An optical PM2.5 result is an inference, not a weighed sample
An optical fine-particle sensor illuminates air in a small chamber and detects light scattered by suspended particles. A model then translates those signals into a reported estimate of PM2.5 mass concentration, normally shown as micrograms per cubic metre. The sensor has not directly weighed a filter. Its estimate depends on particle optics, instrument geometry, airflow, firmware and calibration. Begin interpretation by identifying the original measurement, the conversion model and the conditions for which that model was evaluated.
Relative humidity matters because some particles absorb water, grow in optical diameter and scatter more light without a proportional increase in dry particle mass. But a humid day can also bring genuine changes in aerosol formation and accumulation. A large optical reading during humid weather is not automatically false; the crucial question is how closely the device agrees with an appropriate independent method under comparable conditions.
Collocation means deliberate comparison with a reference monitor sampling sufficiently similar air. It helps distinguish systematic bias, scatter, repeatability and changing performance over seasons. A field correction may improve a sensor’s usefulness for a defined project, but it does not transform a consumer unit into a regulator-designated reference method. The worked numbers here are constructed examples and should not be mistaken for real observed air-quality records.
The hidden meaning of the PM label
PM2.5 refers to a mass fraction defined through aerodynamic size selection, not a claim that every airborne particle is exactly 2.5 micrometres wide. Optical size is also not automatically aerodynamic diameter. Density, shape and optical refractive index influence how different particle types appear to a scattering detector. An irregular soot aggregate, a mineral fragment and a liquid droplet of similar optical intensity can have different dry masses.
Some devices estimate number in size classes before using an assumed density to infer mass. Others treat scattering as a bulk optical signal. Their PM1, PM2.5 and PM10 channels may be derived from one set of optical events. Do not treat these displayed channels as three independently weighed samples, or subtract channels without confirming the device’s definition and internal algorithm.
The mass-number distinction is essential: ten thousand tiny particles may contribute less mass than a much smaller number of relatively large fine particles. A jump in mass concentration does not prove the same percentage jump in particle count. Read the axis units, size definitions and manufacturer’s method before telling a story about what changed in the atmosphere.
Why scattering changes with aerosol composition
Scattered-light intensity depends on particle size relative to the illuminating wavelength, shape, refractive index, absorption and detector angle. The relationship is not a simple linear conversion from particle brightness to mass. A factory calibration aerosol can provide an orderly standard in a test chamber, yet the outdoor air may contain mixtures of smoke, marine salt, mineral dust and chemically formed fine particles that behave differently.
A sensor calibrated on dry particles may overestimate mass for water-swollen particles in one weather regime and underestimate optically absorbing soot in another. The exact direction depends on instrument design and composition; there is no reliable universal sign of bias. A high correlation with a reference across common conditions can hide poor response to a rare smoke plume or winter mixture.
When a project compares sites, collect context about likely source mixtures and meteorology, but avoid claiming that the optical detector identified chemical compounds. A scattering signal can suggest changes and support spatial screening; direct chemical identification requires other methods. Explaining the boundary of observation is a strength, not a weakness.
Relative humidity: distinguish aerosol response from environmental coincidence
Relative humidity is a ratio involving water vapour and saturation at the relevant air temperature. It is not itself the mass of fine particulate matter. Many atmospheric salts are hygroscopic and take up water at conditions governed by composition and chemical history. Water uptake increases particle dimensions and can increase scattered light, while optical sensors often translate intensity using a dry-mass relationship.
The sensor chamber may be warmer than the outdoor air. A weather-station RH value does not guarantee that every particle inside the measurement chamber has the same moisture state. Fan heat, enclosure heating, residence time, condensation and dryer configurations can change sampling conditions. Check what environmental sensors report and where those sensors are located.
To test whether humidity influences error, compare matched optical and reference values across humidity ranges. Plot the difference against RH while also examining concentration, time of day and source changes. A correlation between humidity and high PM does not establish instrument artefact: stagnation and secondary aerosol production can generate real concentration changes under the same weather pattern.
