Talk:Number needed to treat

From Wikipedia, the free encyclopedia

Number needed to treat is part of WikiProject Pharmacology, a project to improve all Pharmacology-related articles. If you would like to help improve this and other pharmacology articles, please join the project. All interested editors are welcome.
Start This article has been rated as start-Class on the quality scale.
Top This article has been rated as top-importance for this Project's importance scale.
WikiProject Medicine This article is within the scope of WikiProject Medicine. Please visit the project page for details or ask questions at the doctor's mess.
Start This page has been rated as Start-Class on the quality assessment scale
High This article has been rated as High-importance on the importance assessment scale

Contents

[edit] OR and RR

I have problems with odds ratio and relative risks being munged as they presently are. Why not present the OR=1/6 and the RR=1/4 as different measures? Johannes Hüsing (talk) 09:34, 18 April 2008 (UTC)

[edit] Hey

--Hey. I have no idea how to edit a table, but I think the worked example table could use a little --clarification. If you agree, could you make the change? --CER control event rate = events / subjects in control group --EER experimental event rate = events / subjects in experimental group --"Events" should be stated as "events in control group" and "events in experimental group", --respectively. Using "events" alone makes it sound like it's using the total events between groups. —Preceding unsigned comment added by 76.69.33.145 (talk) 22:08, 4 February 2008 (UTC)

I took out this example:

the drug gabapentin, frequently used for neuropathic pain, only gives appreciable symptom relief in about 33% of all cases; the NNT is therefore 3.

I think we need to know how often the pain goes away without treatment before we can compute the NNT here. NNT always depends on quality of the drug and likelihood of the defined endpoint without the drug. For example, if 10% of patients get better by themselves, then the NNT would be slightly bigger than 4. AxelBoldt 22:43, 30 May 2005 (UTC)

I simply added a number from my own experience. I should have sourced it from the Bandolier pain site. The NNT is not affected by spontaneous improvement. JFW | T@lk 08:32, 31 May 2005 (UTC)
From reading the Bandolier article, I believe NNT is affected by spontaneous improvement; in fact their first migraine example explicitely compares a placebo control group's outcome with the treatment group's outcome. With treatment 5% recur, without treatment 30% recur, so the NNT is 1/(30%-5%)=4. AxelBoldt 19:58, 31 May 2005 (UTC)
Number needed to treat refers to the number of individuals that need to be treated to appreciate a beneficial result in addition to those in whom the result would be observed without the treatment. It is derived in studies which have endpoints, therefore the endpoint for the outcome is implicit. The statement "drug x provides relief in only 33% of cases" isn't sufficient to determine a NNT. You need data on how many subjects got relief without geting drug x. See the difference between a case-control study and a cohort study.Bakerstmd 03:33, 25 May 2007 (UTC)
Shall we find a better example? It seems you're correct about the spontaneous recovery rate, but spontaneous recovery rate is subtracted from the response in the treatment group before calculating the NNT. JFW | T@lk 01:49, 2 Jun 2005 (UTC)


[edit] Statins

I disagree with the removal of the NNT data for statins. It is a concrete example written in plain English for non mathematicians and non science majors. If you have another contradictory value, add it and add the source. I am restoring the value. Business week, is a reliable source. I came to the article after reading the article and was surprised it wasn't mentioned. The example is written so that the average non scientific reader can understand the concept. --Richard Arthur Norton (1958- ) (talk) 11:40, 31 March 2008 (UTC)

As I indicated on your talk page, you are misrepresenting the source. People take statins for various indications, and you are not distinguishing between primary prevention sec, primary prevention in diabetics, and secondary prevention after numerous cardiovascular pathologies. I urge you to revise your contribution and discuss here first before reinserting this content again. JFW | T@lk 15:53, 31 March 2008 (UTC)
  • I think your confusing verifiability with truth. The number is verifiable from a reliable source, if you have information to modify it or disparage it, please fell free to add it. The encyclopedia is for everyone not just statisticians. You haven't presented any data that disparages the number used, and the source for the reference is given. ultimately its boils down to is Business Week a reliable source, and what evidence do have that disparages the number. You may notice that my information is sourced, the rest of the article except the lede is not sourced. No one is served by removing data from a reliable source. I suggest keeping the data, and you provide information disparaging the number, and they can exist side-by-side. I am not a clinician, I am a medicinal chemist and I found the BW explanation the best example in plain English. --Richard Arthur Norton (1958- ) (talk) 16:02, 31 March 2008 (UTC)

