A wearable dashboard shows a higher daily average on Tuesday than on Monday. Before calling that a change in the person’s movement, check when the device supplied usable observations. A day recorded mainly in the evening is not automatically comparable with one recorded mainly in the morning.
The short answer is that dividing by wear time does not, by itself, remove differences in which hours were observed. A wear-time average describes the observed minutes. To compare days fairly, retain their time-of-day coverage, define the intended comparison and distinguish a common-window result from an estimate about unobserved hours.
This specialist guide extends The Gold Standard of Wearable Technology. Its fictional motion-index values demonstrate data reasoning, not a product evaluation, activity recommendation or medical interpretation. The same habits can help a student inspect many kinds of incomplete sensor record.
eduKateSG · Wearable Technology
Find your next learning step
Compare wearable records using time-of-day coverage, common windows, standardised weights and explicit limits on missing-data inference.
Chapter contents
Understand the question · Chapters 1–4
Work the first cases · Chapters 5–8
Test the boundaries · Chapters 9–13
Transfer and repair · Chapters 14–18
Report and continue · Chapters 19–23
CHAPTER 1 OF 23 · Understand the question
1. Identify the number the dashboard is averaging
A wearable can report steps, acceleration summaries, activity labels or other processed outputs. These are different quantities. Before interpreting a daily average, identify the underlying variable and the time unit over which each observation was formed.
For this guide, imagine a fictional device supplying one motion-index value for each usable minute. The index has arbitrary units and no supplied health interpretation. If three usable minutes have values 1, 2 and 6, their ordinary minute-weighted average is 3 index units.
The value three does not establish an activity category or a physiological state. It is a summary of a specified processed variable over a specified set of minutes. Keeping that scope visible avoids turning a mathematical comparison into an unsupported statement about health or behaviour.
Record whether the device averages equal-duration epochs, combines unequal intervals or reports a cumulative count. A mean of interval totals can change when interval lengths change. Likewise, a mean of per-minute index values is not the same quantity as a total daily step count.
The wearable root follows the chain from physical event to sensor, processing and display. This article begins after that chain has produced a defined minute-level variable and asks what happens when the set of observed minutes changes between days.
CHAPTER 2 OF 23 · Understand the question
2. Separate usable observations, non-wear and uncertainty
A record may have usable observations, known non-wear, missing transmission, invalid measurements or uncertain wear classification. These states should not all be represented by the numerical value zero.
A valid zero is an observed output under a stated measurement method. A missing value supplies no output for that minute. A non-wear flag describes an interpretation of device use, potentially based on an algorithm or an external record. Those distinctions affect the denominator and the meaning of the result.
The research study Validation of Accelerometer Wear and Nonwear Time Classification Algorithm examines classification against an observed wearing reference. Its existence is a reminder that wear classification itself needs evidence; it is not automatically known perfectly from a displayed movement value.
Our initial calculations assume the supplied usable-minute flags are correct. Later chapters relax that assumption. This lets the first worked case isolate coverage composition without confusing it with every possible sensor error.
For a clear data table, retain a value column and a status column. An empty value accompanied by “known non-wear” means something different from an empty value accompanied by “upload pending.” A summary that drops the status history may remain numerically tidy while losing the explanation of what was observed.
CHAPTER 3 OF 23 · Understand the question
3. Define morning and evening before comparing them
Split the fictional analysis period into two predefined windows called morning and evening. Their labels are convenient teaching names, not universal research categories. A real analysis needs exact start and end times, date boundaries and an appropriate time reference.
Suppose the analysis period contains four hours in each window. A day with three observed morning hours and one observed evening hour has four hours of total coverage. A different day with one morning hour and three evening hours also has four hours.
Equal total wear time does not make their coverage patterns equal. If the variable differs systematically between the two windows, changing their observed proportions can change the daily wear-time average without any within-window change.
Write the windows before looking at which choice produces a preferred conclusion. If the boundaries are changed after inspecting the results, preserve that change and its reason. A comparison should not silently replace its question midway through the analysis.
The same issue appears with weekdays and weekends, work and leisure periods, or other predefined conditions. The labels should follow the actual research or user question. They should not be invented simply to divide the record into groups with attractive-looking averages.
CHAPTER 4 OF 23 · Understand the question
4. Work a case with no within-window change
Assume the fictional minute-level motion index averages 1 in the morning and 5 in the evening on both days. For this first exact model, every observed and unobserved minute within a given window has its stated value.
