ASA Connect

Dear ASA community,

I will be presenting short (very short) lecture (5-10 mins) on the most fundamental and commonly encountered concepts in Biostatistics during monthly seminar series. My audience will be medical residents, fellows and other clinicians, with presumably little exposure to statistics. The goal is to clarify/explain some of the concepts (misconceptions) they could encounter or could possibly have (e.g. while reading research articles). Based on your experience, and intuition, could you suggest some of the topics that we biostatisticians, think are important for medical researchers to comprehend? I am looking for 12-15 topics. I plan on starting with "P-value(s)" and weave in some of the points mentioned in the latest ASA statement about p-values.

Thank you in advance :).

Two issues that I would bring up are:

1. the sample size calculation

2. the fact that the analysis depends on the fact that the data is discrete or continuous or mixed.

Thank you.

Two issues that I would bring up are:

1. the sample size calculation

2. the fact that the analysis depends on the fact that the data is discrete or continuous or mixed.

Thank you.

Rather than start with P-values, I would start with the role of chance in research (using chance to select a representative sample, on the average) and then present hypothesis testing (and P-values) and interval estimation as ways to take that role of chance into account.  Add a short explanation of interpreting P-values and confidence intervals and your 5-10 minutes will be done.

For medical students, important topics include (a) effect size - why it's important and typical measures, and (b) meta-analysis - what it does and how results are presented. Along with other topics already mentioned by other respondents, that's an awful lot to get through in 5-10 Minutes!

Clinicians are generally comfortable with the properties of a diagnostic test and the uncertainty issues surrounding diagnostic tests. I have found that presenting statistical tests as special diagnostic tests is a good way to get started. The 2x2 tables of test vs. truth work well to show the similarities and probabilities.

Hi, Milan:

I think a good concept to touch on is confidence intervals. In my experience most people think that CI indicate a range of individual values, or have no idea what they mean!  I don't think medical researchers need to have a fully acceptable statistical understanding, but at least they should know that they are a margin for error around an estimated value, not where 95% of the subjects fall. I often cover that, then do a gotcha question to make the point.  Say an example with RR = 2, 95% CI 1.5 to 2.8.  "T/F: 95% of the subjects in the sample had RR between 1.5 and 2.8" (Most vote false).  " T/F: 95% of the subjects in the population had RR between 1.5 and 2.8" (Many vote true). 

You can talk about CI for significance, but I often also include a non-significance related example. "In the above example, you would ban the product if the RR was shown to be at least 2. Does this example provide good evidence for a population RR of at least 2?  Suppose you wanted to rule OUT a true RR of 3. Does the example achieve that?"

Not a lot of time in a 10 minute lecture.  I think that is far too short.  I give a lot of intro talks to med residents and such.  If I was going to cover CI and hypothesis testing in any useful way, they would be most of a 45 minute talk.

Dear ASA community,

I will be presenting short (very short) lecture (5-10 mins) on the most fundamental and commonly encountered concepts in Biostatistics during monthly seminar series. My audience will be medical residents, fellows and other clinicians, with presumably little exposure to statistics. The goal is to clarify/explain some of the concepts (misconceptions) they could encounter or could possibly have (e.g. while reading research articles). Based on your experience, and intuition, could you suggest some of the topics that we biostatisticians, think are important for medical researchers to comprehend? I am looking for 12-15 topics. I plan on starting with "P-value(s)" and weave in some of the points mentioned in the latest ASA statement about p-values.

Thank you in advance :).

Milan,

In addition to a discussion on the proper interpretation of p-values, I would also discuss effect size and confidence intervals.  I also give a brief overview of the odds ratio and hazard ratio with a brief discussion on how these are measures of relative risk, not absolute risk.  In particular, I emphasize the importance of considering baseline risk for an outcome when interpreting ORs and HRs.

Recently, I added a brief overview of NNT and this seemed to go over well.

I also end my presentation with a brief overview of diagnostic accuracy measures with emphasis on sensitivity, specificity, and ROC curves.

I hope this is helpful to you!

Mike

You might look at the table of contents for each day's slides posted at http://www.ma.utexas.edu/users/mks/CommonMistakes2015/commonmistakeshome2015.html, and if a topic piques your interest, then look at the relevant pages for more. Also the appendix Suggestions for Readers of Research might have some relevant suggestions. (The website is from a  continuing education course which usually draws several people from the State Department of Health Services.)

I think all of the suggestions were spot oon. I would add one that is replated, and can be presented in the time frame assumed -- specifically, it is choosing the p that you care about.

Traditionally, it is P(data|Ho), but often in medicine you ae more interested in P(Ho|data).

