Understanding Margins of Error¶

The American Community Survey (ACS) is a sample — not a full count of every person in the United States. Because only a fraction of the population is surveyed, every ACS estimate comes with a margin of error (MOE) that quantifies the uncertainty.

This guide explains MOE at three levels of depth, from the basics that everyone needs to understand, through statistical testing for analysts, to the propagation formulas researchers need when deriving new estimates.

Level 1: What is an MOE?¶

Audience: Everyone working with ACS data.

The basics¶

An MOE tells you the range around an estimate where the true value likely falls. The Census Bureau publishes MOEs at the 90% confidence level, meaning there is a 90% chance the true value is within the estimate +/- MOE.

Reading an MOE

Suppose a county’s median household income is reported as:

Estimate: $50,000
MOE: +/- $3,000

This means the true median income for that county is likely between $47,000 and $53,000 (with 90% confidence).

Why do MOEs exist?¶

The ACS surveys about 3.5 million households per year — roughly 2.4% of all US households. This sample is large enough to produce reliable estimates for states and big cities, but for smaller geographies (tracts, block groups, small counties) the sample size drops and uncertainty grows.

Rule of thumb: The smaller the geography or population, the larger the MOE relative to the estimate.

Geography level	Typical reliability
United States	Very precise (tiny MOE)
State	Precise
Large county (500k+)	Good
Small county (<50k)	Moderate to poor
Census tract	Variable — check MOE
Block group	Often very large MOEs

Can I ignore MOEs for mapping?¶

For visualization purposes (choropleth maps, general patterns), it is common practice to map the estimates without displaying MOEs. The spatial patterns are usually meaningful even if individual values are uncertain.

Be cautious with comparisons

MOEs matter most when you are comparing two geographies or tracking changes over time. Two estimates that look different may have overlapping confidence intervals, meaning the difference is not statistically significant. See Level 2 for how to test this.

MOE in PyPUMS output¶

By default, get_acs() returns both the estimate and MOE in tidy format:

SourceResult

from pypums import get_acs

df = get_acs(
    "county",
    variables="B19013_001",
    state="CA",
    year=2022,
)

print(df[["NAME", "variable", "estimate", "moe"]].head())

                       NAME    variable  estimate    moe

0 Alameda County, California B19013_001 122488 1231 1 Alpine County, California B19013_001 101125 17442 2 Amador County, California B19013_001 74853 6048 3 Butte County, California B19013_001 66085 2261 4 Calaveras County, California B19013_001 77526 3875

Notice how Alpine County (population ~1,100) has a much larger MOE relative to its estimate compared to Alameda County (population ~1.6 million).

Level 2: When does MOE affect my analysis?¶

Audience: Journalists, policy analysts, data scientists.

Testing if a difference is statistically significant¶

The most common question is: “Is the difference between two estimates real or just noise?” PyPUMS provides the significance() function to answer this.

from pypums import significance

Function signature¶

significance(
    est1,           # first estimate
    est2,           # second estimate
    moe1,           # MOE of first estimate (at 90% confidence)
    moe2,           # MOE of second estimate (at 90% confidence)
    *,
    clevel=0.90,    # confidence level for the test: 0.90, 0.95, or 0.99
) -> bool           # True if difference is statistically significant

Example: comparing two cities’ income¶

SourceResult

from pypums import significance

# Suppose:
# City A: estimate = $85,000, MOE = $4,000
# City B: estimate = $78,000, MOE = $5,000

is_different = significance(85000, 78000, 4000, 5000, clevel=0.90)
print(is_different)

True

The $7,000 gap is statistically significant at the 90% confidence level.

Confidence levels¶

The clevel parameter controls how strict the test is:

Confidence level	Z-score	Interpretation
`0.90`	1.645	Standard Census threshold — “likely different”
`0.95`	1.960	Common in academic research — “probably different”
`0.99`	2.576	Very strict — “almost certainly different”

SourceResult

# Same comparison at different confidence levels
print(f"90% confidence: {significance(85000, 78000, 4000, 5000, clevel=0.90)}")
print(f"95% confidence: {significance(85000, 78000, 4000, 5000, clevel=0.95)}")
print(f"99% confidence: {significance(85000, 78000, 4000, 5000, clevel=0.99)}")

90% confidence: True 95% confidence: False 99% confidence: False

How the test works

significance() converts 90% MOEs to standard errors, computes the standard error of the difference, and checks whether the observed difference exceeds the critical value:

Convert MOE to SE: $SE = MOE / 1.645$
SE of the difference: $SE_{diff} = \sqrt{SE_1^2 + SE_2^2}$
Significant if: $|est_1 - est_2| > z \times SE_{diff}$

