The Medical student Review

Now we are going to cover 'descriptive statistics'

you take the data, summarize ..

Statistics are made better beyond 'descriptive when it is inferential..
taking information and summing from it what the whole world must be like.. that is how we do science and research:

you want to be able to take the data and apply it to the general population so that every patient isn't a unique event ..
a quick example..
you have taken a test, someone is in a hurry to find the results so they take random samples from the class.. take scores and calculate the mean .. is this mean likely to represent the class as a whole? it is unlikely, but it is likely to be close!
that is what inferential statistics.. we don't have all the data but we guess them and for that we pay a price.

we give up 'precision' for what we call a 'degree of confidence' a certainty about what the number probably is.. this is what we 'believe to be reality for the rest of the world'
a confidence interval!
this is how statistics is done in medicine and it is said that if you don't understand how confidence interval works, you don't understand the articles in the journals.
confidence interval is an intuitive concept.
class starts at 8, you stroll at 5 after 8, the single number isn't reality but around there..
meet a friend at 'sevenish' if you are ten mins late.. you are within the confidence interval.

calculate how much high and how much low.. to come up with the same solution.

confidence interval
confidence interval is the mean +/- Z ( S/N)

x bar is the mean +/- symmetrically (for z see bottom) (the standard error) this is a measure of error.. since error = bad you want this number to be as low as possible .. the standard error is a measure of the quality of your estimate .. based on data you have, if you have alot of error, you have a poor estimate, if you have little error, you have a good estimate . the standard error is made of two things (s) and (N) standard deviation and sample size which is designated (n)

now n in terms of sample size, if you have 200 participants of research subjects, it doesn't make for as good a picture as a sample size of 2000 . more people more reality to work .. so we put sample size on the bottom, since as sample size increases error should decrease .

s.. if everyone gets 80% on a test no exception, pull 5 people at random and compute the mean and now take that mean to represent the class as a whole, is that a good representation? Yes!

test, all score all over the map.. take 5 people at random use it to calculate the mean, we get a poor estimate. thus, the smaller the standard deviation the better my estimate, the larger my standard deviation the poorer my estimate .. thus S goes on top ..

as the standard deviation gets larger the standard error gets larger, as the sample size gets larger the standard error gets smaller!

larger standard deviation more error, the larger the sample size less error..

what you have control over a a researcher is the 'sample size' (N) the sample size is your control to reduce error ..

there is no reality in a small sample size.

standard error gives quality of estimate.

Z = standard score not standard error as in parenthesis
each standard means something very different.

if they give you a 'standard error' don't divide by square root of N, it has already happened -- don't treat it like a standard deviation ..

a standard score is a distribution from the curve with known properties ..
this curve was invented to have a mean = zero and a standard deviation = 1
this is born of the statisticians imagination because these numbers are easy to work with ..

that way if it is a positive number it is on one side of the mean, and if negative then on the other . Whether a number is positive or negative we know instantaneously whether it is to the right or the left of the mean. Standard deviation = 1 means score is in standard deviation units. concretely a z score of +1 you are one standard deviation over the mean .. if it is -1.6 you are 1.6 below the mean .
if we wanted to get 95% of the cases in a z score distribution .. if we go between +/-2 of the z score distribution that gives us 95.5% of the cases .. but if we wants 95% exactly we need to pull the numbers from the right and from the left so in order to get 95% we don't go +/- 2 it is +/- 1.6 giving us exactly 95% of the cases ..

now if we wanted to get 99% of the cases ..+/-3 using Z score we get 99.7%, need to tuck it in for a nice round number tuck it in 2.58.. so all we need to do to complete confidence interval calculation is .. we know how good our data is, now how 'confident' are we with the interval so we use either or those two numbers.

either will have a 95% confidence interval in which case we plug in 1.96 for Z or we are going to be 99% confident in which case we'll plug in 2.58 for Z
z in this case is simply a binary switch ..

how to decide which number to plug in depends on the confidence interval in the question.. but to make it easier for 1.96 use 2 and for 2.58 use 2.5 ..

now we can compute the confidence interval of the mean

let's do question
we have a mean of 60%
a standard deviation of 9%
and a sample size surveyed 9 from several hundreds

what is the 95% confidence interval for that estimate of the mean?