Day A records 180 morning minutes and 60 evening minutes. Its observed average is (180 × 1 + 60 × 5)/240 = 2. Day B records 60 morning minutes and 180 evening minutes. Its average is (60 × 1 + 180 × 5)/240 = 4.
| Day | Morning usable minutes | Morning index mean | Evening usable minutes | Evening index mean | Observed-minute mean |
|---|---|---|---|---|---|
| A | 180 | 1 | 60 | 5 | 2 |
| B | 60 | 1 | 180 | 5 | 4 |
The displayed average doubled, while the stipulated morning and evening profiles were unchanged. The difference came entirely from the weights attached to those windows in the observed record.
The example does not claim that a real person follows a perfectly constant profile within each window. That deliberately strong assumption makes the mechanism visible. Later examples remove it and show why adjusting two broad windows still cannot reconstruct every missing minute.
A useful first conclusion is therefore: Day B’s observed minutes had a higher index average. A broader claim about a changed complete-day profile requires more evidence.
Let m be the observed morning mean, e the observed evening mean and w the fraction of usable minutes that fall in the morning. The combined observed mean is y = wm + (1 − w)e.
If m and e are unchanged between two days, their combined difference is yB − yA = (wB − wA)(m − e). This equation isolates the contribution of changing coverage weights in that restricted model.
In the previous example, wA = 0.75 and wB = 0.25. With m = 1 and e = 5, the difference is (−0.5)(−4) = 2, matching the shift from 2 to 4.
This formula also identifies conditions under which the weight change does not affect the average. If the two window means are equal, m − e = 0. If the observed coverage fractions are equal, wB − wA = 0. Either condition makes this particular composition effect zero.
Neither condition proves that every other measurement issue has been resolved. The formula answers one defined question under one model. Use it to explain a possible mechanism, then return to the actual coverage and measurement evidence before assigning the mechanism to a real record.
CHAPTER 6 OF 23 · Work the first cases
6. Compare the same window before combining windows
A transparent first comparison places the morning means side by side and the evening means side by side. In the constant-profile example, each pair is equal. That immediately reveals that the pooled difference is not a within-window difference.
For real data, preserve the usable-minute counts alongside those means. A mean based on a few minutes has a different observational basis from one based on several hours. The difference cannot be handled merely by displaying both values to the same number of decimal places.
Also inspect where the usable minutes fall inside each window. Two morning summaries may cover different hours or different activities. Matching the broad label “morning” does not establish that all underlying conditions are comparable.
A useful table has one row per day and window, with usable duration, observed mean and relevant status notes. Keep the original minute-level record or a suitably detailed summary available for review rather than retaining only the final daily number.
The practical question is what comparison the data actually support. You may be able to describe a particular common observed period clearly while remaining unable to infer a complete-day change. That is a useful result when its boundary is explicit.
Choose a fixed reference weight q for the morning and 1 − q for the evening. A standardised two-window summary is z = qm + (1 − q)e. Using the same q for both days removes differences caused solely by their observed morning-to-evening proportions.
With equal reference weights q = 0.5, both days in the first example have z = 0.5 × 1 + 0.5 × 5 = 3. The summary answers what the two observed window means would produce under that fixed weighting rule.
It does not automatically estimate the actual complete-day average. To make that broader interpretation, one would need a justified connection between the observed window means and the unobserved minutes they are being used to represent.
Choose the reference weights for a stated reason. Equal weights may be useful for teaching or a balanced comparison of two windows. Duration-based weights may match a defined target period. Neither choice is universally correct without an intended quantity.
Label the result as a common-weight or standardised summary and state the weights. A number that looks like the original dashboard average can otherwise be mistaken for a newly measured complete day. The transformation is an analytical choice, so its meaning should travel with the result.
CHAPTER 8 OF 23 · Work the first cases
8. Work a reversal even when both windows decrease
Consider a second fictional dataset. Day A has observed morning and evening means of 2 and 8, with 180 morning and 20 evening minutes. Its observed-minute mean is (180 × 2 + 20 × 8)/200 = 2.6.
Day B has lower means in both windows: 1 in the morning and 7 in the evening. But it records 20 morning minutes and 180 evening minutes. Its observed-minute mean is (20 × 1 + 180 × 7)/200 = 6.4.
Both within-window means fell by one index unit, while the pooled observed mean rose by 3.8 units. The changing coverage distribution outweighed the within-window decreases in the pooled calculation.
Under equal reference weights, Day A’s summary is 5 and Day B’s is 4. These common-weight summaries preserve the direction of the two stipulated within-window differences in this particular example.
Do not reduce the lesson to a slogan that every pooled increase is false. The pooled result is a valid description of the observed minutes. The problem occurs when it is interpreted as a different quantity without examining the changed observational support. State the observed result, the window-specific results and the intended comparison separately.
