Confidence in Institutions, GSS
1973-2024
Fifty-one years of one question, and the summary
that does not survive it
Democracy’s Data
Last updated September 15, 2026 at 4:42p EDT
Americans have lost faith in everything. You have read that sentence,
or one of its cousins: trust has collapsed, institutions are in crisis,
nobody believes in anything anymore. It is repeated so often that it
reads as a fact rather than a claim. Is it true — did confidence fall in
every institution, in which ones, and since when?
Answering needs something rare: the same question, asked the same
way, for long enough to watch. Survey wording usually drifts, sponsors
lose interest, and a question asked in 1975 is rarely asked in 2024 in a
form anybody can compare. One battery of questions survived half a
century intact, and most of what this brief can do rests on that
fact.
The test
The test is the General Social
Survey, the GSS: a survey of American adults about their lives and
opinions, run since 1972 by NORC, a survey research organization at the
University of Chicago, and funded largely by the National Science
Foundation, the federal agency that pays for basic research. Since 1973
its interviewers have read the same sentence to a national sample of
American adults — a few thousand people chosen at random so that they
stand in for the country — 33 times across 51 years:
“I am going to name some institutions in this country. As far as
the people running these institutions are concerned, would you say you
have a great deal of confidence, only some confidence, or hardly any
confidence at all in them?”
Nothing about the question is clever; its entire value is that nobody
changed it. No election ever grades an answer like this — there is no
true national figure for confidence, which is the standing condition of
the Public Opinion cluster’s data
chapter. When nothing outside the survey can grade the answer, the
method is the whole credential, and this survey’s method held still. A
survey whose product is the comparison across time has every incentive
to leave its wording alone, and the GSS acted on that incentive for
fifty years.
Two features of the wording should be fixed in mind before any number
arrives. It asks about “the people running” the
institution, not the institution itself, so somebody who
reveres the Supreme Court and distrusts its current members is recorded
as distrusting. And there are only three options with no neutral
middle, so a lukewarm respondent has to choose “only some.”
That is why the middle box is the largest almost everywhere, and why the
interesting movement is at the two ends.
The battery also asks about 13 institutions in one interview, of the
same respondents. That makes it an unusually fair test of the claim in
the title: if confidence moved as a bloc, these 13 series should mostly
move together.
The data
The cumulative file — every year’s interviews stacked into one table,
downloadable free from NORC at https://gss.norc.org/us/en/gss/get-the-data.html — is
570 megabytes in the format of Stata, a statistics program, holding
75,699 respondents and about six thousand variables, or columns. One row
is one person interviewed, once, in one year. Here are the first five
people interviewed in 1973, in six of those columns:
| 1973 |
1 |
2 |
2 |
1 |
0.7229722 |
| 1973 |
2 |
1 |
1 |
1 |
1.4825919 |
| 1973 |
3 |
2 |
2 |
2 |
0.7921651 |
| 1973 |
4 |
2 |
2 |
1 |
0.7814606 |
| 1973 |
5 |
2 |
2 |
2 |
0.7412959 |
The meaning lives in the codebook, not the file.
conlegis is confidence in Congress, coded 1 for a great
deal, 2 for only some, 3 for hardly any; wtssps is the
survey weight the respondent carries. A 2 is not a
quantity. It is the middle of three ordered boxes, so every
number below is a share of people who picked a box, never an average of
the codes. Alongside the three answers sit twelve distinct ways of not
answering: don’t know, refused, inapplicable, skipped on web, and more.
All twelve are treated below as the same thing: no answer.
Everything below rests on 427 rows like the one under this sentence,
each of which reduces one year’s interviews about one institution to a
single line. Here is what those 1,504 people become for Congress:
| Congress |
conlegis |
1973 |
1458 |
24.1 |
60.6 |
15.2 |
One row per institution per survey year, in
confidence.csv: the institution, its GSS
variable name, the year, the number who
answered, and the weighted shares great_deal,
only_some and hardly_any, which sum to 100. A
second table, institutions.csv, gives each institution its
first_year, last_year,
survey_years and the change across the series.
What one row stands for has changed, from a person to an institution in
a year, and there is no going back. Nothing below can ask whether the
people who distrusted Congress also distrusted the press, because the
rows that knew are gone.
Four decisions sit inside that one row. The denominator is
who answered, not who was asked. The denominator is the bottom
number, the figure the percentage is divided by: 1,458 of the 1,504
people interviewed in 1973 answered the Congress item, and the rest are
simply not in the percentage. Every percentage is
weighted, and weighting means counting some answers more than
others, to make a sample that does not look like the country look like
it; the targets are the Census Bureau’s figures for how many people of
each age, sex, race and region the country holds. The weight used
throughout is wtssps, the only one spanning every year in
the file, so no seam between two weights lands inside the argument.
An institution-year with fewer than fifty answers is
dropped, not shown small. And twenty columns of six
thousand were kept: everything else about these respondents
stays in the source file and cannot be recovered from these tables.
