Executive Summary

Key Numbers

Source: Girardini, Kan et al., Behavioral calibration of mobile-phone GPS data for population-representative analyses, arXiv:2609.01042 (September 1, 2026)

Location records from mobile phones are not a probability sample. How heavily someone uses a smartphone and whether they consent to share location vary from person to person, so older adults, lower-income areas and particular neighborhoods end up systematically under-recorded. The standard treatment is post-stratification. You compare the share each group holds in the sample against the share the census reports, and give more weight to the groups that were under-recorded.

The place the paper points at is what that correction leaves alone. Matching the composition of a population and matching how people spend their day are not the same task. Within the same age band and the same income bracket, someone who carries a phone everywhere and someone who leaves it at home produce different records of the same day. The sentence the authors put in the Discussion says exactly this.

The check ran on GPS records from three metropolitan areas. Roughly 140,000 users in Phoenix, 96,000 in Boston and 527,000 in New York, counted after preprocessing. The records come from users who opted in to share data anonymously, the collection complies with GDPR and CCPA, and the provider is Cuebiq through its Data for Good program. The observation window runs from January through June 2020, and only the January and February records, from before the pandemic, went into building behavioral profiles.

One point is worth settling up front. The variables this study matched to the census are age and income. Gender was not used. Phone records carry neither the age nor the income of the user, so both were assigned probabilistically from the distribution of the census block group containing the inferred home location. Age comes in four bands, 18 to 24, 25 to 44, 45 to 66, and 67 and over, and income in quartiles of each metro area's income distribution. Multiplied out, that gives sixteen strata. Saying the demographic composition was matched means it was matched at that resolution.

To say behavior is off, you first have to fix what it is off against. The researchers set two reference tables. For demographic composition it is the American Community Survey run by the US Census Bureau, and for behavior it is the American Time Use Survey, a nationally representative survey that collects 24-hour activity diaries. Records from 2004 through 2019 were pulled for each of the three metro areas, giving samples of roughly 2,000 in Phoenix, 2,600 in Boston and 9,000 in New York.

The difficulty is that survey answers and GPS traces are written in different languages. So both were rewritten in one format. A day is cut into 48 half-hour bins, and each bin holds the main activity for that interval. There are four activity categories: home, work, points of interest, and unspecified. Travel time between stops in the GPS data goes into unspecified, and survey activities classified as travel are reassigned there for the same reason. Home and work are inferred from recurring nighttime and weekday daytime patterns, and the remaining stops are classified by linking them to nearby OpenStreetMap points of interest. Sensitive locations were excluded.

Embedding those time use survey sequences and grouping them by cosine distance produces behavioral profiles. The number this study settled on is four, using K-medoids clustering. On the phone side, one day between January 3 and February 29, 2020 was drawn at random for each user, and the roughly 760,000 user-days that resulted went into the calibration. A day is attached to a profile if at least half of its ten nearest time use survey neighbors belong to that profile. If it fails that test, or if its distance from the medoid of the closest profile falls above the 99th percentile of within-profile distances, it is attached to nothing.

The weights arrive in two stages. First comes a factor that brings the share of stratum d, defined by age and income, in line with the census. Then, within that stratum, a factor that brings the share of behavioral profile b in line with the time use survey. Conditioning the behavioral target on the stratum rather than on the population as a whole is what keeps the different rhythms of different groups from being flattened together. In estimating those targets, the respondent weights published by the Bureau of Labor Statistics were applied on the survey side. A final factor scales the sample up to the size of the population, so one finished weight reads as the number of individuals in the target population that user-day stands for.

A day that attaches to nothing receives a weight of zero and drops out. What the authors found interesting was what that residue turned out to be. Opened up, it was mostly low-quality trajectories, or records where home, work or a point of interest had been attached wrongly. Failing to attach to any behavioral profile was doing the work of a quality signal.

The evaluation metric is the Jensen–Shannon distance between two distributions. The smaller the value, the closer the phone sample's behavioral distribution sits to the time use survey. The comparison splits four ways: time allocation across activities, structural metrics of the sequence, transition proportions between activity categories, and mobility motifs, which summarize a day's activities as a directed network. There are four structural metrics: reciprocity, which asks whether transitions ran in both directions; the number of distinct activities in a day; the turnover rate, meaning how often the bin changes; and normalized entropy, which measures the diversity of the activity mix. The first three of the four lines are what the calibration embedding was built from, and motifs are not. The motif comparison covers the six most prevalent motifs in the time use reference. Every value comes from 50 bootstrap replicates per metro area.

Time allocation closed by 14.80%, structural metrics by 13.69% and activity transitions by 0.65%, and the mean of the three is 9.71%. The three values do not carry equal weight.

Start with transitions. The 95% confidence interval is ±2.21, so as a single total it is hard to call an improvement at all. The passage where the paper plots transitions by category pair says something else, though. There, the calibration improves the alignment towards the ATUS distributions for nearly all transition categories. Each category narrowed a little, and collapsed into one number the movement fell inside the interval. The dimensions that were already aligned before calibration lie elsewhere. Of the four structural metrics, reciprocity and turnover rates are the ones the paper calls already well aligned in the uncalibrated sample, which makes 13.69% a figure carried by the number of activities and by sequence entropy.

Motifs sit at the other end. They were never put into the calibration embedding, and they closed by 17.00%. The authors read that as evidence the method has not been overfitted to the particular measures it was tuned on.

The number this article turns on comes next. Swap the comparison from an uncalibrated sample to one matched on age and income, and a gap survives. BePop sat 10.44% closer to the time use survey overall, and 17.10% closer on motifs. A sample whose demographic composition has already been matched still carried that much behavioral divergence.

