Science Education Systems · Article 83. Maya, Jia Jun, Hana and Ethan remain fictional Punggol learners. This article follows the missing-data layer: how Science treats absent observations as part of the evidence system rather than pretending blank cells do not matter.
The 50-second parent route
Missing data is not merely less data.
Why a value is missing can change the conclusion.
The route is:
intended measurement → missingness → missingness mechanism → pattern → bias risk → recovery strategy → imputation or model → sensitivity analysis → transparent interpretation
The key question is:
Would the conclusion change if the missing observations were systematically different from the ones we still see?
This article extends How Scientific Data Quality Works, How Scientific Provenance Works and How Scientific Longitudinal Studies Work.
1. Missing data changes the evidence base
Ten measurements were planned.
Seven remain.
The sample is smaller.
But the deeper question is:
Why are those three missing?
2. Missingness can be harmless or dangerous
If values disappear for reasons unrelated to the quantity being studied, the damage may be mostly loss of precision.
If disappearance depends on the outcome, exposure or hidden variables, bias can appear.
3. Maya’s missing-data error is deleting blanks silently
Her spreadsheet function ignores missing rows.
She assumes nothing changed.
Her repair:
count missingness and report who or what disappeared.
4. Jia Jun’s missing-data error is replacing everything with zero
Blank does not mean zero.
His repair:
distinguish absence of measurement from an observed value of zero.
5. Hana’s missing-data error is treating every missing value the same
Instrument failure, participant refusal and outcome-related dropout have different implications.
Her repair:
model the reason for missingness, not only the blank cell.
6. Ethan’s missing-data error is trusting sophisticated imputation automatically
A complex model fills every gap.
He assumes the problem is solved.
His repair:
imputation inherits assumptions and cannot recover information that the data fundamentally do not contain.
7. Missing completely at random is the strongest simple assumption
Missingness is unrelated to observed or unobserved values relevant to the analysis.
Under this condition, complete-case analysis may remain unbiased for some estimands, though less precise.
8. Missing at random uses observed information
After conditioning on measured variables, missingness does not depend on the unobserved value itself.
Models can use observed predictors to recover part of the lost information.
9. Missing not at random is harder
Even after accounting for observed data, missingness depends on the unobserved value or related hidden quantities.
Strong assumptions and sensitivity analyses are usually required.
10. These labels describe models, not visible facts
Researchers cannot usually prove from observed data alone which missingness mechanism is true.
Scientific judgement and sensitivity analysis matter.
11. Primary Science can learn missing-data thinking simply
A plant was not measured on Day 4.
Was the learner absent?
Was the plant damaged?
Was the ruler unavailable?
The reason changes interpretation.
12. Primary 3 can distinguish zero from missing
Zero leaves observed is a measurement.
No leaf count recorded is missing data.
They are not interchangeable.
13. Primary 4 can mark missing values explicitly
Use NA or another defined symbol.
Do not leave ambiguity between “not measured” and “measured as zero.”
14. Primary 5 can ask whether missingness is patterned
Are values missing only on rainy days?
only from one instrument?
only from weaker plants?
Pattern can reveal bias risk.
15. Primary 6 can compare complete and incomplete datasets
Calculate a trend with all observations.
Then remove a critical time point.
Students see how missingness can change the story.
16. Secondary Science can formalise missing-data mechanisms
MCAR.
MAR.
MNAR.
complete-case analysis.
imputation.
weighting.
sensitivity analysis.
Students can see that missingness is a modelled process.
17. Complete-case analysis is simple
Use only observations with all required variables present.
This can waste information and introduce bias when complete cases differ systematically from incomplete ones.
18. Pairwise deletion uses more available data
Different analyses use different subsets depending on which variables are present.
This can create inconsistent sample bases and awkward interpretation.
19. Single imputation fills each missing value once
Mean substitution.
last observation carried forward.
regression prediction.
These methods can underestimate uncertainty because the imputed values are treated too confidently.
20. Mean imputation can distort variance
Replacing every missing value with the same mean makes the dataset look less variable than reality.
Correlations can also be distorted.
21. Last observation carried forward assumes stability
In longitudinal data, carrying the last observed value into the future implies no further change.
That may be unrealistic and can bias treatment comparisons.
22. Multiple imputation represents uncertainty better
Create several plausible completed datasets.
analyse each.
combine estimates and between-imputation uncertainty.
The method remains only as good as its imputation model and missingness assumptions.
23. Imputation models should include useful predictors
Variables related to missingness.
variables related to the missing value.
important outcome or exposure variables.
Richer observed information can make MAR assumptions more plausible.
24. Imputation should respect variable type
Binary.
count.
continuous.
bounded score.
time-to-event.
Using the wrong model can generate impossible values.
25. Structural missingness is different from accidental missingness
A question applies only to participants who answered yes previously.
The later blank is by design.
It should not be treated as an accidental non-response.
26. Censoring is related but distinct
An exact event time is not observed because follow-up ended first.
We still know the event occurred after a certain time.
Special time-to-event methods use that partial information.
27. Detection limits create another partial-information problem
A concentration is below the instrument’s quantification limit.
It is not simply zero.
The measurement contains bounded information.
28. Dropout is longitudinal missing data
A participant contributes early measurements but later disappears.
If dropout is related to worsening outcome, observed trajectories can look falsely favourable.
29. Attrition can destroy randomised comparability
Groups begin comparable.
More high-risk participants leave one group.
