Approach and contribution
30-second case brief
From household installations to interpretable environmental records.
- Problem
- How can residential IAQ monitoring work within a whole-home improvement program, and what can its environmental records reveal?
- My contribution
- Led monitor preparation, household installations, data extraction, analytical development and the first-author field-study paper.
- Outcome
- The published field study documents 54 targeted homes, 11 installations and five usable datasets. A reproducible exploratory analysis retains 1,404 intentionally overlapping six-hour windows and describes three environmental-profile groups under the primary analysis. Household composition and sensitivity checks define the limits of that grouping.
- Explore the evidence
ApproachCommunity-engaged mixed methods · multisensor monitoring · coverage-gated time series · regression-slope detection · standardized features · PCA and k-means
Project description and responsibility
- Context
- Drexel University with Philadelphia Energy Authority (PEA), Energy Coordinating Agency (ECA) and Esperanza
- My role
- Project lead for the research workflow; field installations, data preparation, modeling, interpretation and first-author writing
- Collaboration
- PEA program access, ECA household coordination and retrieval, Esperanza community relationships; publication with S. Hoque
- Instrumentation
- Awair Element and IQAir AirVisual Pro; four-channel AirVisual exports for the event analysis
- Methods
- Ten-minute medians; six-observation regression; six-hour windows; 12 standardized features; PCA; k-means
- Outputs
- 2024 RBDCC field-study paper and presentation; October 2026 analytical manuscript, reproducible five-household analysis and pseudonymous result figures
Interactive tool · fieldwork planner
Stress-test an occupied-housing IAQ deployment.
Adjust four field conditions. The planner identifies the weakest link and the next preparation priority before monitoring begins.
Priority before deployment
Plan for offline data continuity.
Connectivity is the current bottleneck. Add local buffering, visible device-status checks, and a recovery plan before relying on remote transfer.
The 54 → 11 → 5 sequence reports the documented study. The adjustable planning signal is illustrative and does not estimate recruitment or completion probability.
Whole-home improvement and occupied housingBeginning with the community project
Connecting weatherization to the indoor environment
The study began within Philadelphia’s Built to Last program, led by the Philadelphia Energy Authority. The program coordinates home repairs, energy conservation and healthy-home services for low-wealth homeowners. Energy Coordinating Agency’s weatherization work connects home evaluation and energy education with measures such as air sealing, insulation and ventilation. This created a practical setting for asking how the indoor environment could be monitored alongside improvements to a home.
The research question was operational as well as environmental: which monitors could be installed, how could data be retrieved, and how could residents’ experiences be collected alongside the measurements? I led the research workflow from monitor preparation and household visits through data extraction, analysis and writing. PEA connected the study to its applicant pathway, ECA supported contact, visit coordination and retrieval, and Esperanza contributed established community relationships. The fieldwork paper was published with S. Hoque.
PEA: Built to Last program context ECA: weatherization servicesMeasurement channels and deployment requirementsChoosing monitors that could support the fieldwork
A common environmental core, with complementary measurements
Awair Element and the field-era IQAir AirVisual Pro both provided PM2.5, carbon dioxide, temperature and relative humidity. These shared variables made it possible to examine particle changes together with the surrounding thermal and CO₂ conditions. Awair also provided a TVOC channel, adding another view of the indoor environment. The event analysis uses the four shared AirVisual export channels, with PM2.5 as the trigger variable.
The selection also had to work in occupied homes. A readable display supported communication with residents, while setup, power, Wi-Fi access, monitor placement, data retrieval and maintenance determined whether a device could produce a usable record. Preparation and side-by-side comparison therefore came before deployment. The photograph shows that preparation arrangement, while the channel table identifies the measurements carried into analysis.
