Uncertainty: distinguish precision from missing people
A sample estimate varies across repeated samples. A confidence procedure is evaluated by how often its intervals cover the population value in repeated use under its assumptions. More precision does not repair a biased frame or an ambiguous question.
Your task: Calculate and draw the interval on a fixed 0–100% scale. Repeat for a standard error of 4 points, keeping the estimate fixed. Label sampling uncertainty separately from nonresponse, coverage and measurement errors, and state the assumptions needed for each conclusion.
Open the investigation · Read the explanation · Your saved evidence
Starting recap and vocabulary
Distinguish the record, the interpretation and the conditions under which the interpretation holds. Identify who produced the record and who is absent.
- Context
- The relationships, circumstances and purpose needed to interpret a record.
- Validity
- Evidence supporting a particular interpretation and use, within stated conditions.
What does doubling the standard error do here?
Calculate and draw the interval on a fixed 0–100% scale. Repeat for a standard error of 4 points, keeping the estimate fixed. Label sampling uncertainty separately from nonresponse, coverage and measurement errors, and state the assumptions needed for each conclusion.
Complete the specified diagram, record or table and explain one assumption in plain language. All supplied evidence can be examined without collecting personal information.
Make a provisional claim
Loading saved evidence…
Inspect evidence and test an explanation
The investigation controls appear when your local evidence is ready.
Connect evidence and mechanism
Uncertainty: distinguish precision from missing people
Predict with context
A sample estimate varies across repeated samples. A confidence procedure is evaluated by how often its intervals cover the population value in repeated use under its assumptions. More precision does not repair a biased frame or an ambiguous question.
Predict how the conclusion changes when the missing context or competing account is added. State your reason before reading feedback.
Inspect the authentic source
BLS explains a 90% confidence interval using an estimate plus or minus 1.645 standard errors. Standard errors quantify sampling variability under a survey’s design; they do not automatically account for every source of error.
U.S. Bureau of Labor Statistics — source note. CPS methodological factsheet, consulted October 2026. Authored paraphrase of an authentic official or educational source. This note is not a verbatim primary excerpt. Its description, purpose and limits must stay attached.
Worked investigation
Synthetic estimate is 40%, with standard error 2 percentage points. Margin = 1.645 × 2 = 3.29 points, giving 36.71% to 43.29%. This is not a claim that 90% of people lie in that interval. If an important group was excluded from the sampling frame, the narrow range cannot establish coverage of that group’s experience.
Produce an inspectable result
Calculate and draw the interval on a fixed 0–100% scale. Repeat for a standard error of 4 points, keeping the estimate fixed. Label sampling uncertainty separately from nonresponse, coverage and measurement errors, and state the assumptions needed for each conclusion.
Keep the original record available alongside your analysis. Label invented details, attributed descriptions and your own interpretations separately. For numeric work, show the intermediate arithmetic and units; for interpretive work, point to the specific feature supporting a claim. A reader should be able to identify where another interpretation could enter.
Transfer
A huge opt-in poll has a tiny model-based interval. What concern remains?
Write the answer and its reason before opening feedback. Identify one similarity that makes transfer useful and one difference that limits it. State a specific next observation rather than saying only that more information is needed.
Limits and responsibility
These are supplied teaching standard errors, not estimates computed from a complex survey. Comparing two estimates needs the standard error of their difference, including dependence; interval overlap alone is not a general significance test.
A source can be authoritative about its own procedure or announcement without establishing every outcome that someone might infer from it. Ask whether the record describes an action, expresses a principle, documents an observation or interprets another source. Those are different evidential roles.
A difference needs its own uncertainty
Use three historical 2015 ACS one-year estimates reproduced in the Census Bureau's 2018 handbook. The measure is the share of all households consisting of a householder age 65 or over living alone. It is not the percentage of older people who live alone, and it does not measure health or care needs.
Predict: Does a larger point estimate establish a population difference? Can overlapping individual intervals rule a difference out? Write an initial claim before comparing cases.
Writing is temporary; copy the report before leaving. Existing lesson work is unchanged. All records are supplied public aggregates or explicitly fictional; do not collect personal household information.
