Skip to content
Analyst work

Data study · November 2026

The case for third places

Question

Is there a real case for third places in Sacramento, and where should FourthSpace launch first?

Headline finding

Time with friends fell from 60.0 to 26.5 minutes a day between 2003 and 2025. Sacramento County is lonelier than 95% of U.S. counties, and 36 of its neighborhoods (census tracts) pair high loneliness with low walk access to any indoor place to gather.

My recommendation

Launch in south Sacramento, with the library branches near the loneliness hot spots as the first anchor venues.

Tools
  • Python 3.13 and pandas
  • statsmodels and SciPy for the models
  • Survey weights with Taylor-linearized variance
  • geopandas and networkx for the walk network
  • esda for the hot spots
  • matplotlib and Plotly for the first charts
  • JavaScript on Node for the numbers on this page
  • React and SVG for the charts on this page

The question

The client

FourthSpace, the community app I’m building for finding third places that are open right now. It launches in Sacramento, and the findings also go into its pitch deck and landing page.

The decision they face

Where FourthSpace starts in Sacramento and who it starts with. I weighed five launch options against the findings.

My starting hunch

I came in thinking tech use, phones and social media most of all, drives people into isolation and loneliness.

How the hunch held up

I was partly right. Time with friends fell by more than half, and heavy social media users reported worse mental health in 485 of 540 models. Bullying and weak ties at school went with far larger gaps in teen sadness. None of it shows cause.

The data, 10 tables

  • atus_with_whom_by_year_age.csv161 rows

    Weighted minutes a day alone and with friends or family, by age group, for every American Time Use Survey year from 2003 to 2025.

    Where it came from
    BLS American Time Use Survey microdata, weighted with the survey’s final weights.
    Why I chose it
    Finding 1. The diary records who a person was with through the day, so friends time is measured directly.
    What it cannot tell us
    It covers one diary day per person. The 2025 sample has 6,146 diaries against 20,720 in 2003, so recent years carry wider error.
  • benchmarks.csv52 rows

    Adjusted gaps in persistent sadness and poor mental health that go with social media use, bullying, sleep, sports and ties at school, in 2023 and 2025.

    Where it came from
    CDC Youth Risk Behavior Survey microdata, read from the fixed-width files and weighted with its strata and clusters.
    Why I chose it
    Finding 2. It puts social media on the same scale as the rest of a teen’s life.
    What it cannot tell us
    Cause. Each wave surveys different students once.
  • spec_curve_results.csv540 rows

    All 540 models of the social media link: six definitions of heavy use crossed with outcomes, control sets, samples and survey years.

    Where it came from
    My own runs on the YRBS 2023 and 2025 files, each a survey-weighted linear probability model.
    Why I chose it
    The other side of the debate. One model can be tuned toward an answer, so I ran every reasonable one and report the spread.
    What it cannot tell us
    Settle direction. Sad teens may turn to social media as much as social media makes teens sad.
  • places_sacramento_county.csv4 rows

    Sacramento County’s loneliness rate and two related measures, with the U.S. and California county medians and its rank.

    Where it came from
    CDC PLACES, 2023 data in the 2025 release, which models BRFSS answers down to each county.
    Why I chose it
    Finding 3, and the adult count behind my Sacramento estimate.
    What it cannot tell us
    Stand as a direct count. PLACES rates come from a model, and the 477,000 figure is my estimate built on one.
  • cbp_third_places_1998_2023.csv520 rows

    Establishments in 9 kinds of gathering place for the U.S. and Sacramento County, every year from 1998 to 2023, with a rate per 10,000 residents.

    Where it came from
    Census County Business Patterns county files, with industry codes mapped across the 2002, 2007, 2012 and 2017 revisions.
    Why I chose it
    Finding 4. It shows which kinds of place grew and which shrank. Sacramento County went from 206 coffee shops and snack bars to 501.
    What it cannot tell us
    Visits. A count of businesses says nothing about how many people go or how long they stay.
  • atus_minutes_by_place.csv23 rows

    Weighted waking minutes a day by location, from home to restaurants and places of worship.

    Where it came from
    BLS American Time Use Survey activity file and its location codes.
    Why I chose it
    The time behind finding 4. Fewer gathering places matter more if people also spend less time out.
    What it cannot tell us
    Who was there. A minute at a restaurant can be spent alone.
  • imls_sacramento_public_library.csv27 rows

    Sacramento Public Library visits and service population by fiscal year.