60 +/- 2 (9 / square root of 9) = 60+/- 6

so the data tells us that the truth lies somewhere between 66 and 54.. we are not 100% that the mean is within that range.. we are only 95% certain. we bought 95% certainty with out z score.. we don't know where the mean is in that.. it is a flat probability distribution.. any number inside the range is as good as the other.
we need to project from this data reality, and this range is our window to reality .. given precision up for confidence.

*error of measurement and confidence interval are terms that can be substituted for one another.

if there is an overlap in data means there is no difference.
try to use that for poll data for political campaigns when they tell you someone is ahead by but within the margin of error!

remember that 99% confidence interval gives you a broader not a narrower picture of reality

confidence interval for 'relative risk'

if you look at a number, how can you tell statistically whether it is a relative risk or not? you can't tell, what you need is to look at the confidence interval .. confidence interval in research will be given as numbers in parenthesis an example for instance (1.12-2.5) from the low number as a possible low to the high number as a possible high .. the critical number if we are trying to determine if something is statistically significant is (1.0)
look inside the confidence interval and see if it contains the number 1!
if one is inside the interval then not statistically significant, if outside the interval then significant.

why is the number one a critical value? what do we do to compute a relative risk? we divide.. a number divided by itself give us 1 so if a risk in group A is the same as the risk in group B the output for the relative risk is one .. so one means the two groups are the same.. now how is it that you can prove that one group is different? you must prove that the output of the relative risk isn't one!
Exclude one and prove that you have a significant number ..

so using the above the lowest possible risk is 1.12 the highest possible value is 2.25 .. is that statistically significant? Yes, because one occurs outside the one.. we have excluded one as a possibility even in the worst case scenario therefore we do have statistical significance, and this indicated increased risk, why? because the relative risk is > 1

1.65 is that statistically significant relative risk? move your eyes to confidence interval and we have .. the CI has a low of 0.89 and a high of 2.34 .. not a statistically significant relative risk.. why? because on occurs directly inside the interval .. thus we haven't excluded one and therefore we can't assume we have statistical significance.
the second relative risk even though larger than the first isn't significant .

another sample relative risk of .76 a sample CI (.53-.93) excluded one therefore significant .. this one denotes 'decreased risk'
the difference between the 'pros' and the 'rookies' when reading these articles .. the pros are looking at the confidence interval, the rookies are staring at the relative risk .. the relative risk doesn't matter unless the confidence interval tells you it matters.

this is also true for odds ratio .. if you cross out 'relative risk' and write in its stead 'odds ratio' follows the same, look at the confidence interval one in means not significant, one out means it is significant .

if you read the abstract and understand the tables then that is it..

probability and probability theory.. is the very foundation of statistics..

how to combine probabilities?

how do we combine independent events and mutually exclusive events?

now using a quarter ..
take the coin and flip it, it comes out heads, and again, heads, and again heads, and again heads .. four coin flips four heads in a row .. flipping the coin one more time .. you are not due for tails '' the gambler's fallacy'' .. the fallacy assumes that what came before has anything at all with what is about to happen and the response to that is that it doesn't .. the coin flips are 'independent events' the chance for a next flip being one or the other is always 50-50 as it has always been!


joint probability of independent events:

what is the chance of getting two heads in a row?
(.5%) x (.5%) = 25%
chance therefore of two heads in a row? 25%
3 heads in a row = 12.5%

_________________________
mutually exclusive events:

same coin flipped, comes up heads.. we know two facts. one it is heads and the second fact is it wasn't tails.
heads and tails in a signal trial event are mutually exclusive that is if one occurs, the other one can't occur.
the chance of getting a head or a tail is 100%
heads (.5) tails (.5) and the 'or' stands for add!

mutually exclusive means if you are A you can't be B and vice versa, there is no overlap!
nonmutual A and B with overlap .. when we have an overlap we have to account for the joined area which can't be A+B it is AxB
what is the chance of A, B or both?
chance of A or B but not both?
to get rid of the combined area subtract 2

now some real application
the chance of being obese is 30% the chance of having diabetes is 10% .. what is the chance of being obese, have diabetes but not both?
obesity and diabetes are non mutually exclusive
we start by adding A & B together so that is 10% + 30% = 40%
the overlap area is characterized by A x B = 3%
so our correction factor is 3%
we wanted to do A and B but not both
so it is A + B = 40 % - 2AB (2x3)= 6% (as a correction factor)
therefore the correction factor here is 34%
if obese/diabetes or both the answer is 37%
we'll not subtract two occurrences of AB but one occurrence

the logic of statistics is you can never prove anything all you can do is disprove something!
if I want to prove a drug works, I can't, but what I can do is disprove that the drug doesn't work! Ahhhhhh the (double negative) that is the way to get to where you wanted!

