PSLE Science Reality Lab Vol No.044 | “The Model Fits the Data” — Does That Prove the Explanation?

PSLE-SCI-REALITY-0044

Wait, What? A line can match every point you gave it and still fail the next point.

A science report shows a neat graph. Black dots are measured results. A coloured line passes close to the dots. The caption says:

“The model fits the data extremely well, proving our explanation.”

The graph looks convincing. The line almost seems to become the evidence itself.

But a scientific model and the observations it was fitted to are not the same object. The measurements came from the world. The fitted line was constructed to summarise or explain those measurements. A close fit can be useful evidence that the model is compatible with the data. It does not automatically prove that the model’s explanation is the only possible explanation—or that it will keep working when conditions change.

This distinction matters far beyond advanced mathematics. A Primary 5/6 learner already has the core tools: separate observation from inference, ask what evidence supports a conclusion, keep alternative explanations alive, and test a prediction rather than admiring a picture.

Quick Answer

When someone says “the model fits the data”, first identify the measured points and the modelled line separately. Then ask how the model was chosen or adjusted, whether the remaining differences show a pattern, whether another model could fit too, and whether this model correctly predicts observations it was not simply tuned to match. Good fit strengthens a model. It does not turn fit into unique proof of mechanism.

Reality Lab rule: Matching the old evidence is one test. Surviving new evidence is another.

The Learner Job This Reality Lab Owns

This article owns one real-world transfer job: how to evaluate a public scientific claim that a fitted line, curve or model “proves” an explanation because it matches the observed data. It does not replace eduKate’s existing PSLE Science owners of graph reading, scientific models, evidence evaluation, extrapolation, alternative explanations or measurement. It applies those skills to a common scientific communication object: observations displayed together with a fitted representation.

The Original Case: The Cooling Box Model

A fictional company tests a small insulated box. The box begins at 60°C inside and is placed in the same room each time. The team records the internal temperature every ten minutes.

Time	Measured temperature
0 min	60°C
10 min	54°C
20 min	49°C
30 min	45°C
40 min	42°C
50 min	39°C

An analyst draws a smooth cooling curve through the points. The line is very close to all six measurements. The advertisement says, “Our model proves the box loses exactly the same fraction of its remaining heat difference every ten minutes.”

What has actually been shown?

The measurements show how this box’s temperature changed during these tested fifty minutes under these conditions. The fitted curve shows that a particular mathematical pattern can represent those observations closely. That is useful. But the word proves goes farther. Another mathematical relationship might fit almost as well over the limited range. A different room temperature, starting temperature, airflow or box geometry might expose where the chosen model stops working. And even a perfectly shaped curve would not by itself identify every physical mechanism causing heat transfer.

Three Objects, Three Jobs

Object	Scientific job
Measured points	Record what was observed under stated conditions.
Fitted line or curve	Summarise a relationship according to a chosen model.
Mechanistic explanation	Explain why the relationship occurs in the physical system.

A fitted curve can support an explanation, but it does not collapse these three jobs into one. This is the same distinction you use when a PSLE Science diagram represents a system: the representation helps you reason, but it is not the system itself.

Why Fitting Is Useful

Scientists often need models because individual measurements contain variation. A model can summarise the main pattern, estimate relationships, make predictions and expose observations that do not behave as expected. A good fit is not a trick. It is evidence that the model captures something important about the data.

The mistake happens when compatible with the data becomes uniquely proven by the data.

The Leftover Difference Has Information Too

For each measured point, compare what the model predicted with what was actually observed. The difference is often called a residual. You do not need the word for PSLE Science, but the idea is powerful.

Time	Observed	Model predicts	Observed minus predicted
10 min	54°C	53°C	+1°C
20 min	49°C	48°C	+1°C
30 min	45°C	45°C	0°C
40 min	42°C	43°C	−1°C
50 min	39°C	41°C	−2°C

If the differences bounce around without an obvious pattern, the model may be a reasonable summary. But if early results are always above the line and later results always below it, the leftover pattern may be telling you the curve is missing something systematic.

NIST’s engineering statistics guidance makes this point explicitly: a high fit statistic such as R² does not guarantee that a model fits well, and residual patterns can reveal problems that a single headline number hides.

A Model Can Fit Because It Was Allowed to Bend

Imagine connecting six measured dots with a line that is allowed to bend exactly at every dot. Of course it fits the six known measurements perfectly. But has it learned a durable scientific relationship, or has it simply memorised the shape of those points?

This is why the question “How well did it fit?” is incomplete. Ask “How much freedom did the model have to adjust itself to these exact data?”

A very flexible model may fit old observations closely yet predict a new observation poorly. A simpler model may leave slightly larger differences on the original graph but work better when the system is measured again.

Two Explanations Can Share One Pattern

Suppose a plant grows taller as daily light exposure increases. One explanation says extra light increases photosynthesis enough to support more growth. Another says the brighter location was also warmer, and temperature contributed to the difference. The same upward graph could be compatible with both explanations if light and temperature changed together.

A curve through the plant-height points cannot by itself decide which mechanism is responsible. You need a design that separates the competing variables or a follow-up test where the explanations predict different results.

Fit tells you a model can describe the pattern. Discrimination tells you why one explanation deserves more confidence than another.

The Strong Move: Ask the Model to Predict Something New

Return to the cooling box. Imagine the model was built using only the first forty minutes of data. Before the 50-minute temperature is revealed, the model predicts 40°C. The actual measurement is then 39°C.

That new point is more informative than simply redrawing the line after seeing all six points. The model had to commit first.

This connects with Reality Lab Vol No.035: a prediction recorded before a result is known is different from an explanation adjusted after the result appears. Both activities can be scientifically valuable, but they are different evidence jobs.