    Where it came from
    IMLS Public Libraries Survey.
    Why I chose it
    Option B. The branches have space to fill and a visit gap to close.
    What it cannot tell us
    Branch detail. The survey counts visits for the whole system.
  • atus_workday_company_by_type.csv4 rows

    Minutes alone, with coworkers, with friends and with anyone outside the household on days with 4 or more hours of work, split by where the work happened.

    Where it came from
    BLS American Time Use Survey activity, location and who files, 2010 to 2019 and 2021 to 2025.
    Why I chose it
    Finding 5.
    What it cannot tell us
    Choice. People who work from home may differ from office workers in ways a diary can’t see.
  • acs1_wfh_share.csv30 rows

    Workers who usually work from home, for five areas from the whole U.S. down to the city of Sacramento.

    Where it came from
    Census ACS 1-year table B08301, read from the summary files.
    Why I chose it
    It sizes the remote worker group in Sacramento County.
    What it cannot tell us
    Hybrid days. The ACS asks how people usually get to work, so someone home two days a week counts as a commuter.
  • priority_tracts.csv36 rows

    The 36 Sacramento County census tracts that pair higher loneliness with low walk access to an indoor gathering place.

    Where it came from
    CDC PLACES tract estimates and OpenStreetMap venues pulled in November 2026, measured on a walk network built from Census TIGER 2024 roads.
    Why I chose it
    Finding 6, and option A.
    What it cannot tell us
    Every venue. OpenStreetMap misses places, and walk access counts street intersections near a venue as a stand-in for homes.

The process

  1. Questions first

    I scoped ten questions and ran seven. Each one ties to a choice FourthSpace has to make about its message or its launch.

  2. Microdata

    I read the raw files for the ATUS diaries and the two CDC surveys, BRFSS and YRBS. Each estimate uses the survey’s own design weights and clusters. My helper computes Taylor-linearized variance, the same method R’s survey package uses for a linear model.

  3. Check against the source

    Before writing any finding, I checked my weighted rates against the figures the CDC and Census publish for the same year.

  4. Every model

    For the social media question I ran all 540 combinations of five analyst choices and kept the full spread, so no single model carries the result.

  5. On foot

    For Sacramento I built a walk network from TIGER roads with freeways and ramps removed. Each tract’s access is the share of its street intersections within 800 meters of an indoor gathering place. Local Moran’s I found the loneliness hot spots.

  6. One stat bank

    Every figure went into a stat bank with its year, geography, source, link and caveat, plus a flag for whether it is safe for the landing page. This page reads 1,397 rows from 10 of the clean tables behind it.

  7. Into the product

    The findings became a 12-slide pitch deck for FourthSpace and a ranked list of venues to call.

U.S. bookstores in County Business Patterns, before and after the code fix.

First run, the 2022 and 2023 files read with the new industry code
year,est
2022,0
2023,0
After, three rows of cbp_third_places_1998_2023.csv (three of eight columns shown)
year,naics,est
2021,451211,4670
2022,451211,4706
2023,451211,4647

The 2022 and 2023 files still file bookstores under the 2017 industry code. Reading them with the 2022 code matched nothing, so bookstores dropped to zero. I kept the 2017 code for every year.

