Uncited Press › Star Wars Open the interactive journal →
Wars · Political Economy & Social Structures

Counting the Undercity: A Three-List Capture-Recapture Estimate of the Residents of the Lowest Inhabited Levels of One Coruscant District, 27–25 BBY

Dr. Kallum Drayce1
1 Coruscant Academy of Letters and Sciences, Faculty of Social Sciences
Received 18 Sep 2026 · Revised 21 Oct 2026 · Accepted 1 Nov 2026

Abstract

Residents of the deepest inhabited levels of Coruscant are the least likely to be counted, yet no estimate of how many the civic census misses has been published for any district. I linked three administrative lists, the civic census returns, the registers of four clinics and the ration-distribution roll, for the six lowest inhabited levels of one district (given the working name Ostrel) and fitted a log-linear capture-recapture model to the overlap table in three depth bands, with records read from 27 to 25 BBY and a reference day in 26 BBY. The three lists named 24,866 distinct persons. The model estimates 27,903 residents (95% interval 27,521–28,366), so that the census, which listed 18,267, missed about 9,600 residents, or 34.5% (33.6–35.6). The undercount rose with depth, from 15.7% in the shallowest band to 31.6% and then 48.2%. The lists are positively dependent, the clinic registers and the ration roll most strongly, and treating them as independent would lower the estimate by 7.9%. These figures hold only if no three-way interaction exists among the lists, an assumption the data cannot test; allowing it to vary, for scale, over a range comparable to the differences between bands in the pairwise associations moves the total between 26,900 and 29,400, a spread wider than the stated interval. The estimate is best read as a conditional figure. Unrecorded shared avoidance of all three lists would make it a floor; linkage error could move it either way.

1. Introduction

The capture-recapture approach to counting a population was built for animals marked on one visit and sought again on another. I apply it here to the residents of the lowest inhabited levels of one Coruscant district, who are named on three administrative lists that overlap only in part. The analysis was carried out at the Coruscant Academy of Letters and Sciences in 24 BBY, which is the vantage of every statement below. The records themselves run from 27 to 25 BBY.

Coruscant is a city of many levels, and the civic enumeration reaches its lower ones with difficulty. Kolsten (26 BBY) argued from enumerator route sheets that the walkways and lifts which carry enumerators stop short of the oldest dwellings, so that completeness should fall with depth. Lorrant (27 BBY) reached a related conclusion for medical care, finding that distance to a clinic grows with depth and that attendance falls with it. Lythe (26 BBY) described how the ration office registers households and what proof of residence it accepts. None of these studies put a number on the residents the census omits, and none showed whether their omission is concentrated in particular levels.

Counting people from several incomplete lists has a statistical literature of its own. A log-linear model fitted to the pattern of overlap among the lists can estimate the number named on none of them (Kessaly, 28 BBY), provided the lists do not behave as independent samples in a way the model ignores (Lindrevan, 27 BBY). Profile likelihood intervals for the unobserved cell are preferable to symmetric ones (Kerrowan, 26 BBY).

This paper asks how many residents of the district the civic census misses, how that undercount changes with depth, and why the three lists cannot be treated as independent.

2. Methods

Estimand. The quantity estimated is the number of persons, of any species, who slept in a dwelling or berth in the district's six lowest inhabited levels on the night before the civic enumeration opened there in 26 BBY, the reference day. Droids are owned property and are outside the estimand and all three lists. Ostrel is a working name; the boundary is the enumeration office's. The levels are numbered here from L1, the lowest inhabited, to L6, and are grouped in pairs into three depth bands: Band I (L5–L6), Band II (L3–L4) and Band III (L1–L2).

Lists. The census list is the household schedules of the civic enumeration for the district, as filed with the Galactic Senate Legislative Research Office, with amendments filed to 25 BBY (Galactic Senate Legislative Research Office, 25 BBY). The clinic list is formed from the registers of the four clinics serving the district, here C-1 to C-4, and includes a person registered at least once in the 180 standard days centred on the reference day (Republic Health Service, Lower District Clinics, 25 BBY). The ration list is the roll of the ration-distribution office as it stood on the reference day, as filed with the Galactic Senate Legislative Research Office, with strike-offs entered to 25 BBY. It names persons by household (Galactic Senate Legislative Research Office, 24 BBY).

