What to expect from your statistical reviewer

What to expect from your statistical reviewer


I hope the stuff that I’m gonna present
today is useful to you as a clinician and I will try to avoid talking about
anything in terms of formulas or anything complicated in terms of
statistics okay so I’m basically gonna start with just a brief overview of the
review process and then walk through each section of a manuscript with you
and talk about some of the common mistakes or things that I’m looking for
when I’m reviewing papers and then finish with a brief discussion so just
to give you some background on myself in addition to my responsibilities here at
Columbia I’m also a deputy statistical editor for The Journal of thoracic and
cardiovascular surgery and I’m also a statistical editor for the American
Journal of Drug and Alcohol Abuse I’ve reviewed a lot of papers in the last few
years and I’ve reviewed four other journals as well and I just wanted to
state that these are not my these are my personal thoughts not necessarily the
thoughts of the journal this is from my own personal experience and I wanted to
make a mention of the fact that when we review papers it’s a service to the
field I don’t get paid to do the reviews and often there’s a quick turnaround so
14 days 3 weeks so as someone who’s writing a paper you want to make the
review process as easy as possible for the reviewers and if you’re asked to
review a paper I encourage you to take that opportunity and do a good job
provide useful feedback so some things to keep in mind
pretty much all top-tier journals will require a statistical review of a
manuscript and many other lower-tier journals also require this in fact some
journals also require what’s called a quantitative author so so on with a
methods background there’s different criteria for this sometimes they require
one of the authors to have an mph or a PhD in the field but at the same time
the statistical reviewers may not be statisticians so there’s still some
seats in the front if you want to come right so
again not all statistical reviewers or statisticians will talk about that later
in the talk and the statistical reviewer often works in the field but is not
necessarily an MD so I’m not an MD I don’t often know all the terminology so
you should make sure that your paper is readable by a general audience so have
someone who’s not working on the project review the paper to make sure that it’s
accessible right so the advice I’m going to talk about stay in the context of
writing papers is also applicable to grant writing and then one final point
second to final point is that the devil is really in the details I’ve found
statisticians to be among the most detail-oriented people almost to a fault
so if there’s a mistake in your paper we’re probably gonna find it right so
pay attention as closely as possible to the details and then the last thing and
I take this really seriously you know as a statistician I feel like almost like
an ethical gatekeeper right so are the interpretations in line with the results
in this study design are the data being analyzed appropriately
are they acknowledging what limitations there are to their their study and I
don’t want a paper to go out into the world unless I’m sure that it’s
ethically sound um so this is a huge responsibility that I take on so we’ll
start first with the abstract so the abstract is the most important part of
the manuscript why anyone want to shout out an answer in reality what are people
gonna read in your manuscript probably just the abstract right so this is
really where I’m gonna spend a lot of attention as a reviewer to make sure
that what’s presented there accurately portrays the research that was done and
tells the story of the paper so keep it simple and whether it’s structured or
unstructured it should have the same main parts of the manuscript so what
should I be able to identify when I read the abstract a statistical reviewer well
I want to know what’s the background what gap in the literature is this study
filling what are the study aims what are the study design what is my main
exposure very what is my main outcome variable how are
they measured and what is my population of interest who is my study sample there
should also be key findings for each aim so I realize you have a limited amount
of space so this is difficult to do but ideally you should have findings for
each of your aims and then lastly something about the impact of the
findings so why should I care how is this gonna change practice and to be
honest I don’t really care so much about the statistical details in the abstract
right I can read the method section for that later I’m much more interested in
these other issues as the reviewer alright so don’t try and fit in you know
I said a multi-level model blah blah blah blah blah
alright just say something simple about how you analyze the data and just so you
know we will be posting these slides so you don’t need to take notes right now
if you don’t want to make questions on the abstract alright so again most
important part of the paper you really want to tell your story there so
introduction so these are some of the key questions that I asked myself as I
read an introduction what do we already know about this problem is the problem
clearly stated I can’t tell you how many times I’ve reviewed papers and I’m not
