We know without a doubt that taking a drug (statin) that is intended to lower LDL lowers cardiovascular event probability even for mildly elevated LDL.
Wouldn’t that viewpoint be to set aside the existing pleiotropies of statins? Or that in the few studies with isolated high LDL it seems that the effects are other, as in 4s post hoc, or TNT for example?
Could we be facing a Simpson’s paradox with the information available so far? Is Mendelian randomization still subject to this bias? Yes, from my point of view, but you are much more aware of these limitations and in-depth knowledge.
Thank you @ESMD for referring to my previous post and alerting me to this discussion. In addition to specific questions about LDL-cholesterol as a modifiable risk factor for atheromatous disease, you @Novato_Deseoso seem to be raising issues about the nature of causation and how to design experiments to test for it. You must not forget that the link between lipids and vascular events was established because hypotheses about formation of the underlying atheroma suggested the link in the first place. An important text about casual inference is: Hern´an MA, Robins JM. Causal Inference: What If. Boca Raton: Chapman & Hall/CRC, 2022. There is a free downloadable version here: Causal Inference: What If (the book) | Miguel Hernan's Faculty Website | Harvard T.H. Chan School of Public Health
According to my understanding of causal inference, an important concept is that of a ‘modifiable risk factor’. A RCT on statins should show not only a lowering in incidence of vascular events but also a corresponding reduction in the LDL-cholesterol if the latter reflects the underlying causal mechanism. We can expect the risk reduction to be proportional to the pre-treatment LDL-cholesterol value as suggested by @f2harrell. However, we would not expect to see a reduction of other risk factors for vascular events such as age or the blood pressure. When we calculate the effect of reduction in vascular events due to statins, we should only do so via the lowering effect on LDL-cholesterol and what the new risk is due to the new, lowered LDL-cholesterol.
If there are other risk factors (e.g., old age, high BP) then we can’t expect a statin to reduce these and their effect on risk of vascular events as well. However, this is what some tools do such as: https://statindecisionaid.mayoclinic.org 3. These multivariable risk factors work by adding individual risks where patient’s risk score = intercept + (bsex×sex) + (bage×age) + (bBP×BP) + (bLDL×LDL) etc.) when bsex, bage, bBP, and bLDL are regression coefficients that describe how a patient’s values of the predictor variables affect risk. Treatment only affects the relevant values of sex, age, BP and LDL (only the LDL in this case). By applying the risk reduction from statins to the overall risk we are also assuming that statins change sex, reduce age and BP as well as the LDL-cholesterol. If you wish to base the risk of vascular events on other risk factors related to LDL-cholesterol, I suppose you need to build these into your multivariable risk calculation and estimate the effect of statins on them…
Thank you very much for your valuable contributions.
In the triangulation that is often made of observational, RCT and Mendelian, I think we make the mistake of ecological associations, which leads us to a very biased perspective of what may actually be happening, from my limited ability to assess the limitations of the studies compared to you.
We have studies where polygenic indices completely eliminate any association with LDL, which is very telling.
Could we be looking at a simple projection that picks up all the time people with birth matches that if you pour gasoline on them they burn, and conversely those that you don’t pour gasoline on them don’t burn, and mixing it with people with lipid triad so extraordinarily present in society?
For example, Velican’s studies of very slight morphological modifications could be among others, those matches, which I understand is undetectable by any type of study, since it would only be the match of the group where we pour more gasoline.
Than you
I’m not sure about this statement…
Most physicians were not trained to interpret regression equations (me included). So most of us have to trust that many eminently qualified statisticians were involved in the development of risk calculators that have been used globally for decades, and in advising physicians about how best to “transport” the results of RCTs to the patients we see in clinic. While not impossible, it seems, to me at least, quite implausible that we’ve all been doing things wrong for the past 25 years.
For whatever it’s worth, this is how I conceptualize the causal path:
I’m thinking of all this using the framework shown in figure 10 of this publication:
I’m really trying to understand your concerns, Huw, and I wonder if the underlying causal diagram is the main point of contention. I sense that you perceive an inconsistency between how CV risk factors are treated in the process of estimating an individual patient’s future MI risk, versus how we conceptualize the potential “relative treatment effect” of statins.