we have a double blind randomized design to prove that a new wonder drug works, one half of the participants gets the wonder drug, the other half gets placebo
follow both groups and see when they will have relief from their symptoms,

1- the research design isn't flawed

now let's discuss the null hypothesis ..the 'null hypothesis is the opposite of what you want to find. group A fails to get over the sx. faster than group B.
with the null hypothesis we state it and then leave it alone .. we pass out pills and collect data next .. take the data feed it into the computer ..
what comes out T as in a t test, x^2, f etc etc.
we should only focus on the P value is key for making statistical decisions
we make decisions by putting a standard in place and comparing empirical evidence to it .. so that is what the P value is, the standard and the summary of the data .
this alpha criterion is something you decide before you make your research, you can set the value high or low, it is your discretion
people put the value of P at less than or equal to .05
a confidence interval of 95% means one is correct 95% of the time .. the other five percent is the time when one is wrong
a confidence interval of 95% corresponds exactly to a P value of </=0.05
95% chance of being right, 5 % chance of being wrong

outcomes for our study p=.02
.02 is under the bar which is very good because it means we get to 'reject the null hypothesis' because the null hypothesis is the opposite of what we are looking for .. if we reject the null hypothesis, the drug works!

is it possible that the drug works in the study but not out in the world?
yes possible though unlikely..

what this means is that we have made a type one error, or alpha error, this type of error basically states we rejected the null hypothesis but we shouldn't have . You'll never know for sure if you have made a type one error, all you know is the chance that you made a type one error that chance is found in the p value a .02 i.e a 2% chance .
p value is type one error
if the number for the p value gets too low, we'll take that chance

2nd outcome for the study, p=1.3 we are now above the bar, we can't reject the null hypothesis, we fail to reject the null hypothesis..
you never 'accept the null hypothesis'
same as 'jury logic' not that you are innocent, just that there isn't enough evidence to convict you . the chance for a type one error here? = 0 why? because to make a type one error you must first reject the null hypothesis.
however in this case we could have made a beta type error means, I didn't reject the null hypothesis but I should have .. in other words in the study the drug is crap, but out in the real world, it works well..
chance of making a type two error? we don't know.. can't look at P value because P value only tells us a type one error ONLY!

type one error is considered worse ..
which is worse looking at you and lying or simply forgetting to tell you something?
lying is worse, that is a type one error a 'sin of commission' because first do no harm is a physician's oath.

now you are giving this new drug because it has been approved and works great, the patient is now asking this drug works great in research, what is the chance it will work for me? best response is 'I don't know' this gives you statistical significance not clinical significance!

you can answer the patient by looking at the table the one that tells you, who got the drug, who didn't get the drug, got better, didn't get better
got drug got better 70%
got drug didn't get better 30%
no drug better 30%
not better no drug 70%

pt.s chance of getting better on drug
the answer here is 70% chance of getting better out of one hundred people that got the drug 70% of them got better!

statistical power:
power is capacity to find a difference if there is a difference ...

like looking at a cell in small magnification can't tell what it is, so increase the power, to get more information :
1-power = beta
in other words one minus statistical power gives us a type 2 error
this is the only way to calculate a type 2 error .
the simplest thing to increase power is to increase sample size.

researcher conducted and got his research to equal to .15 set the study to have statistical power = 70%
what is the chance of a type 2 error? it should be equal to 30%.
the P value is irrelevant..