I wish you'd have the common courtesy to discuss before you revert! Your source is doubtful, but even so you are misrepresenting it. Your professional background has no bearing on this. JFW | T@lk 16:24, 31 March 2008 (UTC)

Neither might like what I have to say, but I am use to that, and perhaps you might find common ground using this as a start. First of all, the 1 or 2 sentences make for a great example in plain English - better even than the table, which of course is hardly surprising, as it the table more technical. Also, you will find it hard to find a technical paper that deals simply with the issue, as it a tool taken-as-granted. However, the context is terrible, specifically the conclusion that the NNT of 100 is not worthwhile (statins are harmless drugs, going generic too, so cheap) to save 1 person a heart attack - in fact the relative risk (3% / 2%) is a staggering 50% which is surely why they recommend 40m Americans should be on them. Furthermore the writer then immediately compares them to antibiotics for h-pylori, with a lower NNT, etc, and somehow cannot see the point - which is that very 1-in-a-100 person. But the reason is, of course, the journalist set out to write a "story", and so he adapted science to his end (consciously or not) - and an encypledia should not reference stories. Can I suggest instead that you simply adapt the text of that paragraph into the Wiki, and leave out the reference? The only other idea I can think of is that there are a number of on-line statistics book, and they may come close to what is needed as a replacement - but really it does not need a reference at all, it is fairly fundamental.....io_editor (talk) 17:19, 31 March 2008 (UTC)
In general, I have no problems with including information like this as a practical example. This issue is discussed frequently, with different drugs and different sources. See this magazine article for another set of numbers. However, rather than appearing like we're bashing statins to promote our point of view, I'd prefer to see this used as an opportunity to highlight some of the limitations of the NNT. For example, every study will produce a different number (at least slightly); different statins have different NNTs; different groups have different NNTs; every endpoint has a different NNT; combinations of drugs have different NNTs. The NNT for statins is different for a diabetic, a tobacco user, an obese man, a woman, etc. Furthermore, the NNT for heart attacks prevented over three years is different for the NNT measured for deaths prevented over ten years. There are a few scholarly papers addressing the issue, such as this one.
However, it might be simpler to switch to a different drug class: perhaps aspirin to prevent heart attacks, but perhaps osteoporosis would be better. There are some recent Cochrane reviews [1][2][3] that we could use for reliable numbers. Using osteoporosis treatments as our example appeals to me because the number of confounding factors is substantially lower compared to heart disease, and it gives us an opportunity to present the difference between "treating the lab report" and "treating the patient". WhatamIdoing (talk) 17:30, 31 March 2008 (UTC)

If the purpose of this article is to help readers understand number needed to treat, why not avoid controversial examples entirely? Primary prevention numbers are often high, and tend to have wider variability between studies. Maybe we could choose a secondary prevention example that isn't as open to various interpretations (I like the previous suggestion from WhatamIdoing of ASA for Myocardial Infarction).Sisyphus (talk) 19:02, 31 March 2008 (UTC)

Alansohn (talk · contribs) thought it would be nice to use the statin data as an example. I don't actually mind (I am not averse of statin-bashing when there is bashing to be done), but I object to the lack of clarity that flows from regurgitating the Business Week story without qualifiers. I have now rephrased the addition in a way that I think makes reasonable sense: the absolute risk of cardiac events in the ASCOT-LLA study was low, therefore to get them lower still with a small dose of atorvastatin was quite a feat.
My version avoids the classical "big pharma versus the public interest" cliché and obviously drives home the message that primary prevention with statins is always going to be an expensive business (see the bandolier summary, about as good science as you're ever going to get). Comments invited. JFW | T@lk 19:59, 31 March 2008 (UTC)
If the statin example remains controversial, I propose PMID 17943857 as a clear case of the use of NNTs to guide treatment decisions - a Cochrane Review about the use of antidepressants in the treatment of chronic pain. JFW | T@lk 20:09, 31 March 2008 (UTC)

