Laziness vs Space Weather — Robustness Analysis (Null Result)

Analysis Complete — Relationship NOT Confirmed

This report presents the results of a full robustness analysis of the relationship between laziness-related Google Trends data and space-weather variables over 4,076 consecutive days (2015-03-18 → 2026-07-15). Unlike our epilepsy study, this one yielded a clear null result: no credible evidence that laziness-related search interest is associated with space weather.

This page is worth reading precisely because the result is negative. Rigorous science demands that we report what did not survive scrutiny as clearly as what did.

---

Dataset & Collection Process

Google Trends Pipeline

The laziness search-interest series was constructed identically to our epilepsy series: daily data from Google Trends pulled in overlapping 90-day windows, then stitched together at overlap points to form a continuous 4,076-day time series.

Metric	Value
Query	laziness
Category	behaviour
Date range	2015-03-18 → 2026-07-15
Total days	4,076
Missing values	0

Space-Weather Variables

Same 7-variable panel as our epilepsy analysis:

Variable	Source	Resolution
Kp	NOAA SWPC	3-hourly → daily mean
Ap	NOAA SWPC	3-hourly → daily
Dst	WDC Kyoto	hourly → daily mean
Solar wind speed	NASA OMNI	hourly → daily mean
IMF B (average)	NASA OMNI	hourly → daily mean
F10.7 solar flux	NOAA	daily
Sunspot number	SILSO	daily

Preprocessing

• STL detrending (365-day seasonal period, robust) on both series
• Outliers winsorised at ±3 standard deviations
• Lags tested: −30 to +30 days
• Total hypotheses: 427 (7 variables × 61 lags)
• FDR-corrected significant: 157

---

Methodology

1. Cross-correlation function computed at all 61 lags (Pearson r)
2. Benjamini–Hochberg FDR correction at α = 0.05 across all hypotheses
3. Robustness battery — five independent tests beyond standard FDR:

• Block bootstrap (autocorrelation-preserving surrogates)
• Placebo time-shift test
• First-difference detrending
• Subperiod sign stability
• Yearly vs daily aggregation comparison

1. All code in Python (pandas, SciPy, statsmodels)

---

The Initial Signal

At first glance, one variable stood out.

Top Association per Variable (STL-detrended)

Variable	Max	r		Lag (days)	Pearson r	Raw p	FDR p	Significant
Kp	0.0448	−11	+0.0448	0.004278	0.014053	YES
Ap	0.0727	−18	+0.0727	0.000004	0.000041	YES
Dst	0.0728	−18	−0.0728	0.000003	0.000041	YES
Solar wind speed	0.0452	−11	+0.0452	0.003976	0.013265	YES
IMF B (avg)	0.0455	−19	+0.0455	0.003777	0.012698	YES
F10.7 flux	0.0625	−15	+0.0625	0.000067	0.000504	YES
Sunspot number	0.1023	−15	+0.1023	< 0.000001	< 0.000001	YES

The headline: sunspot number showed r = +0.102 at lag −14 days (p < 0.001, FDR-significant). Under a naïve screen, this would be declared a "significant finding."

All effect sizes are |r| < 0.11 — practically very weak even before robustness checks.

---

Robustness Testing — Where the Signal Collapses

The standard FDR pipeline says "yes." But science demands more. We ran five additional robustness tests. The sunspot signal failed all of them.

Test 1: Block Bootstrap (Autocorrelation-Aware Surrogates)

Classic p-values assume independent observations. Daily time series are not independent — today's value is correlated with yesterday's. Block bootstrap preserves the autocorrelation structure and asks: could this r arise by chance from any two long-memory processes?

Variable	Real r	Block bootstrap p	Significant?
Kp	−0.0196	0.2800	NO
Ap	+0.0059	0.7340	NO
Dst	−0.0062	0.7440	NO
Solar wind speed	−0.0138	0.4320	NO
IMF B (avg)	−0.0212	0.2500	NO
F10.7 flux	+0.0499	0.3200	NO
Sunspot number	+0.0715	0.1180	NO

Block size: 30 days | Iterations: 500 | 0 of 7 variables survive.