Two-window adjustment can leave important gaps. Suppose a morning window contains a quiet hour with index 1 and a busier hour with index 7. A device worn only during the quiet hour reports a morning mean of 1. A device worn only during the busier hour reports 7.
Both records may truthfully say “one hour of morning wear.” Their broad time-of-day coverage matches, but their selected minutes differ. Treating either observed mean as representative of the entire morning requires an additional assumption.
This is why a common-weight adjustment is not a universal missing-data correction. It can address one level of composition while leaving selection inside each group unresolved.
For a classroom exercise, draw the morning as two labelled blocks and shade the observed block on each day. Ask which comparison is directly observed and which would require information about the unshaded block. The drawing keeps the missingness visible.
In a real analysis, finer time strata, matched intervals or a justified model may be considered according to the actual question and evidence. Increasing the number of strata also creates practical limits: some groups may contain few observations or no overlap. More detailed categories do not create measurements where none exist.
CHAPTER 10 OF 23 · Test the boundaries
10. Use matched observed minutes for a narrower question
Suppose two days both have usable records during a predefined 07:00–08:00 interval. Comparing that interval can be more direct than comparing daily means formed from different hours.
The resulting statement concerns the shared observed window. It does not automatically describe the rest of either day. If the source variable or wearing conditions differ, those still need attention even when the clock interval matches.
Consider a fictional common-hour mean of 3 on Day A and 2 on Day B. That supports a lower observed index during the matched interval under the supplied measurement conditions. It does not establish that Day B’s complete-day total or average was lower.
Preserve how the shared interval was chosen. Selecting whichever common hour happens to show the strongest difference after examining many possibilities changes the interpretation and can exaggerate the apparent evidence.
A useful report distinguishes a planned common-window comparison from an exploratory one. Both can be informative, but they should not be presented as if they had identical evidential roles. The immediate goal is clarity about the observed support, not a claim that one simple matching rule solves every data-quality problem.
CHAPTER 11 OF 23 · Test the boundaries
11. Keep complete-period totals separate from observed means
A mean over usable minutes and a total over an entire period answer different questions. Even if a common-weight mean can be constructed, it does not provide a measured total for the missing minutes.
Suppose a fictional counter records 3000 events during six usable hours of an eight-hour target period. The recorded total is 3000. The observed rate is 500 events per usable hour. Multiplying that rate by eight gives 4000 only under an explicit assumption that the missing two hours had the same average event rate.
The actual missing period could have a different rate. The extrapolated 4000 is therefore a model-based quantity, not an additional observation. Do not silently replace the recorded total with it.
For the arbitrary motion index used elsewhere in this article, summing index values also needs a definition. Not every processed variable has a meaningful cumulative interpretation. An average of an index may be useful while its sum has no agreed practical meaning.
The first repair is to identify the target: observed-period average, common-window comparison, complete-period event total or another specified quantity. Then choose a method that addresses that target and state where assumptions enter. A dashboard label such as “daily activity” is too broad to resolve these decisions on its own.
CHAPTER 12 OF 23 · Test the boundaries
12. Bound a missing contribution only when a bound is justified
Suppose the target is the mean of eight equally weighted hourly index values. Six are observed with total 18 index-hours in this teaching convention. Two are missing. If each missing hourly value is independently known to lie between zero and ten, the full-period mean lies between 18/8 = 2.25 and (18 + 20)/8 = 4.75.
These are conditional bounds based on the supplied range. They do not establish that zero or ten occurred, and the midpoint 3.5 is not automatically the best estimate.
If no justified upper bound is available, do not invent one because the resulting interval looks useful. A device’s displayed scale limit may also differ from a bound on the underlying quantity the analysis intends to discuss.
The role of this calculation is to make the missing information explicit. A comparison with another day may remain unresolved when their possible full-period values overlap. That uncertainty is part of the result, not a defect to hide by choosing a convenient fill value.
Keep this bounding approach separate from time-of-day standardisation. One constructs a range under stated limits; the other applies a fixed weighting rule to available window means. Neither is automatically interchangeable with an estimate from a validated missing-data model.
Imputation assigns plausible values to missing observations using a stated method. It can be useful in a suitable analysis, but the assigned values remain estimates rather than measurements made by the device during non-wear.
Research on simulation-based handling of accelerometer non-wear examines methods under defined conditions. The important reading habit is to ask which assumptions connect observed information to missing values and how uncertainty in those estimates is represented.
In a fictional example, an analyst fills missing morning minutes with the observed morning mean. That assumes the observed morning minutes adequately represent the missing ones for the intended calculation. If the device was removed during a distinctive activity, that assumption may fail.
