Reference

Glossary

Plain-English explanations of the statistical and method terms you will meet around Strata. No prior stats background assumed — if a definition is doing its job, it should make the term feel smaller, not scarier.

Each entry pairs a precise definition with a plain-language example (and the occasional lighter aside) to make the idea click — the examples illustrate the term rather than restate the formal methodology.

95% confidence interval

A confidence interval is a range around a survey estimate that reflects sampling uncertainty — the fact that we surveyed a sample, not every American. A 95% confidence interval is constructed so that, across many repeated samples, about 95% of such intervals would contain the true population value. Wider intervals signal more uncertainty (often from smaller samples); narrower intervals signal more precision. It is not the probability that the true value falls inside this one particular interval.

In plain terms

If we estimate 62% of adults use a platform with a 95% interval of 59%–65%, read it as: "our best guess is 62%, and the true number is very plausibly somewhere between 59 and 65." It is the estimate wearing a margin of safety rather than pretending to be a single exact number — the honest version of "about 62%, give or take."

Effect size (correlation magnitude)

An effect size measures how large a relationship is, separately from whether it is statistically detectable. For the correlations in Strata, the relevant effect size is the magnitude of ρ — how far it sits from zero, regardless of sign. Strata reads |ρ| in four bands: negligible (below 0.1), weak (0.1 to 0.3), moderate (0.3 to 0.5), and strong (0.5 or above). Negligible associations are drawn in muted grey so that noise-level relationships do not read as findings. In practice, correlations in survey data like this tend to be modest: most here fall below 0.4, so a “moderate” association is often the strongest you will encounter — which is normal, not disappointing.

In plain terms

Direction tells you which way two things move together; effect size tells you how much. A ρ of 0.05 is technically positive but so faint that it may just be a mirage in the data. A ρ of 0.45 is a real, noticeable pull. With a big enough sample, even the mirage can become “statistically significant” — which is exactly why we also look at magnitude. Significance says “it is probably not zero”; effect size asks “but is it big enough to care about?”

Effective sample size

When responses are weighted, they no longer all count equally, so a weighted estimate behaves as if it came from a smaller sample than the raw headcount. The effective sample size (sometimes written n_eff) is that adjusted number — it captures how much statistical precision remains after weighting. The more uneven the weights, the further the effective sample size falls below the actual number of respondents.

In plain terms

Picture a tug-of-war where some players are far stronger than others. Even with 1,000 people on the rope, if a handful do most of the pulling it "effectively" behaves like a smaller, lopsided team. Weighting can do the same to a sample: 1,000 respondents might carry the statistical weight of, say, 700. Strata reports the honest ~700, not the flattering 1,000.

Margin of error

The margin of error is the "± value" you add to and subtract from an estimate to get its 95% confidence interval. It packs sampling uncertainty into a single number: a result of 62% with a ±3-point margin of error means the 95% interval runs from roughly 59% to 65%. Margins of error shrink as sample size grows, and they say nothing about non-sampling problems such as poorly worded questions.

In plain terms

It is the "give or take" attached to a poll: "62%, give or take 3 points" is a margin of error of 3. When two numbers sit within each other's margins of error, treating one as clearly bigger than the other is exactly how people end up confidently wrong on election night.

Point-biserial correlation

A point-biserial correlation measures the association between a binary variable (e.g., uses a platform: yes/no) and a continuous or ordinal one (e.g., a wellbeing score). It is algebraically the Pearson correlation with one variable coded 0/1, and it ranges from −1 to +1. It answers a simple question: do the "1"s tend to score higher or lower than the "0"s, and how consistently?

In plain terms

Split people into coffee-drinkers (1) and non-drinkers (0) and compare their typing speeds. A positive point-biserial means drinkers tend to type faster; negative means slower. Same idea as any correlation — one of the two variables just happens to be a plain yes/no switch.

Reverse-coded

Where the underlying scale is reverse-coded, the flip is applied at the cleaning stage; the UI does not need to re-flip.

