Ok, so I’m down on the NHST analysis of allergic rhinitis. And, I probably am not going to get ahold of the data to re-do the analysis, but what kinds of things SHOULD be done if the data were actually available?

First off, let’s acknowledge that the goal of all of this is to make people feel and function better. There is therefore a latent “how good do I feel” variable which we can only measure imperfectly which is our real subject of interest. There are two ways they tried to measure this variable. The first is the RQLQ (symptom) score which says how bad the person perceives their symptoms, and the second is the RMS (medicine usage) score which tells how much the person used medication to combat the symptoms. However, the RMS score also is a measure of a kind of forcing function to reduce symptoms. The medications available were Cetirizine and oral steroids, and a diary was kept by each patient so we actually know how much medicine of what types was used on every day of the study..

So, let’s imagine a function S(t) for “symptoms”. We have some causal process we think is going on in which S increases through time when exposure to pollen occurs but eventually saturates (fortunately pollen rarely actually makes your whole body combust). We have some treatments thought to reduce the symptoms, including medications which are taken according to the patients own choice, based on a tolerance for symptoms, and in addition, sham acupuncture, or real acupuncture. Real acupuncture is given either in period $$t \in [0,8]$$ or $$t \in [8,16]$$ weeks and symptoms assessed at t=0, t=8, t=16, t=60. Furthermore, we have individual patient beliefs about whether sham or real acupuncture was being applied, which were assessed after their 3rd session with needles.

The model should therefore be that S(t) is observed with errors using the RQLQ assessment, and furthermore that the choice to take various medications at time t is based on a combination of S(t) and individual tolerance for the symptoms (so that M(t) gives us information about the patient’s S(t)). If M(t) is the medication dosage function (dose per day) then $$M(t) = M(S(t), Tol)$$ and also $$dS/dt =F(M(t), Ac(t), Sa(t), B, P(t))$$ for each patient separately, where M is medicine dose, Ac is acupuncture dose, Sa is sham acupuncture dose B is belief in whether they are getting sham or real acupuncture, and P(t) is pollen exposure which we might imagine as a constant for each clinical center.

Put all of those ideas together with some simplifying mathematical assumptions, analyze with a Bayesian computational software program like Stan, and discover how strongly Ac, Sa, and M affect symptoms. Compare the effect Ac/M and Sa/M averaged across people to get an answer to the question “what is the average effect of Acupuncture measured relative to the average effect of taking Cetirizine on patient Symptoms, and how does it compare to sham acupuncture?”

One Response leave one →
1. June 20, 2016

If you allow for acupuncture to adjust tolerance for symptoms (ie. a psychological effect of being poked with needles rather than say an immunological effect that reduces mucus and itching etc) then I think we find that it’s difficult to figure out which is going on, are symptoms adjusting downwards so that at constant tolerance then less medication is used and better RQLQ scores are obtained? Or is tolerance changing so that at constant symptoms the medicine usage is going down and the RQLQ reporting is improved? Or what’s maybe most likely, perhaps BOTH at the same time. We could assess this in part by using the assessment of the patient’s belief, but it’s plausible that the psychological effect could exist independent of the patient’s belief (ie. perhaps just seeing a needle stuck in your body would produce some similar effect regardless of whether we think we’re getting real or sham acupuncture) it would be even better if we had some objective measure of immune system function, such as cortisol levels, blood counts, measurements of secreted molecules in the mucus, histamine response to a scratch test… whatever.