Another analyst might use comparable times on other days. That adds assumptions about similarity across days. The resulting estimate can be reasonable under a justified model, but the choice should remain visible and should not be described as if the missing minutes had been directly recorded.
Keep an indicator for each imputed value and retain the original missingness pattern. A later reviewer should be able to distinguish observed-only summaries from model-completed summaries. Replacing the original record permanently with filled values makes that distinction harder to recover.
CHAPTER 14 OF 23 · Transfer and repair
14. Investigate wear classification before interpreting quiet periods
A long run of small or zero movement values can reflect several situations. The device may be worn during little movement, removed and left still, or affected by a recording or processing issue. The available channels and validation evidence determine how well those possibilities can be distinguished.
The wear and non-wear classification validation study provides a primary research example of testing a classification method against an observed reference. It does not establish one universal rule for every placement, device or population.
For our arithmetic exercises, usable-minute status was initially supplied as correct. If that status is uncertain, both the numerator and denominator of the observed mean can change when classification changes.
Imagine ten minutes of genuine low-index wear incorrectly removed as non-wear. The remaining observed mean may rise because low values were excluded. Conversely, including a stationary unworn device as valid wear can pull the mean down. Neither direction is guaranteed without examining the actual values and classification errors.
An honest review therefore treats coverage as an evidence product, not a perfect label inherited from software. Record the method and version, relevant auxiliary information and any adjudication rules. Do not infer sleep, a medical condition or a person’s intentions merely from an unexplained quiet interval.
CHAPTER 15 OF 23 · Transfer and repair
15. Align clock windows and dates consistently
The words morning and evening require a time reference. A record may use local clock time, coordinated universal time or a device clock that has been changed. Comparing windows without checking those conventions can align different parts of the day.
Suppose one export labels an observation 08:00 in local time while another labels the same instant in a different reference. Treating both displayed labels as identical local hours can misclassify observations into different windows.
This guide does not prescribe timezone software or a specific jurisdiction’s clock rules. The general requirement is to retain enough timestamp information to establish which actual intervals the records represent.
Date boundaries matter too. A summary called Tuesday might cover midnight to midnight, a waking period spanning two dates or a rolling twenty-four-hour window. Those are different supports. The daily label alone does not identify them.
Before calculating a time-of-day comparison, document the target interval, timestamp convention and treatment of any clock changes. If the record does not contain enough information to align the windows, preserve that uncertainty rather than shifting data into a convenient pattern. Correct arithmetic on misaligned intervals still answers the wrong comparison.
CHAPTER 16 OF 23 · Transfer and repair
16. Do not let a valid-day label replace the coverage table
A study or application may use a minimum usable-duration rule to decide which days enter a particular analysis. Such a rule can support a defined workflow, but passing it does not establish identical time-of-day coverage or suitability for every question.
Two fictional days might both exceed a four-hour teaching threshold while one covers only morning and the other only evening. The label valid would not resolve their composition difference. The threshold itself is an invented classroom rule here, not a recommended standard.
The primary study Quality Control Methods in Accelerometer Data Processing: Defining Minimum Wear Time illustrates why minimum-duration decisions belong to an analysis context. Do not transplant a number from one protocol into a different device or question without justification.
Excluding days can also change which people or conditions remain represented. If missingness is associated with the activity or circumstances of interest, a clean-looking subset may differ systematically from the target population or period.
Retain inclusion decisions and their reasons. A report can state which days met its rule while still presenting the coverage distribution relevant to the comparison. A binary quality label is useful for organisation; it is not a substitute for the evidence needed to interpret a specific average.
CHAPTER 17 OF 23 · Transfer and repair
17. Build a record that preserves the comparison
Keep one row per defined epoch or interval, with timestamp, duration, value, status and relevant device or processing version. If a summary is supplied instead, retain enough grouped counts and totals to reconstruct its denominator.
Create a separate coverage table by day and predefined window. Include usable duration, missing duration, observed mean and whether the window contains any imputed observations. The table should distinguish no usable observations from a valid observed mean of zero.
For equal-duration epochs, the pooled mean can be reconstructed from the group mean multiplied by the group count, divided by the total count. If intervals have unequal duration and the target is a duration-weighted mean, use durations instead.
Record which values are excluded and why. A missing upload should not quietly become a measured zero, while a rejected measurement should not reappear later through an undocumented default fill.
Finally, keep a plain-language target statement beside the calculation. For example: “Compare the recorded index during the common observed morning window.” That sentence makes it harder for a later chart title to expand the result into a claim about complete-day activity without additional evidence.
