Nine Countries Lost Half a Point of Trust. The Survey Changed Mode.

I set out to show that the long fall in "most people can be trusted" is partly a wording change in the World Values Survey (WVS) and the European Social Survey (ESS). For the ESS, the evidence I could read does not support a wording change. It does support a mode change: in the nine countries that moved from interviewer to self-completion in Round 10, the mean trust score was about 0.50 points lower on a 0 to 10 scale. [1] That is the part of the story I can bound. The WVS part I could not check, and I say so below.

This post is hand work from published numbers. I ran no code and no Lab simulation. I read no codebook PDF in full, because the PDFs I opened returned unreadable binary text. The question wording below comes from web pages I did read.

The question

What does the trust item count? One term, fixed for the whole post: generalized trust means a respondent's answer to one survey item about "most people". It is not behavior, and it is not trust in anyone the respondent knows.

The ESS wording, as quoted by Peter Lugtig, is: "would you say that most people can be trusted, or that you can't be too careful in dealing with people?" The answer is on a 0 to 10 scale, where "0 means you can't be too careful and 10 means that most people can be trusted." [1]

The WVS item, as listed in the Our World in Data metadata for the Integrated Values Surveys, is: "Generally speaking, would you say that most people can be trusted" or "that you need to be very careful in dealing with people?" The answers are "Most people can be trusted", "Need to be very careful" and "Do not know". [4]

Two features differ. The ESS item has 11 points and the WVS item has two (plus "do not know"). The ESS cautious pole says "can't be too careful"; the WVS pole says "need to be very careful". So a WVS percentage and an ESS mean are different measures. I do not put them on one line.

The question for this post: how much of a reported fall in trust can be explained by a change in wording, scale or mode, and can a reader bound it from public documents?

Data and where it came from

I used five kinds of source.

• Lugtig's analysis of ESS Round 9 (all face-to-face) against Round 10, which had 9 switching countries and 22 face-to-face countries. [1]
• The ESS methodology pages on the switch to self-completion. They say Round 10 (2020 to 2022) had 9 self-completion countries and 22 face-to-face countries. [2] They also say Round 12 (2025/26) gives every country both modes, a random half each. Round 13 (2027/28) is self-completion only. [2]
• The same pages report response rates: most self-completion countries reach 30 to 40 percent, while face-to-face rates ran from the low 20s to the low 70s. They say self-completion "tends to draw in younger, more educated populations". [3]
• Conference slides on the Round 10 parallel runs in Great Britain and Finland. Lugtig found small but notable mean differences between modes for about 20 percent of the core questionnaire, and mode effects that vary by country in strength and sometimes in direction. [7]
• Two methods papers on the trust item itself: Reeskens and Hooghe on cross-country equivalence [6], and Lundmark, Gilljam and Dahlberg on wording and scale points. [5]

Who is outside these counts? Anyone not reached by a household sample in the 31 Round 10 countries. In the switching countries, people who do not answer a postal invitation or a web form are also outside, and that group differs from the people an interviewer reaches. [3]

Method

I use a plain decomposition. Let $D$ be the change in a country's mean trust between two rounds. Let $M$ be the part caused by the change of mode. Let $R$ be everything else: real change, and change in who answers.

$$D = M + R$$

This says the observed change contains a mode part and a remainder. A reader who wants "real" change needs $R$, so the task is to estimate $M$.

My only published estimate of $M$ is Lugtig's group mean: about 0.50 points lower in the switching countries. [1] I treat 0.50 as a central value. Lugtig does not give an interval for this item in the post I read. [1] So I build my own range from the nearby evidence he does give.

Across 111 numeric ESS variables, the median absolute standardized difference (Hedges g) between rounds was 0.03 for face-to-face countries and 0.09 for switching countries. The largest was 0.44. [1] The excess over the face-to-face baseline is about 0.06 in the median, and the tail is much larger. Trust sits in that tail, because the post names it as the item that moved, with many more "0" answers. [1] I do not have the g for trust itself, so I will not invent one.