even sure what they’re trying to do right so if that’s the case it doesn’t
matter what your methods look like later if I don’t know what question you’re
trying to answer I can evaluate the rest of the paper
so what gap is the study filling in the literature why is this important why
should I care and then what are your study aims and aren’t they congruent
with the problem that you just built up so be as explicit as possible ideally
state a hypothesis right so not all research is hypothesis driven in that
case you should state that this is exploratory analyses and then one last
sort of reminder is that make sure that the aims that you list in your
introduction match the results you’re going to present later so often I’ll
read the aims move to the next section get to the results and I find things in
there that were not stated in the introduction so you want to have
congruence between the two so to get a little bit more technical when I’m
reading your introduction it should be clear
me what X, Y and P are. This is gonna be the extent of my algebra here so X is my
exposure variable Y is my outcome and P is my population of interest these are
the three key pieces of information that I need to know to move forward with the
paper and again I should also have some idea of the study design so you don’t
have to provide all the details but I should know is this a coke or is this an
RCT is it cross-sectional because that’s gonna frame the rest of the review
process for me right so you’re gonna have plenty of space later to describe
the design but include one or two sentences on the design in the
introduction to frame the rest of the paper questions about introduction we’re
also gonna have time at the end for questions so if you want to save them
for them that’s fine as well right so tell me what the story is why are we
researching this particular topic and then give me some more details what is X
what is Y what is P and how are we gonna examine those things so for the methods
section we’re gonna walk through the several different parts so the first
part is the study sample so I need to know who you included in your study who
did you exclude and are those reasonable inclusion and exclusion criteria you
should also tell me when you conducted your study so when did you collect your
data what is your sample size this seems obvious but it’s often forgotten so how
many people did you include in your study even if it’s in the table later
put it up front so people know what they’re working with you should also
include some sort of study flow chart so for a clinical trial there’s very
standard on reporting guidelines called the consort so I have a picture there of
what a typical concert it looks like you can go to their website to download the
template but I also wanted to say that the concert is a lot more than just a
diagram so there’s a lot of reporting criteria that you need to meet with an
RCT that websites a really great resource for that
and then this equator network is another great resource for other study designs
so they’ve tried to take what’s presented in the concert for RCTs
apply it to other study designs so if you’ve done a cohort case-control all
sorts of designs they try to give reporting criteria there for them as
well and when you’re working through this
diagram make sure all the numbers add up again seems like common sense that often
it’s not the case that that happens so I will go through and count up the people
in time you know the ends don’t match and then the last part related to study
sample is you need to tell me something about this this group of people so we
want to always have that table one with study demographics um who are the
participants in this study in the study design section you should also say
something about IRB approval and whether or not you had approval or if you had a
waiver you have to say something about that pretty much all of the journals I
review for required me as a reviewer to check a box saying that I checked for
that piece of information in the paper and if it’s not there your papers not
going to get accepted if it’s an RCT we also have to make sure that it’s
registered in clinical trials comm or some equivalent actually that should be
clinical trials.gov but basically your RCT needs to be registered so now we’re
going to come back to X Y and P so again what is your exposure what is your
outcome what is your population of interest so
in terms of your exposure in your outcome things that I want to know is
how are you measuring that how often are you measuring it what types of variables
are they are they binary are they continuous what is your primary endpoint
so maybe you’ve collected multiple measures over time but one particular
assessment is of interest you need to provide all of that information to the
reader and if you’re doing some sort of longitudinal study with follow-up you
need to tell us about Vanessa follow-up so how many people dropped out why did
they drop out and how that might impact your results these are the sort of
things that I’m starting to think about when I read the study design section so as a reviewer my main goal when
reading the design is to answer the following question so is the design
appropriate to answer the questions they are interested in related to the design
what biases might be introduced by using such a design so I’m gonna be thinking