Collocation is an experiment in comparability
Put the optical units and reference in a position where their inlets see similar outdoor air. Record inlet height, obstruction, sun and rain exposure, nearby exhausts and any protective enclosure. A unit under a roof or next to a vent may sample a different microenvironment from an open-air reference, making the difference partly geographical rather than instrumental.
Use replicate units where possible. If three identical sensors agree with each other but consistently exceed a reference, investigate shared calibration or environmental response. If one sensor diverges while two agree, inspect that unit’s fan, inlet, optics and software. Similar outputs among replicas demonstrate repeatability, not necessarily absolute accuracy.
EPA’s testing guidance for non-regulatory supplemental and informational monitoring describes collocation evaluation, recommended field periods and replicates. Read the source protocol for the relevant purpose rather than borrowing a convenient distance or duration as a universal certification test. Document decisions so that another analyst can reconstruct which instruments and conditions produced the findings.
Clock alignment can produce a fake disagreement
An optical sensor may report minute values while a BAM reports hourly results. An hourly value could refer to the hour beginning or the hour ending at the stated time. A cloud app might timestamp data when uploaded rather than when observed. Matching the wrong 60-minute interval can create a false sensor lag, especially during rapidly changing smoke or traffic events.
For an hour with twelve equally long valid five-minute observations, take their arithmetic mean to compare with a matched hourly reference when that aggregation is appropriate. If readings represent unequal time intervals, weighting by duration may be needed. Do not pad missing minutes with zero: zero is a measured concentration, whereas missing means no valid observation was available.
Keep time zones and daylight-saving conventions explicit. An outdoor event can look two hours late if one file uses UTC and the other local time. Before fitting a correction, graph the two raw series with matching time boundaries. A surprisingly weak regression sometimes disappears once the clocks and averaging windows are repaired.
Numerical example: a one-hour optical difference
Suppose a matched, quality-accepted reference hour reports 24 µg/m³ and the optical unit reports 31 µg/m³. The optical-minus-reference difference is +7 µg/m³. Divide by the reference to describe relative error: 7/24 × 100 = approximately 29.2 percent. Using 31 as denominator instead gives 22.6 percent, a different statistic. Always state the denominator in prose.
One data pair cannot establish whether the discrepancy arose from humidity, calibration, composition, siting or timing. Subtracting seven from every optical observation would make this single point fit perfectly but could produce negative values during clean hours. A valid correction needs multiple matched observations across the range relevant to the study and should be assessed on independent data.
An important inverse check is to recompute the absolute difference from the reported percentage: 0.292 × 24 ≈ 7.0 µg/m³. This simple habit protects meaning when reporting rounding, percentages or units. The result concerns agreement between instruments during one matched hour; it is not a measure of personal inhaled dose.
Numerical example: bias and error are not the same
Consider four hypothetical reference readings, 5, 10, 20 and 40 µg/m³, alongside optical readings, 8, 14, 26 and 50. Paired differences are +3, +4, +6 and +10 µg/m³. Mean signed bias is (3+4+6+10)/4 = +5.75 µg/m³. All four readings exceed the reference, so the example reveals systematic overreading, not merely random error.
Relative errors using the reference denominator are 60, 40, 30 and 25 percent. Percentage error decreases even as absolute error increases. A plot of differences against reference concentration reveals information that a single mean statistic hides. A high correlation would simply show that the instruments rise together, not that the optical values equal the reference.
Now change the paired residuals to +6, +6, −6 and −6. Mean signed bias is zero, yet mean absolute error and root-mean-square error are both 6 µg/m³. Reporting zero bias as proof of accuracy would be misleading because positive and negative errors cancel. The metric should answer the intended question, not merely flatter the device.
A correction is a model with boundaries
A common empirical relation is corrected PM = a + b × raw optical PM, where a is an offset and b a scale factor estimated from collocation. Record the fitted dataset, site, period, aerosol conditions, humidity range, reference method and model version. An equation becomes meaningful only when the conditions of its use and independent performance are documented.
