Long COVID appears to be less frequent after Omicron infection, but there are a bunch of caveats. Saturday, the Washington Post, in collaboration with Epic Systems, a medical electronic records company, reported that, for those whose first COVID infection is Omicron, the probability of having a long COVID symptom attributable to COVID infection has declined. Before I get to the story and analysis, I want to put this in the context of two major themes.

First, this will likely be discussed by pundits and other self-proclaimed experts*, and there is a difference between how pundits and scientists will approach this: pundits will discuss the implications of the work without kicking the tires–they won’t dive into the methods and the data, and that’s essential (albeit hard and often beyond their ken). Second, trying to assess frequencies of rare events–and long COVID is a rare event (of severe magnitude)–is very difficult to do.

With that as prelude, let’s discuss the findings. Being scientists, not pundits, let’s start with what they did (boldface mine):

The Washington Post worked with the electronic health records company Epic Systems and with input from Kaiser Family Foundation to design a study on who is most likely to report long-covid symptoms.

The study looked at 4.88 million de-identified people of all ages in the national Epic Research Cosmos patient-record database who were diagnosed with covid-19 for the first time between March 2020 and January 2022. The patients studied were separated into categories corresponding to the major coronavirus variant circulating at the time they became ill. The original variant was from March 2020 through June 2021. The delta variant was from August 2021 through November 2021. The omicron variant was in January 2022. July and December 2021 were omitted because of transitions between major variants during those periods.

Epic used a multistep process to identify patients reporting new symptoms. Epic analyzed each patient’s electronic health record going back to 2017. Using that history, Epic identified whether each patient had sought care for the first time for at least one symptom that the Centers for Disease Control and Prevention has listed as a potential indication of long-term covid, including fatigue, difficulty breathing, cough, chest pain, brain fog, headache, sleep problems, dizziness, depression, muscle pain, rash and stomach pain. Only symptoms for which a person had not sought care since 2017 were classified as new symptoms.

The review established whether each patient sought care for any new symptoms from one month to six months after the coronavirus infection. A second step established whether each patient had reported any new symptoms in the six months before receiving a covid diagnosis.

The share of patients with new symptoms before experiencing coronavirus infections established a baseline rate of how often these symptoms appear even without covid. The share of patients with new symptoms in the period after infection constituted the rate after covid.

Baseline rates and post-infection rates were calculated separately for the overall group and for each wave, as well as for demographic groupings by sex, age and race, and for groups of patients with various preexisting conditions (comorbidities), and with different severities of covid infection. The baseline rate was subtracted from the post-infection rate to establish the change, expressed in percentage points.

Data shared with The Post was aggregated at the national level in accordance with the Epic Research standards to protect patient privacy.

Patients who had been hospitalized in intensive care units were excluded from most of the long-covid analyses because the severity of their illness as well as post-ICU syndrome could cause symptoms that are indistinguishable from those of long covid.

Patients may have been constrained from seeking care for new symptoms during the pandemic, especially in its early phases. That may have affected patients’ reported rates of new symptoms before they had coronavirus infections. The duration of symptoms or how many symptoms each patient had — or their severity — were not measured in this study.

This is a clever idea, and the scale of the data are massive–we don’t have to worry about large confidence intervals**. That said, we’ll return to these methods. So what did they find? Here’s a figure!

Note that severity or number of symptoms isn’t presented, so we’re using the gain of one or more symptoms as a proxy for ‘long COVID’, which can be a problem. Also, this is a lower bound–people as it ignores whose symptoms get worse (e.g., someone with asthma and thus shortness of breath who then has worse/longer bouts of it due to COVID).

Anyway, you’ll notice, as some asshole with a blog has been writing for quite some time, that one to two percent of people develop long COVID. You’ll also notice that Omicron infections have a lower rate, around 0.3% overall. To me, this appears to be an underestimate, and it should likely be 0.4-0.6% higher. That’s still better than the initial wave or the Delta period–which is a good thing!

The reason I think the Omicron long COVID rate is too low has to do with the baseline rates–they’re higher for the Omicron wave, as this breakdown of baseline rates by age cohort and COVID wave shows (the colors don’t mean anything, they’re just there to make the age cohorts clearer):

You’ll see that the baseline rates for the Omicron era increase by 0.37% to 0.66% compared to the original era. That doesn’t seem like much, but, remember, the stated Omicron long COVID effect is 0.3%. And these are large (YOOGE!) samples. So why would the baseline be higher for the Omicron era across all ages?

One possible reason is many of the people who are assumed, based on their Epic medical records, to have not had COVID previously, did have COVID. If many people either were asymptomatic, had minor symptoms they didn’t think were COVID (‘allergies’), didn’t get tested, or didn’t have this information incorporated into their medical records (given the patchwork that is the U.S. medical system, that’s not trivial), then these data are incorporating previously infected people. Importantly, some of these ‘unidentified’ previously infected people might have long COVID symptoms, but would not be considered to have long COVID for purposes of the analysis.

The baseline surges in the group that had the largest increase in infections during the Delta period. According to national seroprevalence data, kids had the largest surge in prevalence during the Delta period; older groups had smaller increases, but are more likely to develop long COVID in the first two periods. In absolute terms, it would be a small effect, but, unfortunately, we’re looking for a small effect.

So, my hunch is that, if we use the Original baseline, we’re probably seeing a more realistic effect of Omicron. The good news is that it’s less around 0.5% for those under 45 (1-2% for older people), but that is higher than the Washington Post is reporting. Mind you, this doesn’t mean the work is bad (it’s impressive!), but the interpretation of their results needs to be more nuanced than is reported–and, as mentioned above, we need a much better deep dive into the particulars of the data. To be blunt, this is not peer-review**** ready. Nonetheless, this might be still good(-ish) news, especially for younger age cohorts.

Anyway, I eagerly await lazy pundits fucking the interpretation of these results up too.

*It takes five to six years to finish a Ph.D., at which point one hopefully is an expert in a sub-(sub-, sub-)discipline. The idea that, after three years of the pandemic, anyone believes they’re an expert is silly.

**Large confidence intervals are bad; we want small ones!

***It’s also interesting to note, that two age cohorts, 30-44 (parents who were busy), and the elderly (who were avoiding unnecessary medical appointments), had a ‘Delta dip’, while the other cohorts did not.

****All the recent drama regarding eLife notwithstanding…