speak for yourself
Short and sweet, as many have said, yes this is an incredibly valuable skill and knowledge base. Once you start learning you will see that getting the "results" is the simplest part (clicking buttons or running a script in R). The hard work comes before that. Anyone on here telling you it's a waste is, and I would put money on it, overconfident in his or her abilities, by far, and probably doesn't even know the most basic concepts in statistics (yet they strongly believe they do). People saying "just kick it to the stats people" clearly have zero understanding of the very thing their whole discussion section hinges upon. The time to involve the statistics team isn't once you've collected the data-- it's before you've solidified the research question. Again, if you develop an understanding of statistics you'll see how many studies are terrible because the MD or MD/PhD thinks the stats is just number crunching after data collection- so they collect the data in an inefficient and sometimes unworkable manner.
To be fair, there are many MDs and MD/PhD people who fall into the same category, yet they lead research every day. People think "research" background from basic science without statistics makes them qualified to act as a statistician on their own project. Even worse is the "I've published X articles, so I know research"...Some of these docs are very good, recognizing they don't know statistics just because they took a "methods" class or two in the PhD coursework or because they have published without a statistician. They're the good ones who know that good research involves specialization in each aspect of clinical and statistical knowledge. Fortunately, I've had the pleasure of working with people who want things done right rather than fluffing their ego with the "I can do it all" attitude.
To say that a practicing physician doesn't need to know some undergraduate statistics is to say they don't need to know how to read and write. Trusting others to interpret their research for you is a faulted approach to consuming research and possibly impacting patient care. To say you won't have to use it is also neglecting the amount it would be used in day-to-day practice for most physicians simply by reading a few journal articles. Physicians are used to being the most knowledgeable and hardest working in many respects, but they don't seem to want to take on the challenge of properly arming themselves to better use or create the research that pushes the limits in the field...(Dunning-Kruger is incredibly strong in MD, MD/PhD, MD/MPH cohort when it comes to applying statistical ideas to research).
How would you suggest I go about learning research-pertinent stats in 2 months? As an MS4 I've got a lot of time on my hands and wanted to equip myself would the necessary tools to be able to do efficient research during residency.
They are direct replies in that you mention "MDPhD/scientist" or something as if it necessitates being more proficient than a medical student at statistics with further implication that this "formal training" of MD PhD wouldn't lead people on more of a goose chase (which it might not clinically) but once stats are involved it often does because the "formally trained" doesn't know more than a med student in the realm of statistics, and this is a huge part of the study. The other reply was similar because you said you'd never trust it to a med student, but then go on to say learn it when and if it's needed which implies not everyone needs it-- and I firmly disagree that not every physician needs it, as I mentioned.
But let's make sure we don't conflate an MD PhD, MD MPH or something as adequately rounding out the research skills of that person. The PhD is almost always basic science and has nothing to do with stats (even in the case of a "methods" class or two, which when called "methods" is usually trash). The MPH isn't a stats program, either. So they may be expert in some of these things, but still likely inadequate on average at the stats. These are common misconceptions I see held, almost always by those with those degrees. Again, the stats part is important because the rest of the paper can be beautiful from a clinical or basic science perspective, but the stats often break a paper either with misapplication or misinterpretation and this squeezes by many top journals depending on whether an actual statistician reviewed the paper.
Agreed, but let's recognize that clinical research and epi largely involves methodologies that are statistical just applied in a clinical or epidemiological frame work, so the key is learning stats, not these peripheral fields that are easier to pick up properly after you learn statistics. Learning and practicing good statistics is like medicine and takes a while and then reveals there is a great danger in what you don't know, so you need to know when to seek help (as you said).
That's totally fair for your projects. I completely agree. If a student has an idea, though, I'll certainly listen and ask for a proposal, basically, and have the right experts on board if it ain't me (assuming it goes through what I am comfortable with).
I agree you shouldn't give them free range, but why not find out what they know? What if they know at least as much statistics as you know?
So you automatically assume the student knows less than you do in an area for which you have probably the same level of education (statistics)? How often do you get someone with an MS or PhD in statistics or biostatistics?