______________________________________________

which test should I run?

scaling, how to convert the world into numbers?

types of scales to focus on, nominal or categorical or interval
nominal or categorical
nominal is this or this or this like doing the laundry, whites, colored, knits, that is nominal data
what is the most common nominal variable? Gender! groups mutually exclusive
male and female, how many nominal variables? the answer is one. gender isn't a variable .. only the comparison of the two that gives us the variation .
so single nominal variable with two groups to it. Things in a classification.

interval data is within a scale, when you measure with a measuring device with equally graded intervals , lights years, weights, meters, mph. you can get means, and standard deviation. What is the mean of gender? for instance computing heights, computing sd of wights.

nominal.. groups, clusters. clumps.. how much.. what something is= nominal, how much it is= interval!

measuring rainfall, interval because there is a measurement

passing is nominal either pass or fail.

decide between nominal or interval you can tell which test to run.

statistical analyses

interval, nominal or combination.

all interval data two interval variables run Pearson correlation

all nominal data, all things in categories, run a chai square test

nominal/interval data combined together, run a t test.

heights of men and women? run a nominal, height interval

compare height to weight, interval/interval pearson correlation

give a drug to one not the other, see who is better or not, chai square. two nominal variables.

Sensitivity:

about detection of disease.. perfect sensitivity means detecting 100% of diseased individuals.. diseased people are a little sensitive!
the rule for the formula of this one is trues on top and divide by everything ..
we only care about the left hand side of the table with this... left side is where the diseased individuals lie.. that is the secret on the two by two table.. just decide on the part of the table that matters..

true positives/ true positives + false negatives

detected disease divided by all the disease..
80% sensitivity is good.. for every ten finding 8, better than chance which gives us 50/50
sensitivity = detecting disease..


Specificity: detecting health.. concerned with right side of the table, only the healthy people..
specificity again, trues on top

true positive/true positives + false positives



everything in specificity is about a healthy person..
if we have a 60% specificity, out of every 10 healthy we were able to find 6, still better than chance but not very good...

a test with good sensitivity and specificity is difficulty, that is why we run two or more screening tests.

what are we trying to achieve with the test is most important thing. if trying to find disease, need a sensitive test even if specificity sucks..

or have a treatment that is horrible, but good, you'll rely more on specificity than sensitivity.

pick what matches problem presented.

you get a patient who wants to test for HIV, the test comes back positive, he is in denial he asks you, are you sure? you say let me find out for you..
finding out is called positive predictive value:
positive predictive value, is set so find out whether you should believe the test results or not!

we now just us the top row of the table :
__TP___ X 100 = Predictive Value of a Positive Result (%)
TP + FP

trues on top.. true positives divided by true positives + false positives .. all you need to answer the question is to know all the positives..
80/120= 67% is the PPV for this patients test...

so now proceed on this basis because there is a 1/3 chance the patient doesn't have this disease..

another scenario:
female patient wants to become pregnant, comes in for a pregnancy test..
results are negative..
patient wants to know if you are sure..
let's find out..

now we are talking about negative predictive value ..

now we are talking about the bottom row where all the negative values lies..
the rule is always trues on top in this case true negatives / all the negatives
60/80= 75%

__TN___ X 100 = Predictive Value Negative Result (%)
FN + TN


so one in every four is wrong..
test is not bad.. good sensitivity ..

accuracy .. again uses all the trues on top..

tp+tn/N (which stands for everything) 70% accuracy!

screening tests are about secondary intervention which is concerned with prevalence not incidence..


accuracy = sensitivity + specificity / 2

After discussion with τhε ṿαlε'ṡ lïlÿ I have decided to add some comments on research and statistics here. I will begin by setting out some thoughts.

1. Medical research is a difficult area because on the one hand it offer so much value and the other when things go wrong so much damage. One can point at huge advances made and the practical eradication of many diseases. However, things have gone wrong and one only has to consider thalidomide or say the use of cardiac anti-arrhythmic drugs which are known to have cost more American lives than the Vietnam war. (See "When Doctors Kill: why and how", by Joshua Perper and Stephen Cina)

2. It is also sadly true that in scientific research there have been many many cases of fraud or misconduct. In fact Professor David Goodstein from California Institute of Technology in a recent book shows that most scientific fraud or misconduct cases involving biological science with medical doctors disproportionately represented in these cases. (See "On Fact and Fraud" by David Goodstein, Princeton Press.

3. There are many reasons for what I have said in item 2 but for the purposes of this thread I will point out three and in subsequent posts elaborate on them.

Scientific Method - over centuries experimental methods and principles have been developed and these must be thoroughly learned and it takes a long to to learn and practice them. Indeed it is only when you do real work that its main points begin to sink in and getting to that point with humility is central to developing your research potential.

Ethics - in all science there is an ethical dimension and it has to be thought through with considerable care. In medical research it is paramount for obvious reasons. Indeed a large number of both fraud and misconduct cases can be traced to poor ethical standards.

