Does Google search reflect disease burden?

The project's goal was to compare health searches with where conditions actually occur. Here that is tested against CDC survey data for every state, then the structure of health searching is examined what varies, where, and how stable it is.

1 · Validation against CDC data

Do searches match where illness actually is?

Each state's average Google interest (2011–2017) against the share of its adults who report the condition in the CDC's Behavioral Risk Factor Surveillance System (BRFSS). The last two columns ask whether the link survives once we allow for how much each state searches about health in general (partial correlation).

In plain words. Do states that google a disease more actually have more of it? For diabetes, depression, cancer and heart disease, yes moderately. But much of that is because states with more illness search more about health in general. Obesity is the surprise: the most obese states don't search more about obesity.
New to statistics? Key terms in plain words
Google Trends score (0–100)
Not a number of searches. For each topic and year, Google gives 100 to the place where that topic made up the largest share of all searches, and scores everywhere else relative to it. 0 means too few searches to report.
Metro area
One of Google's 210 US media markets a city plus its surrounding region (Nielsen DMA). Some span two or three states.
CDC prevalence
The share of adults who report having a condition in the CDC's yearly telephone survey (BRFSS).
Age-adjusted death rate
Deaths per 100,000 people, recalculated as if every year had the same age mix so an ageing population doesn't distort the trend.
Correlation (r)
How closely two things rise and fall together, from −1 to +1. 0 means no link; around 0.3 is a moderate link and 0.5 or more a strong one. A correlation does not prove that one thing causes the other.
p-value
How likely a link this strong would appear by pure chance if there were really none. Below 0.05 is called statistically significant unlikely to be a fluke.
95% confidence interval (CI)
The range the true value very probably lies in. A narrow range means a precise estimate.
Controlling for / adjusted
Taking another factor out of the comparison, so the link that remains is not just that factor in disguise.
Mean vs median
The mean is the ordinary average. The median is the middle value when everything is sorted, so a few extreme values don't pull it.
Partial correlation
A correlation after controlling for a third factor here, how much a state searches about health in general.
Principal component analysis (PCA)
A way to boil many related measures down to a few underlying patterns. Here nine search topics reduce to two main patterns.
Percentile
The share of places this one beats. 50 is typical; 90 means more interest than 90% of metro areas.
Correlation between state search interest and CDC prevalence
Searches forCDC survey measureCorrelation r [95% CI]p-valueStrengthr after allowing for overall searchingSpecific to this condition?

Search vs prevalence

Overall: search intensity vs disease burden

Burden = average z-score of six CDC prevalence measures. Each point is a state.

Specificity: every search against every prevalence measure

Caveats. 48 states (plus DC) with their own metro areas; BRFSS figures are self-reported, crude (not age-adjusted) prevalence; heart-disease data covers 2015–2017 and flu vaccination ages 65+ from 2013. These are links between states, not people they don't show that the person searching has the condition.

2 · Structure

Two dimensions of health searching

Principal component analysis (PCA) a standard way to boil many related measures down to a few underlying patterns applied to the nine search scores of every metro area with complete data (2004–2017 averages, put on a common scale).

In plain words. The nine topics mostly follow two patterns. The first is simply how much a place searches about health overall highest in smaller Southern and Midwestern cities. The second is a North–South split: the South leans toward rehab, heart disease and diarrhea; the North toward depression, vaccines and obesity. Metro areas don't fall into neat separate groups.

Variance explained

What each dimension measures

The north–south axis (PC2) on the map

3 · Co-occurrence

Which health worries travel together?

Correlation between each pair of topics across metro areas (14-year averages): 1 = always rise together, 0 = unrelated. Select a cell to see the scatter plot.

In plain words. Places that search a lot for one health topic tend to search more for related ones too. Depression and obesity are the closest pair; rehab goes with few others. Darker cells mean a stronger link.

Scatter

4 · Stability

Is the geography of health searching stable?

How similar the ranking of metro areas was in 2008–10 and in 2015–17 (Spearman rank correlation ρ: 1 = same order). Rankings aren't affected by Google's yearly rescaling.

In plain words. The places that searched most for a topic around 2009 are mostly still the places searching most around 2016. Diarrhea is the exception its hot spots move around, as outbreaks do.

5 · Mortality

How US death rates changed, 2004–2017

CDC deaths per 100,000 people for the ten leading causes, adjusted for age so years compare fairly. Log scale: equal steps on the axis mean equal percentage changes.

Reading. Heart disease and cancer still the two leading killers fell steadily; unintentional injuries (largely drug overdoses), Alzheimer's disease and suicide rose. These national series can't be matched to metro search data directly, which is why section 1 uses state-level CDC survey prevalence instead.

Original findings, revisited

What changed from the first version

The first dashboard added up Google scores across metro areas, years and topics. Because Google scores each topic and year on its own 0–100 scale, adding them up mixes different scales, so several conclusions needed revising.

Original claims and corrected findings
Original claimCorrected finding