Two irregularities are the file’s own. The survey ran yearly into the
1990s and every other year after that, so the longest gap between
readings is 3 years, across the pandemic. And one institution, banks and
financial institutions, joined the battery in 1975, so its series is
shorter than the rest.
Before you read on, commit to three answers. Of the
13 institutions in the battery, how many fell between their first
reading and their last — all 13? Ten? Seven? How many rose? And which
single institution fell the furthest? Write the three down. A guess you
have not written down is one you will quietly revise once you see the
answer.
What it says about democracy
Start with the series everybody quotes. Confidence in the people
running Congress fell from 24.1% with a great deal of confidence in 1973
to 5.6% in 2024. The more striking number is at the other end:
“hardly any” went from 15.2% to 51.9%, from about one
respondent in seven to a clear majority.
Figure 1. Confidence in the people running Congress, drawn as
a composition rather than a level. The three answer shares at
every survey year, stacked to 100% — the right form here, because every
respondent picks exactly one of three boxes, so what moves is not a
level but a mix. “Hardly any” overtook “a great deal” in 1974 and passed
half of all respondents in 2014, marked by the dashed line. Even the
lukewarm middle fell, from 60.6% to 42.5%.
At this point the summary is sitting there waiting to be written:
Americans have lost confidence in their institutions. But the
claim is about institutions in general, and only one has been looked at.
The file holds 13 of them, asked of the same respondents. Here they are,
ranked by how far each moved between its first reading and its last:
| Medicine |
1973 |
33 |
-28.6 |
| Organized religion |
1973 |
33 |
-20.2 |
| Congress |
1973 |
33 |
-18.5 |
| Executive branch of federal government |
1973 |
33 |
-17.5 |
| Education |
1973 |
33 |
-17.4 |
| Banks and financial institutions |
1975 |
31 |
-15.7 |
| Major companies |
1973 |
33 |
-15.7 |
| Press |
1973 |
33 |
-15.7 |
| U.S. Supreme Court |
1973 |
33 |
-13.4 |
| Television |
1973 |
33 |
-10.4 |
| Scientific community |
1973 |
33 |
-4.8 |
| Organized labor |
1973 |
33 |
1.3 |
| Military |
1973 |
33 |
10.6 |
The largest decline is not a political institution at all.
Medicine is down 28.6 points, further than Congress.
Organized religion, the press, television, major companies, banks and
education all fall by double digits. And the military rose 10.6
points, the only substantial increase anywhere in the file.
Most readers name Congress for the steepest fall, and most put the
number of risers at zero; neither holds.
A ranked list of thirteen differences invites a second mistake, which
is reading the small ones as movements at all. Every number in that
table is an estimate from a finite sample, and the difference between
two estimates carries the uncertainty of both.
Figure 2. First reading against last, with the uncertainty on
each. Share with a great deal of confidence in 1973 (hollow
circle; 1975 for banks) against 2024 (filled), one row per institution,
with a margin of error on each single reading — the thin rule, a range
that would contain the true share nineteen times in twenty if the survey
were repeated. Rows whose change does not clear the uncertainty of its
two readings are drawn pale and marked ns, not significant. The
intervals treat each year as a simple random sample, as if every
respondent had been picked independently; in fact the GSS interviews
people in clusters of neighbourhoods, so the true intervals are wider,
which makes the small changes less secure rather than more.
Written out honestly: eleven institutions fell, one rose, and
one did not move. Take organized labor, the row that did not
move. In 1973, 16.2% had a great deal of confidence in the people
running it; in 2024, 17.5%. The change is 1.3 points. But each of those
readings came from a sample, and a different draw of people would have
moved each one by a point or two: the first reading is good to about
±1.9 points and the second to about ±2.3. Put the two wobbles together
and a difference of anything up to about 3.0 points could be nothing but
the luck of who was asked. 1.3 is inside that, so the right description
is no detectable change. That wobble is sampling
error, the uncertainty that comes from asking a sample rather
than everyone. Everything else moved by more than it can explain,
including the scientific community’s 4.8-point fall, the smallest real
decline in the set.
The exception deserves its story, because a single rise could be a
fluke of the two endpoints Figure 2 happens to compare. Follow the
military’s actual series:
| 1973 |
32.7 |
17.2 |
| 1980 |
29.5 |
17.5 |
| 1986 |
32.3 |
13.3 |
| 1991 |
63.1 |
5.6 |
| 2000 |
42.0 |
9.7 |
| 2008 |
51.6 |
10.9 |
| 2018 |
62.2 |
6.8 |
| 2024 |
43.3 |
11.6 |
It is not a steady climb. It peaks at 63.1% in 1991, the year of the
Gulf War, falls back, then rises again. That cuts directly against the
long-slow-decline framing the Congress series invites:
confidence responds to events, and it went up sharply
when an institution did something visible and widely regarded at the
time as successful. Whatever drives the other twelve series is capable
of moving in both directions.
