NNT has some shortcomings for psychiatric practice, as these medications work on a gradient, rather than a binary outcome like a cardiovascular event. Obtaining remission from depression, a dimensional illness, is distinct from meaningful improvement (which I know can be folded into response), whereas for say a statin LDL lowering is almost of no consequence except as for reducing the event of a stroke. It is also true that there can be improvements in person's life that are not captured by HAMS-D 17, and some things that don't matter, such as "insight into being ill."
I think the real interesting dive is into the specific measures that see improvement from the respective antidepressants - i.e. if agomelatine's effect size via is primarily reduced through sleep on HAM-D, then that is an important mediator for antidepressant selection. It can help one move into thinking about the type of depression and the network symptomatology for the person that is keeping them in a depressed state. It can also probably yield the true lack of meaningful efficacy for some of the agents we use - for example, if a participant is experiencing significant insomnia, this could correspond to a total of 6 points added to their HAM-D, as there are three questions about this with points 0-2. You could give a medication that improved insomnia dramatically just by sedating them (is this amitriptyline's secret sauce?) and move their HAM-D down by 3 points (1 point for each insomnia question), which corresponds to an effect size of 0.5. Their HAM-D would be improved but most people would not say this is a real targeting of depression, and could end up causing long term harm as the trial is only 6 weeks and the harm of disrupted sleep architecture, potential cognitive dulling, may only manifest over the long term.
It stinks that we lack a better or more valid construct - MDD is so poor and so it is no wonder that medication effects are going to be extremely noisy. At his point I have little curiosity for more MDD analysis; would much prefer to see analysis of the individuals in these trials who seem to get a high treatment response to which drug, looking at their specific depression profile along with what the drug modifies, in addition to other moderators and mediators.
Great article, thanks for writing and diving into the details of the research, Nils. Interesting to read from the perspective of a resident psychiatrist.
Interesting write-up! I'd love to see some discussion of treatment efficacy for anxiety vs. depression. As somebody with a couple of anxiety disorders and zero issues with depression, I feel like depression always gets center-stage, with anxiety being the red-headed stepchild.
I'll add it to the ever growing list of essay topics. Interestingly enough the "antidepressants" seem to help more with anxious symptoms than depressive symptoms when you go symptom-by-symptom
I agree. The more I practice, the less and less impressed I become with antidepressants for depression. Better results with anxiety and definitely better results with OCD and related things.
I wouldn't call the scores meaningless, you just need to know your tools and what they're measuring.
I didn't talk about it in this essay, but the HAM-D 17 has particular problems in that its questions are not very depression specific - there's a ton of overlap with anxious symptoms. The MADRS is a better tool for research. I think the Beck Depression Inventory does a decent job for clinical work.
Brilliant article, thanks so much.....I suspect we'd do better with fluoxetine esp started at half the dose....a lot of neurodivergent ppl are very sensitive to medication and do better starting slowly/low dose
As a patient of long standing, I say that if you laid all the psychiatrists that have treated me end-to-end, I'm sure the would be grateful but they'd still point in different directions. My best experience was with an empiricist who was able to accurately evaluate self-reports on efficacy.
It's an odd choice, but from a statistical point of view it has no effect on the noisyness of the data since 1 is always the "not applicable" answer and all of the questions have the same score range.
Edit: To expand on this. You can see that it doesn't introduce any real noise into the analysis because you could easily shift the scale to 0-6 by subtracting 30 from the composite score.
I’d have to disagree, because in many analysis they look for >X% change which would be a greater change if the scaling was from 0. Maybe I’m misreading that.
I wouldn't call that statistical noise, but I see what you're saying.
I think you would have a good argument if the % changes they were looking for were just arbitrarily set because 25% and 50% are nice round numbers that looked good, but those thresholds are based on clinical effect. i.e. 25% is right around a CGI-I of 3 and 50% is right around a CGI-I of 2.
Really your point is a (legitimate!) caution against taking percentages at face-value, and highlights why such thresholds need to be informed by clinical correlations. I think that's the solution as opposed to just making the absolute value of the range smaller, because you can do that infinitely. Even if you knocked the scale down from 0-6 you could then argue that the scale should use half-point steps instead of whole point steps to further reduce the distorting effect of percentages.
NNT has some shortcomings for psychiatric practice, as these medications work on a gradient, rather than a binary outcome like a cardiovascular event. Obtaining remission from depression, a dimensional illness, is distinct from meaningful improvement (which I know can be folded into response), whereas for say a statin LDL lowering is almost of no consequence except as for reducing the event of a stroke. It is also true that there can be improvements in person's life that are not captured by HAMS-D 17, and some things that don't matter, such as "insight into being ill."