Statistics - when one begins statistics it can seem quite easy but this is a false assumption and unless you really know what you are doing you can make horrendous blunders. These days we have SSPS and Excel so given a set of data one can generate a whole raft of statistics with zero effort. However, like any science, all statistics are hedged about with condition and limits so interpretation of what you have been given is likely to be very hard EVEN if you are expert. Statistics is ultimately based on probabilities and everyone have difficulty in that area.

Sadly, the literature is legion with cases of scientific blunders because researchers do not understand what they are doing. For example, there are many cased where a researchers confused correlation and regression, were selective in what data points they used, collected the wrong data and so on. So what we have to say here is NOT simple and if your are to get any benefit you will have to work hard. To give a simple example, suppose my risk of stroke is assessed as 12% and my doctor tells me that if I take a statin it will reduce my risk by 16% (we will not complicate it by adding in side effects). Almost no one outside of a numerate discipline can explain why your new level of risk if you take the statin is 10%.

It is also uncomfortably true that even the best researchers sometimes get over-confident, not to say arrogant, and try to go it alone and do not get advice from a competent statistician - that is unforgivable and may amount to misconduct

^^Thanks for that..

under the heading of neurology, three syndromes amongst many seem to creep in when no one expects it.. here is a quick read from third part source, just know the artery affected and ipsilateral/contralateral symptoms:

• Weber's syndrome
• Wallenberg's syndrome
• Moritz Benedikt

Once I start my pharmaceutical chemistry and pharmacology modules next academic year, I'll be a lot more active in this thread inshaAllah . I'll also me starting immunology - is it a tough subject?

format_quote Originally Posted by Banu_Hashim

Once I start my pharmaceutical chemistry and pharmacology modules next academic year, I'll be a lot more active in this thread inshaAllah . I'll also me starting immunology - is it a tough subject?

Great.. khyer insha'Allah.. (I think this is the most viewed thread on this forum?)

Immunology and microbiology are sister subjects they go hand in hand and they are light and fluffy.. part of the easier subjects in the sciences .. it is mostly memory not many concepts involved..

This is my second post on the theme of evidence based research and statistics and here I discuss two basic ideas which I shall call Research Style and Research Type. Please be aware I am INVENTING data here.

RESEARCH STYLE – at this stage you need to consider if your style is quantitative or qualitative. It is easy to confuse these two and simply think of them as describing data types but to do so means you are missing the whole point. In general, if you outcome is in some way intended to be predictive then your style is likely to be quantitative whereas if it is intended to be mostly descriptive then it is almost certainly qualitative. For example, suppose I decide to study infection rates after surgery.

Quantitative - here I would for simplicity chose two elements: the procedure and the infection rate. So over time and with a sample of patients I record relevant data. Now using this data I could process it statistically and predict say in knee surgery that the infection rate is likely to be 25% of patients.

Qualitative - knowing that 25% patients become infected after knee surgery is interesting but not of much use in deciding what to do about it. So my next research task could be to study surgical procedures used in knee surgery with a view to constructing a check-list which when used by the surgical team will lead to fewer infections. So here I am ending up with a description, as a series of questions (usually not more than 8), of what to do both at the start and end of an operation to minimise or prevent post-op infection.

Thus you can see that the terms Qualitative and Quantitative are to do with the kind of OUTCOME you want not primarily the data itself - it is VITAL that you understand this point.

RESEARCH TYPE – broadly speaking there are two types; the first is interventionist where you deliberately make a change in a situation and then study its consequences and the second is observational where you simply record what is currently going on. Using the same example as above.

Observational - as for the above example, I do nothing except record which surgical procedure was used and post-infection rates. I don't interfere in any way and make no changes.

Interventionist - as for the examples above, let us suppose I develop the necessary check-list to be used by the surgical teams - that is my intervention, that is the change I make. Now I start recording the data exactly as before about procedure and post-op infection rates. Therefore at the end of these two studies I CAN decide if my intervention made any difference at all.

You can also see in this example how studies can and in some cases must, as in this one, be linked.

1. Start by getting raw data on post-op infection rates
2. Next a study to develop a check-list
3. Finally, get the check-list into use and collect infection data again and see if there has been a reduction of significance (or of course an increase but we hope not). Here we might, indeed should, employ a statistical text of significance to be sure the change has made a real difference.