Specifically, you seem to be viewing CV risk factors as “prognostic” factors, such that non-lipid risk factors (e.g., smoking/HTN/DM2/age) would not be expected to biologically “modify” the relative treatment effect of statins (e.g., the HR) in a statin RCT. In contrast, my conceptualization above treats the non-lipid risk factors as both “prognostic” and “predictive” [a view that seems, to me, to be more in keeping with our current biologic understanding of the interactivity of various CV risk factors].
I sense that you are troubled by what you perceive to be a “bait-and-switch” with regard to how the medical community estimates potential statin benefits. I think you’re saying that if we are going to treat non-lipid risk factors as strictly “prognostic” for the purpose of estimating a patient’s future MI risk (i.e, no arrow from the other CV risk factors toward " plaque stability/ plaque progression"), then we shouldn’t act like treating a patient with a “good” LDL with a statin could, under any circumstances, be expected to confer much CV protection (?) In contrast, as discussed previously, I tend to view the underlying known biology as strongly supportive of a view of LDL as interactive with the other risk factors, as though some patients will just “tolerate” a given LDL level better than others, depending on the presence/absence of other risk factors. While you seem to view the proposed interactivity of CV risk factors as post-hoc rationalization that is being used to bolster the perceived benefits of statins in primary prevention contexts, I view the interactivity as strongly supported biologically.
I don’t think you’ll find any serious scientist who doubts the role of LDL in atherogenesis. There are many “bad actors” (including a very small number of medical professionals who have, unfortunately, very loud voices) who have built lucrative careers around sowing doubt about the role of LDL in atherogenesis by appealing to the conspiracy-minded segment of the population (which is not small). They are, without a doubt, responsible for countless premature deaths from coronary disease globally. Primary care practitioners and cardiologists fight hard to counter these highly pernicious forces every day in clinic. These grifters should be assigned to console the children of patients who have dropped dead prematurely from an MI that might have been prevented if only they hadn’t read a screed from “that famous doctor (insert bad actor MD or PhD of choice) on the Internet.”
Very much in agreement, that is why recent findings are so relevant and many recent discussions and perspectives.
My point has nothing to do with what you indicate in this sentence. But there are many, many points that outweigh already at this point LDL which is clearly off under APO B or Non-HDL-c, not to mention Lp(a) or high draft opinions like Libby’s who is far from being a denialist.
But my question is more about statistical plausibility and evaluation of the existing science and the limits of each type of study, than about the existing evidence which I know quite deeply and follow on a daily basis. Thank you very much for participating
I try to go deeper into these questions about the existing gaps in the types of research available to us in this specific subject.
Hi again Erin
I agree with your conceptual diagram. However, if the risk from young age, non-diabetic status, non-smoking status and normotension was 4% and that from pre-treatment LDL-C was 5% then on the additive scale the total risk would be 4%+5% = 9%. On a statin the new risk would be still 4% from the 4 risk factors but 5% x 0.6 = 3% from a risk reduction of 0.6 from high dose statin. Therefore, the new risk by applying a risk reduction on the additive scale would be 4%+3% = 7%. However, by applying the risk reduction to all risk factors, the new risk on statin treatment would be 0.6(4%+5%) = 4% x0.6 + 5%x0.6 = 2.4%+3% = 5.4%, lower than the 7% and therefore exaggerating the risk reduction in that patient.
I am not saying that statins are not important in preventing vascular events or do not work. Far from it. If I was presented with someone with a high risk of vascular event and a high HbA1c, a high BP but a low LDL-cholesterol, then I would persuade the patient to allow me to help to improve diabetic and BP control and focus on this at least initially. I do not believe that much would be gained from offering a statin as the risk reduction would be small when calculated on the additive scale (unless there is a RCT that contradicts this). I would not burden the patient with another pill. However, if the LDL-cholesterol was also high or the patient was already known to have had a vascular event, then I would add a statin because the expected risk reduction would be greater. Also, if a 75-year-old person with an inevitably high risk of a vascular event in the next 10 years had a low LDL-cholesterol, low BP, and no diabetes etc, I would not recommend medication.