about this as I read the paper and hopefully when I get to your limitation
section at the end of the paper you’re gonna say something about the study
design and how that might limit the generalizability or the interpretation
of your findings I’m also looking for something about sufficient number of
patients and/or events to answer the study questions all right so I’m
thinking about power and sample size so if necessary and depending on the
journal and the type of study you’re doing you might include your power and
sample size calculations in the manuscript and if you’re gonna do this
you need to give me all of the information I need to replicate the
results and I have in fact done this myself whenever I review papers
especially for clinical trials I’ll plug the numbers into my software and make
sure I get the same numbers as you and one sort of big warning make sure you
don’t confuse the total sample size with the per arm sample size so a lot of
software will compute the sample size for you and it’ll tell you you need 50
people but really what it means is that you need 50 controls and 50 treatment
people this has happened in real life I reviewed a paper they had the protocol
everything was in line but they misinterpreted their power and sample
size so they had half the amount of people they needed and they had enough
finding so it was really hard to interpret that paper given that mistake
so they had carried out all of this research spent all this money and now
we’re not sure what to do with them so please be careful make sure you double
and triple check your software if you’re using it so any questions about study
design and how to present that [waiting for questions in the audience] yes so if you check out this equator
network they tried to basic to come up with alternatives to the concert for
those other types of study designs so you can go in there and look for case
control studies what should you be including so they’ll give you more
details on that but you’re right it might not be exactly the same as the
concert diagram but there is an equivalent if you’re interested in
looking at it and not all journals require you to do this that I do find
it’s a useful resource as you’re preparing your paper to think about what
sorts of things you need to end up in the paper even if it’s not required any
other questions yep sure so let’s say you’re you’re conducting a study you’re
following people up for a year it would be useful to them for me as a reviewer
to know well how many people let’s say made it to three months how many people
made it to six months nine months and a year so if you did your study for a year
and you only had 10% of your sample left at the end of the year is that really
interesting and relevant so that’s what I mean by that is just to provide some
details on when are people dropping out and at what rate and of course you’re
gonna have to now deal with that missing data later you could also report like
you’re doing a survival study so median follow-up time something like that just
to give a sense of how long people are remaining in the study great questions
so the statistical methods section is probably the section I spend the most
time reading and the goal when you’re writing this section is that another
person should be able to read it and replicate your results I realize that
you have a limited amount of space your manuscript and I’ll give a
suggestion for how to handle this at the end of the talk but specifically what am
I looking for so I’m looking for methods that are cited for each objective so if
you have three aims what are the analyses you’re going to use for aim one
into a name three what tests are you gonna use and then how are you gonna
handle the missing data all right so missing data is a reality
of doing research you always try to minimize missing data but even given the
best methods to sort of ensure complete data we’re always gonna have things that
come up right so again report how much missing data there is if you’re gonna do
a complete case analysis so what that means is you’re gonna only include
people who actually gave you data or completed the trial you need to think
about what are the potential biases and mention them in your limitations are you
gonna use imputation methods right so there’s a whole sort of universe of
imputation methods what’s appropriate for your study and I just wanted to warn
against last observation carried forward so in our CTS a lot of people tend to
use this last observation carried forward method which means if someone
drops out you take the last available measurement that you have on that and
carry it forward to their endpoint there’s a lot of bias associated with
doing that there are much better methods out there to handle missing data so do
not use last observation carried forward and the other thing I will say relating
to missing data is don’t pretend like you didn’t have missing data in your
paper that’s a huge warning sign to me that you had something major go wrong
with your study it’s if you just completely sort of sidestep the issue as
a reviewer I’m gonna be like wait what happened why are they not mentioning
this so you have to be upfront and have to be explicit and be as detailed as
possible and again you may not necessarily be able to do some of these
things to imputation methods might be complex work with a statistician who can