Sample flow. The census schedules held 19,102 person entries for the six levels. I removed 211 duplicate entries of persons already listed and 624 entries marked as departed before the reference day, leaving 18,267. The clinic registers named 13,214 distinct registered persons, of whom 1,519 gave no address in the six levels, leaving 11,695. The ration roll named 21,940 persons, of whom 902 had been struck off before the reference day and 639 were enrolled after it, leaving 20,399.

Linkage. Entries were matched across lists on name, allowing for spelling variants, on species, on stated age within three standard years, on household or berth address, and on clinic and ration numbers where both appeared. A second clerk re-read 400 linked pairs and 400 near-miss pairs that I had left unlinked. She disagreed on 11 of the linked pairs (2.8%), a rate we treat as representative of two-list persons, and on 14 of the near-miss pairs (3.5%). I did not revise the links. The linked pool was the 11,199 persons named on exactly two lists. The near-miss pool was 999 pairs of single-list entries on different lists that agreed on species and address but not on name or age, allocated in proportion to the smaller single-list count. A disagreement on a linked pair indicates a false link, which lowers the estimate; one on a near-miss pair indicates a missed link, which raises it. Both rates are per re-read pair, so I applied them to the pools to obtain person counts.

Model. Each person falls into one of eight cells according to presence on the census, the clinic registers and the ration roll. The cell for presence on none is unobserved. Within each band I fitted a log-linear model to the seven observed cell counts, with a term for each list and for each pair of lists,. With seven observed cells and seven parameters, the fit is exact by construction and carries no information about adequacy. The unobserved count is the value that makes the three-way interaction zero, which has a closed form: the product of the three single-list-only counts and the all-three count, divided by the product of the three two-list-only counts. A zero three-way interaction means that the odds ratio between any two lists is the same among persons on the third list as among persons absent from it. The data cannot test that, since the cell needed to test it is the one that cannot be seen.

Intervals and dependence. The interval for each band's unobserved count is a profile interval: it contains the values for which the completed table departs from the no-three-way model by a likelihood-ratio statistic below 3.84, the 95% point of a chi-squared distribution on one degree of freedom, with counts treated as Poisson. It is conditional on the zero three-way assumption and on the linkage. The all-band interval adds the lower and the upper distances of the three bands separately in quadrature, which assumes the bands are independent. Dependence between lists was described by odds ratios, computed in the slice of the table where the third list is present, with Wald intervals from the reciprocal cell counts. The nine intervals are not adjusted for multiplicity; the six for census–ration and clinic–ration lie well above 1. I also fitted three simpler models by iterating the unobserved cell to convergence, and compared them by the likelihood-ratio statistic summed over bands.

Sensitivity. The three-way interaction was varied by setting its log odds ratio to −0.4, −0.2, 0.2 and 0.4 in every band, where a positive value raises the odds ratio between two lists among persons on the third and lowers the estimate. As a yardstick for scale, the conditional log odds ratios differ between bands by 0.55 for census and clinic, 0.45 for census and ration and 0.50 for clinic and ration, so a three-way term of ±0.4 is of the same order as the variation between bands that the table does show. Linkage error was examined in both directions. Correcting false links, I split 11 in 400 (2.8%) of the persons named on exactly two lists into two single-list persons, which raises the estimate. Correcting missed links, I merged 14 in 400 (3.5%) of the near-miss pool into two-list persons, which lowers it. Moved counts were rounded to whole persons in each cell. Finally, I collapsed the bands into one table and re-estimated.

3. Results

The three lists named 24,866 distinct persons. Table 1 gives the overlap in each band. The census listed 18,267 of them, the clinic registers 11,695 and the ration roll 20,399. About 29% of the named persons (7,148 of 24,866) appeared on all three lists, and 26% were on only one.

Table 2 gives the estimates. Band III alone holds 2,006 persons on no list (1,660–2,422) among 11,643 residents (11,297–12,059). Summed over the bands, the model estimates 3,037 persons on no list (2,655–3,500) and 27,903 residents (27,521–28,366) in the six levels.