As far as I understand, this is how most physicians would have been reasoning over the past 25 years. In this sense they would not assume that they should treat a high risk of vascular event with a statin irrespective of its cause(s). You seem to assume that there is a very strong interaction between non-lipid based risk factors (e.g. age) and lipid-based risk factors, which should be modelled by an effect on all the risk factors (which if untrue would potentially increase sales of statins unjustifiably). However, as far as I can understand, the way that multivariable risk factors are calculated assumes that there is no such interaction at all so that there is independence between risk factors (maybe @f2harrell could comment on this). Perhaps the truth lies in between. Establishing such a truth would be very difficult as we discovered by looking at the literature previously (Risk based treatment and the validity of scales of effect - #12 by HuwLlewelyn).
Hi Huw
I would put this patient on a statin and would prioritize getting him on a statin and controlling his BP over controlling his blood sugar (though I would also put him on an SGLT-2 inhibitor and/or GLP-1 agonist if he needed something to bring his sugar down).
I think we’ll have to agree to disagree on this point.
I would not disagree strongly with you as the evidence is very flimsy on these issues. I would give my reasoning and suggestions to the patient, who might well ask for a statin for a low LDL-cholesterol as you would recommend. However, if we accept the assumptions and reasoning used to construct multivariable risk scores, then to apply them logically, we should estimate the risk reduction based on the statin’s effect on lipids alone and not also on an assumption that statins reduce age, BP etc (an assumption that the risk calculators do not make). If the risk calculation were to be changed to take into account some way that age, BP etc modify the effect of statins on vascular risk, then that would be a different matter.
https://pubmed.ncbi.nlm.nih.gov/15325833/
"Interpretation: Atorvastatin 10 mg daily is safe and efficacious in reducing the risk of first cardiovascular disease events, including stroke, in patients with type 2 diabetes without high LDL-cholesterol."
https://www.bjd-abcd.com/index.php/bjd/article/view/310/472
“For primary prevention, most trials and meta-analyses have demonstrated a significant benefit of statin therapy in reducing cardiovascular events in those with diabetes.”
https://guidelines.diabetes.ca/cpg/chapter25#bib0165
https://ccs.ca/app/uploads/2022/07/2022-Lipids-Gui-PG-EN.pdf
https://www.ncbi.nlm.nih.gov/books/NBK554923/
Primary prevention for people with and without type 2 diabetes
1.4.17.
offer atorvastatin 20 mg for the primary prevention of CVD to people who have a 10-year QRISK3 score of 10% or more. [2023]
Statin Treatment Recommendations
- The following are guideline recommendations for statin treatment:
- Patients ages 20-75 years and LDL-C ≥190 mg/dl, use high-intensity statin without risk assessment.
- T2DM and age 40-75 years, use moderate-intensity statin and risk estimate to consider high-intensity statins. Risk-enhancers in diabetics include ≥10 years for T2DM and 20 years for type 1 DM, ≥30 mcg albumin/mg creatinine, eGFR
“This meta-analysis is the largest in terms of pooling results from the largest number of articles and sample size. The main results indicate that statin use in patients with diabetes is associated with a reduced risk of CVD events and ischemic stroke in primary and secondary prevention, but is not associated with reduced all-cause mortality in either group”
Esential “The mechanism by which statins reduce the risk of CVD and stroke in diabetes is related to the potential protection of endothelial cell injuries induced by hyperglycemia”
“Diabetes may increase systemic inflammation, and increased inflammation and coronary calcification may counteract the effects of statins in patients with diabetes. Therefore, it seems reasonable to explain the negative results for mortality in patients with statin use and diabetes”
https://www.sciencedirect.com/science/article/pii/S0939475322003210
This will be my final contribution to this thread. Since I don’t sense that anybody here is open to having his/her mind changed at this point, continuing to lob citations back and forth likely isn’t productive. More importantly, this discussion isn’t about statistics (the purpose of this forum).
From the paper you cite:
“Elevated inflammation and coronary plaque progression have been proposed in statin users [[42]”
I don’t find this statement credible. A more widely accepted explanation:
“Statin therapy appears to accelerate the process of transforming a potentially highly metabolically active plaque to a more inert state. Specifically, statin therapy is associated with a decrease in low-attenuation and fibro-fatty plaque volumes and an increase in high-density and 1K volumes. Higher calcium density is associated with slowed plaque progression…
This present analysis supports findings from the above literature, i.e., suggesting a role of statin therapy in accelerating plaque transformation from noncalcified to calcified content and thus aiding plaque stabilization.”
I asked statistical questions, because it was not my intention to debate evidence in lipidology which I would be happy to do because it is a subject I study, but as you say it is not the right forum.