help you so sort of the big question is do the methods match the type of outcome
data all right so this is not meant to be an exhaustive list but just to give
you a sense of some of the analyses I’m looking for depending on the type of
outcome you have right so your outcome variable is going to drive the analysis
so if you have a continuous outcome so something like
blood pressure or or weight and you’re comparing two groups right so you want
to compare Group A – Group B you could do something like a t-test if you’re
comparing three groups so Group A – Group B – Group C you want to use
something like an ANOVA you can also use a linear regression model in this case
and I have this under the column of unadjusted analyses because all we’re
interested in doing right now is comparing Group A – Group B if we want
to compare Group A – Group B adjusting for some other information so adjusting
for age or gender or whatever then we have to move towards linear regression I
just wanted to point out that we can use linear regression for both unadjusted
and adjusted analyses a lot of people think you have to do a t-test or an
ANOVA and then move to linear regression later you can use linear regression for
both sets of analyses so Bruno came with what I mean by unadjusted versus
adjusted we need any clarification all right so again it’s our outcome variable
that’s driving the analysis so if we have a binary outcome so death or no
death infection no infection and we’re comparing two groups there’s a whole
host of methods we can use the most common or chi-square test Fisher’s exact
test and logistic regression and again we can use logistic regression for those
unadjusted analyses and the adjusted analyses and then the last most common
type of outcome is time to event data so time to heart attack time to recovery
time to reassure to remission these are all time to event outcomes we want to
use survival methods so we have Kaplan Meier which gets us our survival curves
that’s a nonparametric estimate of survival so it gives us those figures
you’re probably very familiar with we’re gonna see one in a couple of slides but
if we want to compare two groups or two or more groups on those Kaplan-Meier
curves we have to use a log ring test we can also fit a Cox proportional hazards
model both in the unadjusted and the adjusted case and there are lots of
other survival methods out there these are just the most common so this is
again not exhausted list these are the most
traditional methods and there’s also a lot of nonparametric methods that are
not listed here that might be of use to you
all right so for me when I’m reading the paper I really need to know what the
outcome is in order to assess whether the appropriate method was used any
questions on this sort of table or the analysis technique so I’m gonna spend
the most time talking about the common mistakes in the methods section so the
first is when conducting an RCT for the most part its standard to carry out
what’s called an intent to treat analysis which means you’re gonna
randomize effort you’re gonna analyze everybody that was randomized whether or
not they completed this study whether or not they actually completed treatment so
if they’re randomized to group a you’re gonna analyze them in group a and if
they’re randomized to group B you’re gonna analyze them in Group B missing
data is a huge issue if you want to do an intent to treat analysis so when I’m
reviewing an RCT I want to make sure that the authors actually carried out an
RCT or an intent to treat analysis you can
carry out other types of analyses as secondary analyses so if you want to do
a per protocol analysis that’s fine but I’m expecting that to be a secondary
analysis not the primary analysis another really common mistake is
analyses that are used in the results are presented in the results are not
mentioned in the methods section so a really common example is the log-rank
test so people write in the methods that they’re going to use Kaplan-Meier curves
and then they report about log-rank test and the results they don’t mention
the log-rank test in the actual methods section all right so if you’re going
through your results section make sure everything you report there shows up in
your methods at least somewhere I think this is next one is at the bane of many
statisticians existence so there’s this word multi-variable versus multivariate
so what do we mean by multi variable and what do we mean by multivariate so multi
variable means we have lots of predictors in our model
so we’re adjusting for covariance when we write multivariate it means we have
many outcomes that were modeling at the same time so multivariable as many X’s
multivariate is many Y’s and more often than not people are fitting
multivariable models not multivariate models so make sure you’re not sort of
interchanging those two they mean they’re those two things they mean
different things to statisticians so when in doubt I would say go with multi
variable related to model building you should be explicit on how you carried
out your model building how did you select what variables were going to be
in the model did you care use some sort of selection procedure did you use
clinical judgment all of that is fine but you need to be explicit about it in
the manuscript and you should also clearly see what variables are in your