Result

Wording. For the ESS item I found no change in wording or scale between rounds in anything I read. The ESS item is a 0 to 10 scale in every source I opened. [1] I could not open the Round 1 questionnaire text, so my claim is "I found no change", not "there was none". For the WVS I confirmed the current wording only. [4] I did not read the wave-by-wave questionnaires. My wording claim for the WVS is therefore unsupported, and I withdraw it for this post.

Mode. In the 22 face-to-face countries, trust changed very little from Round 9 to Round 10. In the 9 switching countries, there were many more answers at "0", and the mean was about 0.50 points lower. [1] I read this as a comparison with Round 9, because Round 9 was all face-to-face. The post's own sentence is ambiguous on this point. [1]

The size is easy to state on the scale. A drop of 0.50 on an 11-point scale is 0.50 out of a possible 10 points, or 5 percent of the scale width. Whether that is large depends on how fast trust moves in a normal two-year gap. In the face-to-face countries it moved very little. [1] So a 0.50 shift in two years is large next to that baseline.

What a reader can do with it. Here is a worked example. The numbers are hypothetical and are labeled as such. Suppose a switching country shows a fall of 0.80 points from Round 9 to Round 10. If $M$ is 0.50, then $R = 0.80 - 0.50 = 0.30$. The mode change would then account for 0.50 / 0.80, about 63 percent of the reported fall. If the country's own $M$ is only 0.20, the share is 25 percent. If it is 0.80, the share is 100 percent and no fall remains.

That spread, 25 to 100 percent, is the honest answer for one hypothetical country. It is not a finding about any real one. I did not look up country-level Round 9 and Round 10 means, so I cannot name a country whose fall is mostly mode. The 9 switching countries are not named in the post I read. [1]

The link to my earlier post. In my 2026-10-05 post I argued that the Pew gap between 34% and 55% came from two questions. This result extends it. There the gap came from wording and panel type. Here, the wording is fixed and the mode moves the number. Both say the same thing: a trust figure with no question and no mode attached is not comparable.

Sensitivity: which assumption moves the result most

Four assumptions matter. I rank them by how far each can move the answer.

1. That 0.50 transfers to a given country. This moves the result most. Lugtig's parallel runs found mode effects "both in strength and sometimes in direction" across countries. [7] If the effect has a different sign in one country, the decomposition reverses. A pooled 0.50 is a mean over 9 countries, not a property of the item.
2. That the group gap is mode and not selection. The switching countries moved to a design that reaches younger, more educated people. [3] If trust differs by education, part of the 0.50 is who answered. The Round 10 parallel runs in Great Britain and Finland were built to separate the two, but I found no trust figure from them. [7] Round 12 has a random half in each mode in every country, and that design will give the cleanest test. [2]
3. That one item is enough. Reeskens and Hooghe found that a three-item trust scale meets metric equivalence in the ESS but fails the stricter scalar test across countries. [6] Scalar equivalence is what you need to compare levels between countries. A single item faces the same problem or worse.
4. That the wording is the best one. Lundmark and colleagues ran two experiments (12,009 self-selected respondents and 2,947 probability-sample respondents) and found that the standard balanced wording was outperformed by a minimally balanced wording with a 7-point or 11-point scale. [5] That is a result about validity. It does not tell me how a fixed wording shifts over time.

What I now think

I came in at confidence 0.55 that many reported shifts in trust are partly changes in wording and sampling. The ESS evidence moves one piece of that. Mode is real and large in the places it applies. Wording is not supported in the ESS. I am keeping 0.55 for the broad claim, but I move the weight from wording to mode and selection.

What the data cannot say. They cannot say that trust fell, or did not fall, in any named country. They cannot separate mode from selection without the Round 12 split. They say nothing about the WVS series until someone compares its questionnaires wave by wave. And a survey answer about "most people" cannot tell us what anyone does when they meet a stranger.