E.g., P(+mammogram|cancer) = .90, but what we care about is P(Cancer|+mammogram) which is about .04

A great place to start is Carl Morris' 1 page discussion in JASA

Morris, C.N. (1987). Discussion. Journal of the American Statistical Association,Vol.82, No.397, 131-133.

and for a fuller discussion

Skorupski, W. & Wainer, H. (2015). The Bayesian Flip: Correcting the Prosecutor’s Fallacy, Significance, 12(4), 16-20.

HW

I have been lecturing in clinical department for few years.  In my experience, hypothesis test will be a good start point.  One sample t-test, two sample t-test, paired t-test, non parametric test, F test, chi-square test, Fisher's exact test, McNemar test.... are the concepts which clinicians are often confused.  Type I and Type II errors, power calculation, effect size and sample size are the topics that they are interested.  The survival analysis skills, especially Kaplan-Meier curves, are also the things that they often see in manuscripts but don't understand well.  They are also interested in study design as well such as case control study, cohort study, randomization, matching....You might want to focus on one small topic each time but examples will be very helpful.  Good Luck!

Jun

Obviously there is no way to cover even the simplest statistical topics in such a short time.  However, I would NOT start with p-values.  I recommend starting with the ideas of uncertainty, definitions of type 1 and 2 errors and the importance of sample size, replication, control and clear hypotheses.  Since they are docs/medical students, you can also talk about statistical significance vs clinical/medical relevance.  P-values, like CIs and effect-size are all just tools for evaluating the evidence obtained from studies.  Good luck.

With the given tight time constraint a single topic would seem most indicated. One gem that could be covered is the importance of randomization in conducting a study assuring as much as possible that the groups are as congruent except for the variable(s) being tested between them, thereby controlling for ambiguity due to bias. I recall one professor that had us students in pairs conduct a survey at a local mall. One of the pair would approach people and ask the questions. After a number of surveys, the other student in the pair instructed the surveying student to survey another batch of the same number but this time assuring that they systematically approached EVERYONE that came by a certain point. It could be instructive to statistically compare the results of these two surveys.

Propensity matching could carry the lesson further by pointing out that such a procedure is an attempt to amend a study not benefiting from randomization. That the latter automatically accounts for all the variables that you include in propensity scoring as well as those you do not including latent variables.

All good comments..

That said, I would start out with the concepts of describing data v. drawing inferences.  In the many years I've taught residents, typically in preparation for them to do a research project, the most important thing is how to define the research question. Too often I hear (prior to any didactic teaching) 'I will compare this number with that', rather than rates, proportions, means, etc. The concept of 'relative to what', though it seems intuitive to us, is not to medical students and residents who spend most time memorizing rather than thinking critically. Certainly introduction of levels of measurement imply what descriptive and inferential statistics to use as well as impact on sample size methodology.  Finally, the concept of study design is important because it not only impacts on sample size and statistical methodology, but examples make the subject matter come to life.

Katherine Freeman, DrPH

President and Founder, EXTRAPOLATE llc

Professor, FAU

Many of the important topics have been suggested in the discussion chain. Two of the very important issues, that should be brought out before any trial is conducted are:

1. Randomization of appropriate type consistent with the design.

2. Type II (false positive) error rate to assign power of detection of any significant effect size.

Without the above two, p-value and the rest become meaningless.

Ajit K. Thakur, Ph.D.

Retired Statistician

Wow....5-10 days isn't even enough time.  I'd focus on the fact that level of measurement should drive analysis and probably mention that just because one calls something a "scale" does not make it interval-ratio (e.g. VAS scale).  Representative samples, certainly,  And confidence intervals.  I teach anywhere between 100-150 healthcare students every year and read their journals all the time.  So many authors take huge leaps from correlations (sometimes even ignoring significance) to relevant medical effect.  Teach them that there is no way that looking at correlation alone, even if significant, can give them the implication that something is medically important!

Joe

Students should be taught that not only should they look at significance, but sample size (a very large sample will have lots of significant effects, especially at lower levels of significance) and ultimately, does the result make sense, can it be explained by someone with Subject Matter Expertise. This final point is one that even experienced stats guys miss. It is often an indication of bad data or bad design.

Like Sex, one of the most important tools we have in stats is our mind and common sense.

Dr. Bimali and Others:

Perhaps one thing the clinicians should leave your short lecture with is a recommendation for a good book.

The biostatistician Andrew Vickers's 2010 book titled "What Is A P-Value Anyway? 34 [Short] Stories To Help You Actually Understand Statistics" should be helpful.

HTH