EVIDENCE · METHODS

Make the limits visible.

Different evidence sources answer different questions. This register explains how the study keeps them separate.

One sample. Different eligible populations.

946

Base records

433 Airtel / 513 Jio

599

Support users

271 Airtel / 328 Jio

520

Failure reporters

240 Airtel / 280 Jio

Support N–R questions require support contact; failure and recovery U–Z measures require a significant failure. Populated answers outside those populations remain in the source workbook but do not enter those analyses. Multi-select proportions use respondents, so totals may exceed 100%.

Ordinal questions, explicit methods.

Descriptive results

Counts, 1–5 distributions, median and all modes are primary. Means, sample standard deviation and sample variance are additional summaries.

Provider comparisons

Two-sided Mann–Whitney U, tie/continuity correction and Cliff’s delta. Positive differences mean Airtel higher. Inference requires 20 eligible records per provider and nonconstant pooled ratings.

Uncertainty and multiplicity

95% percentile intervals use 2,000 respondent resamples within provider. Holm correction uses the family of five primary proxies. Arbitrary filtered exploration remains exploratory.

Associations

Spearman correlation uses pairwise eligibility and 4,999 deterministic permutations. Chi-square/Cramér’s V requires expected cell counts of at least 5.

A nonsignificant difference does not establish equivalence. Statistical association does not establish cause. No population weights or representative-sample claims are used.

Why there is no overall SERVQUAL score.

No paired expectation questions are present. Physical tangibles were not measured; the digital-experience item is a limited proxy. The candidate reliability group has α = 0.474 and the support-eligible responsiveness group α = 0.401. These weak internal-consistency diagnostics support an item-first analysis.

Inspect scale diagnostics

The discrepancy register.

Source issue	Decision
Slide 8/9 survey overlays conflict with editable network/framework content	Survey and independent network evidence remain separate.
Slide empathy: 3.22 / 3.25	Does not reproduce. All-row R means: 3.420 / 3.415. Support-eligible means: 3.450 / 3.424.
TRAI broadband denominator	Label wired + wireless, including M2M; not unique people or households.
Reliability Experience scale	100–1000 points, not 0–1000.
426 incident placeholders and repeated text	Exclude placeholders from incident NLP; expose repeated-text sensitivity.
“Call drop Billing”	Preserve the ambiguous compound category; do not invent a split.
Jio AirFiber versus national FWA counts	Withhold cross-source comparison until definitions reconcile.
Original PowerPoint absent	All 15 experiences follow the supplied Markdown reconstruction.

Text is evidence, with context.

Incident, best-aspect and improvement prompts are analysed separately. VADER labels are automated estimates. A frozen 120-response semantic review assesses sentiment and themes; model-assisted review is not independent human validation. Aggregate sentiment requires macro-F1 ≥ 0.75. The Text Explorer reports the actual result.

Original identifiers are omitted from the web data. Text is scanned for identifying information before export. No raw workbook, customer names or email fields are bundled.

Open the Text Explorer