model don’t just say we adjusted for common covariance give the details questions on those first few common
mistakes so as long as they haven’t been randomized yet it’s okay so if they if
you enroll them in the study and the statistician hasn’t randomized them yet
and they drop out that’s okay because they’re not randomized as soon as you
press that button and you assign someone to a treatment group you need to keep
them in that treatment group for the duration of the study they can drop out
they could stop taking the medication they can die and how do you handle that
well that’s a difficult issue and you need to work with a statistician to
figure that out but if it’s before randomization it’s okay no so it’s
really from randomization forward that’s where the clock starts and in terms of
the difference between the ITT or the intention-to-treat in the protocol the
per protocol one example would including people who actually take the
full dose of treatment and excluding people who either drop out or switch
arms or something like that and those are useful and they have practical
implications but usually the primary analysis should be intent to treat there
are exceptions to that rule but um where most generally speaking ITT is the way
to go so I’m hesitant to say yes personally I would prefer that the
process is driven by your knowledge of the problem so I think a good
distinction is whether this is hypothesis generated work or is this
exploratory work if it’s exploratory work and let the data choose the model
for you but even then there’s some caveats so you probably don’t want to
choose the model and fit the model on the same data so this is idea of you
know prediction and validation so having training data and then testing data but
if you’re doing hypothesis-driven research hopefully you’ve already sort
of worked out some of those details in previous studies or you have pilot data
that informs the research process so if it’s hypothesis driven your your message
to sort of be specified ahead of time or you should be going into the paper
already knowing how you’re gonna analyze the data if its exploratory then you can
sort of choose you know several different models see which one does a
good job but then you have to be careful when you interpret it right don’t over
interpret those results then so you’re a priori analysis plan could say that
you’re gonna check for a quadratic effect of whatever and then if that’s
not significant work so you can write that all up in your plan as long as you
by all the steps you’re gonna take up hurry before you sort of start digging
into your data I think it’s okay but again if it’s hypothesis-driven
hopefully you have some previous research that shows you you know the
shape of that thing is not linear it’s some sort of strange function of time or
something whatever the example is but again even in that exploratory world we
want to be careful about sort of using our data too much and overfitting our
data that’s always something we want to be careful about are there questions on
this side of the room sorry you guys are in my blind spot so great questions so
[waiting for questions in the audience] another really common mistake is that
information is not given on how things are modeled so this might be obvious to
you when you’re doing the analysis but one common example is time so there’s
lots of ways to treat time and by time I mean historical time so if your study is
from 2002 to 2010 how are you taking time into account in
your modeling strategy are you treating it linearly are you categorizing it are
you dichotomizing it there are lots of different ways there’s an infinite
number of ways that you can include time in your model so you have to be explicit
on how you included it and potentially why you included it that way another
common mistake related to variables and how they’re modeled is categorical
variables so when you’re modeling a categorical variable you need to choose
a reference group right so if you have let’s say treatment a B and C maybe you
choose treatment a is the reference group it doesn’t matter from a modeling
perspective but in terms of interpreting the results that the model produces it’s
really important to know what the reference group is I can’t interpret the
results if you don’t tell me that group a was the reference group and related to
this if you’re choosing Group A is your reference group then you might give
results for B versus a and C versus a right that’s sufficient information for
me then so you give me the two comparisons when you have three levels another really common mistake is that
often we have data that are matched repeated or correlated so we have
before/after data we have data collected over time so let’s say you know every
month we collect data on the same we can’t analyze this data using those
traditional methods that I presented in that table so all of those methods so
linear regression logistic regression t-test they all assume that we have
independent observations this is not the case when we have longitudinal or mashed
or repeated data um so if that’s the case you have to use a whole different
group of methods um so there’s the paired t-test the MacRumors test those
are sort of the simplest but then there’s more complicated things like
conditional logistic regression mixed effects modeling GE so definitely work
with the statistician if you have repeated measures data don’t just use