The code for the key steps

The weighting helper every survey estimate ran through. survey.py
def svy_lm(df, y, X_cols, weight="weight", strata="stratum", psu="psu"):
    d = df.dropna(subset=[y, weight, strata, psu] + X_cols)
    X = np.column_stack([np.ones(len(d))] + [d[c].to_numpy(float) for c in X_cols])
    w = d[weight].to_numpy(float)
    yv = d[y].to_numpy(float)
    A = X.T @ (X * w[:, None])
    Ainv = np.linalg.pinv(A)
    beta = Ainv @ (X.T @ (w * yv))
    e = yv - X @ beta
    u = X * (w * e)[:, None]
    U = pd.DataFrame(u)
    U["h"] = d[strata].to_numpy()
    U["j"] = d[psu].to_numpy()
    z = U.groupby(["h", "j"]).sum()
    meat = np.zeros((X.shape[1], X.shape[1]))
    for h, zh in z.groupby(level=0):
        n_h = len(zh)
        if n_h < 2:
            continue
        dev = zh.to_numpy() - zh.to_numpy().mean(axis=0)
        meat += n_h / (n_h - 1) * dev.T @ dev
    V = Ainv @ meat @ Ainv
    se = np.sqrt(np.diag(V))
    n_psu = len(z)
    n_strata = z.index.get_level_values(0).nunique()
    return {"beta": beta, "se": se, "n": len(d), "df": n_psu - n_strata, "names": ["const"] + X_cols}
The spread of all the social media models behind the verdict. derive.mjs
export function specCurve(rows) {
  const summarize = (sample, group) => {
    const estimates = group.map((row) => Number(row.estimate_pp));
    return {
      sample,
      models: group.length,
      positive: estimates.filter((value) => value > 0).length,
      above_zero: group.filter((row) => Number(row.ci_low_pp) > 0).length,
      below_zero: group.filter((row) => Number(row.ci_high_pp) < 0).length,
      median_pp: median(estimates),
    };
  };
  const samples = [...new Set(rows.map((row) => row.sample))];
  return [summarize("All models", rows), ...samples.map((sample) => summarize(sample, rows.filter((row) => row.sample === sample)))];
}
The count of lonely adults in the priority tracts, behind finding 6. derive.mjs
export function priorityTotals(rows) {
  const lonely = rows.reduce((sum, row) => sum + Number(row.est_lonely_adults), 0);
  return [
    {
      tracts: rows.length,
      lonely_adults: lonely,
      no_indoor_venue: rows.filter((row) => Number(row.n_indoor) === 0).length,
      in_hot_spot: rows.filter((row) => row.lisa_cluster === "High-high hot spot").length,
    },
  ];
}

Dead ends

  • The Census API wanted a key

    The ACS calls in my fetch script came back as a page asking for an API key. I switched to the summary files on the Census file server, which need none.

  • Library visits before 1998

    The IMLS files from the 1990s name the visits field ATTEND, and those years didn’t line up with the rest. The library series starts in fiscal 1998.

  • An institution at the top

    A special land-use tract, likely an institution, topped my first priority ranking. Tracts numbered 98xx and 99xx are now out of every ranking.

  • Curvy roads counted twice

    My first access measure counted every point along a road’s shape, which favored winding streets. It now counts intersections only.

  • A wrong place name

    My first map caption put the hot spots in the northern suburbs. Matched to ZIP codes, they sit in south and north Sacramento, so I rewrote it.

  • A title the data didn’t back

    One chart title said teens who skip social media are not the happiest. That held in 2023 and failed in 2025. The title now describes both years.

The 6 findings that change the decision

  1. Finding 1

    26.5 min

    a day with friends, ages 15 and up, 2025

    Americans 15 and up spent 26.5 minutes a day with friends in 2025, down from 60.0 in 2003. They spent 49 more minutes a day alone.

    Dataset atus_with_whom_by_year_age.csv
    Query derive.mjs, friendsChange()

    Time with friends fell to 26.5 minutes a day

    Averages from American Time Use Survey diaries, weighted to the population. Ages 15 to 24 fell the most, from 155.1 to 53.6 minutes. Hover, tap or use the arrow keys for each year.

    • Ages 15 to 24
    • Everyone 15 and up

    Source: atus_with_whom_by_year_age.csv, derive.mjs, friendsByYear().

    Show the numbers as a table
    YearAges 15 to 24Everyone 15 and up
    2003155.160.0
    2004145.160.0
    2005150.256.4
    2006123.051.5
    2007113.251.8
    2008116.150.2
    2009111.351.1
    2010133.556.4
    2011126.354.7
    2012125.154.7
    2013123.254.8
    2014104.550.2
    2015104.748.8
    201699.147.2
    201790.341.5
    201891.240.7
    201968.634.3
    202043.819.7
    202149.423.3
    202251.024.1
    202355.825.9
    202446.425.7
    202553.626.5
  2. Finding 2

    +17.3

    points of persistent sadness for students not close to people at school, against +8.3 for heavy social media use

    In 2025, students who don’t feel close to people at school reported persistent sadness 17.3 points more often than other students. Using social media more than once an hour went with 8.3 points.