Against that total the census listed 65.5% of residents (64.4–66.4) and missed about 9,636 (9,254–10,099), an undercount of 34.5% (33.6–35.6). The undercount rose steadily with depth, from 15.7% in Band I (14.5–17.3) to 31.6% in Band II (30.6–32.7) and 48.2% in Band III (46.6–50.0). Band III held 41.7% of the estimated residents and 58.2% of those the census missed. The intervals for the three bands do not overlap.

Within the slice where the third list is present, the odds ratio between census and ration roll was 3.31 (2.55–4.30) in Band I, 2.12 (1.78–2.51) in Band II and 2.85 (2.43–3.35) in Band III. For the clinic registers and the ration roll it was 3.00 (2.58–3.49), 2.34 (2.04–2.68) and 3.85 (3.34–4.45), the larger of the two in Bands II and III. Census and clinic were weakly dependent, and the sign of the association differed between bands: 1.11 (0.94–1.31), 0.74 (0.67–0.82) and 1.28 (1.17–1.40).

Table 3 shows what the dependence does to the estimate. A model with three independent lists gave 25,706 residents, 7.9% below the 27,903 of the full model, and it fits very poorly (908.8 on 9 degrees of freedom). Adding only the census–ration term left the estimate at 25,918. Adding the clinic–ration term as well brought it to 27,655, with a statistic of 62.8 on 3 degrees of freedom; the fit was acceptable in Band I (1.4 on 1, p = .24) and not in Bands II and III (32.0 and 29.5). Pairing the census with the ration roll alone, as a two-list design would, gave 25,414 across the bands.

Collapsing the bands into a single table gave 27,613 residents, 1.0% below the banded estimate, with an acceptable fit (0.1 on 1 degree of freedom); it cannot show the gradient with depth.

Table 4 gives the sensitivity analysis. The assumption about the three-way interaction moves the estimate more than sampling error does: across log odds ratios from −0.4 to 0.4 the total runs from 29,397 to 26,902, and the undercount from 37.9% to 32.1%. In Band III alone the total runs from 12,630 to 10,982.

Census share stayed ordered by depth throughout, from 82.1%, 66.1% and 47.8% at −0.4 to 85.8%, 70.1% and 54.9% at 0.4. Correcting false links (308 two-list persons split into 616 single-list persons) raised the estimate to 29,515 and the undercount to 38.1%; correcting missed links (35 near-miss pairs merged) lowered them to 27,703 and 34.1%. The changes are 5.8% and 0.7%, because the near-miss pool is small beside the linked pool.

4. Discussion

This estimate is conditional, and its conditions matter more than its interval. Under a zero three-way interaction the district held about 27,900 residents in the six levels, and the census listed roughly two in three. If the three-way term departs from zero by up to 0.4 in log odds ratio, comparable to the 0.45 to 0.55 by which pairwise log odds ratios differ between bands, the total could lie anywhere from about 26,900 to 29,400, and the interval that sampling alone gives does not cover that range. I therefore report 27,903 as the figure the method returns on its stated assumption, not as the population.

The direction of the error may be knowable even where its size is not. Persons who avoid every official list, because they fear eviction or have no standing to claim a ration, are likely to be absent from all three together, and that kind of shared absence tends to pull the three-way term negative, which raises the estimate. I could not measure this, and the reasoning is a hypothesis. If it holds, 27,903 would be a floor and the census undercount of 34.5% a minimum. Linkage error, however, acts in both directions (Table 4). False links push the estimate down and missed links push it up; at the pool sizes I set the first effect is larger, but those sizes are my choice, so the net is unknown and the floor reading concerns one source of bias only.

Of the results, the gradient with depth is the best supported, although it, too, is conditional. The census share fell from 84.3% to 68.4% to 51.8% across the bands. A reading of the route sheets that Kolsten (26 BBY) described might account for part of the gradient, but I have not tested any mechanism.