For illustration, suppose a locally fitted model is Ccorrected = 0.80 Craw + 1.5 µg/m³. A raw value of 30 gives 25.5 µg/m³. Applying it to 3 gives 3.9 µg/m³. These computations are algebraically correct, but do not prove that the coefficients apply to a future wildfire, wet sea-salt episode or a sensor with new firmware.
Separate fitting from validation. Testing only on the data used to estimate coefficients gives overoptimistic performance. Hold out periods or sites, with a defensible strategy for temporal dependence. Randomly splitting adjacent minute readings can leak almost identical air and weather conditions across training and test groups. A correction that works on a new season provides more convincing evidence than one that only reproduces last week’s pattern.
Correction diagnostics that catch failure
Plot corrected-minus-reference residuals against relative humidity, temperature, reference concentration and date. If a correction is systematically positive at high humidity and negative during dry smoke, a good overall average may conceal two important failure regimes. Report performance by conditions that matter to the intended decision rather than relying on one all-period summary.
Do not silently truncate negative model predictions to zero or cap high predictions at an arbitrary maximum. Such transformations change the statistical properties of the output and can bias averages. If clipping is required by a published application rule, document it and quantify its effect. Preserve raw and corrected fields separately, with correction version and processing time.
A linear relation may be inadequate when optical response varies nonlinearly or saturates, but a more complicated model can overfit sparse data. Complexity is not a substitute for independent evaluation. Prefer a transparent correction that holds up under realistic new conditions over an elaborate equation with a beautiful training fit.
Meaningful site-level and regional conclusions
A properly maintained optical network can help reveal temporal patterns and local gradients. A high reading next to a delivery yard may be a real local exposure signal without representing an entire district. A monitoring site should therefore be interpreted in relation to siting objective, nearby sources, height and airflow. A sensor map does not prove that every point between instruments has a smoothly interpolated concentration.
Do not treat an indoor sensor and an outdoor reference as interchangeable. Cooking, cleaning, filtration, open windows and ventilation can alter indoor aerosol populations substantially. If the question concerns indoor conditions, define that question and validate the sensor appropriately; do not force an outdoor regulatory comparison onto an indoor measurement without considering the different air masses.
Similarly, an optical PM estimate cannot identify whether a factory, vehicle, sea-salt aerosol or a distant fire caused the observed signal. Source attribution requires more information than an elevated mass estimate. Write what was measured, which hypotheses fit, and what additional evidence would distinguish those hypotheses.
Practical interpretation of missing, stuck and suspect data
A run of identical displayed values could indicate very stable air, but it may also reflect a frozen output or communications problem. Repeated zeros during obvious smoke deserve checking for blockage, sensor saturation or software suppression; however, unusual-looking values are not alone sufficient grounds for deletion. Use quality logs and independent evidence.
Humidity sensors can fail or be mounted in a position that does not represent the optical chamber. Fans and inlets accumulate dust or insects, and firmware updates may alter reporting without obvious changes to display design. A sudden long-term change in bias after service should trigger documentation and, where warranted, fresh collocation.
Retain a provenance trail: raw output, sample time, processing time, firmware version, calibration model, acceptance decision and reason for changes. Data can be reprocessed as science improves, but readers must be able to tell a revised estimate from a freshly observed concentration.
Transfer exercise: aggregate without inventing clean air
A device records twelve five-minute values forming an arithmetic sequence from 10 to 32 µg/m³ in steps of two. For equal intervals the full-hour mean is (10+32)/2 = 21 µg/m³. If the last three samples are missing, the nine available values run from 10 to 26 with mean 18 µg/m³. Do not pretend 18 is necessarily the true full-hour mean.