The observed survivors are no longer comparable in the same way.
30. Prevention is often better than imputation
Good follow-up.
backup instruments.
clear forms.
real-time data checks.
participant support.
Designing against missingness preserves more information than repairing it afterward.
31. Missingness should be monitored during data collection
Which site has more blanks?
which instrument?
which time point?
which subgroup?
Early patterns can reveal procedural failures while they are still repairable.
32. Missing-data provenance matters
Why is the value missing?
When was it marked missing?
Was it ever observed and later removed?
Who changed the record?
Data lineage prevents false blanks and silent deletion.
33. Missing data and bias are inseparable
If missingness depends systematically on the phenomenon under study, the visible sample becomes biased.
34. Missing data and statistical power are connected
Fewer observed outcomes mean less information.
Effective sample size falls.
Confidence intervals widen.
35. Missing data and effect size are connected
If high responders remain while low responders drop out, the observed treatment effect can inflate.
Missingness can change magnitude, not merely precision.
36. Missing data and external validity are connected
If some population groups respond less often, the final dataset may no longer represent the target population.
Transportability weakens.
37. Missing data and construct validity are connected
If difficult items are skipped more often, total scores may increasingly represent easier subskills.
The measured construct can shift.
38. Sensitivity analysis is essential under uncertain missingness
Assume missing outcomes are slightly worse.
Then substantially worse.
Does the conclusion change?
Sensitivity analysis maps how dependent the result is on unverifiable assumptions.
39. Pattern-mixture models are one advanced approach
Different missingness patterns can be modelled separately and combined under explicit assumptions.
This helps explore MNAR scenarios.
40. Selection models are another advanced approach
The outcome process and missingness process are modelled jointly.
Again, identification often requires assumptions not testable from observed data alone.
41. Weighting can compensate for differential observation
Units with a lower probability of remaining observed can receive greater weight if that probability is estimated appropriately.
Inverse-probability weighting depends on measured predictors and correct models.
42. No method makes MNAR disappear magically
When missingness depends on unseen outcomes, the data cannot reveal the missing values uniquely.
The honest solution is explicit assumptions and robustness checks.
43. Missing data matters in AI training
A dataset may lack examples from certain languages, environments or rare cases.
The absence can become a capability gap.
44. Missing labels are not random automatically
Hard examples may be less likely to receive reliable labels.
Training only on easy labelled cases can bias the model.
45. AI systems can confuse missing with negative
Blank medical field.
no diagnosis recorded.
That does not necessarily mean the diagnosis is absent.
Data semantics matter.
46. AI can help audit missingness
Useful prompts:
“Show missingness percentage by variable and subgroup.”
“Find whether missingness increases over time.”
“Compare complete cases with incomplete cases.”
“Run a sensitivity scenario where missing outcomes are worse.”
47. AI-generated imputation needs validation
A model may fill gaps plausibly but invent unsupported certainty.
Imputed values should be marked as inferred and evaluated against appropriate statistical methods.
48. Generative models can hide the distinction between observed and imputed
If synthetic completion overwrites original blanks without provenance, future users may mistake guesses for measurements.
Observed and inferred data must remain distinguishable.
49. Parents can teach missing-data reasoning through school records
A child misses two tests.
The average of remaining tests rises.
Did performance improve?
Or were the missing tests harder?
Absence can change interpretation.
50. Small-group tuition can run missing-data puzzles
Give three incomplete datasets:
random missing readings.
missing only high temperatures.
dropout after poor performance.
Students decide which conclusions remain trustworthy.
51. A compact missing-data checklist
- How much data is missing?
- Which variables and time points are affected?
- Which groups have more missingness?
- Why did values become missing?
- Could missingness depend on the unseen value?
- Is complete-case analysis defensible?
- Would imputation preserve uncertainty?
- Does the imputation model respect variable type?
- Could weighting help?
- How much statistical power was lost?
- Could missingness bias effect size?
- What sensitivity analyses test MNAR scenarios?
- Are observed and imputed values distinguishable?
52. Frequently asked questions
Why is missing data a scientific problem?
Because missing observations can reduce precision and, when missingness is systematic, bias estimates and change which population or construct the data represents.
Is a blank the same as zero?
No. Zero is an observed value. A blank means the value is unavailable unless the data dictionary explicitly defines otherwise.
What is multiple imputation?
It creates several plausible completed datasets, analyses each one, and combines the results so uncertainty about missing values contributes to the final estimate.
Can missing data be fixed perfectly?
No. Methods can reduce damage under assumptions, but information never observed cannot always be recovered uniquely.
How does missing-data thinking help PSLE Science?
It teaches learners to notice incomplete tables, distinguish zero from unmeasured values, and avoid conclusions that silently ignore missing observations.
How does it change in Secondary Science?
Students can reason more formally about dropout, bias, imputation, uncertainty and how missingness changes statistical inference.
53. Continue the Science Education Systems series
- How Scientific Construct Validity Works
- How Scientific Measurement Reliability Works
- How Scientific Measurement Invariance Works
Conclusion: A blank cell is part of the scientific story
Maya sees the missing value.
Jia Jun asks why it vanished.
Hana tests whether the missing units differ.
Ethan asks how strong the conclusion remains under worse-case plausible scenarios.
Science needs all four.
Count the missingness.
preserve its provenance.
model it carefully.
test the assumptions.
Then let uncertainty grow honestly where evidence disappeared.
