The Post That Refutes Itself

What happened when a patient brought me a graphic, an AI agreed with it, and the paper underneath said something else.

A woman sat down in my office with her phone already open.

Someone had sent her a post saying the pill she had been taking for two years raises her risk of a blood clot. She had stopped it that morning. Not tapered. Stopped.

Then she did what most of us do now. She checked. Not with a doctor, with an AI, and it agreed with the graphic. So by the time she got to me it was two against one, and I was the one.

Here is the part that stayed with me. She did not ask whether it was true. She asked who she was supposed to believe.

That is the wrong question. It is also the only question anybody teaches her to ask. So I want to give you the other one, and the four small tools that answer it. They are not hard. You can use them tonight, on any paper, in any field, about anything.

Tool one: "a study" is not one thing

When a post says "studies show," ask which kind. There are six different things wearing that one word, and they are not equal.

A randomized trial flips a coin to decide who gets the treatment, which is the only design that reliably separates the drug from the woman taking it. A prospective cohort follows people forward without assigning anything. A database review counts prescriptions somebody filled. A case-control starts with people who already had the event and looks backward. A tiny mechanistic crossover measures what a drug does inside the body, often in a handful of people. And a meta-analysis is not a rung at all: it pools whatever went into it and inherits that quality, so a pooled analysis of weak studies is a weak study wearing a suit.

None of these is worthless. They answer different questions with different confidence. When somebody shows you a number without telling you which kind of study produced it, they have left out the thing that determines what the number means.

Tool two: if the interval crosses 1.0, that comparison found nothing

Almost every risk number you see in a graphic is a ratio. 1.0 means no difference. Above 1.0 means more. Below 1.0 means less.

But the ratio is only an estimate, so it comes with a range around it, called a confidence interval. Look at the range, not the headline number.

If that range includes 1.0, then "no difference" is one of the answers the data is compatible with. The honest reading is that this comparison did not find a difference. Not that it proved safety, and not that it proved harm. It did not settle it.

You would be amazed how many frightening posts are built on an interval that quietly crosses 1.0.

Tool three: "of what?"

A ratio without a base rate is a rumor.

"Doubles your risk" sounds enormous. Doubles it from what? If something happens to 1 woman in 10,000 a year, doubling it means 2 women in 10,000. That is a real change and it deserves an honest conversation, but it is not the same sentence as the one your imagination just wrote.

So always ask: of what? If the post cannot tell you the underlying rate, the post cannot tell you what the risk is.

Tool four: compared to what?

This is the one that does the most damage when it is missing.

Every risk number is measured against something. Against a woman taking nothing. Against a woman on a different drug. Against a woman on the same drug by a different route. Two numbers that were each measured against nothing are not a comparison of those two treatments with each other, but they get printed side by side constantly, as though they were.

When you see two bars next to each other, ask what each one was compared against. If the answer is different for the two bars, the picture is not showing you what it appears to be showing you.

Now the post

The graphic my patient brought me was, honestly, beautifully made. Clean, well designed, more readable than most of what my profession produces. It cited a real paper, and the paper is free to read, which is more than most posts manage.

The chart showed clot risk by route: the patch below the line, the pill above it. Route is the story. Pills are the problem. That is the entire message, and it is the message she acted on before breakfast.

Then you open the paper.

The same table that produced the chart has more rows than the chart drew. It is a large study, more than 20,000 people who had clots compared with more than 200,000 who did not. And in that table, oral estradiol, the bioidentical molecule, taken with a progestogen, comes out at 1.14 with a range of 0.95 to 1.37.

Look at the range. It crosses 1.0. By tool two, that comparison did not find a difference.

That row was in the same table as the chart. It was not drawn.

And one table over is the comparison that actually tests the post's thesis. Same route, both taken by mouth, different molecule: conjugated equine estrogen against estradiol, 1.33 with a range of 1.02 to 1.72. The risk moved when the molecule changed, with the route held still.

If route were the whole story, that number should not exist.