Awair Element channel documentation Field-era AirVisual Pro user guide| Channel / unit | Awair Element | Field-era AirVisual Pro |
|---|---|---|
| PM2.5 · µg/m³ | Measured | Measured |
| CO₂ · ppm | Measured | Measured |
| Temperature · °C | Measured | Measured |
| Relative humidity · % RH | Measured | Measured |
| TVOC · ppb | Additional channel | Not listed in the 2021 guide |
Awair support (2024) and AirVisual Pro user guide (2021)
Project conference presentation
Mixed-methods fieldworkMaking the monitoring plan work inside homes
Installation, participant communication and retrieval
ECA contacted eligible households about their interest in monitoring and their access to Wi-Fi, then coordinated installation appointments. I installed and prepared monitors within this partner-supported process, checking power, connectivity and a practical location with the household. Placement mattered because a monitor’s record depends on airflow, proximity to sources and obstructions, as well as on the device itself.
The environmental measurements were paired with participant communication through in-person surveys, phone calls and online surveys, including respiratory-health and wellbeing questions. Follow-up, retrieval and extraction completed the monitoring loop. The published field record documents 54 targeted households, 11 installations and five usable datasets. These are separate stages of the field process, rather than interchangeable sample sizes.
Why technical and social decisions had to be made together
The field experience organized itself into four connected domains. Participants needed clear explanations of the study, privacy and what monitoring involved. Collaborators provided access and trust, but their responsibilities and schedules had to remain coordinated. Logistics connected appointments, electrical outlets, monitor locations, survey timing and retrieval. Technical continuity depended on Wi-Fi, device maintenance and successful extraction.
Connectivity was especially consequential: the published study reports that more than 27 of the 54 potential homes lacked stable usable Wi-Fi. That experience linked the fieldwork directly to the next analytical stage. A long apparent monitoring period could contain substantial gaps, so the analysis had to establish what was actually observed before interpreting changes in indoor conditions.
Five-household analytical extensionTurning irregular exports into comparable records
Preserving measured coverage before creating a common time grid
The analytical extension reconciles 49 AirVisual exports from five household folders. Exact duplicate records were removed and conflicting timestamps excluded. The retained export spans run from September 2022 to June 2023 for H01 and H03 and from March to June 2023 for H02, H04 and H05. These spans describe supplied records; deployment and device-movement logs were not sufficient to confirm that every retained timestamp represented an occupied-home deployment. The four channels retain their named physical units, with household pseudonyms H01 to H05.
A ten-minute grid provides a common resolution for comparison. Each observed bin uses the median of its available readings. Linear interpolation is restricted to gaps of at most two missing bins and requires observations on both sides. Longer outages remain missing. A six-hour event window is retained only when all four channels are available and at least 80% of its bins are measured. The coverage figure keeps those measured values separate from the short filled gaps.
Measured coverage ranges from 2.10% in H02 to 95.09% in H01. H01 has a median 60 source readings per nonempty ten-minute bin, compared with two in each other home. The primary PM2.5 screen excludes values above 1,000 µg/m³ as an analysis convention. Sensitivities retaining or capping those readings change the fitted partition, so screening and concentration accuracy must be considered separately. Missing records are not assumed to be random.
| Household | Measured 10-minute bins | Measured share |
|---|---|---|
| H01 | 40,919 | 95.09% |
| H02 | 341 | 2.10% |
| H03 | 34,907 | 82.10% |
| H04 | 2,815 | 17.32% |
| H05 | 12,333 | 75.96% |
Same coverage-gated analytical run
Change-rate detection and temporal contextFinding concentration changes, then keeping their context
A regression slope identifies sustained changes in PM2.5
The detector fits a line to six consecutive ten-minute PM2.5 values, spanning 50 minutes between labels. A trigger requires a positive slope greater than 4 µg/m³ per ten-minute index step, equivalent to 0.4 µg/m³/min, and a local range of at least 20 µg/m³. At least two of the trigger and next six values must exceed the six-point local minimum by more than 20 µg/m³. Those two values need not be consecutive or measured.
After a qualifying trigger, repeat triggers are suppressed until two consecutive measured bins return to within 20 µg/m³ of the local baseline after an elevation, a PM2.5 outage occurs, or a 72-hour cap is reached. The coverage gate then determines whether a complete four-channel context is retained. Look-ahead, interpolation and the post-trigger context make this a retrospective analysis; real-time detection performance has not been established.