[edit] Business Week redux

I am adding back in the BW reference. It seems you prefer citing journals written by statisticians for other statisticians. The BW is the only RS definition written in plain English. This here is a great article if you already know the subject, BW is great for me, and I am a scientist. One more reference does no harm. --Richard Arthur Norton (1958- ) (talk) 20:26, 31 March 2008 (UTC)

I had the impression that my addition was reasonly plain English. If it is not, please change the phrasing to make it more understandable. I have preserved the source as a compromise, but I have removed the quote because it adds nothing to the summary already provided. I hope you can live with that. JFW | T@lk 21:11, 31 March 2008 (UTC)

You expressed NNT as a percentage, which did not seem correct, but I didn't change it for fear of having it reversed. I had already trimmed the quote to satisfy your objections. --Richard Arthur Norton (1958- ) (talk) 22:11, 31 March 2008 (UTC)

That was indeed a mistake, which io io editor (talk · contribs) has already corrected. I still find the quote terribly imprecise (it's not 3% vs 2%, those are rounded figures), and the article remains in my view a shabby source. It doesn't quote the name of the trial, let alone provide a full reference, and it presumes that just because Pfizer paid for it, it was therefore "conducted by Pfizer" (which is incorrect and casts aspersions on the integrity of the researchers).
I did ask you directly to correct my version to make it sound more like plain English. I really don't understand why you prefer a quote in a footnote over an opportunity to make the Wikipedia content easier to understand. JFW | T@lk 07:49, 1 April 2008 (UTC)

I don't think the Business Week article serves as a reliable source for purposes of this article. However, I take the point from Richard's argument that the Business Week article explains NNT in a way that is more intelligible than most. So how about modeling the explanation here after that article? --Una Smith (talk) 15:34, 1 April 2008 (UTC)

I have no problem with the style used by BW, although I note again the several problems with the example chosen by the journalist. JFW | T@lk 15:48, 1 April 2008 (UTC)
There is no ruling from Arbcom on using quotes in footnotes. A Supreme Court agreeing to hear a case, that you have an opinion on, does not mean your opinion has been validated. It only means that Arbcom agreed to hear the opinions of the involved parties. Deleting the reference that you have already expressed a negative opinion on is premature. I know you don't like references from non medical journals, but this is an encyclopedia for everyone, not just people who have degrees in statistics. I have already agreed to move the information from the body to a footnote. --Richard Arthur Norton (1958- ) (talk) 02:45, 16 May 2008 (UTC)
The Arbcomm is not an arbiter on style; there is consensus against the use of quotes in the way that you insist on doing.[4] My point in the edit summaries was that perhaps you should not be undoing my edits while being involved in that particular Arbcomm case. Until I saw that case I was unaware of the fact that you had been editing against consensus on WP:CITE all along. JFW | T@lk 17:19, 16 May 2008 (UTC)
Incidentally, I think the way the page deals with the example at present does not need "degrees in statistics" (you keep on using the same point, which I have addressed numerous times). If people are desperate for a plain-language summary written in newspaper style by a journalist busy trying to prove a point, they can click the URL to the Business Week article. Strictly speaking, the Bandolier source is relatively plain-language and is "fit for purpose" vis a vis the Business Week source. JFW | T@lk 17:22, 16 May 2008 (UTC)

[edit] Let reader choose examples

Although I understand the concept of NNT, this article still is rather opaque to me. Probably also to many other readers. So, how about making it more concrete for the reader by adding links to other Wikipedia articles on applications of NNT? Also edit those articles to include the NNT data. --Una Smith (talk) 05:07, 1 April 2008 (UTC)

This search shows a moderate-sized list of articles linking to NNT. Some of them might be appropriate, although the few I checked tend to mention it in passing, instead of explaining it. WhatamIdoing (talk) 05:51, 1 April 2008 (UTC)
Good idea. Looking at those links, I find inconsistent usage: NNT is expressed as "1/N" or as "N". --Una Smith (talk) 06:02, 1 April 2008 (UTC)
The article already gives examples. We'd be better off providing more examples rather than using other content elsewhere (which may change) to illustrate our points. NNT is a number larger than on 1 by definition, so 1/N probably refers to 1/ARR (absolute risk reduction). JFW | T@lk 07:49, 1 April 2008 (UTC)