Scaling MOE to different confidence levels¶

The Census API publishes MOEs at 90% confidence. If your analysis requires 95% or 99% confidence intervals, use the moe_level parameter:

# Get data with MOE scaled to 95% confidence
df_95 = get_acs(
    "county",
    variables="B19013_001",
    state="CA",
    year=2022,
    moe_level=95,
)

The scaling formula divides by the 90% z-score and multiplies by the target z-score:

$$ MOE_{new} = MOE_{90} \times \frac{z_{new}}{1.645} $$

Target level	Multiplier	Effect
90%	1.000	No change (default)
95%	1.192	MOE increases ~19%
99%	1.566	MOE increases ~57%

The MOE-to-SE relationship¶

All MOE math starts from converting to standard errors:

$$ SE = \frac{MOE}{z} $$

Confidence level	Z-score
90%	1.645
95%	1.960
99%	2.576

Since the Census publishes at 90%:

$$ SE = \frac{MOE_{90}}{1.645} $$

Level 3: MOE propagation for derived estimates¶

Audience: Researchers, statisticians, advanced analysts.

When you create a new estimate by combining Census variables — adding them up, computing a ratio, calculating a proportion — you need to propagate the MOEs through the arithmetic. PyPUMS implements the standard Census Bureau formulas from the ACS General Handbook, Chapter 8.

`moe_sum()` — combining margins of error for sums¶

When adding estimates together (e.g., summing age groups to get a total):

from pypums import moe_sum

Formula:

$$ MOE_{sum} = \sqrt{\sum_{i} MOE_i^2} $$

Example: Combine two age groups to get total population 18-34.

SourceResult

from pypums import moe_sum

# Age 18-24: estimate=5000, moe=800
# Age 25-34: estimate=7000, moe=600

total_est = 5000 + 7000  # 12,000
total_moe = moe_sum([800, 600])  # sqrt(800^2 + 600^2) = 1000.0

print(f"Population 18-34: {total_est} +/- {total_moe:.0f}")

Population 18-34: 12000 +/- 1000

Note

This formula assumes the estimates are independent (uncorrelated). If the variables come from the same table and are subsets of the same total, the formula still provides a conservative approximation.

`moe_ratio()` — margins of error for ratios¶

When dividing one estimate by another where the numerator is not a subset of the denominator:

from pypums import moe_ratio

Formula:

$$ MOE_{ratio} = \frac{\sqrt{MOE_{num}^2 + \left(\frac{num}{denom}\right)^2 \times MOE_{denom}^2}}{denom} $$

Example: Ratio of renters to homeowners.

SourceResult

from pypums import moe_ratio

# Renters: estimate=3000, moe=400
# Owners:  estimate=7000, moe=500

ratio_moe = moe_ratio(num=3000, denom=7000, moe_num=400, moe_denom=500)
ratio_est = 3000 / 7000  # 0.4286

print(f"Renter-to-owner ratio: {ratio_est:.3f} +/- {ratio_moe:.3f}")

Renter-to-owner ratio: 0.429 +/- 0.065

`moe_prop()` — margins of error for proportions¶

When the numerator is a subset of the denominator (e.g., percent of population with a bachelor’s degree):

from pypums import moe_prop

Formula:

$$ MOE_{prop} = \frac{\sqrt{MOE_{num}^2 - \hat{p}^2 \times MOE_{denom}^2}}{denom} $$

where $\hat{p} = num / denom$.

Negative radicand fallback

When $MOE_{num}^2 < \hat{p}^2 \times MOE_{denom}^2$ (the expression under the square root is negative), the proportion formula is undefined. In this case, moe_prop() automatically falls back to moe_ratio(), which uses addition instead of subtraction under the radical.

This is the recommended approach from the Census Bureau’s handbook.

Example: Proportion of population with a bachelor’s degree.

SourceResult

from pypums import moe_prop

# Bachelor's holders: estimate=15000, moe=1200
# Total population 25+: estimate=50000, moe=800

prop_moe = moe_prop(num=15000, denom=50000, moe_num=1200, moe_denom=800)
prop_est = 15000 / 50000  # 0.30

print(f"Bachelor's rate: {prop_est:.1%} +/- {prop_moe:.4f}")

Bachelor’s rate: 30.0% +/- 0.0235

`moe_product()` — margins of error for products¶

When multiplying two estimates together:

from pypums import moe_product

Formula:

$$ MOE_{product} = \sqrt{est_1^2 \times MOE_2^2 + est_2^2 \times MOE_1^2} $$

Example: Estimated total income (households x median income).