If the missing three samples are replaced by zero, the reported value becomes 13.5 µg/m³, which is an artificial underestimate under this constructed trend. A student should report nine valid values, the observed-partial-period average and the missing interval. Whether enough data exist for a valid hourly result depends on the project’s published completeness rule.
Now compare with a reference valid hour at 20 µg/m³. The apparent agreement of the full twelve-point mean is much stronger than the nine-point partial average, but neither comparison justifies falsifying missing readings. Data quality is part of the interpretation, not a cosmetic note under the graph.
Transfer exercise: test a correction where it matters
Suppose Ccorrected = 0.75 Craw + 2. For raw sensor values 8, 24 and 64 µg/m³, the corrected estimates are 8, 20 and 50. Independent reference readings are 7, 22 and 58, giving residuals +1, −2 and −8 µg/m³. Mean bias is (1−2−8)/3 = −3 µg/m³ and mean absolute error is (1+2+8)/3 ≈ 3.67 µg/m³.
The model underestimates the highest point, which could be important if the project aims to detect severe episodes. A correction can improve average results and still fail precisely where decisions are most sensitive. Evaluate error conditional on high-end concentration as well as the overall mean. Three observations cannot establish a robust model; their value here is to teach calculation and the need for independent testing.
Questions to ask before sharing a graph
Does the graph show raw optical output or a corrected estimate? Are units and time windows displayed? How many observations were missing or invalid? Was reference collocation actually performed, or are two consumer sensors simply compared? Do data cover the humidity and concentration conditions being discussed? Is a citywide claim being made from a single locally influenced site?
The reader should leave knowing how a particle-derived optical signal becomes a mass estimate and why that estimate needs context. Responsible interpretation is neither unconditional trust nor dismissal of every low-cost device; it is using the right evidence for the right question.
The three layers of uncertainty in an optical estimate
Instrument uncertainty is only the first layer. Photodetector noise, fan fluctuations, optical contamination and electronic drift affect the raw signal. Conversion uncertainty is the second layer: assumptions about particle size, density and refractive index determine how a scattering observation becomes a mass estimate. Representativeness is the third layer: a precisely functioning sensor may correctly describe its chamber’s sampled air while still failing to represent an entire neighbourhood.
These layers need different evidence. For raw-signal problems, examine maintenance and diagnostics. For model problems, collocate across mixtures and humidity conditions. For representativeness, check siting and compare spatial patterns across independent locations. If a public report merely states ‘the instrument has 5 percent uncertainty’ without saying what component was tested, the number might not describe the real uncertainty of the derived concentration.
An independent evaluation may report precision, bias and error under specific controlled conditions. Those results are informative but not interchangeable with uncertainty at every outdoor site. State the test aerosol, humidity range and concentration range. If a device is deployed in a salt-laden coastal location after testing mostly in dry laboratory smoke, explain why extrapolation matters.
A worked example of fixed and proportional bias
Consider two hypothetical sensors tested at reference concentrations of 10, 20 and 40 µg/m³. Sensor A reports 15, 25 and 45: it appears to have a roughly fixed +5 µg/m³ offset. Sensor B reports 12, 24 and 48: it appears to have a proportional 20-percent overreading. If a simple correction is justified on independent observations, A might use raw minus five while B might use raw divided by 1.2. These are pedagogic patterns, not recommended coefficients.
At a reference concentration near 2 µg/m³, a fixed −5 correction would make a raw Sensor A value of 7 become 2, but a raw value of 3 would become −2. That is a warning that calibration relationships may not remain valid at very low concentration. Sensor B’s proportional correction keeps zero at zero but may misrepresent baseline noise or detection limits. Neither model should be adopted without testing outside the three calibration pairs.
Ask which errors matter to the research question. A fixed offset can substantially distort a clean-air baseline; proportional bias can distort extreme episodes. A graph that simply overlays two similar time series is less revealing than plotting residuals against reference concentration. Good interpretation explains which error structure the data suggest and how to verify it.