CHAPTER 18 OF 23 · Transfer and repair
18. Test the summary against alternative missing profiles
Suppose the observed portion of a day is fixed. Construct two possible missing profiles: one in which the unobserved evening is quiet and another in which it is relatively active. Both profiles can agree perfectly with every observed minute.
If the two completed profiles give different complete-day conclusions, the observed data alone do not identify that conclusion. This is an identifiability check, independent of how precisely the observed mean was calculated.
For a small example, four observed hourly values sum to 12 and two hours are missing. Filling those hours with 1 and 1 gives a six-hour mean of 14/6. Filling them with 7 and 7 gives 26/6. The recorded four-hour mean remains 3 in both stories.
The exercise does not endorse either completion. It demonstrates what has been left unresolved. Additional observations or a justified model are needed to prefer one over the other.
This is a useful learner task because it reverses the usual direction of calculation. Rather than finding one answer from a complete table, students ask how many unseen situations could produce the same incomplete table. Recognising that multiplicity is essential to reading sensor summaries honestly.
Day C has morning mean 3 and evening mean 9, based on 150 and 50 usable minutes. Its observed mean is (150 × 3 + 50 × 9)/200 = 4.5. Day D has morning mean 4 and evening mean 8, based on 50 and 150 minutes. Its observed mean is 7.
Under equal reference weights, both common-weight summaries are 6. The pooled observed means differ, while the specified balanced comparison does not. Within windows, the morning increased by one and the evening decreased by one.
Now choose reference weights 0.75 for morning and 0.25 for evening. The summaries become 4.5 for C and 5 for D. The target weighting changed, so the resulting comparison changed as well.
Neither fixed weighting is automatically the true complete-day answer. The appropriate choice depends on the stated question and the assumptions connecting observed window means to the intended target.
For a final check, ask whether equal total wear minutes were enough to make the original daily means directly comparable as complete-day behaviour. They were not. Explain the different window weights, the mixed within-window changes and the unresolved missing minutes separately.
“The daily average doubled, so movement doubled.” The average may describe a different set of observed hours. State the measured variable and inspect coverage before extending the comparison.
“We divided by wear time, so missingness is corrected.” Division defines an observed-period average. It does not automatically make the observed minutes representative of missing minutes or another day’s hours.
“Both days are valid, so their coverage is equivalent.” A duration threshold can admit different time-of-day patterns. Preserve the coverage table and the target-specific quality checks.
“A filled record is a fully observed record.” Imputed values depend on a method and assumptions. Retain their status and distinguish measured information from estimated completion.
For each repair, identify one useful next check: the exact interval, window-specific means, wear classification, common support or the imputation model. A good correction should help the reader proceed, not merely replace confidence with vague doubt.
CHAPTER 21 OF 23 · Report and continue
21. Write a conclusion with the observation window attached
A useful conclusion might say: “The mean index over usable minutes was higher on Day B, but its recorded minutes were concentrated more heavily in the evening. The two window-specific means were unchanged in this example.” Each part can be checked against the table.
If a common-weight analysis was performed, add its target weights and state that it is an analytical summary. If complete-day activity remains unknown, preserve that boundary. This wording gives the reader a result and a reason, while avoiding an unsupported judgement about the person or the missing hours.
Is a wear-time average wrong?
Not when it is correctly calculated and described as an average over usable observations. The problem is interpreting it as a different quantity without the assumptions or evidence that broader interpretation requires.
Does equal wear time solve the problem?
No. Equal totals can cover different hours. The distribution of usable time matters when the measured variable differs across those hours.
Can standardisation recover missing activity?
It applies common weights to available group means. Interpreting that as a complete-period estimate requires additional assumptions about representation within groups. It does not directly observe the missing minutes.
Do these examples establish a health judgement?
No. The fictional index has arbitrary units. The calculations concern data coverage and comparison, not diagnosis, activity targets or the accuracy of a particular wearable product.
CHAPTER 23 OF 23 · Report and continue
23. Continue through measurement and evidence
Return to Wearable Technology for the full sensor-to-display system and its grouped engineering pathways. Use Sensor Fusion Engineering for combining uncertain measurements, and Wireless Communication when the missing record concerns delivery rather than wearing.
Primary research routes include wear/non-wear classification validation, minimum wear-time quality control and simulation-based evaluation of non-wear handling. Their methods have particular study contexts; the original arithmetic cases here are not claims about their participants or results.
The wider route is Learning That Transfers into Life within How X Works. Take two incomplete days, draw their observed windows and write the narrowest comparison the evidence supports. That is a useful next step before interpreting the dashboard’s largest number.