Skip to main content
Category: Effects
Type: A visibility rule meant to open work across groups, followed by a safe public version and a costlier hidden adjustment
Origin: Ethan S. Bernstein, Administrative Science Quarterly 57(2), 2012, pages 181–216. The paper’s name is the transparency paradox. It does not print Dixon’s Reversal Effect
Also known as: transparency paradox. William J. Dickson, on the Hawthorne page, is a different person
Quick Answer — Dixon’s Reversal Effect, on this page, is what follows a rule that opens work so everyone across the old walls can see it: people keep a safe sentence in the open channel and move the real adjustment into a channel that is harder to see and costs extra work. The papers read here do not print this name. They print the transparency paradox. In one Chinese mobile-device factory, four curtained lines led defect-free output by as much as 10–15 percent. That pair of figures is units per hour on those lines. It is not a count of private messages.

​
What is Dixon’s Reversal Effect?

Dixon’s Reversal Effect is a safe public version of the work after a rule made that work visible across groups, with the real adjustment moved somewhere harder to see.
“maintaining observability of workers may counterintuitively reduce their performance by inducing those being observed to conceal their activities through codes and other costly means”
That sentence is Ethan S. Bernstein’s abstract in Administrative Science Quarterly 57(2), 2012, pages 181–216. He names the pattern a transparency paradox. He does not print Dixon’s Reversal Effect. This page uses Dixon’s Reversal Effect for the organizational form of that move: an open channel, often imposed so departments will talk straight to each other, fills with sentences no one can punish, and the bargain shifts to a side channel that takes more effort to maintain. The picture is a shop window. The display has to look right to anyone walking past. The price is settled in the back room. Keeping both versions in step is the extra cost. Social Desirability Bias is a pull toward the answer that sounds acceptable. These papers did not give that questionnaire. Spotlight Effect is an overestimate of how much one person is being watched. Bernstein measured lines, not that overestimate.

​
Dixon’s Reversal Effect in 3 Depths

  • Beginner: If the shared thread is full of praise and the decision shows up later already made, the open channel is the display. Ask which smaller thread carried the constraint.
  • Practitioner: Before you require every cross-group sentence to be visible, write who can punish a real constraint in that room. Give the people doing the work one place to adjust before they owe the whole audience an explanation.
  • Advanced: The counted gain is defect-free units per hour on four curtained lines in one factory, described as as much as 10–15 percent, beside all-shift means of 216.013 and 240.027 against 219.071 and 218.158. It is not a law about every open meeting.

​
Origin

Dixon’s Reversal Effect is the name on this page. Bernstein’s article does not print it. The citation is Ethan S. Bernstein, “The Transparency Paradox: A Role for Privacy in Organizational Learning and Operational Control,” Administrative Science Quarterly 57(2), 2012, pages 181–216, DOI 10.1177/0001839212453028. The journal page lists first publication online on 21 June 2012. A footnote places him at Harvard University. His definition is the transparency paradox. Keeping workers observable can lower performance, because the people being watched conceal their activities through codes and other costly means. A zone of privacy can, under some conditions, raise performance. The factory he studied was built so the floor could be seen. Seeing the floor did not produce an accurate view of how the work was done. William J. Dickson is not this name. Hawthorne Effect is the page for the Western Electric studies. Bernstein cites F. J. Roethlisberger and William J. Dickson, Management and the Worker (1939), only as background, through later shop accounts by Donald Roy and Michael Burawoy, for restriction of output that grew worse when managers were in sight. That 1939 book was not re-read. No count from it, from Roy, or from Burawoy is on this page. The spelling Dixon on this page is not Dickson.

​
Key Points

Four readings keep the open sentence, the hidden adjustment, and the factory’s units per hour apart. A polite meeting is not a productivity rate.
1

Seeing the floor is not seeing the work

Bernstein’s first study placed three student operators on the lines of a contract manufacturer in southern China. He calls the site PrecisionMobile. Only the general manager, the head of human resources, and the head of operations knew who the students were. In the first days, peers showed them a “ton of little tricks” that kept production going or made it faster, easier, or safer. Then they were told that when customers, managers, or leaders came around, they should not do that, because those people would get mad. Under observation, the embeds were trained to appear to follow the posted way. The research team called the arrangement double-sided transparency: the lines could see a manager coming, because managers were easy to spot, and could switch in time. Groupthink is agreement that hardens inside a cohesive group. This switch is a performance for an outsider who can punish the real method.
2