Choosing useful validation statistics
Mean bias tells you the average direction of error. Mean absolute error answers how far readings are from the reference without positive and negative cancellation. Root-mean-square error penalises large outliers more strongly. Correlation measures how closely changes track, but a perfect correlation can coexist with a serious fixed offset. Slopes and intercepts describe calibration tendencies, while replicate spread gives evidence about device-to-device precision.
Suppose residuals are −2, −1, +1 and +2 µg/m³. Mean bias is zero. Mean absolute error is 1.5 µg/m³ and RMSE is the square root of (4+1+1+4)/4, or approximately 1.58 µg/m³. Compare this with residuals of −6, +6, 0 and 0: mean bias is again zero, but MAE is 3 and RMSE is about 4.24 µg/m³. A zero signed average alone fails to distinguish these very different outcomes.
Confidence intervals, bootstrap summaries or repeat evaluations can express variability when supported by enough independent data. Do not attach narrow statistical uncertainty to results with only a handful of strongly correlated adjacent minutes. Effective independence matters. A transparent report states sample counts and how periods were chosen, so readers know whether the reported statistics are robust or exploratory.
A humidity-binned teaching investigation
Divide a constructed collocation dataset into low, moderate and high relative-humidity conditions. Suppose there are equal numbers of valid matched pairs and the mean optical-minus-reference differences are +2, +4 and +12 µg/m³. This pattern is consistent with increasing high-humidity bias, but not sufficient to prove a causal water-growth mechanism. High-humidity hours might also coincide with a different source mixture or unusually high actual PM concentration.
Improve the investigation by comparing errors at similar reference PM concentration but different humidity, if the dataset has enough overlap. Examine particle composition proxies when available and check whether the effect persists after a firmware or housing change. A stratified evaluation can reveal important structure that disappears in a simple all-hour regression. It still cannot isolate every confounder without a carefully planned experiment.
Now imagine the high-humidity bin includes three readings after condensation was observed on the optical window. That may be a separate failure mechanism from gradual hygroscopic particle growth. A correction developed for one should not be assumed to repair the other. Record condensation events as distinct quality concerns, and avoid masking their effect with a general humidity regression.
Why response time deserves its own test
Optical sensors can produce frequent readings, yet fast reporting does not guarantee immediate, accurate response to every concentration change. The chamber has airflow and residence time, firmware may smooth multiple samples, and network uploads can add delay. A reference reported as an hourly average cannot resolve every sub-minute plume. Performance should be judged at matching timescales.
In a controlled teaching example, suppose a source pulse lasts five minutes at 100 µg/m³ while the next 55 minutes remain at 10 µg/m³. The true time-weighted one-hour concentration is (5×100+55×10)/60 = 17.5 µg/m³. A sensor that reports 100 for the peak is not necessarily disagreeing with a validated hourly reference of 17.5; they describe different averaging windows. Conversely, if the sensor’s own five-minute estimate remains at ten, a possible response or timing problem deserves investigation.
Plot high-frequency raw values and the exact rolling or fixed-window aggregation. Check whether a moving average is centred, trailing or otherwise defined. Comparing a peak to a mean without explaining this difference can create a false alarm about accuracy or, worse, conceal a real short-lived event.
A complete example of an independent holdout
A community project collocates sensors for four weeks, fits correction coefficients using the first three weeks and reserves the last week for evaluation. That is a useful first separation, but the held-out week may share much of the same weather and aerosol composition. State its limits. A stronger transportability question might reserve a later season or a different site, provided the new environment is comparable enough to address the intended application.
Suppose raw holdout absolute errors average 8 µg/m³ and corrected errors average 5 µg/m³. That indicates improvement in that holdout, but it does not prove universal accuracy. Inspect the error on the highest-concentration days, low-end noise, completeness and humidity dependence. A reduction in one mean error can coexist with deteriorating performance for episodes that matter most to the reader.
Do not retune coefficients after viewing the test results and then continue calling the same period independent. If changes are made based on holdout performance, create a new untouched evaluation period for the revised model. This distinction between choosing a model and verifying it is foundational to reliable data translation.