I will base my next few posts largely on an article which can be found at the following site. Please be aware though that the terms themselves are not absolute and you therefore will find variations in the literature and when you join a research team so focus on understanding the ideas. http://www.pharmj.com/pdf/cpd/pj_200...encebased1.pdf

Evidence-based research means extracting good evidence to make sound clinical decisions because valid evidence of clinical benefit and cost-effectiveness is what will influence purchasers of health care. There are two fundamental questions that one needs to answer: how do we assess the evidence itself and how do we know that the piece of research that generated it is good enough to be relied on? The point is that the evidence may be overwhelming BUT unless we know exactly how it was obtained we cannot rely on it.

Often we speak of an ‘evidence pyramid’ and what that means is we must look in detail at the data and the methods used to see that they were ethical, appropriate and well-executed. For example, if I trial is described as ‘randomised’ then we need to know exactly how the randomisation was carried out because if it was not done with the necessary scientific and mathematical rigour the results may be totally invalid. Indeed many studies have eventually been rejected because of poor randomisation processes – everyone can invent such a process and almost none of them will be valid.

The hierarchy of evidence beginning from the most trustworthy/reliable to the least is: systematic Reviews, Meta analysis, cohort studies, case control studies, cross sectional surveys, case series and case reports. In later posts I will explain each one in more detail as well as explain how methods can be evaluated. But here starting from the top.

Meta Analysis – before research can be readily accepted it must be reviewed by competent specialists both medical and statistical. One way of doing this is to gather the data from many trails together and pool them as results into one big spreadsheet (or whatever system you are using). This can be very useful if there have been many small trials but each one too small to be conclusive. So if you had 10 trails with a paltry 40 patients in each then do a Meta analysis and you have a 400 person trial to work with. In fact there have been several cases where a treatment previously believed to be ineffective turns out to be rather good but because each individual trial was too small to detect the benefits no one spotted it.

You might like to look into the case of New Zealand doctors having the idea of giving a short and cheap course of steroids to help improve premature babies between 1972 and 1981. Two trials showed benefits and five failed to detect any and the idea was dropped. Eight years later in 1989, a meta-analysis done by pooling the trials was carried out and when this was done a clear benefit emerged showing a 30-50% reduction in risk of babies dying from complications of immaturity. One must always remember the human cost of the abstract numbers: babies died unnecessarily because they were denied a life saving treatment for a decade.

By the same token a meta analysis can show that something that was thought to work does not in fact work and without going into details there have been several such cases.

Here is a couple of questions for Hugo or anyone who has been reading notes in the statistics/epidemiology portion of this.. or who simply wants to offer a stab...

format_quote Originally Posted by Banu_Hashim

Once I start my pharmaceutical chemistry and pharmacology modules next academic year, I'll be a lot more active in this thread inshaAllah . I'll also me starting immunology - is it a tough subject?

I love microbiology. Microbiology is too broad of a discipline and includes bacteriology, virology, genetics and is also very closely intertwined with immunology because immunology studies those things which are related to "microbes" or their pathogenicity within mammals, more specifically humans, monkeys, rabbits, goats, mice etc.

I personally found virology to be interesting but quite heavy on memorizing. You are just supposed to know which virus causes which diseases. You'd be surprised to see that most viral infections show similar symptoms! Headache, fever, myalgia, dizziness and what not. That is because these symptoms are related to immunological reactions/weaknesses. Althought each virus has its own unique symptom. For example Hepatitis viruses have jaundice etc. Khayr, its a long long discussion and I can go on for days ...

Immunology by itself is not tough, thought it has tons of memorization. You will be required to memorize how T cells work and how they are activated. You will be required to memorize how leukocytes mature. You will look at immunology from a broad cell perspective and then go into nitty gritty molecular details as to which cellular receptors are involved and what chemicals are formed and on what enzymes they act, all within just one cell, so on and so forth. As I said previously, after you finish your studies, you'll be left in awe after realizing that we are walking talking universes in our own selves .... probably more intricate than the cosmos.

Pharmaceutical chemistry is also very interesting. It is related to medicinal chemistry as well. Its research focused and is concerned with drug design, specifically the chemical reactions, conformations of molecules etc required to produce a specific outcome. Very strong emphasis on organic chemistry but recently inornganic chemistry and transition metals and their complexes have shown to have strong anti-cancer activity as well so inorganic can be a part of it too.