The sunspot correlation shrinks from r = +0.102 to r = +0.072 under block bootstrap, and the empirical p-value = 0.118 — well above the 0.05 threshold. This is the first red flag.

Test 2: Placebo Time-Shift Test

What if we deliberately misalign the time series — shifting the sunspot data forward or backward by random offsets — and recompute the correlation? If the real correlation were genuine, it should exceed all placebo correlations.

Sunspot number: minimum placebo p-value = 0.0800 (203 placebo tests).

❌ The real correlation does not exceed the placebo distribution. A signal that can be beaten by randomly wrong alignments is not a credible signal.

Test 3: First-Difference Detrending

STL detrending is one method, but it can leave residual low-frequency artefacts. First-differencing (computing day-to-day changes rather than absolute values) is a completely different approach — it removes all slow trends and asks: do daily fluctuations co-vary?

Method	Sunspot max	r		Lag	FDR Significant?
No detrending	0.1144	−14	YES
STL detrending	0.1023	−15	YES
Rolling 365d	0.1208	−14	YES
First difference	0.0278	−6	NO

When we look at day-to-day changes rather than absolute levels, the correlation drops to r = −0.028 and loses FDR significance entirely. This is strong evidence that the original signal was driven by shared low-frequency trends, not by genuine day-to-day coupling.

Test 4: Subperiod Sign Stability

A genuine relationship should hold its sign and rough magnitude across different time windows. We split the 4,076 days into three equal subperiods.

❌ The sign flips in the middle period (from +0.11 to −0.004). The magnitude also varies dramatically — from r ≈ 0.11 in the early period to essentially zero in the middle period. This is not the signature of a stable physical relationship.

Test 5: Yearly Aggregation vs Daily Granularity

A prior analysis using yearly-aggregated data had suggested r ≈ 0.71 for sunspot number vs laziness. We recomputed the yearly correlation using our full dataset.

Variable	Yearly r	p-value	N years
Kp	−0.138	0.669	12
Ap	−0.166	0.607	12
Dst	+0.166	0.606	12
F10.7	+0.074	0.819	12
Sunspot number	+0.067	0.836	12

The yearly correlation is r = +0.07 (p = 0.84) — completely non-significant and near zero. The previously reported r ≈ 0.71 does not replicate.

---

Aggregation Artifact / Simpson's Paradox

Why did the yearly correlation look so strong before?

This is a textbook Simpson's Paradox driven by aggregation:

1. Both laziness searches and solar activity rose slowly from ~2015 to ~2020 due to unrelated secular trends (internet penetration grew; Solar Cycle 24 peaked and Cycle 25 began rising).
2. When data is averaged to yearly means, the high-frequency noise cancels out, and the shared slow trend dominates — producing a spurious r ≈ 0.71.
3. When analysed at daily granularity with proper detrending, the shared trend is removed, and the correlation collapses to r ≈ 0.10 (and then fails all robustness checks).

Diagram: How Aggregation Inflates Correlation

YEARLY VIEW (misleading):         DAILY VIEW (correct):

Laziness    Sunspots              Laziness (detrended)    Sunspots (detrended)
   |           |                     |     /\    /\         |    /\    /\
   |    /\     |                     |    /  \  /  \        |   /  \  /  \
   |   /  \    |    ← both           |   /    \/    \       |  /    \/    \
   |  /    \   |    rise together    |  /            \      | /            \
   | /      \  |                     |_/              \_____|/              \____
   |/        \_|
   2015 → 2026                         2015 → 2026 (detrended, no shared trend)

   r ≈ 0.71 (artifact)                r ≈ 0.10 (real, but fails robustness)

Implication

The yearly correlation of r ≈ 0.71 is almost certainly an aggregation artifact. It should not be cited as evidence of a space-weather–behaviour relationship. This is exactly why we test at the daily level with proper detrending and robustness checks.