In plain terms

Some survey questions are worded backwards on purpose — "I feel calm" mixed in among "I feel anxious" — so people do not just autopilot the same answer down the page. Before analysis, those flipped items are turned back around so that "higher" always means the same direction (say, more distress). Reverse-coding is simply un-flipping the questions that were flipped on purpose. Yes, researchers do this to themselves voluntarily.

Spearman's ρ

All pairwise correlations in Strata are Spearman's ρ, computed per wave. Spearman handles ordinal Likert and skewed count variables better than Pearson and does not assume a linear relationship. Where the underlying scale is reverse-coded, the flip is applied at the cleaning stage; the UI does not need to re-flip. Correlation is not causation, and small samples or small ρ values should be interpreted with caution.

In plain terms

Think of ρ as a number between −1 and +1. Positive means the two things rise together: as one score goes up, so does the other (more coffee, faster typing — allegedly). Negative means they move in opposite directions (more coffee, fewer hours of sleep — definitely). Near 0 means they are basically ignoring each other. And "ρ" is just the Greek letter rho; statisticians like Greek letters because they make ordinary ideas look more intimidating than they are.

Suppression (n < 30)

Cells with fewer than 30 respondents are suppressed by design. This is because statistics based on small samples are at a higher risk of being unreliable and misleading the viewer. Given the broad interest in social media and the contentious debates around the effects of social media on people, we chose to suppress cases from the graphs where the sample size was too small for us to trust. Throughout Strata, these instances are marked with an “insufficient n” label.

In plain terms

Imagine describing "the average opinion" of a group when only four people answered — one person having a weird morning could swing the whole result. To avoid big conclusions from tiny groups, Strata simply hides any cell with fewer than 30 people instead of showing a number it does not trust. Better an honest blank than a confident wrong answer.

Tertile / fixed split

Tertiles split a measure into three groups. With sample-based tertiles, the cut points are chosen so each wave’s respondents divide into roughly equal thirds. With a fixed split, the cut points are held constant across waves so the groups mean the same thing over time, even if their sizes differ. These types of splits are often more informative than simply splitting the data on a fixed threshold (e.g., the middle point of the distribution). Strata notes which approach a chart uses, because it changes how comparisons across waves should be read.

In plain terms

Like sorting a class into bottom third, middle third, and top third by score. "Tertile" is just the fancy word for those thirds — quartiles cut into four, tertiles into three. No one will judge you for picturing three buckets.

Wave

For longitudinal panel surveys, the same group of people are surveyed at multiple points in time. Each point in time represents a wave. Currently, Strata covers six waves of the Understanding America Study (UAS514–UAS519), collected between 2023 and 2025. Each wave has its own field dates and sample size; the About page lists the full table.

In plain terms

Imagine marking a child’s height on the same doorframe every birthday. Each mark is one wave — the same person, the same measurement, taken again at a new point in time — and lining the marks up is how you see growth. Strata does this with attitudes instead of height: the same panel of people answer the same kinds of questions at six points between 2023 and 2025, so we can see what shifted and what held steady. The key is that it is the same people each time, which is what lets us track change rather than just compare different crowds.

Weighted estimate

Every precomputed row carries weighted estimates (weighted_value, with matching standard errors, confidence intervals, and effective sample sizes). UAS provides probability weights to adjust for panel design and non-response; weighted estimates are generalizable to U.S. adults at the time of each wave.

In plain terms

A survey sample is rarely a perfect mini-America — maybe it has too many night owls and not enough early risers. Weighting nudges each response up or down so the final picture matches the real population. It is like adjusting a recipe when you accidentally bought the giant eggs and you may need to reduce the total number of eggs you use (because each giant egg contains more than regular-sized eggs) so the cake still comes out right. Not that I have ever done this myself. Okay, you caught me — I have done this myself. But I promise I only do it with cake, not survey data.