Does Office Temperature Affect How Well You Work?

The famous answer is that performance peaks around 22 degrees Celsius, and that answer has not held up. A 2006 analysis produced the widely quoted curve. A 2021 reanalysis using roughly twice as many data points found the original model’s predictive accuracy was very low and that no relationship between temperature and office work performance could be established.

So the number everyone cites is weaker than its reputation, and its replacement does not exist.

The number you have seen

The original work is Seppänen, Fisk and Lei’s 2006 analysis, produced at Lawrence Berkeley National Laboratory. They calculated from a body of studies the percentage change in performance per degree of temperature increase, and analysed measured work performance against temperature.

Their result: performance increases with temperature up to 21 to 22 degrees and decreases above 23 to 24 degrees, with the highest productivity at around 22 degrees. As an illustration, at 30 degrees performance is only 91.1 per cent of the maximum, a reduction of 8.9 per cent.

That is where the office thermostat argument comes from, and it is reasonable-looking work: real data, a clear curve, an intuitive shape.

What happened when someone checked

Fifteen years later, a team published a reanalysis in Building and Environment, examining 35 studies on the effects of temperature on office work performance.

They tested the 2006 relationship directly, describing it as probably the best known of the published temperature and performance relationships. Their verdict: even when significant, its prediction accuracy is very low, with an R-squared of 0.05, a mean absolute error of 1.9 per cent and a root mean square error of 3.1 per cent.

An R-squared of 0.05 means the model accounts for about five per cent of the variation. The curve exists; it just does not predict much.

Then they tried to build a better one, normalising data from 35 studies using the original approach and exploring regression models, models based on the maximal adaptability framework, and machine learning.

Their finding: they could not find a relationship between temperature and office work performance.

What the reanalysis covered

Worth being precise, because scale is what makes the null informative rather than a failed attempt.

The complete dataset included 358 performance measures from 35 studies published between 1946 and 2020, roughly doubling the data points in the 2006 analysis. Of those, 215 data points between 20 and 30 degrees were used to test for a new relationship. Eleven models and algorithms were tested.

A serious attempt to find the effect, with more data than the original, using multiple approaches. Coming up empty under those conditions is a different result from a small study finding nothing.

What this does and does not mean

It does not mean temperature has no effect on anything. It means that across this body of studies, a stable predictive relationship between air temperature and office work performance could not be established.

It does not mean you should be indifferent to being cold or hot. Thermal comfort is measured separately from task performance, and none of this addresses how it feels to work in an uncomfortable room.

And it does not settle the question forever. The reanalysis authors note that some more recent work suggests cognitive performance may stabilise over a wider temperature range, or even at high temperatures, because of psychological and physiological adaptability. That is a hypothesis about why the relationship is hard to find, not a demonstration that temperature is irrelevant.

What it does mean is that the specific claim, that there is a known optimal office temperature with measurable percentage penalties for departing from it, is not currently supported. That pattern repeats across the indoor environment literature: the effects of stuffy air and office lighting both turn out smaller or more contested than their headline numbers suggest, while office noise is the one environmental factor with a mechanism that holds up.

A note on comfort and health

Nothing here is about safety. Extreme heat and extreme cold are genuine health risks and this article is not about them.

If you work somewhere persistently uncomfortable, that is worth raising on comfort grounds alone and does not need a productivity justification. If heat or cold produces symptoms, dizziness, headaches, breathlessness, or if you have a condition affected by temperature, that is a matter for a doctor and for your employer rather than for this literature.

Common questions

Is 22 degrees the ideal office temperature?

It is the figure from a 2006 analysis, and a 2021 reanalysis found that model does not predict well. The original reported that performance increases up to 21 to 22 degrees and decreases above 23 to 24, with highest productivity around 22, and that at 30 degrees performance falls to 91.1 per cent of maximum. The 2021 reanalysis tested that relationship and found its prediction accuracy very low, with an R-squared of 0.05. It then used 358 performance measures from 35 studies and eleven models and algorithms to look for a better relationship, and could not find one. So 22 degrees is a widely repeated number rather than an established optimum.

So temperature does not matter at all?

That is stronger than the evidence. The reanalysis established that a stable predictive relationship between air temperature and office work performance could not be found across 35 studies, not that temperature is irrelevant to everything. Thermal comfort is a separate measurement from task performance, and nothing here addresses how it feels to work in an uncomfortable room. The reanalysis authors also note that some recent work suggests cognitive performance may stabilise across a wider temperature range due to psychological and physiological adaptability, which is a proposed explanation for the difficulty rather than proof of no effect.

Why did this take fifteen years to catch?

Because the original result was plausible, useful and specific, which is a combination that tends to get adopted rather than tested. The 2021 team describe the 2006 relationship as probably the best known of the published models in this area. Their reanalysis roughly doubled the data points, using 358 performance measures from 35 studies published between 1946 and 2020. It is worth noting what they did not find rather than what they disproved: they tested eleven models and algorithms including regression and machine learning approaches, and could not establish a relationship. That is a failure to find, on more data than the original.

Should I argue with my employer about the thermostat?

On comfort grounds, that is a perfectly good argument and needs no research support. On performance grounds, the evidence is thinner than most people assume: the widely quoted optimum comes from a model whose prediction accuracy was later measured at an R-squared of 0.05, and an attempt to build a better relationship from twice the data across eleven modelling approaches came up empty. If temperature at your workplace produces symptoms rather than just discomfort, that is a health and safety matter for your employer and potentially a doctor, not a productivity debate.

When a curve becomes furniture

What is interesting here is not the temperature. It is how a single analysis became the answer.

The 2006 curve had everything a useful finding needs: a specific optimum, a percentage penalty per degree, and a shape that matched intuition. It got quoted, then designed around, then cited by people quoting people. Fifteen years passed before anyone assembled twice the data and asked whether it predicted anything, and the answer was that it accounts for about five per cent of the variation and cannot be improved on.

The lesson generalises past thermostats. A number that is very specific, very actionable and very widely repeated has those properties because it travelled well, and travelling well is not the same as being right.

Set your office to whatever is comfortable. That was always the better reason, and it turns out to be the only one currently standing.