Right, but these skills don't develop overnight or in a week, contrary to what many medical people think. So saving all the learning for clinical medicine or residency is foolish because your need and time horizon at that point will definitely keep you from learning it effectively. Also, good mentors know how to take good suggestions from students/residents and toss out the bad, so a student who learns can often positively contribute toward a project.
The sad fact is, most med students, residents, and attendings don't know or understand basic statistics. And I agree they don't absolutely need to be an expert (read as PhD Statistics) nor doing programming but those things will improve their career opportunities (I know people who this has happened to) because most people in medicine aren't good at this stuff, at all.
I think you can make a lot of headway there if you want to, especially given the breakneck pace that you've become accustomed to in med school. If you want, we can PM so I can ask more about your background and goals and all of that to give some more pointed advice. Shoot me a message if you're interested!
Not trying to crap on anyone else, but the one thing you need to be careful of is learning coding from "data science" or "big data" people because they often don't understand the stats as well as they know the coding. Below I give strong resources in both the programming and stats. After you see their website a bit, you'll realize how much they have (UCLA and Frank Harrell with Vanderbilt). Including data sets to download they have output where they walk through the syntax and the meaning of what you're seeing and it's pretty thorough and accurate from a statistical sense (UCLA, because it's done by actual stats people). So, be wary of the statistical accuracy you get from other things that are from a lot of "learn R" or whatever online.
Personally, I would avoid data science/big data and epi/public health for any stats knowledge (coding would be okay), because these aren't usually taught by PhD statisticians. The coursera stuff with the Hopkins biostats people should be good since they're legit statisticians and I think Brian Caffo does some of their teaching and he's had some important papers in theoretical and applied stats if I recall.
Fair enough. We'll have to agree to disagree on the importance of statistical literacy for physicians. However, I agree that the benefit isn't spending your time learning programming, it's in learning statistics correctly which necessitates having skills to program to practice applying the theory and applied knowledge you read about. If you're doing stats in SPSS or something else that's point and click, it's likely not very good, and you probably don't know as much as you think-- this isn't my opinion but a really simple fact because SPSS and other point and clicks use, for example, algorithms that are not recommended or don't allow you full access to a full "work up" for a statistical question. They make it much harder to check and verify assumptions for a particular test or model which is like taking tires away from someone trying to ride a bike. The programming interfaces make it much easier to get what you actually need once you have some experience in the syntax. So my emphasis is on learning stats with the programming to facilitate learning and that will make you a big asset.
This is actually easily done by just permuting the Y values without doing anything with the IVs/explanatory variables—if I remember correctly, your model is overfit if you permute the dependent variable values randomly but still see some indicator of “signal” or good performance (because you should see nothing special when you know the y values are unrelated to the Xs due to the random permutation). I will read later once I am home, but the title makes me think it is related to that.
SAS is also how I found the UCLA site and saw there is some on Penn. and yeah the courses on coursera have a lot of real biostatisticians on board!
See above.
I think an even better way is to create problems for yourself to solve with a data set so you’re forced to look up the syntax, write it, trouble shoot it (a lot at first usually) and then have to redo it for another problem. Looking at a more experienced coder’s work can help show efficiency and good structure for coding, too.
Just read the stackexchange thread and see what you mean. I was referring to an easy method to study how overfitting can make a nonsense model look good by scrambling data and then using some kind of variable selection/model building on totally unrelated variables.
Now that I look at this not on my phone, I realize this isn't really a "big data" or "data science" problem as much as it is just stupidity. The problems in BD/DS are when statisticians don't teach it because you have the lean six sigma types teaching what buttons to click for a logistic regression and misinterpreting all of the crap along the way.
You can be very valuable because statisticians are at a hub point with all consultees coming to them like spokes on a wheel.
This is actually why I prefer to have an actual statistician look at my data. No matter how much I know, it is far better to have someone whose actual specialization in life is working with statistics to be working with my statistics to ensure that no errors are made and I'm not drinking my own Koolaid via statistical errors