    Dataset benchmarks.csv
    Query derive.mjs, teenGaps()

    School life tracks teen sadness more closely than heavy social media use

    U.S. high school students, 2025. Each bar comes from its own model, weighted to the population and adjusted for demographics. Across all 540 social media models the median gap is 5.0 points, 7.6 for girls and 3.0 for boys. These are associations from one survey wave.

    Extra points of persistent sadness

    Source: benchmarks.csv, derive.mjs, teenGaps(), spec_curve_results.csv, derive.mjs, specCurve().

    Show the numbers as a table
    FactorExtra points of persistent sadness95% intervalShare of studentsStudents in the model
    Bullied electronically31.327.9 to 34.614.6%9,562
    Bullied at school27.824.1 to 31.619.7%9,570
    Does not feel close to people at school17.314.2 to 20.436.8%8,468
    Sleeps under 8 hours on school nights14.110.9 to 17.275.7%7,888
    Plays on no sports team10.76.9 to 14.542.4%6,392
    Uses social media more than once an hour8.35.6 to 10.928.4%8,886
  3. Finding 3

    ≈477,000

    adults in Sacramento County who feel lonely at least sometimes, my estimate

    Sacramento County’s loneliness rate, adjusted for age, was 39.5% in 2023, higher than 95% of the 2,300 U.S. counties with an estimate and second of 58 in California.

    Dataset places_sacramento_county.csv
    Query derive.mjs, sacramento()

    Sacramento County is lonelier than 95% of U.S. counties

    CDC PLACES model-based estimates for 2023. My count of about 477,000 adults multiplies the unadjusted rate of 38.9% by 1,226,535 adults, with a range of 426,000 to 532,000.

    Adults who feel lonely at least sometimes, age-adjusted (%)

    Source: places_sacramento_county.csv, derive.mjs, sacramento().

    Show the numbers as a table
    PlaceAdults who feel lonely at least sometimes, age-adjusted (%)95% intervalRank in California
    Sacramento County39.5%35.3 to 44.22 of 58
    California county median36.2%58
    U.S. county median34.0%2,300
  4. Finding 4

    +91%

    coffee shops and snack bars per resident, 1998 to 2023

    Coffee shops per resident grew 91% from 1998 to 2023 and civic and social organizations fell 40%. Americans now spend 74 more waking minutes a day at home than in 2003.

    Dataset cbp_third_places_1998_2023.csv
    Query derive.mjs, supplyIndex()

    Coffee shops per person nearly doubled, and civic clubs fell 40%

    Establishments per 10,000 U.S. residents, 2023 against 1998. A bar at 100 means no change, and all establishments together sit at 99. Time at third places fell from 83 to 67 minutes a day between 2003 and 2025.

    2023 level per resident (1998 = 100)

    Source: cbp_third_places_1998_2023.csv, derive.mjs, supplyIndex(), atus_minutes_by_place.csv, derive.mjs, placeMinutes().

    Show the numbers as a table
    Kind of place2023 level per resident (1998 = 100)Per 10,000 in 1998Per 10,000 in 2023Establishments in 2023
    Coffee shops and snack bars1911.312.5084,105
    Fitness and recreation centers1510.791.2040,307
    Barber shops1430.150.217,231
    Full-service restaurants1106.997.66258,273
    Religious organizations955.845.54186,615
    Bars and taverns621.911.1839,855
    Civic and social organizations601.160.7023,488
    Bookstores310.440.144,647
    Bowling centers280.200.051,820
  5. Finding 5

    57 min

    with anyone outside the household on a home workday, 2021 to 2025

    On days worked from home in 2021 to 2025, people spent 57 minutes with anyone outside their household, against 423 on days at the workplace.

    Dataset atus_workday_company_by_type.csv
    Query derive.mjs, workdayCompany()

    Home workdays hold 57 minutes with anyone outside the household

    U.S. employed people on days with 4 or more hours of work. 2020 is left out. Most of the gap is coworkers, and time with friends is about the same on both kinds of day.

    Minutes with anyone outside the household

    Source: atus_workday_company_by_type.csv, derive.mjs, workdayCompany().