The hidden version costs extra work

The abstract’s cost is concealment through codes and other costly means. Operators described hiding a better method so they would not “bear the cost of explaining better ways of doing things to others,” or “get in trouble” for doing things differently. Table 1 sets a posted sequence for attaching identifying bar-code stickers beside a faster sequence used when nobody in authority was watching. The page does not copy the hand motions. One embed saw one bottleneck draw managers at once, then a two-hour wait for the right process engineer, because a visible problem was read as a failure to follow the chart. That wait is one line. It is not an average. Doing the work twice, once for the audience and once for the job, is the cost the title is pointing at. A private chat can be that second version. Bernstein did not count chats.
3

Privacy here is a boundary around the group

The curtain was not in the original plan. Bernstein had asked for a curtain between one experimental line and a neighboring control line. An operator beside an embed said it would be nice if curtains hung all around the line, so they could be closed off and more productive. He replaced the planned increase in transparency with that idea. For five consecutive months, four experimental lines worked inside the equivalent of a hospital-bed curtain. Inside it, the paper says, tweaks that used to be hidden from other operators on the same line became easier to share. The zone is privacy from people outside the working group. It is not a secret kept from the people doing the task together.
4

The gain is defect-free units, on those lines

Performance was defect-free units per hour. Bernstein writes that each of the four curtained lines increased by as much as 10–15 percent after the first week and kept a lead over the 28 control lines for the remaining five months. Table 2, all shifts, prints the means. Treatment lines: 216.013 before the curtain (SD 16.212) and 240.027 after (SD 24.924). Control lines: 219.071 before (SD 12.679) and 218.158 after (SD 16.359). In the grouped difference-in-differences model, the interaction is 24.9283 units per hour (SE 2.7269), marked significant at the 1 percent level, on 741 grouped daily observations. A collapsed check, his model 7, is 21.2457 (SE 9.5666), significant at the 5 percent level, on 46 observations. Do not divide 24.9283 by 216.013 and call the result the 10–15 percent. Goodhart’s Law is what happens when a measure becomes the target. The posted chart had become the thing outsiders could see. These models do not test that law.

​
Applications

Four checks keep an open channel from being scored as if it were the work.

Write who can punish the sentence

Before you require cross-group talk to happen in one visible room, write who will read it and what a person can lose by naming a real constraint there. If the answer is a boss, a client, or another department’s target, expect a safe sentence in that room.

Leave one adjustment inside the group

Give the people doing the task a bounded place to change the method before they owe an explanation to everyone outside. Bernstein’s curtain raised defect-free output on four lines. It is not a template you can hang in an office and score at 10–15 percent.

Ask where the decision already happened

When the open meeting is all agreement, ask which smaller channel already settled the constraint, and write that channel down. Unanimity in the visible room is a cue. It is not evidence that the hidden thread was wiser, and it is not the factory’s units per hour.

Score the rework, not the tone

Do not grade a team by how agreeable the open record sounds. Compare one visible commitment with the rework that shows up later in that team’s own log. Campbell’s Law warns that a score used for judgment gets corrupted. Politeness is an easy score to corrupt.

​
Case Study

Bernstein studied PrecisionMobile, his name for the largest plant of a global mobile-devices division, in southern China. The plant had about 14,000 people and a floor of about one million square feet. At the time, he describes that division as the second largest producer of mobile devices in the world. The first visit used three embedded students on the lines for eleven working days, sometimes twelve hours a day, after they had gone through orientation and lived as operators. Monthly line-operator turnover at the site was 6 percent, which is why three new recruits were ordinary. The product mix later shifted. The experiment, about ten months after that first visit, followed lines making the same 3G USB data cards. Methods: on an average day about 16 lines ran across two shifts, 32 line-shifts, with capacity on the order of half a million of those cards a week. Do not merge that figure with the plant’s separate capacity of up to two million mobile devices a week. Two lines, four line-shifts, were randomly selected. Operators had been randomly assigned when the lines were staffed, and they rarely switched lines for good before a product’s run ended. The results and the figure call the curtained set four experimental lines and the rest 28 control lines. The curtain went up for five consecutive months. The pre-curtain window in the figure is 55 operating days, 1 May 2009 through 27 June 2009. The post-curtain window is 142 operating days. Table 2’s descriptive statistics run from 1 May through 30 November 2009, with shift rotations on six listed dates. The paper does not print a single calendar day as “the hour the curtain was hung” beyond “several days” after the operator’s remark. The night shift used the same curtained space but was under the impression that the curtains were for something the day shift was doing. Bernstein writes that the research team heard this impression and neither created nor disproved it. He uses it to separate the gain from a feeling of being singled out for a study. Any change on the night shift, in his reading, ties to less transparency from the curtain. He also writes that the Hawthorne account itself has been questioned. This design note is not a number on the Hawthorne Effect page, and that page’s counts do not belong here. The all-shift means are the ones above: treatment 216.013 then 240.027, control 219.071 then 218.158. The author’s summary of the lead is as much as 10–15 percent on each of the four lines after the first week. Operators, in 15 interviews after the experiment plus the embeds’ first month on the line, split the boost about equally across three kinds of privacy: tweaking the line when a temporary problem appeared, trying an idea before explaining it to management, and avoiding outside interruptions without hiding. “Roughly equally” is their account. It is not a second table. The lesson stays inside those sentences. A floor designed to be seen produced a performed method when outsiders approached, and a modest privacy boundary around the line preceded a sustained lead in defect-free units per hour. The paper does not count departmental meetings, flattering sentences, or private messages. A company that opens every channel and then finds the real bargain in a side chat is the situation this page is named for. It is not what the 32 line-shifts measured.