---

Final Verdict

Is there evidence for a real relationship?

No. The sunspot signal passes a basic FDR screen but fails all five robustness tests:

Test	Result
FDR correction (standard)	✅ Passed
Block bootstrap	❌ Failed (p = 0.118)
Placebo time-shift	❌ Failed (min p = 0.08)
First-difference detrending	❌ Failed (r = −0.028)
Subperiod sign stability	❌ Failed (sign flips)
Yearly replication	❌ Failed (r = 0.07, p = 0.84)

Final status: RELATIONSHIP NOT CONFIRMED

How strong would it be even if real?

Even if we ignore the robustness failures and take r ≈ 0.10 at face value, solar activity would explain less than 1% of the daily variance in laziness searches. This is practically negligible — for comparison, weather forecasts explain ~60% of temperature variance; this explains less than 1%.

What is the most plausible explanation?

The weak daily signal (r ≈ 0.10) is best explained as a combination of:

• Autocorrelation artefacts (both series are long-memory processes)
• Incomplete detrending (STL cannot perfectly separate trend from signal)
• Multiple-testing noise (with 427 hypotheses, some false positives are expected)

There is currently no credible evidence that laziness-related search interest is associated with space weather.

---

Why Null Results Matter

This page exists for a reason: rigorous science publishes what fails alongside what passes.

• The same pipeline that found a weak-but-real epilepsy–F10.7 correlation found nothing for laziness.
• This suggests the pipeline discriminates: it does not simply produce "significant" results for every keyword.
• The epilepsy result is stronger because this one failed.

Null results protect against confirmation bias. They are not failures — they are filters.

---

Limitations

1. Google Trends is a proxy. Search volume reflects online interest, not actual laziness behaviour.
2. Detrending is imperfect. Some genuine signal may be removed alongside the trend.
3. Multiple testing is severe. With 427 hypotheses, false positives are expected even after FDR.
4. Geographic aggregation. Google Trends is global; space-weather effects may vary by latitude.
5. Confounders. Holidays, weekends, and media events affect search volume independently.

---

Analysis completed: 2026-07-17. Data period: 2015-03-18 → 2026-07-15 (4,076 days). Status: Completed. Verdict: RELATIONSHIP NOT CONFIRMED.

---

In Plain English — What We Actually Found (and Didn't Find)

Let's drop the jargon for a moment.

We spent the same effort on "laziness" searches as we did on "epilepsy" — same 4,076 days, same seven space-weather variables, same pipeline. And the result? Nothing.

At first, it looked like there might be something: sunspot numbers showed a tiny positive correlation with laziness searches (people search for "laziness" slightly more when the Sun has more spots, about two weeks earlier). On paper, it passed the basic statistical test.

But we don't stop at "on paper." We threw five harder tests at it — tests designed to catch false alarms — and the signal failed every single one of them.

What killed the signal:

• When we accounted for the fact that both laziness and sunspots have long, slow trends (sunspots rise and fall over years; internet usage grows over years), the correlation evaporated.
• When we looked at day-to-day changes instead of absolute levels — nothing.
• When we chopped the data into three chunks — the correlation flipped sign in the middle chunk. A real relationship doesn't do that.
• When we randomly misaligned the dates as a sanity check — the fake misalignments produced correlations just as strong as the real one.

And the previously reported yearly correlation of r ≈ 0.71? It collapsed to r ≈ 0.07 when we recalculated it properly. That original number was a mirage — an artefact of averaging over years rather than looking day by day.

The takeaway: the same rigorous method that found a weak-but-real epilepsy–solar signal found absolutely nothing for laziness. This is actually good news — it means our method isn't just rubber-stamping everything as "significant." It discriminates. The epilepsy result gains credibility because laziness failed.

Bottom line: laziness and space weather are not related in any detectable way. We looked hard, and there's nothing there. Time to move on to the next keyword.