# Analysis Plan

## Clean the data first
- remove obvious setup errors
- separate baseline readings from peaks
- annotate any source events
- standardize timestamps and locations
- keep ELF magnetic-field data separate from any RF or audio/other exploratory data
- keep source state, visible infrastructure, and route-access notes as separate fields from the actual measurement value
- keep AMI meter exchanges or other active utility work separate from ordinary energized/idle/maintenance state
- keep electric AMI module installs, gas AMI module installs, electric meter exchanges, gas meter exchanges, and door-hanger/no-access visits as separate operational subtypes
- keep construction_context separate from source_state so AMI exchanges, system-upgrade work, and routine service are not merged into one operational bucket
- keep `cleanup_phase` separate from both `construction_context` and `source_state` so late-2026 mobilization, active construction, monitoring, and completed phases do not get collapsed into one category
- keep access_constraints separate from both exposure and source-context fields so blocked driveways, lane closures, locked gates, or no-access stops do not look like higher exposure
- keep `workzone_location` and `access_method` separate from exposure values so public right-of-way work, yard work, street work, and behind-the-gate stops do not get mistaken for EMF signal
- keep preliminary utility work such as underground marking, locates, or property-state documentation separate from both exposure and construction_context
- keep Hobart Marsh Conservation Area NIPSCO ROW stops separate from both exposure values and restoration activity context; treat Duck Creek Tributary work as managed-corridor context, not signal
- if you later compare against historical or literature-based EMF sources, keep source family, frequency band, and dosimetry type separate and only use quality-screened source references
- keep stationary spot readings separate from walking scans and body-worn samples
- keep walkthrough screening summaries separate from spot/personal exposure estimates; treat route-level continuous-monitoring screens only as a screening layer to rank sites and choose follow-up points
- compare route screens with personal measurements only as separate method families; do not fold the screen into the final exposure distribution
- keep route-screen metrics separate from exposure distributions so a screening pass does not get pooled into the final exposure estimate
- keep NIOSH method families separate too: spot measurements, walkthrough surveys, area monitoring, source surveys, and personal monitoring should not be pooled unless you have an explicit transformation or sampling design
- keep `sampling_interval_seconds`, device class, and peak-capture mode attached to each observation so maxima from different logging intervals are not pooled as if they were interchangeable
- keep meter response class (true-RMS, linear-response, flat-response, waveform-capable) attached to each observation, and only compare unlike response classes after an explicit correction or a clearly stated screening transformation
- compare peak readings only within the same meter family or after an explicit interval adjustment; a coarser peak logger is a screening tool, not a directly comparable peak estimate
- if a spot reading used a single-axis probe on a stand, keep that measurement mode distinct because it suppresses motion-gradient fields and should not be merged with moving measurements without an explicit transformation
- keep waveform-characterization readings separate from standard ELF magnitude spots; if the pilot adds a waveform-capable check at repeated hotspots, store it as a diagnostic field-characterization sample, not as part of the main exposure series
- when waveform data are available, summarize them separately with static-field magnitude, harmonic/non-harmonic content, and orientation or waveform-shape notes instead of folding them into the ordinary broadband field column
- when waveform follow-up is used, keep the exact meter model and usable frequency range attached to the record so the diagnostic layer remains interpretable
- keep AMI-related meter-integrity or billing anomalies separate from exposure values; treat them as utility-ops context when they appear
- keep AMI subtype, install duration, advance-notice lead time, visit duration, door-hanger status, access-request status, temporary shutoff, and relight notes separate from exposure values
- keep AMI advance-notice windows separate from exposure values: electric AMI installs are typically noticed about 30 days ahead, gas AMI module installs are typically noticed about 30 days ahead, gas meter exchanges are notified at least 14 days in advance, and immediate-maintenance / door-hanger-only cases should be treated as workflow context, not exposure
- keep access-outcome notes such as inside-meter, locked-gate, no-access, or door-hanger-only separate from both exposure and source-state fields
- keep AMI RF metadata separate from ELF magnetic-field values
- keep electric AMI installs separate from gas AMI installs and keep meter exchanges separate from module-only visits
- keep electric AMI module installs separate from electric meter exchanges, and gas AMI module installs separate from gas meter exchanges
- keep gas AMI relight-required stops separate from simple install or door-hanger-only stops
- keep door-hanger, access-request, temporary shutoff, and relight notes as route metadata, not exposure values
- keep advance-notice lead time and visit duration separate from exposure values; use them as route-context covariates when interpreting access success or temporary interruption patterns
- compare AMI results by exact workflow subtype first — electric module install, electric meter exchange, gas module install, gas meter exchange, no-access, door-hanger-only, or immediate-maintenance — before pooling anything into a broader utility-work bucket
- keep probe height, body location, and calibration reliability attached to the observation so confidence can be judged later
- keep service-workflow metadata such as advance notice, door hanger, access requested, temporary gas shutoff, and relight separate from source-state and exposure values
- keep route-screen summary metrics such as mean, max, or Hi-5-style peak screens separate from the main exposure distributions
- keep waveform-diagnostic samples separate from the ordinary exposure tables and use them only to explain repeated hotspots
- keep waveform-diagnostic samples with failed motion/artifact correction out of the ordinary exposure tables entirely; use them only as flags that the capture window needs a cleaner repeat
- for waveform diagnostics, keep the raw capture, any motion-corrected version, and the correction method separate; annotate out-of-range, baseline-shift, and non-integral-cycle windows before any correction is treated as analysis-ready
- if waveform diagnostics are added to a repeated hotspot, summarize the diagnostic layer separately with sample window length, sampling frequency, meter model, usable frequency range, and whether the capture was accepted, corrected, or rejected