Keep the source margins separate from the selected interval
| State | Estimate | 90% margin |
|---|---|---|
| Florida | 12.6% | ±0.2 |
| Arizona | 10.5% | ±0.3 |
| Indiana | 10.8% | ±0.2 |
This selected pair compares distinct states in the same ACS year. Florida: 12.6% ±0.2 points; Arizona: 10.5% ±0.3 points. These supplied margins are always 90% margins, even when you select 95% intervals.
Recover each standard error: SE = source 90% margin ÷ 1.645. Here SE(A) = 0.121581 and SE(B) = 0.182371 percentage points. The selected multiplier is 1.645; estimate ± multiplier × SE gives 90% individual intervals.
Florida: [12.4, 12.8]%; Arizona: [10.2, 10.8]%. These intervals do not overlap. Check the difference directly rather than treating overlap as a general test.
Test the design assumption before using the formula
The observed states are nonoverlapping. Unchecking is a hypothetical design challenge: imagine one group contains the other, or periods share observations. It does not change the source states. Without appropriate covariance/design information, this workbench refuses to calculate difference uncertainty.
Point difference A − B = 2.1 percentage points, not a percent change. For the selected comparison, uncertainty about the difference uses SE(difference) = √(SE(A)² + SE(B)²) = 0.219182 points.
Difference interval = 2.1 ± 0.360555 = [1.739445, 2.460555] points. It excludes zero at the selected 90% level. The method distinguishes these estimates at this threshold; this does not identify a cause or practical importance.
Repeated-sampling confidence describes a procedure's long-run coverage under its assumptions. It is not the percentage of households inside these bounds or a probability assigned to a fixed unknown difference. Sampling uncertainty does not measure all coverage, nonresponse or measurement bias. Repeatedly selecting comparisons also changes error risks; this activity applies no multiple-comparison adjustment.
Transfer and defend a qualified result
Try Indiana–Arizona, then the separate fictional pair (52% ±1.2 and 50% ±1.2). At 90%, inspect both individual overlap and the difference interval; then change to 95%. Explain what changed, what did not, and why a threshold crossing is not evidence that the underlying population changed. Finally challenge independence and name the missing design information.
Compare with the reasoning guide
Florida–Arizona has a 2.1-point difference, SE about 0.219182 and z about 9.581. Indiana–Arizona differs by only 0.3 points with the same SE and z about 1.369; its 90% difference interval includes zero. That is insufficient evidence of difference under this procedure, not proof of equality.
For the fictional pair, the individual 90% intervals overlap, but the difference interval is approximately [0.302944, 3.697056] points and excludes zero. At 95% it is approximately [−0.022024, 4.022024]; use the live calculation for the exact endpoints. The point difference stays 2. Practical significance requires context; no threshold or aggregate comparison identifies causation.
Sources and reproducibility
Census 2018 handbook chapter 7, printed pages 47–48, supplies the three observed estimates, 90% margins and comparison method. Page 45's DP02 table and 2015 DP02 metadata establish the total-household denominator (DP02_0012PE/PM). The handbook's abbreviated wording about older householders is not a conditional rate among older people.
These selected factual values were transcribed from the official handbook and independently checked. They are not a newly retrieved API dataset, microdata or replicate-weight analysis. The handbook cautions about overlapping groups/periods and nonsampling errors. The published source precision is retained; results are approximations from rounded inputs. Federal source facts are public domain with attribution; source illustrations are not copied.
Copy your uncertainty report
Guidance exposure persists through reset during this mount, clears on reload, and cannot verify an unseen first attempt. Writing remains learner-authored and ungraded.
A five-year estimate is not its final year
The previous workbench refuses an unsupported dependent comparison. Here a documented historical Census approximation supplies a bounded method for one overlapping-period case; it does not authorize the independent formula for every dependent dataset.
Predict: “The 2010–2014 estimate is 0.5 percentage points below the 2006–2010 estimate, so the measure fell by 0.5 points from 2010 to 2014.” Identify what the dates actually describe and which observations overlap.
Writing is temporary. Copy the report to retain it; existing lesson experiments are unaffected.
Reconstruct the actual source comparison
| Period | Estimate | 90% margin |
|---|---|---|
| 2006–2010 | 13.6% | ±2.3 points |
| 2010–2014 | 13.1% | ±2.3 points |
Measure: share of all households consisting of a householder age 65+ living alone. It is not a conditional percentage among older adults, a health measure or an observation about individuals. The two periods share 2010: one of five years, so C = 1/5 = 0.2.