Questions that show the difference between a map and a sample
Suppose two sensors 500 metres apart report 17 and 42 µg/m³ during a busy hour. The average, 29.5 µg/m³, is a valid arithmetic midpoint of two values. It is not a measured concentration at the street halfway between them. A third location may be sheltered, traffic-filled or affected by different wind flow. Spatial interpolation requires assumptions and evidence about geography, not just the mean of two displays.
Now imagine the second sensor is next to an idling bus. That does not automatically make it faulty: it may be accurately recording an intended hotspot. For estimating regional background, however, that placement may be unsuitable. Define the goal before interpreting disagreement. The same data can be useful for local-source screening and poor for producing a regional average.
Explain this in public language by naming each station’s context and using phrases such as ‘measured at this location’ and ‘estimated between stations’. Do not draw a smooth map that hides missing coverage or silently assigns a citywide meaning to a single monitor.
Why a sensor with few missing data can still be misleading
High availability is attractive, but continuous output can be continuously wrong. A clogged inlet, stale firmware or sensor whose fan has stopped may keep transmitting plausible-looking numbers. A different device may have several documented maintenance gaps but high-quality valid measurements when operational. Data completeness and measurement validity must be reported as separate dimensions.
Consider two invented devices over 100 hours. Device A reports 100 values, but an independent review invalidates 15 because of a persistent fan fault. Device B reports 92 values, all of which pass the project checks. If the target is valid hours, A has 85 accepted readings while B has 92. Simply awarding A ‘100-percent uptime’ would invert the quality comparison.
Be explicit about denominators: 85 of 100 scheduled hours is 85-percent validated coverage, while 92 of 100 is 92 percent. Do not report the percentage of delivered messages as if it were the percentage of trustworthy observations.
An instrument update can impersonate an environmental trend
A firmware change might revise the PM mass conversion, internal averaging or low-end treatment. If a multi-year dataset joins measurements before and after the update, a discontinuity could reflect a software transition rather than a genuine step in pollution. Preserve software version history alongside time series. Where possible, conduct a bridge comparison with an independent reference through the change.
An analyst should test whether mean residual and residual-versus-humidity patterns change at the update date, without assuming every observed shift was caused by software. Weather, sources and instrument wear can also change. A controlled check compares periods under similar conditions where feasible and records all maintenance actions in the vicinity of the transition.
For cross-study sharing, identify whether published values are raw, vendor-corrected, locally corrected or reprocessed. A link to a dashboard is not enough if values can change later without versioning. Keep exported datasets with retrieval dates, configuration descriptions and an explanation of the applied model.
Learning to write a measured conclusion
An accurate report can say: ‘During the 16:00–17:00 matched interval, the optical sensor reported 31 µg/m³ and the collocated reference recorded 24 µg/m³, a +7 µg/m³ difference. The cause remains uncertain; humidity, aerosol properties, processing and siting were considered.’ This sentence provides a quantity, period, method comparison and qualified interpretation. It is more informative than saying simply that the sensor was wrong.
A second sentence can explain what evidence would improve the claim: performance across matched intervals and conditions, device diagnostics, firmware history and independent testing. For a lower-cost monitoring project, the goal is not to bury uncertainty under caveats but to make the evidence sufficient for its intended use.
The transferable lesson is the same as in science reading and mathematical problem solving: identify what is observed, identify what is inferred, check denominators and compare alternative explanations. Technical vocabulary becomes a practical tool when it helps people avoid treating an estimate as an observation or an association as a cause.
FAQ: the problems readers most often meet
Why does my sensor jump when a shower begins? Humidity, droplets and real aerosol changes are possible; check chamber conditions, location and independent measurements before attributing the cause. Does a laser count PM2.5 mass directly? No; it detects scattering and applies a model. Are negative corrected numbers physical? No; they can occur from statistical models and need documented treatment rather than silent editing.