Where everything stands now completes the picture:
| Military |
43.3 |
11.6 |
31.7 |
| Scientific community |
36.3 |
12.0 |
24.3 |
| Medicine |
26.3 |
19.8 |
6.5 |
| Education |
19.8 |
25.8 |
-6.0 |
| U.S. Supreme Court |
18.6 |
36.7 |
-18.1 |
| Organized labor |
17.5 |
23.3 |
-5.8 |
| Banks and financial institutions |
17.0 |
24.6 |
-7.6 |
| Organized religion |
15.5 |
34.9 |
-19.4 |
| Major companies |
15.0 |
24.3 |
-9.3 |
| Executive branch of federal government |
12.1 |
46.4 |
-34.3 |
| Television |
8.2 |
48.1 |
-39.9 |
| Press |
7.5 |
55.0 |
-47.5 |
| Congress |
5.6 |
51.9 |
-46.3 |
Sorted by the net column, a pattern appears that the individual
series conceal. The institutions people still trust do something
technical and visible — the military, the scientific community,
medicine. The ones at the bottom are the ones whose entire function is
contestation: Congress, the press, television, the executive branch.
Congress now stands at 5.6%, the lowest figure for anything asked about,
and the Supreme Court at 18.6%.
Move across for a year readout; hover a label to isolate a line, click
to hide it.
Figure 3. All 13 institutions, every survey year from 1973 to
2024. Share with a great deal of confidence in the people
running each. The band narrows as it moves: institutions that began
between 16% and 55% in 1973 end between 6% and 43%, with the military
the only line finishing above where it started.
So the sentence this brief opened with is not right as written.
Something did happen to confidence in American institutions. But it is
not uniform, it is not confined to government or even to politics, and
at least one institution moved the other way. The largest loss belongs
to medicine and the smallest to science, neither of which is a branch of
government. That is a strange result for a story usually told about
politics, and the unevenness is the finding.
What this brief cannot tell you
Whether institutions became less trustworthy, or people became less
trusting, or the phrase “a great deal of confidence” drifted in meaning
across fifty years. The question records a report, not a relationship,
and no amount of this series settles which changed. Citrin and Stoker
add that trust measures move with partisanship: some of what looks
like erosion about Congress is people disliking the party currently
running it, and those people flip when control changes.
Whether the decline was warranted. The 1973 baseline sits immediately
after Watergate and Vietnam, and institutions went on to earn skepticism
through visible failures. Low confidence in a body whose approval rating
has hovered around twenty percent for much of the past decade may be
accurate perception rather than civic decay, and the data cannot tell
those apart.
Whether Americans changed their minds, or the Americans who held the
old view died and were replaced by people who never held it. The GSS is
a cross-section — it samples new people each round; only a panel, a
survey that goes back to the same people, could separate the two, and
this design confuses them.
Who moved. The file built here carries no breakdowns by party, age,
race or education, which is where most of the interesting variation in
this literature lives. And the GSS changed how it collected answers in
2021, moving to web and mail during the pandemic. That is exactly the
kind of break that can look like a trend, and it sits directly
underneath every 2024 reading above.
What you should have learned
- One unchanged question, asked 33 times across 51 years, is what
makes the claim testable at all; almost no other social data holds still
that long.
- The fall is real but wildly uneven: eleven of 13 institutions fell
by more than sampling error can explain, one rose, and one did not
move.
- The steepest loss is medicine’s 28.6 points, not Congress’s — the
decline is not confined to government, or even to politics.
- The military’s series peaks at 63.1% in 1991 and rises overall, so
whatever drives the declines is not an irreversible drift in national
character.
- A percentage is a stack of decisions: who counts as answering, which
weight, which cells are too thin to print. Every one of them was
somebody’s choice.
Extensions
institutions.csv gives every institution a
change across the series. Sort by it. Which institutions
lost the most confidence, and which gained?institutions.csv also has first_year,
last_year and survey_years. Find the
institutions measured for fewer years than the rest. Does a shorter
series make a bigger change easier to claim?confidence.csv keeps great_deal,
only_some and hardly_any separately. Pick an
institution whose great_deal fell. Did those people move to
the middle, or all the way to the bottom?- Find a year in
confidence.csv where many institutions
move together, using the year column. Then go and read what
happened that year. Is the survey measuring the institutions, or the
mood?
- Stretch: Gallup has asked its own confidence
battery since 1973, with different wording and a fourth answer option.
Pull its Congress series from https://news.gallup.com/poll/1597/confidence-institutions.aspx
and set it beside this one. Where the two disagree, which difference —
wording, options, sample — would you blame first?
- Stretch: the GSS Data Explorer at https://gssdataexplorer.norc.org/ can cross any
confidence item by party identification. Pick one institution from
Figure 3 and find out whether its decline belongs to one party or
both.
Sources
Data
Academic literature
The data itself
Every figure above is drawn from these tables. They are plain CSV,
they open in a spreadsheet, and they are yours to take further.
confidence.csv, institutions.csv
