Margin of Error, Confidence Intervals & Sample Size: A Plain-English Guide

The single most important concept in polling is not sample size. It is not margin of error. It is random sampling. Random sampling is what allows a small group to represent a large one. Without it, a poll is just an expensive opinion survey of whoever happened to answer the phone.

In a true random sample, every person in the target population has a known, non-zero probability of being selected. In NigeriaPolls' face-to-face surveys, this means we use stratified random sampling across all 774 local government areas. We randomly select villages and city blocks, then randomly select households within those areas, then randomly select one adult per household. No one gets to volunteer. No one gets excluded because they are hard to reach. The randomness does the work.

Why is randomness so powerful? Because it breaks the link between who you are and whether you are interviewed. In a random sample, rich and poor, educated and uneducated, urban and rural, young and old — all have the same chance of being selected (within their stratum). This means the sample automatically mirrors the population on every characteristic, even characteristics we do not explicitly measure. It is as if the population is holding up a mirror to itself.

The key assumption, of course, is that the "stirring" is thorough. In polling, this means the sampling frame must cover the entire population, and the selection process must be truly random. If you only stir the top layer of the soup, your spoonful will not represent the bottom. Similarly, if you only poll people with smartphones, you will miss the millions of Nigerians who do not own one. This is why NigeriaPolls uses face-to-face interviews as our primary mode: it reaches everyone, not just the digitally connected.

The margin of error (MOE) is the most misunderstood number in polling. It is also one of the most important. The margin of error tells you how far the poll result might be from the true population value, purely due to the fact that you interviewed a sample rather than everyone.

Here is what it means in practice. If a poll shows Candidate A with 42% support and the margin of error is ±3%, then Candidate A's true support in the entire population is probably between 39% and 45%. The pollster is not saying the true value is exactly 42%. They are saying 42% is their best estimate, and the true value is likely within 3 points of that estimate.

The margin of error is calculated using this formula:

Several things are worth noting about the margin of error:

• It is largest when the result is closest to 50%. A result of 50% has the maximum margin of error. A result of 10% has a smaller margin of error because there is less room for variation.
• It applies to each percentage individually. If Candidate A is at 42% ±3% and Candidate B is at 38% ±3%, the gap between them is 4 points, but the true gap could be anywhere from -2 to +10. The candidates could actually be tied.
• It does not include non-sampling error. The margin of error only captures random sampling error. It does not account for bias from bad question wording, non-response, social desirability, or coverage problems.

The "Statistical Tie"

When two candidates are within the margin of error of each other, pollsters call it a statistical tie. This does not mean the candidates are literally tied; it means the poll cannot confidently say who is ahead. In Nigerian media, you will often see headlines like "Candidate A Leads by 2 Points!" without mentioning that the margin of error is ±3%. A 2-point lead with a ±3% margin is not a lead at all — it is noise.

At NigeriaPolls, we always report margins of error alongside our results, and we flag races that are too close to call. We would rather be honest about uncertainty than generate misleading headlines.

If the margin of error is the range, the confidence level is the probability that the range contains the truth. The 95% confidence level is the standard in polling. It means that if you conducted the same poll 100 times, 95 of those times the true population value would fall within the reported margin of error.

A confidence interval combines the point estimate (the poll result), the margin of error, and the confidence level into a single statement. For example: "We are 95% confident that the president's approval rating is between 47% and 53%." This is a more honest and informative way to present a poll than simply saying "50% approve."

Why 95%? Why Not 100%?

You could calculate a 99% confidence interval instead of 95%, but it would be wider. A 99% interval for a poll of 1,000 might be ±4.0% instead of ±3.1%. You would be more "certain," but your estimate would be less precise. The 95% level is a pragmatic compromise between certainty and precision. It means we accept a 5% chance of being wrong — that is, 1 in 20 polls will produce a confidence interval that does not contain the true value. This is considered an acceptable risk in social science and public opinion research.

Confidence Intervals for Differences

When comparing two candidates, the confidence interval for the difference is wider than the margin of error for either candidate alone. This is because both estimates have uncertainty, and those uncertainties compound. If Candidate A is 42% ±3% and Candidate B is 38% ±3%, the 4-point lead has a confidence interval of roughly ±4.2%. This means the true lead could be anywhere from -0.2 to +8.2 — and the race is effectively a toss-up. Many pundits miss this subtlety and overstate small leads.

How do pollsters decide how many people to interview? They use a sample size formula that balances three factors: the desired margin of error, the confidence level, and an estimate of the population proportion. Here is the formula pollsters use:

Example: Calculating for a Nigerian Presidential Poll

Let us say you want to conduct a national presidential poll in Nigeria with a margin of error of ±2.5% at 95% confidence. You do not know the true support for each candidate, so you use p = 0.5 (the most conservative estimate, which maximizes the required sample size). Plugging in the numbers:

The Finite Population Correction

For very small populations (like a single LGA or a small organization), you can apply a finite population correction (FPC)to reduce the required sample size. The FPC accounts for the fact that once you have interviewed a large fraction of a small population, each additional interview adds less new information. For Nigeria's national population of 210 million, the FPC is negligible and can be ignored.

Design Effect: The Hidden Multiplier

The simple formula above assumes a simple random sample. In practice, NigeriaPolls uses cluster sampling(randomly selecting geographic areas, then households within them) because it is logistically impossible to send enumerators to completely scattered locations. Clustering reduces efficiency — respondents in the same cluster tend to be more similar to each other than to the broader population. This inefficiency is captured by the design effect (DEFF), typically around 1.3 to 1.5 for Nigerian surveys. The effective sample size is the nominal sample size divided by the DEFF. To achieve an effective sample of 1,500, we might need to interview 2,000–2,400 people.

Let us compare what different sample sizes actually look like in practice, using hypothetical but realistic Nigerian polling scenarios. This will help you understand the trade-offs between cost, time, and precision.

Scenario 1: A Poll of 500 Respondents

A small Nigerian media outlet commissions a quick poll of 500 adults using phone interviews. The margin of error is ±4.4%. The results show Candidate A at 44%, Candidate B at 40%. The 4-point gap sounds meaningful, but with ±4.4% margins, Candidate A's true support is 39.6–48.4% and Candidate B's is 35.6–44.4%. These ranges overlap significantly. The race could be tied, or Candidate B could actually be ahead. A poll of 500 is useful for detecting large, obvious trends, but it cannot reliably distinguish between close competitors.

Scenario 2: A Poll of 1,500 Respondents (NigeriaPolls Standard)

A NigeriaPolls national survey interviews 1,500 adults face-to-face across all 36 states. The margin of error is ±2.5%. The results show Candidate A at 44%, Candidate B at 40%. Now the confidence intervals are 41.5–46.5% for Candidate A and 37.5–42.5% for Candidate B. There is still some overlap, but it is smaller. We can say Candidate A probably leads, though the race is close. The larger sample also gives us reliable subgroup data: we can credibly report that Candidate A leads by 8 points among men but trails by 2 among women, because each gender subgroup has 750 respondents — enough for meaningful analysis.

Scenario 3: A Poll of 3,000 Respondents

A major international foundation funds a comprehensive pre-election benchmark poll of 3,000 Nigerians. The margin of error is ±1.8%. The precision is excellent: even a 3-point lead is statistically significant. The large sample allows detailed state-by-state analysis, with approximately 80 respondents per state — enough to produce state-level estimates with ±5.5% margins. The trade-off is cost and time: this poll costs more than twice as much as the 1,500-respondent version and takes nearly a month to field. For most purposes, the precision gains are not worth the extra resources.