SourceResult

from pypums import moe_product

# Households: estimate=10000, moe=500
# Avg income: estimate=60000, moe=3000

product_moe = moe_product(est1=10000, est2=60000, moe1=500, moe2=3000)
product_est = 10000 * 60000  # 600,000,000

print(f"Total income: ${product_est:,.0f} +/- ${product_moe:,.0f}")

Total income: $600,000,000 +/- $42,426,407

Practical workflow: deriving a custom estimate¶

Here is a complete example that computes the percentage of housing units that are vacant, with a properly propagated MOE:

SourceResult

from pypums import moe_prop

# Get occupied and total housing units
df = get_acs(
    "county",
    variables=["B25002_001", "B25002_003"],
    state="CA",
    year=2022,
    output="wide",
)

# B25002_001E = total housing units
# B25002_003E = vacant housing units
df["vacancy_rate"] = df["B25002_003E"] / df["B25002_001E"]

# Propagate MOE for each row
df["vacancy_moe"] = df.apply(
    lambda row: moe_prop(
        num=row["B25002_003E"],
        denom=row["B25002_001E"],
        moe_num=row["B25002_003M"],
        moe_denom=row["B25002_001M"],
    ),
    axis=1,
)

print(df[["NAME", "vacancy_rate", "vacancy_moe"]].head())

                       NAME  vacancy_rate  vacancy_moe

0 Alameda County, California 0.059465 0.002781 1 Alpine County, California 0.730983 0.040091 2 Amador County, California 0.163079 0.023066 3 Butte County, California 0.104761 0.009734 4 Calaveras County, California 0.375095 0.020489

Z-score reference table¶

All MOE calculations use z-scores from the standard normal distribution:

Confidence level	Z-score	Use case
90%	1.645	Census Bureau default; `moe_level=90`
95%	1.960	Standard academic threshold; `moe_level=95`
99%	2.576	Conservative threshold; `moe_level=99`

The relationship between MOE and standard error at any confidence level is:

$$ MOE = z \times SE $$

$$ SE = \frac{MOE}{z} $$

Source¶

All formulas in this guide come from the U.S. Census Bureau’s ACS General Handbook, Chapter 8: Calculating Measures of Error for Derived Estimates. This is the authoritative reference for working with ACS margins of error.

Quick reference¶

Task	Function	Formula
Combine estimates by addition	`moe_sum(moe)`	$\sqrt{\sum MOE_i^2}$
Ratio (numerator not subset of denominator)	`moe_ratio(num, denom, moe_num, moe_denom)`	$\frac{\sqrt{MOE_n^2 + r^2 \cdot MOE_d^2}}{denom}$
Proportion (numerator is subset)	`moe_prop(num, denom, moe_num, moe_denom)`	$\frac{\sqrt{MOE_n^2 - p^2 \cdot MOE_d^2}}{denom}$
Product of two estimates	`moe_product(est1, est2, moe1, moe2)`	$\sqrt{est_1^2 \cdot MOE_2^2 + est_2^2 \cdot MOE_1^2}$
Test if difference is significant	`significance(est1, est2, moe1, moe2, clevel=)`	$\|est_1 - est_2\| > z \cdot SE_{diff}$
Scale MOE confidence level	`get_acs(..., moe_level=95)`	$MOE_{new} = MOE_{90} \times \frac{z_{new}}{1.645}$

All functions are available from the top-level pypums namespace:

from pypums import moe_sum, moe_ratio, moe_prop, moe_product, significance

Understanding Margins of Error¶

Level 1: What is an MOE?¶

The basics¶

Why do MOEs exist?¶

Can I ignore MOEs for mapping?¶

MOE in PyPUMS output¶

Level 2: When does MOE affect my analysis?¶

Testing if a difference is statistically significant¶

Function signature¶

Example: comparing two cities’ income¶

Confidence levels¶

Scaling MOE to different confidence levels¶

The MOE-to-SE relationship¶

Level 3: MOE propagation for derived estimates¶

moe_sum() — combining margins of error for sums¶

moe_ratio() — margins of error for ratios¶

moe_prop() — margins of error for proportions¶

moe_product() — margins of error for products¶

Practical workflow: deriving a custom estimate¶

Z-score reference table¶

Source¶

Quick reference¶

See Also¶

`moe_sum()` — combining margins of error for sums¶

`moe_ratio()` — margins of error for ratios¶

`moe_prop()` — margins of error for proportions¶

`moe_product()` — margins of error for products¶