Fit Inside the Tested Range Is Not a Promise Outside It

A model that fits temperatures from 20°C to 40°C may fail at 80°C if the material changes state, the instrument reaches a limit or a different process becomes important. A model that fits small objects may not scale to very large objects. A model that fits one species may not fit another.

This is why “the line fits here” and “the rule is universal” are separate claims.

The Model-Fit Audit

Find the observations. Which points came from actual measurements?
Find the model. Which line, curve, equation or simulation was constructed from assumptions or fitting?
Ask how it was chosen. Was the model specified first, or adjusted after seeing the pattern?
Inspect what is left over. Do the differences between model and observation show a systematic pattern?
Keep alternatives alive. Could another model or another mechanism explain the same observed pattern?
Test a new case. Does the model predict evidence it did not merely fit?
Check the boundary. Is the claim staying within the conditions and range tested?
Separate description from cause. Does the model only summarise the relationship, or is there independent evidence for the proposed mechanism?

Worked Case 1: The Straight-Line Advertisement

A fictional advertisement tests four fan speeds and room-cooling times. The four points lie nearly on a straight descending line. It says, “The straight line proves every increase in fan speed causes exactly the same improvement.”

The data support a roughly linear relationship over the tested speeds. They do not establish that the relationship stays linear at speeds not tested, nor that no other changing condition contributed. A better claim would stay inside the measured range and stated conditions.

Worked Case 2: Two Curves, Same Old Data

A report has measurements at 1, 2, 3 and 4 hours. Two different curves both pass very close to all four points but predict different values at 8 hours. Which curve is proven by the first four measurements?

Neither is uniquely proven. The 8-hour observation becomes a discriminating test because the models make different predictions there.

Worked Case 3: Perfect Fit, Wrong Variable

Five plant pots receive increasing amounts of fertiliser and are also placed progressively closer to a bright window. Plant height rises perfectly with fertiliser amount. A curve fits the pattern with no visible error.

The fit does not isolate fertiliser as the cause because light changed too. A perfect graph cannot repair an unfair comparison.

Worked Case 4: Small Errors, Clear Pattern

A model misses every point by only 1–2 units, but it is always too low in the morning and always too high in the afternoon. Why should you care if the errors are small?

Because the error direction is systematic. Time of day or another changing condition may be missing from the model. Small does not mean random.

Tempting Reasoning That Fails

“The line goes through the dots, so the theory is true.” A close fit supports compatibility, not unique proof.
“The fit number is 0.99, so there cannot be a problem.” A single summary number can hide patterned errors.
“The model predicted the old data.” If it was built using those data, that may be fitting rather than independent prediction.
“A model with more detail must be better.” Extra flexibility can fit noise as well as signal.
“If another model also fits, science cannot know anything.” Follow-up evidence can discriminate between competing explanations.

What Evidence Would Strengthen the Explanation?

the model was specified before key new results were known;
new observations fall close to genuine model predictions;
the leftover differences do not show a systematic pattern;
alternative models make different predictions and perform worse on discriminating evidence;
the experimental method controls important competing variables;
the proposed mechanism is supported by additional observations, not only the curve shape;
the model’s limits and tested range are stated explicitly.

PSLE-Style Transfer Case

A student measures the distance travelled by a toy car released from ramps at four heights. She draws a smooth curve that passes through the points. Her friend says, “The curve proves the car will follow the same pattern at any ramp height.”

Question: Give one reason the friend’s conclusion is too strong.

Strong answer: The curve is based only on the tested ramp heights. It may describe those observations, but the same relationship has not been tested at every other height, and other effects may become important outside the tested range.

Practice Lab

Practice A

A curve matches five measurements. All five were used to choose the curve. What extra evidence would be especially useful?

Answer: A new measurement made after the model gives a prediction, preferably under a condition the model claims to cover.

Practice B

The model’s errors are +3, +2, +1, 0, −1, −2, −3 as time increases. Random or patterned?

Answer: Patterned. The model moves systematically from underprediction to overprediction, suggesting the model shape may be wrong or a changing factor is missing.

Practice C

Two explanations predict the same result in the original experiment. How should you choose a useful follow-up test?

Answer: Find a condition where the two explanations predict different results, then measure that condition fairly.

Delayed Independent Return

When you next see a graph with dots and a smooth line, cover the line first. Describe only what the dots show. Then uncover the line and ask what extra job it performs. Finally ask one question the line cannot answer by itself.

If you can do those three steps, you are no longer treating a beautiful fit as the same thing as reality.

Where to Route Next

Teaching Guide for Parents and Tutors

Use transparent overlays or a sheet of tracing paper. Give the learner six measured points and ask them to draw one plausible line. Then show a second plausible line that also fits the old points but diverges later. Ask: “What new measurement would make these two models disagree?” This turns model comparison into an evidence problem rather than a drawing exercise.

Next, give a fitted line with a clear residual pattern. You do not need to teach formal statistics. Ask whether the model is sometimes too high and sometimes too low without pattern, or wrong in one direction for a whole region. The learner should discover that how a model misses matters.

Finally, separate descriptive success from mechanistic success. A learner is ready when they can say: “This model represents the data well, but I still need a test that distinguishes its explanation from other explanations.”

Authoritative Sources

The Quiet Return

A good scientific model earns trust in stages. First it must respect what has already been observed. Then it must survive the patterns in what it got wrong. Then it should face evidence it did not simply bend itself around.

So when a graph says “the model fits”, do not dismiss it—and do not surrender to it. Ask the next scientific question: what new observation could make this model fail?

Small Group Tutorials