Dependence is the most useful finding for anyone who would reuse these lists. The census and the ration roll move together, with odds ratios between 2.1 and 3.3, and the clinic registers and the ration roll more strongly in Bands II and III, between 2.3 and 3.9. In Table 3 the clinic–ration term did most of the work, cutting the statistic from 850.1 to 62.8 and moving the estimate by about 1,740; the census–ration term alone moved it by 212. The administrative rules of the lists may explain this: the ration office accepted a census receipt as proof of residence, and the clinics may have given charity care on production of a ration card. I found the rules in the office papers and did not test their effect. The odds ratio below 1 between census and clinic in Band II, a negative association, might point to an outreach clinic reaching households that enumerators missed, but the figure is a single estimate in a single band and could be an artefact of linkage.

A reader who wants one number should not take the lower figures: the census and ration roll alone gave 25,414, 8.9% below the three-list estimate, because their positive dependence makes them look more complete than they are.

5. Limitations

The persons on no list are the missing piece of the record.

Their number comes entirely from the zero three-way assumption, which no cell of the table can check. Different assumptions about it fit the observed counts equally well, so the data do not decide among them, and I have no independent count, such as a door-to-door recount of a sample of dwellings, against which to calibrate.

Scope is limited to six levels of one district, and the bands are pairs of levels chosen by me. The pattern may differ elsewhere in the city, and I give no figure for any area outside the district. Within a band, people differ in how likely they are to be listed, by species, household size and length of residence. That unmodelled variation could bias the estimate in either direction, and none of those characteristics was recorded on all three lists.

A reference day is a convenience. Residents come and go, and the clinic window of 180 standard days and the ration roll's strike-off dates blur the closed population the method needs. Linkage is the second weak point. A second clerk disagreed on 2.8% of linked pairs and 3.5% of near-miss pairs, and the sensitivity analysis shows that correcting false links at the audited rate would raise the estimate by 5.8%, while correcting missed links at the audited rate would lower it by 0.7%. These are per-pair audit rates applied to pools whose sizes I set, and they do not give the net person-level error.

Intervals here are narrow because the counts are large, and they omit most of what could be wrong. They assume Poisson counts, no clustering of households, independent bands and a three-way interaction of exactly zero. The ration roll lists persons by household, so clustering is likely, and it would widen the intervals by an amount I could not estimate.

Reach is set by the lists. The method can say, with some confidence, that the census misses more residents with depth and that the lists move together. It cannot say what the district's population is, only what it would be if one unobservable quantity took a particular value.

capture-recapturecensus undercountCoruscant lower levelslog-linear modellist dependenceration rollclinic registers

References

  1. Galactic Senate Legislative Research Office (25 BBY). Civic enumeration returns for the lowest inhabited levels of the Ostrel district, household schedules and amendments, 26–25 BBY. Galactic Senate Legislative Research Office Reports, Report LRO-25/11.
  2. Galactic Senate Legislative Research Office (24 BBY). Ration-distribution roll for the lowest inhabited levels of the Ostrel district, as in force on the enumeration day and with later strike-offs. Galactic Senate Legislative Research Office Reports, Report LRO-24/14.
  3. Republic Health Service, Lower District Clinics (25 BBY). Patient registers of clinics C-1 to C-4, Ostrel district, 27–25 BBY. Republic Health Service Clinical Bulletin, Register extract series C-1 to C-4.
  4. Kolsten, A. (26 BBY). Where the enumerator's route ends. Walkway and lift access in the deep levels. Coruscant Journal of Galactic History, Occasional paper 7.
  5. Lorrant, M. (27 BBY). Distance to care and clinic attendance among residents of the deepest inhabited levels. Republic Health Service Clinical Bulletin, Supplement 4.
  6. Lythe, P. (26 BBY). Household registration and proof of residence in the administration of ration rolls. Journal of Galactic Public Finance, Research note 5.
  7. Kessaly, D. (28 BBY). Log-linear models for the overlap of incomplete lists and the count of the unlisted. Proceedings of Applied Speculative Statistics, Series C, 1(1), 12–38.
  8. Lindrevan, R. (27 BBY). Dependence between administrative lists and the bias of two-list estimates. Proceedings of Applied Speculative Statistics, Series C, 1(3), 44–67.
  9. Kerrowan, T. (26 BBY). Likelihood-ratio intervals for the unobserved cell of an overlap table. Proceedings of Applied Speculative Statistics, Series C, 2(1), 5–29.
Read this article inside the full journal experience — browse by faculty, search across universes, and filter by discipline.
Open in Uncited Press →