Five-household analytical rerun
The six-hour window carries the other variables into interpretation
Each retained context begins one hour before the trigger and contains 36 ten-minute bins over the half-open interval [-1, 5) hours. Nearby windows intentionally share context. In this library, 811 of 1,399 adjacent within-household pairs overlap, and 50,544 window-bin slots use 35,935 unique ten-minute bins. These are dependence diagnostics, not independent sample sizes or counts of household activities.
Returning to the measured curves makes the event concrete. In this example, the particle excursion can be followed beside its CO₂, temperature and relative-humidity context. The channels evolve on their own scales, and their changes are not interchangeable. Plotting them separately preserves those relationships and gives the feature-based analysis a traceable environmental reference.
Actual retained window from the analytical rerun
Features, model selection and original curvesGrouping environmental profiles without losing the measurements
Twelve physical features become a smaller comparison space
Each window contributes the mean, sample standard deviation and range of PM2.5, CO₂, temperature and relative humidity. The 12 features are standardized before PCA. This handles unit scales but does not remove correlated-feature redundancy or give each household equal influence. Household labels and timestamps remain metadata, outside the fitted features.
Six components retain 95.793% of standardized feature variance. K-means candidates from two through ten groups are compared using the archived minimum-size and repeat-seed criteria, then the lowest eligible Davies–Bouldin index. This rule selects three groups. The four-group fit has a lower Davies–Bouldin index but its smallest group has 14 windows, below the required 14.04, meaning at least 15 integer windows. The two-group fit has a higher silhouette score. Three groups is therefore a conditional descriptive choice.
Inspect the groups in feature space and in their measured time series
The first two principal components display 53.941% of standardized feature variance and provide a view of the grouped windows. The grouping itself uses all six retained components, so the two-dimensional scatter is a projection of the full comparison. Each point is one six-hour window, not a household or a single sensor observation.
For interpretation, the closest actual window to each group’s centroid in PCA space is returned to its four measured channels. These representative windows show different particle excursions, CO₂ levels and thermal trajectories. They are observed examples from the records, rather than synthetic centroid curves. Common axis ranges within each channel make the profiles directly comparable.
Five-household analytical rerun
Field outputs and analytical resultsResults: usable records, detected episodes and recurring profiles
The event library now spans all five household exports
The coverage-gated workflow retains 1,404 contextual windows: H01 1,116; H02 6; H03 133; H04 24; and H05 125. H01 contributes 79.49% of the library. These counts reflect available spans, measured coverage and detected PM2.5 changes. They do not estimate household activity frequency or prevalence.
The primary three-group fit has Davies–Bouldin index 1.3055, silhouette 0.3295 and median repeat-seed adjusted Rand index 1.000. These measure internal structure and repeat-run consistency on the same records. They do not establish stable classes or accurate prediction in new households.
| Household | Candidates | Retained windows |
|---|---|---|
| H01 | 1,129 | 1,116 |
| H02 | 8 | 6 |
| H03 | 187 | 133 |
| H04 | 27 | 24 |
| H05 | 153 | 125 |
Five-household result registry
Analytical rerun · 14 October 2026
Household composition changes the interpretation
Group 2 contains 239 H01 windows out of 261, or 91.57%. Group 3 contains 79 H05 windows out of 114, or 69.30%. All five homes contribute to groups 1 and 3, and profiles also vary within homes. The groups describe both environmental variation and strongly unequal household representation.
At fixed three groups, a disjoint-window subset has adjusted Rand index 0.990 against the primary labels on shared windows. Equal-household weighting reduces agreement to 0.344, robust scaling to 0.342 and logarithmic PM2.5/CO₂ features to 0.295. Each sensitivity refits preprocessing, PCA and k-means. Removing overlap alone therefore does not account for the primary partition, while household weighting and representation materially affect it. Equal weighting is a different descriptive target, not an automatic correction for sparse or missing records.
Adjusted Rand index (ARI) measures partition agreement on identical shared windows. The sensitivity chart holds the number of groups at three; the manuscript separately reports the group counts selected again within each sensitivity.
Stability within these records
One hundred seven-day block resamples per weighting scheme refit the full fixed-three-group pipeline. Median agreement is 0.930 for the primary weighting, with empirical 2.5th–97.5th percentiles 0.835–0.973. Equal-household weighting gives median 0.921 with a wider 0.279–0.977 range against its own original weighted fit. These are conditional perturbation ranges, not confidence intervals or validation in new homes.