[edit] Definition

Above, JFW says "NNT is a number larger than on 1 by definition" but I don't see that in the article. On the contrary, the article Derivation section defines NNT as 1/(pB - pA) which looks like a number smaller than 1. And pB - pA in the Derivation is ARR in the Worked Example. It is possible for ARR to be greater than 1, in which case NNT would be smaller than 1. The Derivation is incomplete. Also, the Worked Example has an internal inconsistency: it states that OR is both 4 and 0.25. Finally, I find it more helpful to give a series of abstract equations first in full, then with numbers plugged in, rather than in a sketch parallel as in the table format. --Una Smith (talk) 15:49, 1 April 2008 (UTC)

The NNT is computed as 1/(pB – pA) which is in fact "a number larger than 1 by definition" - because pA and pB are probabilities measured not as percentages but as decimals (of one) - and 1/(decimal-of-one) is always "a number larger than 1 by definition". It is also not possible for ARR (this is not the relative risk, but the difference between those pair of decimals) to be greater than one. You are correct about the OR, it is confusing the way that it is both 4 and 0.25 (in my opinion, I prefer 0.25 here, but it should not be written as .25 as the decimal then is hard to spot). I always prefer tables or bullets above prose when a list of technical things w/numbers or w/attributes are being described, and I think that is the case here; however, I do think that this table could be a little more descriptive; I only wish I had more time but hopefully this helps...io_editor (talk) 23:49, 1 April 2008 (UTC)
Chiming in to agree with io io regarding the math and the use of a table. I'm thinking that the prose leading up to the table could be "de-condensed" a bit. Antelantalk 01:50, 2 April 2008 (UTC)
Progress! If pA is greater than pB, 1/(pB – pA) will be negative, so the absolute value of NNT is still greater than 1 but the sign will be negative rather than positive. I prefer something between prose and the current table, namely blocks of "math" prose a la Diophantine equation. The syntax is very similar to LaTeX, which I can do. --Una Smith (talk) 02:05, 2 April 2008 (UTC)
Yes, what you say technically is true, but it is generally meant in the context of drug v control (or placebo), and (usually, but not absolutely always!) it is known in advance that one should be greater than the other. As to your example, while I love the huge font, in my work I like ready-reckoner stuff like tables (plus my eye-sight tires) - maybe combine best of both worlds or ask Antelan...io_editor (talk) 02:36, 2 April 2008 (UTC)
Can we add a sentence that says if the treatment has the opposite effect (eg, people who take a drug to prevent heart attacks are unexpectedly more likely to have an MI than the untreated group), then you'll get a negative/nonsensical number (that is, a number smaller than zero)? I'm not sure that I can make what's clear in my head be sensible in words. I strongly support including the equations. (Please don't tell my math geeks.) WhatamIdoing (talk) 06:13, 2 April 2008 (UTC)

If a treatment causes harm, then the NNT becomes an empty number (there is no "treating" being done). In fact, a negative NNT is the same as the number needed to harm (a parallel page on adverse effects, more properly named "number to treat to harm"). JFW | T@lk 11:37, 3 April 2008 (UTC)

Yes, but you're talking common sense, and I'm talking math. 1/(0.5-0.75) = –4, even if that doesn't mean anything in the real world. Also, there's a divide by zero problem if there's no difference between the two groups. I think this article should tell readers how to interpret these two cases. WhatamIdoing (talk) 01:56, 4 April 2008 (UTC)
I agree. --Una Smith (talk) 02:01, 4 April 2008 (UTC)
Perhaps we should merge number needed to harm into this article, and mention that although the equations are identical, the label used depends upon the sign of the result. --Arcadian (talk) 00:00, 5 April 2008 (UTC)
I would say needless, as JFW was implying. In such circumstances (no "treating" being done, and possible harm) the tool is not usable - its an irrelevant formula, and is not an NNT of any kind.io_editor (talk) 18:27, 5 April 2008 (UTC)
I would merge, perhaps into Number needed; that would help to explain the bifurcation into harm and treat. --Una Smith (talk) 18:39, 5 April 2008 (UTC)
The bifurcation is only in the formula. In actual pratice there is no chain-of-logic that leads to a bifurcation. I dont mind if you merge them, but I think it will be a waste of your time better spent - and quite difficult too. Other pages need real work. See you.io_editor (talk) 19:49, 5 April 2008 (UTC)