And apply tool two to my own number before you accept it. The same table also compares the two molecules in estrogen-only preparations, and that one comes out at 1.19 with a range of 0.99 to 1.43. It crosses 1.0. By my own rule, that comparison did not find a difference. One result and one tie is what that table gives me, and I would rather tell you that than have you find it.

I want to be careful here, because this is where I could do exactly what the post did. That is not a randomized head to head. It is a comparison between drug categories inside an observational study, "any progestogen" covers a mixed bag of different molecules, and the whole thing counts prescriptions that were filled. The authors say so themselves on page 8: they had no information on whether the medication was actually taken. Those caveats are real and they travel with the number.

But the point stands, and it is not a small one. The paper behind the post contains the comparison that complicates the post's headline, four lines from the bottom of the same table. Nobody hid it. It simply was not drawn, because a chart with one message is a better chart, and a chart with one message is a worse map of the evidence.

The honest number

For oral estradiol, that same paper works out to roughly one additional clot per year for every 2,320 women taking it.

I am giving you that figure first, before anything reassuring, because you should not have to go looking for the number that cuts against me. It is a real risk. It is not zero. It belongs in a conversation with your own clinician, alongside your own history, because a woman who has already had a clot is in a completely different conversation from a woman who has not.

What it is not is a reason to stop a medication in your kitchen at 7am because a graphic frightened you.

What I actually want you to walk out with

Not a position on hormones. A reading skill.

The next time something crosses your feed with a number in it, run the four questions. What kind of study. Does the interval cross 1.0. Of what. Compared to what. Four questions, about ninety seconds, and you will understand that post better than the person who shared it.

And then go read the paper. The one in this episode is free and open access. Every figure I quoted is in the show notes with its source, including the row that never got drawn. Check my work. That is the entire point, and it is the opposite of what the graphic asked of her, which was to believe it.

She did not need somebody to trust. She needed a way to look.

This is part one of a two part episode. Part one is the reading skill. Part two, What Estradiol Actually Does, goes to the evidence on the other side of the ledger: not what the therapy might cost, but what the deprivation costs. It is coming soon.

🎧 Listen to part one: the podcast is out now on Spotify.

📺 Watch the masterclass:How to Read a Study: Estradiol Edition, now on YouTube.

⚠️ Educational content, not individual medical advice, diagnosis or a treatment recommendation. Nothing here tells any woman what to take, and nobody should start, stop or change a prescription because of an article. Your own labs and history belong in a one on one.

🩺 DNP led telehealth hormone optimization. Book a consultation and find everything: withinyou.health/links

Sources

Weller SC, et al. Res Pract Thromb Haemost. 2023;7:e100135. Open access. 20,359 cases and 203,590 controls. Table 3: transdermal 0.70 (0.59 to 0.84); oral estradiol 1.24 (1.09 to 1.40); oral estradiol plus progestogen 1.14 (0.95 to 1.37); CEE/MPA 1.52 (1.25 to 1.84); ethinyl estradiol plus norethindrone 2.35 (1.71 to 3.25). Table 4: CEE versus estradiol 1.33 (1.02 to 1.72) combined, 1.19 (0.99 to 1.43) estrogen only. Page 8: "filled prescriptions without information on adherence."

Vinogradova Y, et al. BMJ. 2019;364:k4810. Oral estradiol alone 1.27 (1.16 to 1.39).

Renoux C, et al. J Thromb Haemost. 2010;8:979-86. 1.49 (1.37 to 1.63).

Smith NL, et al. JAMA Intern Med. 2014;174(1):25-31. CEE versus estradiol 2.08 (1.02 to 4.27), P=0.045, on 68 clots.

Boardman HMP, et al. Cochrane Database Syst Rev. 2015;3:CD002229. 19 randomized trials; p.30, "a 'class effect' of hormone therapy, which may not be warranted."

Chlebowski RT, et al. JAMA. 2020;324(4):369-380. The WHI used conjugated equine estrogens, not estradiol.