sort of the off-the-shelf methods that you might have learned if you took an
intro to bio stats class those are not going to be sufficient in that case and
then another really really common problem is that multiple comparisons is
not accounted for right so let’s say you have four groups and you want to compare
each of the four groups to each other every time you do that comparison you
risk making a type 1 error right so if you do that I don’t know how many
comparisons that is but it’s a lot of comparisons there’s lots of chances to
make a type 1 error so you definitely want to think about that and control
your type 1 error rate again it’s going to depend on whether you’re doing this
hypothesis driven research in that case you definitely want to control your type
1 error or if you’re maybe doing exploratory analysis maybe then you
don’t need to be so concerned about that but in general you should be aware of
multiple comparisons and have a plan to deal with it questions on these mistakes
[waiting for questions in the audience] right so these are all really difficult
questions don’t think that you need to tackle them on your own even the best
statisticians will have a tough time handling some of these things just be
aware that these are things that you have to worry about and that the
reviewer is going to be looking for so another really common issue is related
to sample size or having a small number of events so depending on what type of
analysis you’re doing sometimes the overall sample size is really important
and sometimes the number of events is important so I’ve given several
examples of this so if you’re you’re if you have a binary predictor and a Bryant
binary outcome you can make a 2×2 table you might want to use a chi-square test
that’s sort of the default method that people go to but if you have small
expected cell counts not going to get into what expected cell counts are but
basically if you have small counts in any of those cells you shouldn’t use a
chi-square test your chi-square test is based on some large sample assumptions
in that case you want to use something like Fisher’s exact test that’s
appropriate no matter what sort of sample sizes you have if you’re looking
at a continuous outcome that’s potentially skewed and your sample size
is small you might not be able to use your standard t-test anymore so you
might have to use a nonparametric method like the Wilcoxon rank-sum all right so
if you have small sample sizes you should be thinking about this and making
sure that you’re using appropriate methods and then the last point I want
to make about sample size so when you’re building models people are often tempted
to just throw variables in the model so if you collected it throw it in the
model the number of variables you can include in your model is actually
dependent on either the number of people in your study or the number of events
all right so if you’re fitting a linear regression there’s a rule of thumb that
you can have one variable in the model per ten people in your study just a rule
of thumb but that’s something to think about right so if you have a hundred
people in your sample you could include up to ten co-variants in your linear
regression model it gets a little bit more complicated for logistic regression
but it’s the same idea you’re going to be limited by your sample size so you
can’t put a hundred variables into your model so be careful have a plan for what
variables you’re going to include and why and again always be explicit about
that process so questions about the small sample size okay so the last sort
of common mistake that I wanted to mention is related to time to event data
so the first thing is you want to make sure you’re using survival analysis
methods so if you’re modeling time first heart attack time to readmission
any of those types of things that involve time to the event you want to
use a survival analysis technique alright so cup Kaplan-Meier is the most
common Cox proportional hazards those sorts of things related to how you write
up your methods section you need to be explicit about censoring so what is
censoring in your study how are people censored how are you gonna handle deaths
or drop outs all of that information needs to be given to the reader and then
another really important issue related to survival analysis is competing risk
so let’s say you’re modeling time to first heart attack and someone in your
study dies before they have a heart attack how do you handle that in your
ear methods alright so once they die they’re no longer at risk of getting a
heart attack what do you do about that can you use standard methods again these
are complicated questions you need to work with a statistician to help you
with this but just to be aware of these are the sorts of things you should be
looking out for questions alright so again there’s a whole field of
Statistics just dealing with competing risks so moving to the results section
right so the first thing I’m looking for is results for each of the objectives so
if you wrote in the introduction you have three aims I’m expecting to see
results for three aims in general do the numbers make up or do they make sense do
they add up so what do I mean by this alright so if you’re reporting a
confidence interval and a p-value is there agreement between the p-value
and the confidence interval right so if the confidence interval contains the