    Show the numbers as a table
    WorkdayMinutes with anyone outside the householdWith coworkersWith friendsAloneDiary days
    At home, 2021 to 20255714 min12 min10.7 hours2,505
    At the workplace, 2021 to 2025423374 min11 min5.8 hours6,912
    At home, 2010 to 201911834 min25 min9.3 hours2,408
    At the workplace, 2010 to 2019486423 min26 min4.9 hours25,761
  6. Finding 6

    ≈52,000

    lonely adults in 36 neighborhoods with little nearby to gather, my estimate

    36 Sacramento County neighborhoods (census tracts) pair higher loneliness with low walk access to an indoor gathering place. They hold about 52,000 lonely adults, and 34 of them have no indoor venue mapped at all.

    Dataset priority_tracts.csv
    Query derive.mjs, priorityTotals()

    36 neighborhoods hold about 52,000 lonely adults with little nearby to gather

    The 8 of the 36 that hold the most. Walk access is the share of a neighborhood’s street intersections within 800 meters of an indoor gathering place, about a 10-minute walk.

    Estimated lonely adults

    Source: priority_tracts.csv, derive.mjs, priorityTop(), derive.mjs, priorityTotals().

    Show the numbers as a table
    Census tractEstimated lonely adultsLonely at least sometimesWalk accessIndoor venues mapped
    Tract 70.112,14642.5%9%0
    Tract 96.102,13542.8%0%0
    Tract 93.112,07040.8%0%0
    Tract 48.012,01844.6%9%3
    Tract 93.211,88642.4%0%0
    Tract 93.341,85541.3%6%0
    Tract 93.191,81342.9%2%0
    Tract 641,80942.2%0%0

My recommendation

Launch in south Sacramento’s loneliest neighborhoods, with library branches as the first venues.

  1. A

    Launch in the loneliest neighborhoods

    Start the pilot in south Sacramento along the Florin Rd and Stockton Blvd corridors, where most of the hot spot tracts sit.

    Evidence
    Findings 3 and 6. 22 of the 36 priority neighborhoods sit inside a loneliness hot spot, and the 36 together hold about 52,000 lonely adults.
    Main risk
    34 of those neighborhoods have no indoor venue mapped, so early users may open the app and find few places nearby.
  2. B

    Make the library system the anchor partner

    Pitch Sacramento Public Library branches near the hot spots as the first venues, starting with Southgate and Valley Hi-North Laguna.

    Evidence
    Sacramento Public Library logged 2,010,888 visits in fiscal 2024, 54% of its fiscal 2019 count. Branches are free to use and run their own events.
    Main risk
    Public agencies take time to approve partners, which could push back the launch.
  • Start where the need is. Most of the priority neighborhoods sit inside a loneliness hot spot.
  • The libraries cover the biggest risk. Almost none of those neighborhoods has an indoor venue mapped, and the Southgate and Valley Hi-North Laguna branches are free to use and already run events.
  • If library approval runs long, the branches can still be listed as public places while the partnership is worked out.

Limits and next steps

What the data cannot support

  • Show cause. Every link between loneliness, phones, places and health on this page is an association.
  • The Sacramento rates are model-based. PLACES fills in counties and tracts from a statewide model, so the 477,000 and 52,000 figures are estimates built on estimates.
  • Business counts and diary minutes say nothing about how a place feels or whether people meet anyone there.
  • Walk access depends on OpenStreetMap. A venue missing from the map makes a tract look emptier than it is.

What I would test with more access

  • Ask the California Healthy Kids Survey for Sacramento County results to see whether the school ties pattern holds for local teens.
  • Add two optional questions to the FourthSpace waitlist on loneliness and volunteering, to build a first-party measure over time.
  • Test foot traffic data from Advan or SafeGraph in the priority tracts. It costs money, so it stays optional.
  • Rerun finding 5 with the 2026 ACS once it shows what the return-to-office order did.

TL;DR

  1. Studied

    Where my community app FourthSpace should launch in Sacramento, worked out across seven studies on loneliness.

  2. Found

    Americans 15 and up went from 60.0 minutes a day with friends in 2003 to 26.5 in 2025.

  3. By the numbers

    A day with friends, 2025
    26.5 min
    Lonely adults in Sacramento County, my estimate
    ≈477,000
    Lonely adults in 36 neighborhoods with little nearby
    ≈52,000
  4. The catch

    The Sacramento rates come from a statewide model, so 477,000 and 52,000 are estimates built on estimates.

  5. Next

    Ask the California Healthy Kids Survey for Sacramento County results to test the school ties pattern locally.