The authors attach a reservation here. Which behavioral dimensions come out well aligned depends on how the embedding was designed, and no single embedding optimizes every behavioral measure. Application-specific embeddings consistently performed best for the dimensions they targeted.

What difference a slightly displaced distribution actually makes is the question left standing. The researchers picked two indicators in common use in mobility research and plotted them before and after calibration. One is the daily radius of gyration, the other the number of co-location events at business locations, which serves as a proxy for face-to-face interaction. The observation window straddles the World Health Organization pandemic declaration of March 11, 2020. Their objective, the authors state in advance, is not to estimate the causal effect of the pandemic. It is only to see how far calibration moves a longitudinal estimate.

The two indicators moved in opposite directions. Radius of gyration comes down under calibration, and business co-location goes up in the pre-lockdown period.

Take radius of gyration first. The calibrated estimate lands below the uncalibrated one in both periods, which says the uncalibrated sample had been reading people as moving more than they did. Both approaches capture a similar contraction in mobility after the declaration, but they start that contraction from different heights.

Business co-location is trickier. Before the declaration, the calibrated estimate runs 10.03% higher across the three cities. After it, the two curves fall quickly and converge to similar levels, which is why the lockdown-period change comes to -1.69% in Phoenix and +0.20% in Boston, close to nothing. What the paper draws from this is not the convergence itself but its consequence. Because the calibrated series begins from a higher pre-pandemic baseline and converges to a similar level during lockdown, it implies a larger relative reduction in business-related interactions. Read from the raw data alone, in other words, the drop the pandemic caused in face-to-face contact comes out smaller than it was.

The authors decline to turn this into a one-directional lesson. Calibration can change both the estimated level of a longitudinal indicator and the magnitude of its temporal change, they write, and depending on the outcome considered, uncalibrated data may either underestimate or overestimate changes in population mobility.

4.1How Far These Numbers Can Be Trusted

One assumption sits underneath the longitudinal analysis. A behavioral profile drawn from a single randomly chosen day has to represent a persistent part of that person's behavior. The researchers checked the assumption at two levels.

For a single person the answer is simple. Re-classifying a user's multiple observed days showed the dominant profile accounting on average for more than 70% of daily assignments. The share of days that attached to nothing under daily re-classification stayed below 10% in both the pre-pandemic and pandemic periods.

The population level adds a layer to the procedure. Each user's reference-day profile is held fixed, the profile is then re-inferred from the sequence observed on each subsequent day, and the results are aggregated using that day's weight computed from the fixed profile. In Phoenix, the distribution obtained that way came 9.70% closer to the pre-pandemic time use survey reference than the unweighted data, with a confidence interval of ±0.61. A weekday-to-weekend oscillation remained, which the authors call expected, since the reference survey combines weekday and weekend diaries. After the declaration, the weighted distribution shifted away from its pre-pandemic composition, and the direction of that shift matched what a separate time use survey sample collected between May and August 2020 showed. Even with the reference-day profiles nailed down, the weighted population followed real behavior when it changed.

The pandemic-period time use survey sample, though, was very small, 51 respondents in the case of Phoenix. The distance reduction measured over that stretch was 0.64% with a confidence interval of ±3.86. The resulting large standard errors prevent any determination of whether the improvement persists through the pandemic period, so the authors draw the line themselves: that stretch should be read as evidence of external consistency rather than as exact validation. They also list the assumption of reference-day profile persistence as shaky in periods when behavior changes abruptly.

The longitudinal analysis was also restricted to users observed both before and after March 11. The restriction limits changes in sample composition caused by complete panel dropout. It also leaves a cohort made only of the users who stayed to the end, and the authors note alongside it that this may therefore introduce survivorship bias.

That is where the paper ends. Read from the data quality side, none of it is unfamiliar. What we count when we judge a dataset is mostly the missing-value rate, the duplicate rate, schema conformance, and out-of-range values. Every one of those asks whether the columns were filled in properly. Where a distribution check exists, it usually stops at the distribution of a single column, something like whether the age-band proportions look close to the population.

The dimension this paper adds sits next to that. It asks not whether each row respected the format, but whether the shape of behavior the rows make together matches the reference population. Time allocation, activity transitions, the motif of a day's activities: none of these come out of looking at one cell, and all of them require reading the whole sequence. And when they are off, they leave no trace in a format check. Data with no missing values, no duplicates and a respected schema can push a downstream indicator by more than 10%, which is what the table in section 4 is.

Three questions carry over from the paper without much translation. The correspondence below is this article's reading, not something the paper argues.

• If our checks look at a distribution, whose distribution is it. A column's distribution, or the distribution of a sequence such as one person's day or one case's processing path?
• Is there a reference distribution to compare against. The way this paper brought in the time use survey, is there a document that separately defines the population our data is supposed to represent?
• If we align a distribution, do we align it over the whole population or within groups? BePop conditioned its behavioral targets on each stratum precisely so that different groups' rhythms would not be flattened.

One more thing worth borrowing is how the paper handles a failed match. A day that attached to no behavioral profile was not discarded data here; it was a quality signal. Sparse trajectories and records with home or work attached wrongly collected there. It is worth asking where the records that attach to no reference distribution currently go in our own pipelines, and whether they are piling up anywhere at all.

The paper's closing sentence points the same way. It calls for treating behavioral representativeness as a complementary criterion for assessing data quality, and for incorporating behavioral calibration as a standard component of mobile phone data analysis. The proposal comes out of mobility research, but it can be set down as it stands in the room where we decide what to call AI-Ready.

• 1.Girardini, N. A., Kan, U., López, E., Lepri, B., Lucchini, L., & Centellegher, S. (2026). Behavioral calibration of mobile-phone GPS data for population-representative analyses. arXiv:2609.01042