in front of you.

Sometimes a rough estimate is enough.

Sometimes it is not.

A coach asking whether today's practice was generally more demanding than yesterday's may not require laboratory-grade measurement precision.

A medical team deciding whether an athlete is prepared to tolerate the demands of competition after a major injury may require substantially greater confidence.

Those are different questions with different consequences.

They should not necessarily require the same standard of evidence.

From assumptions to decisions

This is where assumptions begin to move beyond a technical problem and become a communication problem.

We rarely collect sport science data simply because we want numbers.

Eventually, someone has to make a decision.

Train more.

Train less.

Modify practice.

Progress rehabilitation.

Hold an athlete out.

Return an athlete to competition.

Change a program.

Ignore the signal entirely.

The assumptions embedded throughout our measurement and analysis process should influence how confidently we communicate those recommendations.

If the measurement is noisy, the construct is indirect, the processing decisions are uncertain, and the model requires several strong assumptions, perhaps our language should reflect that uncertainty.

"We know this athlete is at increased risk" is very different from "This information may suggest something worth monitoring."

Both statements can lead to action.

But they communicate very different levels of certainty.

And the amount of certainty we require should depend partly on the consequences of being wrong.

When the risk tolerance for a decision is high, we may be comfortable acting with imperfect information. When the consequences of a wrong decision are substantial, we may need better measurement, stronger evidence, multiple converging signals, or simply more information before acting.

Put differently, lower tolerance for being wrong should generally raise the standard of information required to make the decision.

This does not mean we ever reach complete certainty. We rarely will in sport. It means the precision of our measurement, the number of assumptions we are willing to tolerate, and the confidence of our recommendation should be proportional to what is at stake.

This is where data stops being a statistics problem and becomes a decision-making problem.

Our confidence in a recommendation should never exceed our confidence in the information used to create it.

That is the bridge to the next part of this series.

We have talked about the ethics of performance data: what we should collect, how we should use it, and our responsibilities to the athlete.

Here, we have examined the assumptions embedded in how that data becomes information.

Next comes perhaps the most important step:

translation.

Because the ultimate value of sport science is not determined by how much data we collect or how sophisticated our analysis becomes.

It is determined by whether we can take imperfect information, understand how confident we should be in it, combine it with the context and risk tolerance surrounding a decision, and communicate it in a way that actually helps someone decide what to do next.

The data will never be perfect. Our job is to understand just how imperfect it is…and make better decisions anyway.

This blog was inspired by many conversations this summer that got me thinking and working to tie multiple perspectives together. Thanks to Avinash Chandran, Courtney Chaaban, Chris Juneau, and Natalie Hendricks for the excellent discussions this summer.

TOMMY OTLEY

Physical therapist, sport scientist, educator, and researcher working at the intersection of rehabilitation and performance.

QUICK LINKS

Free Downloads

Services

Work With Me

Ask A Question

READ MORE