Documented approximation selected. Standard error of each estimate = 2.3/1.645. Independence would give SE = 1.97732 points. The handbook adjustment multiplies that by √(1 − C), here √0.8.
Selected SE = 1.768569 points; later minus earlier = -0.5 points. Its approximate 90% difference interval is [-3.409295, 2.409295] points, and z = -0.282714. Both the source approximation and the ignored-overlap calculation include zero here. Neither establishes equality or the magnitude of any single-year change.
More overlap lowers the estimated variance of the difference under this approximation because shared information partly cancels. It does not turn either period into a one-year estimate or cure nonsampling error. The overlap fraction is computed from the actual dates, not tuned to obtain a preferred result.
Find the inference failure with a separate fictional record
The following invented annual index values are an equal-weight arithmetic illustration, not ACS observations or its estimation method. Predict whether the same two five-year averages require the same change between their ending years.
Annual values for 2006–2014: 10, 10, 10, 10, 10, 12, 12, 12, 12. Earlier mean = 10; later mean = 11.6; difference = 1.6 index units. But the 2014 minus 2010 endpoint change is 2 units.
Cancel the shared value algebraically: window difference = (sum of 2011–2014 − sum of 2006–2009)/5 = (48 − 40)/5 = 1.6. This compares four entering years with four leaving years; it does not isolate the final two endpoints.
Transfer: compare A and B. Preserve both means, but explain the opposite endpoint directions. Then return to Holmes County: write a supported period-to-period claim and identify the data needed for a genuine 2010-to-2014 annual comparison. Do not infer unreported annual ACS values from these two aggregates.
Compare with the reasoning guide
Both fictional sequences have earlier mean 10 and later mean 11.6, yet their endpoint changes are +2 and −8. Thus matching period means cannot determine endpoint change. The numerical illustration proves an information limitation; it does not reproduce ACS weighting.
The observed Holmes difference is −0.5 percentage points. The overlap-aware historical approximation yields |z| about 0.283, below 1.645; there is insufficient evidence of a period difference at this threshold. The source recommends nonoverlapping periods where feasible. A qualified report names periods, geography, denominator, source precision and method; it does not interpret a nonsignificant result as equality, assign a cause or infer individual experience.
Primary source and scope of the approximation
Census 2018 ACS handbook chapter 7, printed page 49, supplies both Holmes County estimates, 90% MOEs, overlap example and formula. The formula was checked visually against the PDF. Printed page 45 and DP02 metadata support the denominator clarification used in the preceding activity. The source gives an approximation for overlapping ACS multiyear periods; it is not a general covariance estimator for overlapping groups, arbitrary panels or unequal survey designs. We retain rounded published inputs and do not claim a microdata or replicate-weight analysis.
Source federal factual material is public domain with attribution. The two annual sequences are original fictional arithmetic examples. No source illustration is republished; no network or paid API is required.
Copy the overlapping-period report
Reset clears writing and restores the documented correction and sequence A, while retaining this mount's guidance exposure. Reload clears that exposure. Ungraded writing and open feedback do not establish independent mastery.
Extend
High school: Complete the specified diagram, record or table and explain one assumption in plain language. All supplied evidence can be examined without collecting personal information.
College: Add a bounded authentic case or methodological record that tests the weakest assumption. Document version, access, selection and interpretation, and present a justified alternative conclusion.
Same eight records, different sampling uncertainty
A fictional archive contains 16 records in four equal boxes of four. A tag is 1 if a specified catalog feature is present, 0 otherwise. All 16 tags are known in this teaching population: eight are 1, so the true proportion is 50%. This is not a dataset about people and no collection activity is required.
Design A chooses eight individual records uniformly without replacement from 16. Design B chooses two of four boxes uniformly without replacement, then includes all eight records in those boxes. Every individual record has inclusion probability 1/2 in either design. That does not make records within a sampled box independent.
Predict: can both designs be unbiased while having different variability? Does equal sample size imply equal precision? Explain your expected distribution before inspecting all possible box selections.
- Box 1: 0, 0, 0, 0; 0 tagged records of 4.
