From Descriptive Patterns to Strategic Insight: MET Analysis with Indexly

Generate actionable marketing insight for Bellabeat’s Spring product using Fitbit activity data — via Indexly — while critically assessing tool limitations.

We focus on:

• METs (activity intensity)
• Steps (next section)
• Sleep (next section)
• Behavioral timing patterns
• Marketing implications

No advanced causal inference is performed (Indexly limitation acknowledged).

📌 Executive Summary

From structured exploration and inference:

• Activity intensity is low overall
• Distribution is extremely right-skewed
• Activity varies more by time-of-day than weekday
• Weekly patterns are statistically detectable but practically tiny
• Rare high-intensity bursts inflate maxima
• Sleep periods show clear inactivity signatures

Users behave habitually, not episodically.

That insight directly informs Bellabeat’s product positioning.

1️⃣ Descriptive Statistics — What the Data Shows

We begin with:

indexly analyze-csv mets.csv --show-summary

MET Distribution

Min: 0.00   Q1: 0.20   Median: 0.40   Q3: 0.90   Max: 140.00
Skew: 11.57

[0–14]        ██████████████████████████████████████████████████ 99.4%
[14–28]       ███                                                0.4%
[28–140]      ██                                                 <0.2%

Interpretation

• 99% of values are light activity.
• Median (0.40) indicates near-sedentary behavior.
• Maximum inflated by rare bursts.
• Mean alone is unreliable due to skew.

📌 Marketing Direction

Target everyday wellness and habit reinforcement — not elite athletic training.

Weekly Pattern

Range ≈ 0.067 (very small)

Behavior appears stable across days.

Time-of-Day Pattern

Range ≈ 0.48 (much larger than weekday variation)

Evening     ██████████████
Afternoon   ████████████
Morning     ████████
Night       ████

Time-of-day appears far more influential than weekday.

Why Descriptive Statistics Were Not Enough

Visual differences do not guarantee real effects.

We therefore asked structured inferential questions.

🔎 GAP 1 — Are Weekday Differences Statistically Meaningful?

indexly infer-csv mets.csv \
  --y mets_pro_mins \
  --group day_of_week \
  --test anova \
  --auto-route \
  --use-raw

Because skew = 11.57, ANOVA assumptions fail.

Auto-route pushed to Kruskal–Wallis.

Result:

• H = 170.014
• p < 0.0001
• Effect size: extremely small

Interpretation

Statistically significant. Practically negligible.

For effect size formula, see mathematical-foundations.md.

📌 Conclusion: Do not design weekday-targeted campaigns.

🔎 GAP 2 — Time-of-Day Effect

indexly infer-csv mets.csv \
  --y mets_pro_mins \
  --group time_of_day \
  --test anova \
  --auto-route \
  --use-raw

Auto-route again selected Kruskal–Wallis.

• H = 19104.98
• p < 0.0001
• Effect size: small but meaningful

ANOVA + Tukey confirmed pairwise differences:

• Evening > Afternoon
• Afternoon > Morning
• Night significantly lowest

ASCII Logic

Variance Partition
------------------
Time-of-day  >>>>>>>>
Weekday      >
Noise        >>>>>>>>>>>>>>>>>>>>

📌 Conclusion: Timing matters. Day label does not.

Marketing should focus on:

• Evening nudges
• Afternoon micro-movement reminders
• Sleep optimization positioning

🔎 GAP 3 — Weekend vs Weekday

indexly infer-csv mets.csv \
  --y mets_pro_mins \
  --group part_of_week \
  --test mannwhitney \
  --use-raw

Results:

Statistically significant. Effect size extremely small.

📌 Conclusion: No meaningful weekend effect.

🔎 GAP 4 — Interaction Effect (Critical)

Does weekday matter after controlling for time-of-day?

indexly infer-csv mets.csv \
  --x day_of_week time_of_day \
  --y mets_pro_mins \
  --interaction day_of_week time_of_day \
  --test ols \
  --auto-route \
  --use-raw

Results:

• R² = 0.005
• Time-of-day coefficients large & significant
• Weekday mostly non-significant
• Interaction term non-significant

ASCII Interpretation

Activity Level =
    Time-of-day  +++++++
    Weekday      +
    Interaction  (none)

📌 Conclusion: Time-of-day drives behavior. Weekday adds almost nothing once timing is included.

Strategically decisive.

🔎 GAP 5 — Robust Central Estimate

Because skew = 11.57, mean alone is unstable.

indexly infer-csv mets.csv \
  --y mets_pro_mins \
  --test ci-mean \
  --bootstrap \
  --use-raw

Result:

95% CI = [1.14, 1.15]

Bootstrap confirms population mean stability despite skew.

Strategic Synthesis

Final Insight

User behavior is:

• Habitual
• Time-structured
• Low-intensity dominant

Bellabeat should:

• Promote consistency over performance
• Leverage evening engagement
• Reinforce micro-habits
• Integrate sleep & activity messaging

Assumptions & Statistical Notes

All tests assume independent observations. Large N inflates statistical power, making tiny effects significant. Non-normality required non-parametric routing. OLS residual normality and homoscedasticity were violated; robust (HC3) standard errors were used. Multiple comparison corrections were not applied in Tukey. Results are associative, not causal.

Tool Scope & Current Usage Notes

The limitations below reflect what was applied in this MET analysis, not what Indexly is fundamentally incapable of.

Clarifications

Causal Modeling

Not performed. This analysis is associative (observational Fitbit data). No causal identification strategy (e.g., IV, DiD, RCT simulation) was implemented.

Hierarchical / Multilevel Modeling

Not applied in this tutorial. A mixed_effects_model module exists (mixed_effects.py) and supports grouped random effects via statsmodels.mixedlm, but it was not invoked in this workflow.

Mixed-Effects Modeling

Available at module level (run_mixed_effects()), but not used for the defined gaps. Therefore, absence here reflects analytical choice and CLI routing — not structural absence.

Bayesian Inference

A Bayesian t-test implementation (bayesian_ttest() in bayesian.py) exists. However, it was not triggered by the CLI during this session. This appears to be a routing/integration issue rather than a missing capability.

Large-N Sensitivity

With >1.3M observations, statistical power approaches 1.0. Tiny effects become statistically significant. Practical interpretation must rely on effect size, not p-values.

Multiple-Comparison Correction

Tukey posthoc was executed without additional correction layers beyond its internal adjustment. No broader correction framework (e.g., FDR control across modules) was applied.

Transition to Steps Analysis

MET intensity tells us how hard users move.

Next, we examine Steps to understand:

• Volume vs intensity mismatch
• Movement frequency
• Behavioral pacing

This will allow us to refine:

Is low MET due to low effort — or simply walking behavior?

That distinction is critical for Bellabeat’s Spring positioning.