In a random 80/20 window-split diagnostic, an average 72.89% of test windows share source bins with training windows. A chronological split with a 72-hour purge on each side and train-only preprocessing reduces direct overlap leakage, but remains an internal diagnostic because the detector was developed using these same source homes. Five households remain the higher-level sampling units.
An archived signal-reference check
The preserved protocol describes sensor-trace annotations made without viewing detector outputs. Thirteen earlier blocks supported method development and checking; four subsequently selected day-long blocks provide the archived fresh comparison. Re-running the scoring verifies the numerical matches but cannot independently establish the annotation process. These are concentration-pattern references, not participant-confirmed activity labels.
The four blocks contain five rising references, four in H01 and one in H05. All five match retained triggers; H03 and H04 each contribute a negative block with no retained candidate. Three additional downward changes are outside the positive-rise target. The counts remain the same with 0, 10 or 20 minutes of matching tolerance. This selected five-positive check is too small and concentrated to support a population detection-rate claim.
| Measure | Observed count |
|---|---|
| Day-long review blocks | 4 |
| Rising signal references | 5 |
| One-to-one matched rises | 5 |
| Unmatched retained candidates | 0 |
| Negative blocks with zero candidates | 2 |
| All-direction references | 8 |
| Unmatched downward changes | 3 |
One-to-one matches unchanged at 0, 10 and 20 minutes of timing tolerance
Archived signal-reference annotations; scoring reproduced 14 October 2026
The grouped result stays tied to physical conditions
The group summaries quantify what the profiles represent. The largest group (Group 1) contains 1,029 windows across 5 homes. Group 3 has the highest median window-average CO₂ concentration (639 ppm), while Group 2 has the largest median PM2.5 range (484.9 µg/m³). The representative curves show how those feature differences appear over time.
The comparison combines PM2.5-triggered episodes with their wider environmental context. The measured examples show that similar particle changes can occur with different CO₂ and thermal conditions. Keeping the original curves, units and feature summaries together makes that result useful for reviewing the environmental record and selecting periods for closer investigation.
| Group | Windows | Median PM2.5 mean (µg/m³) | Median PM2.5 range (µg/m³) | Median CO₂ mean (ppm) |
|---|---|---|---|---|
| 1 | 1,029 | 31.80 | 135.55 | 473.64 |
| 2 | 261 | 111.82 | 484.90 | 522.82 |
| 3 | 114 | 26.42 | 115.32 | 639.07 |
Each entry is the median of the corresponding window-level feature
The final output connects field design to interpretable data
The published field study translated the installations and participant interactions into a four-domain framework and practical recommendations. The analysis extends that work from securing usable records to inspecting concentration-change episodes, reducing their multivariate features with PCA and returning grouped results to actual measurements.
The contribution is a reproducible exploratory description of PM2.5-triggered contexts in the available five-home records. Original exports, analysis choices, source-linked figures and sensitivity results are documented in the full manuscript and private reproducibility package. Generalization requires additional households, deployment and calibration records, and independent reference observations. The present data do not establish household activity prevalence, health effects or before/after retrofit benefits.
Read the full analytical manuscript (PDF)| Recommendation | Practical purpose |
|---|---|
| Accessible IAQ education | Explain the measurements and the monitoring process |
| Privacy and informed-consent protocols | Establish transparent expectations with households |
| Durable community partnerships | Support access, trust and follow-up |
| Flexible scheduling and communication | Coordinate installations, surveys and retrieval |
| Connectivity and sensing alternatives | Plan for power, data continuity and pollutant coverage |
| Equipment and research funding | Support preparation, maintenance and field capacity |
| An integrated field plan | Coordinate technical requirements with participant and partner needs |
Complete analytical record
Overlapping indoor air quality windows in five low income households
Research manuscript, October 2026; unpublished.
The complete project record
Full book
Read the complete report, thesis, or presentation here. Use the page controls to explore the original figures and methods.
Challenges in Mixed Methods Research: IAQ in Low-Income Households
The original document, page by page.
The reader loads when you open the book.