## Core outputs
- site-by-site summary table
- corridor vs control comparison
- time-of-day plot
- heat map or route map
- list of repeated peak locations
- comparison of readings by source class
- comparison of readings by operating state where the source is known
- comparison of peak readings by sampling interval and meter class
- comparison of readings near active utility work, including AMI meter exchanges or system-upgrade activity, versus quiet utility-adjacent stops
- comparison of readings by construction context, especially AMI exchange versus system-upgrade versus routine service
- comparison of readings by AMI subtype and meter-exchange status
- compare readings by cleanup phase in the Grand Calumet / East Chicago corridor, especially late-2026 mobilization versus active construction versus monitoring
- compare readings by Grand Calumet cleanup subarea, especially Junction Reaches, Eastern 5 Miles, Lake George Canal East, and Lake George Canal Middle, before pooling anything into a broader cleanup-phase bucket
- compare readings by actual cleanup footprint where the route data support it: Junction Reaches should split into the Indiana Harbor Ship Canal south of Columbus Drive, the West Branch between Indianapolis Boulevard and the IHC, and the East Branch between the IHC and Kennedy Avenue; Lake George Canal should stay split into Middle and East
- compare readings by corridor context, especially Grand Calumet AOC / East Chicago mixed-use cleanup areas versus inland controls
- comparison of readings by preliminary-work context when crews are locating utilities, marking lines, or documenting property state
- keep current-year Urban Waters context separate from corridor context so managed-corridor activity does not get mistaken for a source-state effect
- keep Hobart Marsh / Deep River restoration and NIPSCO ROW context separate from exposure tables and compare it only as a route-context layer
- keep Duck Creek Tributary restoration phase and completion timing separate from exposure values; use it as managed-corridor context when the route overlaps the Hobart Marsh / Deep River work area
- keep the Hobart Marsh Conservation Area NIPSCO ROW extent separate from exposure values; treat its acreage / mileage as corridor geometry, not signal, and keep that geometry separate from the broader ~2,000-acre managed/protected landscape
- separate summaries for stationary spot, walking scan, and body-worn data
- compare readings by cleanup-window context in the Grand Calumet / East Chicago corridor when route planning overlaps active remediation or monitoring windows
- compare spot measurements, walkthrough surveys, area monitoring, source surveys, and personal monitoring separately before any pooling
- compare electric AMI installs against gas AMI module changes and gas meter exchanges as distinct utility-op subtypes
- compare gas exchanges that required relight or temporary shutoff against stops that only left a door hanger or required no entry
- if source-distance and source-dwell metadata are available, compare readings by distance bins and dwell-time bins before calling a hotspot real
- if waveform-characterization samples exist, compare them only as a hotspot diagnostic layer and do not mix them into the main corridor-vs-control tables unless you explicitly model the method difference
- when a reading looks odd, ask first whether the sampling strategy, probe placement, body location, source distance, dwell time, or calibration story is weak before calling it a signal
- treat the Hobart Marsh NIPSCO ROW as corridor context when it appears, not as an exposure effect by itself
- treat Grand Calumet / East Chicago cleanup windows as route context first, and only compare them as exposure drivers if the signal survives the context split
- compare readings by specific Grand Calumet cleanup windows: Junction Reaches construction start, Lake George Canal monitoring, and any 2028 follow-on remediation
- compare readings by AMI service-workflow subtype, including electric install, electric meter exchange, gas module install, gas meter exchange, and gas exchange with relight or temporary shutoff
- compare readings by AMI notice window and access outcome when you are trying to explain visit completion or interruption patterns; do not use those metadata as proxies for EMF level
- compare route-screen summaries only as a screening layer for choosing follow-up points, not as final exposure estimates
- compare waveform follow-up only within repeated hotspots or source-adjacent points and keep those diagnostics out of the main corridor-vs-control tables

