How Climate Attribution Works for Nepal's Floods, and What 'More Likely' Really Means

Every big monsoon in Nepal produces a sentence that travels further than the numbers underneath it. Climate change made this more likely. The line turned up again this year, and it will turn up after the next bad August too. It is a fair claim, and it comes out of a real research field, climate attribution. But it means something narrower, and stranger, than most of the reporting around it lets on.

What happens inside a rapid climate attribution study

The first thing an attribution study does is decide what 'this' is. Not the flood, not the landslide, not the bridge that came down. It picks an event class: a multi-day rainfall total, over a defined region, in a defined season, above a defined threshold. Three-day rainfall over eastern Nepal during the monsoon is one version of that. Change the area, the number of days or the threshold and you are studying a different event entirely, which matters a lot when you get to the result and try to match it against what you saw on the news.

Next comes the counterfactual. Researchers run a climate model many times with greenhouse gas concentrations close to observed. Then they run the same model with those concentrations turned down, usually to a world roughly 1.2 °C cooler than today rather than literally 1850. Neither set of runs reproduces the weather of that particular week. They generate many plausible versions of that season in each world, and the researchers count how often the event class shows up.

Divide the rate in today's world by the rate in the cooler one and you get the probability ratio. That single figure is what most coverage is quoting, whether or not the coverage says so. The standard protocol for doing this quickly comes from Sjoukje Philip and colleagues, published in Advances in Statistical Climatology, Meteorology and Oceanography in 2020. For which event types the method handles well and which it does not, the National Academies' 2016 report on attribution is still the reference people reach for. It ranked heat events far above heavy rainfall in how much confidence the method supports, and that basic ordering has not shifted.

Rapid describes turnaround time, days to weeks instead of the years a full study takes. It says nothing about how well reviewed the work is. Plenty of rapid analyses go public before peer review, as consortium reports or preprints, and World Weather Attribution is the group most associated with that format. Read an unreviewed rapid analysis as provisional.

What 'more likely' actually means

Start with the part that gets lost in the retelling. A probability ratio of 2 does not mean the rain doubled. It means the odds of that event class doubled, and if the odds were tiny to begin with, they stay small. Say, for the sake of argument, a rainfall total had roughly a 1 in 200 chance in the cooler counterfactual world and roughly 1 in 100 in ours. The ratio is 2. The event is still, by any ordinary standard, unusual. A headline reading 'climate change made these floods more likely' is usually describing a shift of that size, not the conversion of ordinary weather into routine catastrophe.

A ratio of 1 means the models produced the event class at about the same rate in both worlds, so no change was detected. Above 1 means it was more probable in the warmer world. Below 1 points the other way. Ratios also come with ranges rather than as single figures, and the lower end of that range is often the more informative number.

What none of it means is that climate change caused one particular flood. Attribution speaks to a class of events defined by a chosen threshold, region and season, and the figure moves when the definition moves. One team might define the event as a two-day total over a single basin while another takes a five-day total over a wider area. Both can return defensible but different ratios for the same monsoon. Neither is wrong. They answered slightly different questions.

A null result counts as a finding too. When a study reports no clear human signal for a given rainfall event, that is information rather than failure, and in mountain regions it is often the most honest answer the tools can give.

Why the Himalaya is the hardest kind of place to run this

Three problems stack up. First, the record. Rain gauges are sparse in Nepal and sparser at altitude, and instrumental histories in many Himalayan catchments run well under a century. Attribution depends on knowing how often the event class occurred in the past. Where that history is patchy, the baseline is shaky, and the uncertainty that produces is real rather than cosmetic.

Second, resolution. Global climate models, and most regional ones, chop the atmosphere into grid boxes tens of kilometres wide. Himalayan terrain changes by kilometres. Orographic lifting and the localised convective downpours that actually cause destructive flash flooding in Nepal get smoothed out or dropped into the wrong valley. Model output and gauge readings then disagree for reasons that have nothing to do with greenhouse gases.

Third, variability. Monsoon behaviour swings hard from year to year, which makes any underlying trend difficult to separate from ordinary noise. On flat terrain a long record helps you tease the two apart. Here you usually do not have the long record either. Add the model's uncertain grip on the terrain and the error bars widen to the point where a clear signal often is not there to find.

A rainfall result is not a flood result

Rainfall attribution answers a rainfall question. What turns rain into a disaster in Nepal is usually something stacked on top of it: a glacial lake bursting, a landslide damming a river, a debris flow, a bridge pier never designed for the flow it met, a settlement built on the fan where the water was always going to go. A study of monsoon rainfall says nothing about any of those chains. It says nothing about exposure or vulnerability either, and those decide far more about the death toll than the millimetres do.

So a rapid analysis can be technically sound and still be a poor guide to why one particular valley flooded. It describes the probability of the rain. The rest of the disaster is a different question with a different literature behind it.

When the next rapid analysis lands, three questions carry most of the weight. What event class did the authors define? What was the probability ratio, and what was its range? And had the work been peer reviewed at the time you read it? Answer those and you will understand the finding better than most of the reporting around it.