Can a sensor be useful without regulatory designation? Yes, for applications whose quality needs it meets, such as some educational or supplementary monitoring. Can a very high correlation establish agreement? No; fixed and proportional bias can coexist with correlation. Why are three replicate units useful? They help reveal device-to-device variation and isolated faults.
How often should I collocate? Match the frequency to the project’s risk, duration, environmental changes, maintenance and relevant guidance; do not invent a universal interval. What should I publish alongside a corrected number? Model and firmware versions, reference method, training and test periods, sample counts, environmental range, evaluation errors and missing-data handling.
What to do when apparent improvement contradicts measurements
Suppose a school installs a sensor and then changes filtration. The optical display falls from 35 to 18 µg/m³. That is encouraging but not yet a complete efficacy calculation. Was the before-and-after interval matched for outdoor pollution, ventilation, occupancy, cooking, cleaning and humidity? Did software update at the same time? A sensor might register a real improvement, but the inference that filtration alone caused the full reduction needs a comparison design.
A stronger test could use a nearby outdoor record, indoor baseline, matched occupancy conditions and repeated intervention periods, recording fan settings and door activity. Report the within-room observed decline separately from an estimate of filtration’s causal effect. If an independent reference or gravimetric sampling is feasible, it could help assess optical bias during the study. This method of separating observation and attribution generalises beyond particulate matter to almost every educational field experiment.
Consider the percentage decline: (35−18)/35 × 100 ≈ 48.6 percent of the earlier optical reading. It is not automatically a 48.6-percent reduction in personal inhaled dose or long-term health risk. The same percentage describes only two reported concentrations under stated conditions. Clear nouns protect interpretation: a ‘reading’ decreased; the extent of true dry aerosol mass decrease and its cause remain to be established.
Detecting model leakage with simple thought experiments
An optical correction can appear remarkably effective if it receives information it would not possess in real use. For example, a model might use the reference monitor’s current value as an input while claiming to predict that reference from the low-cost sensor. In a collocation experiment that would be circular. Similarly, adjusting coefficients after examining every validation case makes the announced independent error estimate unreliable.
One useful audit is to draw a timeline: which data existed when the correction was fitted, which observations were excluded from fitting, and what information will be available when the sensor is deployed alone? Test the model under those actual constraints. If a correction requires future reference readings, it may still be useful for retrospective analysis, but should not be advertised as a real-time independent prediction.
When a model’s measured performance is suspiciously perfect, search for leakage before celebrating. Adjacent minutes, duplicated timestamps, pre-corrected reference outputs, circular humidity features or accidentally repeated test days can make a weak method appear strong. A reproducible process should expose these dependencies for a second analyst to verify.
The compact field notebook for independent replication
A useful optical-sensor field notebook records each device’s serial number, model, firmware, inlet position, power supply, protective housing, time setting, averaging interval, raw output units, humidity and temperature channels, collocated reference identifier, maintenance dates and any correction formula. It also records why a reading was marked valid or invalid. This is practical evidence, not bureaucracy added after the analysis is finished.
Suppose two students obtain slightly different results from the same exported file. If one averaged minute readings by clock hour and the other used trailing rolling hours, they may both have applied correct arithmetic to different intervals. Their notebooks reveal the real disagreement. If a reference dataset is later validated and revised, the original retrieval date and data version allow the comparison to be repeated fairly.
The overarching rule is to preserve transformations: raw measurement, flagged dataset, aggregated dataset, corrected estimate and public chart should be separately identifiable. That chain lets someone who did not conduct the original project check the conclusions and find genuine improvements rather than merely trusting a polished graphic.
Primary sources and verified related learning
US EPA’s Enhanced Air Sensor Guidebook explains sensor study design. Its collocation guidance, testing protocol FAQ, quality assurance guidance and siting guide provide first-party technical background.
Continue with eduKateSG’s beta-attenuation instrument explanation, bridging inference vocabulary, epistemic stance and interpreting a research data plan.