null value right so if we’re talking about an odds ratio if the confidence
interval contains 1 then your p-value should not be significant right there
are very rare cases where this might not be true but in general we expect for the
confidence interval and the p-value to align so if I see that sort of
discrepancy I know that something went wrong somewhere and again now that’s
raising red flags my brain about who did this analysis
were they careful I mean that’s not something you want this statistical
reviewer to be thinking about so do the results match those in the abstract and
the tables and figures again this is a sign of sort of sloppy work if you have
one set of results in the results section and then a different set in the
abstract or the tables and figures maybe you have one version you updated it and
then you forgot to update other parts all of this is giving me the sense that
maybe you weren’t very careful and then I start to wonder well were you not
careful with other parts of the analysis section I’m gonna go back and be even
more critical of my you know in the reviewing process so are you reporting
effect sizes and confidence intervals and not just p-values so if you’re
saying that there’s a significant difference between two groups don’t just
give me the p-value tell me what the mean is in Group one and the mean is in
group two and give me a confidence interval for the difference between two
groups a p-value is not very helpful to me other than it tells me you know your
p-values there and then 0.05 or less than 0.05 in terms of having sort of
impact you definitely have to have those effect sizes in your pavement in our
next seminar so a little advertisement is going to be just about p-values and
whether or not we should trust them and how should we report them and whether or
not they should be banned by journals so several journals have now banned key
values outright my personal opinion is that’s a little extreme but you
definitely want to include effect sizes and confidence intervals and this also
applies to Kaplan-Meier curves so what do I mean by that okay
so that doesn’t look so bad at there right so here we’re looking at time to
some events so lung cancer survival so we have time on the x-axis we have
survival on the y-axis the first thing I’d like to point out is that the axes
are clearly labeled which is good and they’re comparing two groups so they’re
comparing that purple group to the blue group so females to males often people
will just include the dark lines they don’t include those confidence bands for
their estimates right and why is why are the confidence bands important why would
I care about those a reviewer all right so what am I
looking at in terms of the confidence bands so do they overlap a lot do they
not overlap I mean I have my moderating pesky value there that’s not labeled but
I’m assuming it’s a log-rank test p-values so it’s telling me those two
groups are significantly different but what do you notice as we move further in
time what’s happening to our sample size how many people are still in our study
as we move further and further out there’s very few so this last estimate
at a thousand I’m not sure if it’s days or months or years because they didn’t
label their axes but we’ll get to that two people left to males and no females
so how much should I trust the end of that survival curve probably not very
much right and hopefully the confidence bands give you that sense right because
you can see how big the confidence bands get at the end right so always provide
those confidence bands that’s part one part two is to always include this table
here with a number at risk right so you need to tell the audience
well should they put a lot of weight at that end result so if there were 200
people less than maybe but with two people left I’m not gonna really pay
much attention to the end of the survival curve I’m much more interested
in the earlier part so again eat honest it’s fine that there’s only two people
left but you have to tell the reader that questions about the survival curves [waiting for questions in the audience] for sure and I think it’s fine to report
the point oh six but then prevent curve provide that confidence interval right
so you telling me that the confidence interval goes from let’s say 0 to 10
that might be a very wide interval that might be a very narrow interval
depending on the thing that we’re talking about
all right so piece of information that would be useful there is well what is
the standard deviation for this particular measure right so it’s a 10
unit change important or is it not important so maybe for cholesterol a 10
point change is not meaningful but on some other scale it is so the clinical
significance part I’m relying on you guys for I can’t do that as a
statistician but that’s a conversation that the statistician and the clinician
need to have together and it’s difficult right so you get your p-value of point O
6 you’re like it’s so close right but just be honest and disclose what you did
provide the key value get the effect sizes and in the discussion you should
say something about whether or not that’s clinically meaningful if it’s a
point O 6 p value and the effect size is really small then what does it matter
right but if it’s point O six and it’s a big effect size then maybe you need to