## Simple statistics
Start with:
- medians and interquartile ranges
- mean comparisons when the data support it
- nonparametric comparisons for small samples
- mixed-effects or repeated-measures models if the pilot grows
- repeat-point differences as a QA/QC check, not as a result to cherry-pick
- repeated-point differences should be interpreted in light of probe placement consistency and orientation
- within each measurement mode, compare corridor and control sites before combining anything across modes
- keep method-family comparisons separate from measurement-mode comparisons so a spot sample does not get treated like a walking or personal sample by accident
- keep meter-response comparisons separate from raw exposure comparisons so linear, flat, and waveform-capable instruments do not masquerade as the same signal
- treat AMI meter exchanges and other visible utility-work stops as transient source-state data, not as the steady-state utility-adjacent baseline
- keep corridor-context comparisons separate from source-state comparisons so location context does not get mistaken for exposure mechanism
- keep construction-context comparisons separate from source-state comparisons so route work and access conditions do not get mistaken for EMF level
- keep `workzone_location` and `access_method` separate when summarizing utility-adjacent stops so access geometry is not mistaken for exposure
- keep preliminary-work summaries separate from construction-context summaries so utility locate/mark activity is not mixed into the active work bucket
- keep cleanup-phase summaries separate from construction-context summaries so route timing is not folded into the operational source bucket
- keep cleanup-window comparisons separate from source-state comparisons so active remediation windows do not become fake exposure drivers
- keep managed corridor geometry separate from cleanup phase and source-state so ROW alignment and restoration timing do not masquerade as exposure
- compare readings by managed corridor context, especially Hobart Marsh ROW versus broader managed-corridor land, before pooling anything into one route bucket
- compare readings by cleanup subarea separately from cleanup phase for East Chicago / Hammond routes
- keep AMI module-install summaries separate from gas exchange summaries so service workflows do not blur into one bucket
- if source-distance is known, use it as a covariate or stratifier before comparing source classes
- for repeatable point sources, prefer pre-set distance bins such as contact, ~30 cm, ~61 cm, and >61 cm before calling a hotspot real; if the reading collapses by the outer bin, treat proximity as the leading explanation until repeated
- keep constrained stand-off readings separate from standard distance-ladder bins; if the ideal ladder was blocked, label the record as `closest_normal_worker_position` rather than pooling it with a clean ladder point
- score unusual readings with a confidence note that reflects sampling strategy, dosimetry type, anatomical location, duty cycle, exposure scenario, calibration reliability, and meter response class
- if peak capture is part of the question, score confidence lower when the logging interval is too coarse to plausibly catch brief spikes
- if waveform-characterization samples are sparse, treat them as confirmatory diagnostics rather than numerically comparable exposure estimates
- if the pilot adds EEG or sleep tracking, thalamic oscillation measurements (delta, spindle, alpha, theta, gamma) belong in their own column, not folded into the EMF field-strength metric — the literature shows external ELF fields alter EEG oscillations (alpha, theta, spindle bands) and those effects are real though variable in direction, so oscillation data is both a plausible exposure biomarker and a downstream health outcome depending on how it is framed

## Decision questions
- Is the corridor consistently different?
- Are the peaks tied to specific operations or times?
- Do results survive a second pass?
- Do utility-adjacent or industrial source classes stand out from inland controls?
- Do readings shift when a known source is energized, throttled, or shut down?
- Are any apparent hotspots actually explained by buried-utility congestion, access limits, or other infrastructure context?
- Do active utility-work stops behave differently from ordinary utility-adjacent background stops?
- Do AMI exchange stops differ from broader system-upgrade or routine-service construction stops?
- Do right-of-way, yard, street, and behind-the-gate stops separate cleanly once access metadata is split out?
- Does any pattern hold within a single measurement mode, or only after mixing unlike sampling methods?
- Do apparent differences persist after removing preliminary utility work such as locates, markings, or property-state documentation?
- Does any corridor difference remain after splitting out cleanup-window context in the Grand Calumet / East Chicago area?
- Does the route still look different after tagging the current Grand Calumet remediation window as context rather than signal?
- Do waveform-characterization hotspots line up with the same locations that keep producing repeated ELF peaks?
- Do gas AMI exchange stops that require relight or shutoff produce a distinct operational pattern from no-access or door-hanger-only visits?
- Do AMI subtypes separate cleanly once module-install, exchange, and no-access stops are split out?
- Does the pattern hold after splitting out Junction Reaches construction, Lake George monitoring, and 2028 follow-on remediation?
- Do AMI workflow subtypes still separate once notice, access, and relight metadata are split out?
- Do readings fall off as source distance increases, or is the apparent hotspot just a proximity artifact?
- Do route-screen maxima just identify follow-up points, or do they repeat as a stable pattern?
- Does the Hobart Marsh / Deep River pattern persist after separating the NIPSCO ROW from the Duck Creek restoration corridor?

## Final memo
Write one short memo with:
- what was tested
- what was found
- what remains unknown
- whether a larger study is justified