I think the real interesting dive is into the specific measures that see improvement from the respective antidepressants - i.e. if agomelatine's effect size via is primarily reduced through sleep on HAM-D, then that is an important mediator for antidepressant selection. It can help one move into thinking about the type of depression and the network symptomatology for the person that is keeping them in a depressed state. It can also probably yield the true lack of meaningful efficacy for some of the agents we use - for example, if a participant is experiencing significant insomnia, this could correspond to a total of 6 points added to their HAM-D, as there are three questions about this with points 0-2. You could give a medication that improved insomnia dramatically just by sedating them (is this amitriptyline's secret sauce?) and move their HAM-D down by 3 points (1 point for each insomnia question), which corresponds to an effect size of 0.5. Their HAM-D would be improved but most people would not say this is a real targeting of depression, and could end up causing long term harm as the trial is only 6 weeks and the harm of disrupted sleep architecture, potential cognitive dulling, may only manifest over the long term.
It stinks that we lack a better or more valid construct - MDD is so poor and so it is no wonder that medication effects are going to be extremely noisy. At his point I have little curiosity for more MDD analysis; would much prefer to see analysis of the individuals in these trials who seem to get a high treatment response to which drug, looking at their specific depression profile along with what the drug modifies, in addition to other moderators and mediators.
Anyhoo, great work man!
Great article, thanks for writing and diving into the details of the research, Nils. Interesting to read from the perspective of a resident psychiatrist.
Interesting write-up! I'd love to see some discussion of treatment efficacy for anxiety vs. depression. As somebody with a couple of anxiety disorders and zero issues with depression, I feel like depression always gets center-stage, with anxiety being the red-headed stepchild.
I'll add it to the ever growing list of essay topics. Interestingly enough the "antidepressants" seem to help more with anxious symptoms than depressive symptoms when you go symptom-by-symptom
I agree. The more I practice, the less and less impressed I become with antidepressants for depression. Better results with anxiety and definitely better results with OCD and related things.
SSRIs are a miracle cure for me, so I'm always surprised to read about how useless they are (for depressives).
Useless on these meaningless depression scores, patients seem to improve more in clinic in real life.
I wouldn't call the scores meaningless, you just need to know your tools and what they're measuring.
I didn't talk about it in this essay, but the HAM-D 17 has particular problems in that its questions are not very depression specific - there's a ton of overlap with anxious symptoms. The MADRS is a better tool for research. I think the Beck Depression Inventory does a decent job for clinical work.
Brilliant article, thanks so much.....I suspect we'd do better with fluoxetine esp started at half the dose....a lot of neurodivergent ppl are very sensitive to medication and do better starting slowly/low dose
As a patient of long standing, I say that if you laid all the psychiatrists that have treated me end-to-end, I'm sure the would be grateful but they'd still point in different directions. My best experience was with an empiricist who was able to accurately evaluate self-reports on efficacy.
Very entertaining thank you
Why did you substract 350 in your AD excel sheet ? Where does that number come from ?
Thank you
Thanks for reading!
The 350 number corresponds to the placebo response rate (i.e. 35% of 1000)
I’m always confused at why the BPRS/PANSS stop at 1 instead of 0, wouldn’t a 0 produce more robust results and have less white noice? Just a thought
It's an odd choice, but from a statistical point of view it has no effect on the noisyness of the data since 1 is always the "not applicable" answer and all of the questions have the same score range.
Edit: To expand on this. You can see that it doesn't introduce any real noise into the analysis because you could easily shift the scale to 0-6 by subtracting 30 from the composite score.
I’d have to disagree, because in many analysis they look for >X% change which would be a greater change if the scaling was from 0. Maybe I’m misreading that.
I wouldn't call that statistical noise, but I see what you're saying.
I think you would have a good argument if the % changes they were looking for were just arbitrarily set because 25% and 50% are nice round numbers that looked good, but those thresholds are based on clinical effect. i.e. 25% is right around a CGI-I of 3 and 50% is right around a CGI-I of 2.
Really your point is a (legitimate!) caution against taking percentages at face-value, and highlights why such thresholds need to be informed by clinical correlations. I think that's the solution as opposed to just making the absolute value of the range smaller, because you can do that infinitely. Even if you knocked the scale down from 0-6 you could then argue that the scale should use half-point steps instead of whole point steps to further reduce the distorting effect of percentages.