Pharmacology is clinical aspect of drugs which are KNOWN. Crazy memorization. Not much innovation. Just memorizing facts for clinical setting and spilling them out when asked on the exam or in the clinic. Though its pretty important for medicine.

format_quote Originally Posted by τhε ṿαlε'ṡ lïlÿ

Here is a couple of questions for Hugo or anyone who has been reading notes in the statistics/epidemiology portion of this.. or who simply wants to offer a stab...

This is a fairly advanced sort of idea which calls on quite a lot of statistics and mathematics. The ROC was used initially to help radio/radar operators decide whether something was just noise, natural and usually small variations or whether there really was a signal there - one does not want to shoot off a missile just because there just might be something there. The idea was taken up by the medical profession because typically you take a test sample from patients and you have to decide when it is significant. There is a good tutorial on this at http://www.anaesthetist.com/mnm/stats/roc/Findex.htm

Take for example PSA, well it has a range of values and there is a cut off - above X there might be cancer and below this, healthy tissue. But of course human beings are biological animals so there is no exact figure for a given individual where one should start to worry, instead there is a range of values. Roughly speaking when the PSA test was mooted someone took a large number of measurement from a representative sample of men. When you do this you in effect get two samples: one of healthy men and one representing those with the condition and of course these two sample overlap.

The reason they overlap is that some of the men the test shows healthy have in fact the disease (false negative) and similarly some of those the test shows have the disease, in fact do not, a false positive. Therefore if we process the data and a with a knowledge of each patients history one can suggest a cut off value. Usually of course one has the test and if its in the cut off range one is sent to a specialist for further tests and that might end as a positive or negative overall result. Two simple statistical ideas are used:

Sensitivity - how good the test is at picking out patients with the condition, so sensitivity gives you the proportion of cases picked out by the test, relative to all cases who actually have the disease usually expressed normalised between 0 and 1.

Specificity - the ability of the test to pick out patients who do NOT have the disease again expressed as fraction between 0 and 1.

I can't draw a ROC curve here but it is constructed by plotting sensitivity on the y-axis and specificity on the x-axis and any good statistical package will do this for you - BUT you really do need to know what your doing so take advice from statistician.

Essentially, the closer the ROC curve is to a diagonal, the less useful the test is at discriminating between the two populations. The more steeply the curve moves up and then across, the better the test. A more precise way of characterising this "closeness to the diagonal" is simply to look at the AREA under the ROC curve. The closer the area is to 0.5, the more lousy the test, and the closer it is to 1.0, the better the test. Essentially, the diagonal is a line that means you might as well just toss a coin because it would be just as good as the medical test at predicting a particular condition.

If you find this hard to see just think of the ROC curve as residing in a square with sides 1, if you then draw a diagonal it is obvious the area under the diagonal line is 0.5 and the area of the whole square is 1. If you then think of the test as being very sensitive the the curve will shoot way up above the diagonal and gradually move across hence getting closer to the square whereas if its not all that sensitive it will rise only slowly and be nearer the shape of the diagonal. Obviously, confronted with a ROC curve it will not be easy for you by hand to find the area under the curve and that is why you MUST use a package.

Please be aware that what I am describing here is what one does to establish cut off points. In the example given in the previous post the curve looks as if it being used as a trauma diagnostic tool and in that setting it should ONLY be used by a competent physician and in concert with other patient data. Be aware that I am speaking mathematically here and have no competence as a physician.

Greetings.. what is your answer for the two q's?

the choices should correlate to points on the graph or 'none of the above'
what is your response and rationale?

thanks

format_quote Originally Posted by τhε ṿαlε'ṡ lïlÿ

Here is a couple of questions for Hugo or anyone who has been reading notes in the statistics/epidemiology portion of this.. or who simply wants to offer a stab...

Cant answer

statistics is not my forte, as of yet. Only took one course entitled "biometrics" in college and hated it to death.

format_quote Originally Posted by mad_scientist

Cant answer

statistics is not my forte, as of yet. Only took one course entitled "biometrics" in college and hated it to death.

well now is a good time to get acquainted.. it won't go away unfortunately for even when you are done, you'll need to keep your CE by reading medical journals which requires understanding that crap... if you don't know what the numbers mean then you don't understand what the journals are about.. all you need are numbers and abstracts.. the extraneous details are just to wow other people..