- Box 2: 0, 0, 0, 0; 0 tagged records of 4.
- Box 3: 1, 1, 1, 1; 4 tagged records of 4.
- Box 4: 1, 1, 1, 1; 4 tagged records of 4.
Selected records: 0, 0, 0, 0, 0, 0, 0, 0. Tagged=0 of 8; estimated proportion=0%. Error from the known population 50% is -50 percentage points. One selected sample's error is not the design's standard error.
Enumerate the sampling distributions
There are six equally likely box pairs. Individual SRS has 12,870 equally likely eight-record subsets. The table counts how many yield each tagged count; these are exact finite possibilities, not simulated repetitions. Pair choice above does not change the design distribution.
| Tagged /8 | Estimate | SRS subsets /12870 | SRS probability | Box pairs /6 | Box probability |
|---|---|---|---|---|---|
| 0 | 0% | 1 | 0.00777% | 1 | 16.666667% |
| 1 | 12.5% | 64 | 0.49728% | 0 | 0% |
| 2 | 25% | 784 | 6.091686% | 0 | 0% |
| 3 | 37.5% | 3136 | 24.366744% | 0 | 0% |
| 4 | 50% | 4900 | 38.073038% | 4 | 66.666667% |
| 5 | 62.5% | 3136 | 24.366744% | 0 | 0% |
| 6 | 75% | 784 | 6.091686% | 0 | 0% |
| 7 | 87.5% | 64 | 0.49728% | 0 | 0% |
| 8 | 100% | 1 | 0.00777% | 1 | 16.666667% |
Mean across possible samples: SRS=0.5, box design=0.5, both compared with population 0.5. Exact design variance is probability-weighted squared error from 0.5: SRS=0.016667, box=0.083333, in proportion-squared units. Standard errors (square roots) are 12.909944 and 28.867513 percentage points.
Design effect=box variance/SRS variance=5. This is a variance ratio, not the ratio of standard errors. It applies to this estimator, population and pair of designs. Both sampling fractions are 1/2, and calculations already account for sampling without replacement. An independent-with-replacement binomial variance would answer a different question.
Why the individual counts have this shape
To obtain k tagged records, choose k of the eight tagged records and 8−k of the eight untagged records. Multiply those numbers of choices. For example k=4 gives 70×70=4900 subsets. Across k=0..8 the counts sum to 12870. Box counts come from directly listing all six pairs. No normal approximation or guessed effective sample size is needed.
Transfer: reorganize the same tags
Explain why mixed boxes preserve the 16-record population and individual-SRS distribution but eliminate the variation among whole-box estimates. Does that mean cluster sampling is always better? Identify the unrealistic balance that creates this example. Name a coverage or recording error that could remain even with zero sampling variance.
Reasoning guide
Homogeneous boxes yield estimates 0, .5, .5, .5, .5, 1; variance 1/12 and SE 28.867513 percentage points. Individual SRS variance 1/60, SE 12.909944 points. Design effect 5, standard-error ratio √5. Mixed boxes always yield .5, so box variance and design effect are 0; this is exact balance in a contrived finite population, not a general survey recommendation.
These are exact repeated-selection variances computed using a fully known population. A real survey usually has only its sample and must estimate uncertainty from its actual design, weights and cluster/stratum information. This packet does not estimate a confidence interval from one pair or replace CPS/ACS design variance. Equal inclusion probabilities do not erase joint selection dependence.
Method source and original population
University of Michigan Cross-Cultural Survey Guidelines, sample-design Appendix C supplies context for finite-population sampling and design effects. The archive, tags and exact enumerations are original teaching material, not source data or a real survey result. No person-level or clinical data are used.
Copy the sampling-design report
Temporary writing is limited to 3,000 characters per field. Reset restores homogeneous/first pair and blank writing, retaining known guide/report exposure this mount; reload clears it. Guided exploration is not independent assessment.
Worked example
Synthetic estimate is 40%, with standard error 2 percentage points. Margin = 1.645 × 2 = 3.29 points, giving 36.71% to 43.29%. This is not a claim that 90% of people lie in that interval. If an important group was excluded from the sampling frame, the narrow range cannot establish coverage of that group's experience.
Check the reasoning, then transfer it
Loading response history…
Keep your evidence and choose a next step
Loading saved lesson status…