Gaslighting and Professional Gatekeeping

So I've been studying for the MRCPsych again, after passing the first of three exams. The standard resource overworked and time-crunched psych-trainees (i.e. all of us) use for this is called SPMM. SPMM costs money. A lot of money. If you pay extra, you get mocks. If you feel desperate, you can pay even more money, for a "stats crash course,".

The MRCPsych Paper B is heavy on statistics. Most of my colleagues find this difficult because most of my colleagues went to medical school, where the stats curriculum is "p < 0.05 means good." I find it less difficult, which is not the same thing as enjoyable.

The problem is that SPMM's stats teaching is bad. I read a passage, frown, screenshot it, send it to ChatGPT or Claude, and they often go "yeah, this is wrong." This has happened enough times that it has stopped being surprising and started being a sort of recurring bit. The remaining percentage of the time, the notes are oversimplified to a degree that would make an actual statistician put their head in their hands, which I'll grudgingly accept as the cost of doing exam prep.

Now, I present the question that broke me today.

There's a study, presumably real, looking at childhood trauma and hallucinations in adulthood. There's a graph. The y-axis is the probability that someone has childhood trauma somewhere in their history. The x-axis is "number of hallucinatory modalities," running from 0 through 5, because apparently the authors decided five was enough senses and shipped it. Sure, I guess we can skip proprioception, vestibular positioning, and the other minor crap.

The question: "The predictor that emerged as the statistically significant variable is most likely to be..."

I pick ratio variable.

Let's examine my reasoning. The variable is a count. Counts have a true zero, where zero modalities is a meaningful and non-arbitrary absence rather than an arbitrary point on some scale. The intervals between values are equal: going from 1 to 2 modalities is the same conceptual distance as going from 3 to 4. And the ratios mean things, in that a patient hallucinating across four modalities is doing so across literally twice as many modalities as someone hallucinating across two. This is the textbook definition of a ratio variable. Stevens 1946. Every introductory stats book ever written.

SPMM informs me that the correct answer is "ordinal." I wanted to die.

I tried to be charitable. Maybe SPMM is operating from some idiosyncratic but defensible framework I'm not seeing. Maybe there's a niche position that count variables with low cardinality should be treated as ordinal because of how they behave in small-sample inference. Maybe somebody, somewhere, has a real argument.

I went to ChatGPT and laid out my case. ChatGPT said my logic was textbook, my answer was correct, and SPMM was wrong, but I should still pick whatever SPMM said in the actual exam because the exam doesn't care about being right, it cares about agreeing with whoever wrote the answer key. This is very sound career advice and also makes me want to lie down on the floor.

I pushed back. I said no, I have standards, I want to know whether there is some technical sense in which SPMM could be right and I am wrong. ChatGPT politely declined to manufacture one. After some pushing, it concedes, and says that I should go with my original answer if this specific question comes up in the actual exam. I screamed internally, because screaming externally would wake my neighbours.

The really insulting part is the upsell.

SPMM, having taught me statistics with the precision and care of a man hammering a nail with a banana, also offers a paid crash course in statistics. I will let you guess how I feel about paying them additional money to clear up confusions they themselves introduced.

There are two hypotheses.

Hypothesis A: SPMM is teaching stats badly on purpose, so that you have to buy the crash course to fix what they broke. This is the version I want to be true, because at least it would mean someone, somewhere, is in charge.

Hypothesis B: at some point, an underpaid registrar was handed a brief that said "write 500 stats questions, here's £500, you have a weekend". They paid peanuts, and the monkey just offered me a banana.

Hypothesis B is overwhelmingly more likely. It is also worse, because at least under Hypothesis A there's a coherent villain. Under Hypothesis B there's just a pipeline of tired people producing slightly wrong content for other tired people, who then sit in their flats at midnight wondering whether they're losing their minds or whether the material is in fact wrong, and there is no one to be angry at, because everyone in the chain was doing their best with too little time and not enough money, and the result is diffuse ambient wrongness that lives in PDFs forever, gaslighting trainees into doubting whether ratios still mean what they used to mean.

Anyway. Number of hallucinatory modalities is a ratio variable. It will continue to be a ratio variable on the day of the exam, and if it doesn't, the number of homicide victims in a future study will remain a ratio variable.

---

PS:

The study is "Association between childhood trauma and multimodal early-onset hallucinations" in the BJPsych. The authors ran a hierarchical binomial logistic regression with childhood trauma as the binary outcome and number of hallucinatory modalities as a numeric predictor, reporting OR = 2.24 (95% CI 1.16 to 4.33) per additional modality. You can't meaningfully report a per-unit OR for an ordinal predictor. Reporting that OR is exactly what you do for a count variable on a ratio scale. The original authors treated it as ratio. SPMM marked me wrong for agreeing with the people who wrote the study.