​
Boundaries and Failure Modes

The pattern fails where the open record is the work, not a performance of it. A shared log of a decision that the same people just argued through, with no second audience who can punish the constraint, is not this reversal. Inside Bernstein’s curtain, people on the same line became more open with each other. Broad visibility was the condition that fed hiding. Group privacy was the condition tied to the gain. A second boundary is the night-shift note. It argues against one reading, that the lines improved because they felt studied. It does not measure every way observation can change behavior, and it does not license a curtain in a hospital, a classroom, or a trading desk. Model 7 remains significant at 5 percent after the time series is collapsed, with a smaller interaction, 21.2457 rather than 24.9283. The other models are not printed here. Do not promote the 1 percent mark onto the collapsed check. A common misuse is to move the numbers onto the meeting that sounds too nice. The 10–15 percent, the four means, and both interactions are defect-free units per hour on USB-card lines in that plant in 2009. They are not the share of employees who said only pleasant things, not a messaging-app result, and not a loss figure you can bill to “department walls.” Hiding a defect so a customer never sees it is not the finding. The outcome Bernstein tracked was defect-free output. The faster sticker sequence in Table 1 is an illustration of hidden variance. It is not a method this page recommends.

​
Common Misconceptions

Three retellings travel with the open channel. One invents a discoverer. One turns a line rate into a count of polite sentences. One treats every shared document as the failure.
No paper read here prints Dixon’s Reversal Effect. The counted paper prints the transparency paradox and names Ethan S. Bernstein. William J. Dickson is a coauthor of the 1939 Hawthorne volume that Bernstein cites as background. That book was not re-read. No cell from it is on this page. The two surnames are not one finding.
Bernstein’s “as much as 10–15 percent” is his summary of the lead in defect-free units per hour on each of four curtained lines. The all-shift means beside it are 216.013 and 240.027 against 219.071 and 218.158. The interaction in the grouped model is 24.9283 units per hour. None of these is a head count, a share of meetings, or a transcript of a side chat.
The reversal on this page is a safe public version plus a costlier hidden adjustment, under an audience that can punish the real sentence. A record the working group itself uses is the opposite of Bernstein’s result: inside the curtain, sharing on the line increased. The 10–15 percent and the means belong to one product, one plant, and those months. Model 7’s interaction is 21.2457, at the 5 percent level. German, French, office, and classroom copies are not in the paper.
Dixon’s Reversal Effect sits beside a change in behavior when people are studied, beside public agreement inside a cohesive group, beside answers chosen because they sound acceptable, and beside a measure that becomes the target. None of those pages owns the 24.9283.

Hawthorne Effect

A change tied to being studied. Bernstein designed the night shift so that shift would not think the curtain was for them. His units per hour are not that page’s counts, and William J. Dickson is not this page’s name.

Groupthink

A cohesive group stops testing its own plan. Dixon’s Reversal Effect is a safe sentence performed for an outsider, with the real adjustment moved elsewhere. The open room can look unanimous for a different reason.

Social Desirability Bias

A pull toward the answer that sounds acceptable on a question. The factory switch was a method performed when a manager could see the line. It was not that questionnaire.

Goodhart's Law

A measure that becomes the target stops measuring what it measured. The posted chart was what outsiders could check. The defect-free units are not a test of this law.

​
One-Line Takeaway

Before you make every cross-group sentence visible, write who can punish a real constraint, and leave the people doing the work one place to adjust before they owe that audience the safe version.