My questions remain open in the sense of the gaps or limitations of each of the evidences we consider to causality (observational, Mendelian and RCT), and the plausibility of the existing evidences showing the ideas I commented.
How do these three lines of evidence eliminate the possibility that it is one or more characteristics present in both groups that react with exposure and are inert in the absence of exposure?
Thank you all very much again
I’m sorry Erin. It was the above statement of yours that I was focussing on. If as you suggest, you replace ‘smoking’ with with ‘no smoking’, you get a risk reduction by repeating the ‘Mayo’ calculation. If you similarly replace the untreated lipid profile with a new improved lipid profile after treatment with a statin, you usually get a smaller risk reduction by repeating the calculation than that provided by the Mayo clinic site by clicking on ‘intervention’. Also a young person with a very low risk lipid profile but severe hypertension gets a surprising 40% risk reduction on a statin. None of these calculators include the risk from the presence of coronary artery plaque.
None of the studies on diabetics (e.g., Colhoun et al) summarise the absolute risk reductions for different baseline risks for each LDL-cholesterol value or of the overall kind provided by the Mayo clinic site. The whole thing is rather muddied (which is why I think that the evidence for making individual decisions is flimsy), leaving physicians to have to muddle through or over-simplify matters.
My understanding is that a statin would reduce the risk of a CV event in the treated group so that it is, for example, 40% lower than in the control group. If you look at all the subgroups (e.g. those with high LDL-C > 200mg/dl , diabetes, smoking, SBP > 150, age > 70, etc.) then you get 40% reduction in CV events in all these groups and their complements. However, if you look at these predictor variables after treatment it will only be the LDL-C that reduces in line with the drop in frequency of a CV event, suggesting a causal connection. The other predictor variables will not change and will therefore be non-causal associations. Does this answer your question?
I’m not sure that certifies causality.
Assuming the existence of unknown pleiotropies and for example unmeasured vascular effects correlated with LDL-c, I think we could be in trouble.
But the main thing is that as an analogy if I have a pipe with holes in it and I remove the oil (LDL-c) from the pipe, it will improve as long as I remove the LDL-c and linearity would be expected. But if I remove the holes from the people in both group and placebo the effect will disappear. Match or gasoline?
This I think is embodied in studies of PRS ( polygenic risk ), where the value of LDL-c disappears, or the lack of associations if CAC=0 is seen recently, or the adjustment in some studies by APOC-III and ApoB blurring.
Hence the doubt as to whether Mendelian randomization overcomes these limitations that I see in RCTs and more evident in observational studies. Or maybe my reading is wrong at some point that I fail to see, and maybe I am making some basic statistical knowledge error.
Thank you Mr. Llewelyn
Good point. However, by ‘causal connection’, I did not intend to imply a direct causal effect but that LDL-c was connected in some way to the causal effect. The latter may or may not be a direct causal effect. There may be some other mechanism (e.g. that a statin acts directly on an atheromatous plaque and at the same time reduces the LDL-c that does not directly cause the plaque). However, from a predictive point of view we can use the lowering of LDL-c to predict a lowering of vascular event risk but as age is not lowered by a statin, we cannot use age in the same way. When we press the ‘Intervention’ button in the Mayo calculator and apply a 40% reduction from a high dose statin to the overall risk due to age, lipid profile , SBP etc we do assume that all the predictor variables, including age, have a ‘causal connection’.
We are getting closer to my point, that the existing evidence does not guarantee us that LDL is the end of the road or a physiological cause, and that we are still on the limitation side of mechanistic knowledge of the state of the art, i.e., would you agree with these statements?
-
The existing evidence strongly suggests to us a causal relationship, from a probabilistic group perspective, but not an individual one, since we are not always limited on mechanistic grounds?.
-
None of the three existing lines of evidence ( MR, RCT, Epy ) can free us from the bias of mechanistic knowledge to establish physiological and therefore purely individual causation.
I am finding the discussion very enriching.
I think that it is generally agreed that we can only ever have a partial understanding of disease mechanisms or any other natural process (e.g., climate change). The best we can do is to apply this imperfect knowledge to choose items of information to allow us to estimate as accurately as possible the probabilities of important outcomes with and without interventions in order that we can make decisions that are as well informed as possible.