do future research to confirm that result it’s a it’s a hard balance but I
think the most important message that I could give you is just to be honest
don’t try to hide things from the reviewer again you don’t want to sort of
send up red flags because then they’re gonna go back and review everything that
much more carefully so I tried to make this as to sort of warn you of this as
much as possible because I see this all of the time right so don’t carry out
redundant analyses so what do I mean by this so I have a simple example here so
suppose you’re interested in the relationship between smoking status so
binary and heart disease so yes no binary first you do a chi-square test to
see if there’s an association between the two then in your next step you fit a
logistic model with heart disease as your outcome and smoking status as your
predictor and then you do a hypothesis test based on that logistic regression
model those two things are testing the exact same hypothesis so without
adjusting for any covariance is there an association between smoking
and heart disease generally the results are gonna agree there might be slight
discrepancies but you should state ahead of time which one is your primary
analysis but don’t report both as if there are two different findings again
that’s a huge red flag to me that says statistical author on the paper wasn’t
heavily involved right because a statistician would know that those two
things mean the same thing and therefore you shouldn’t include both so just be
careful that you’re not including redundant analyses it will raise a red
flag if you do so tables and figures so after the abstract probably the most
important why if people are gonna read anything other than your abstract what
are they gonna do they’re gonna probably look at the tables and figures all right
so you want to spend a lot of time on these and make sure that they’re perfect
so make sure that they’re labeled if you’re using abbreviations throughout
the manuscript don’t assume that someone has read the manuscript spell things out
on the tables and figures again don’t just give me p-values I want to see
effect sizes I want to see test statistics degrees of freedom it’s going
to depend on the journal but not just the p-value if appropriate report column
or a row percents people often interchange these and get them confused
make sure you’re reporting the right kind of percent and if you’re gonna have
plots make sure that the plot has a title make sure it has a y-axis and an
x-axis label make sure you include units right so we were just looking at that
Kaplan-Meier curve I didn’t know if time was days years months so make sure that
the reader knows what you’re talking about
even if they haven’t read your paper questions about tables and figures okay
[waiting for questions in the audience] so last section discussion so what am I
looking for here so are the conclusions in line with the data are you over
selling what you found are you being honest are the claims supported by the
results presented how are these new findings fitting into the literature do
they agree what other people found are these new findings what are the
limitations this is to me as the reviewer the most important paragraph in
the discussion so what are biases that you have to be worried about this might
be related to study design and then specifically related to statistical
yours there’s statistical issues talk about
things like low power how did you handle missing data I want to know how all of
those decisions might impact these results and avoid presenting new results
or data so if you have new results or data you’re putting them in the
discussion stop go back to the results section and update that first don’t put
new findings in the discussion section so sort of ran through all of the
sections of the manuscript I want to give you just some overarching guidance
in terms of responding to reviews so hopefully you submit your paper they
want a revision now you have to go point by point and address the reviewers
concerns I think it’s usually easy to identify the statistical reviewer is
that true and you guys usually tell when you read the reviews which one was the
statistician I would assume it’s pretty easy start by reviewing the criticisms
with your statistical collaborator hopefully you already have won it so
hopefully this is not the point where you’re going and finding a statistician
to work with you don’t have to make all of the changes that the reviewers
suggest if you are gonna go against their advice be respectful and try to
provide references for why you’re doing what you’re doing
the statistician isn’t always right I put here misinformed I was trying to be
polite right so again sometimes the statistician or the statistical reviewer
is not actually a statistician by training
there are clinician who has a lot of experience and methods
so they might not know all there is to know about statistics they might not be
up-to-date on what’s the most current method again it’s fine to disagree be
respectful and try and provide support for why you’re making those decisions
and then just sort of as an aside from my experience a good editor will get a
second opinion so if the reviewer and the authors are
going back and forth about what is the best method or the best approach they
will often call in a second statistician to review the paper and give their
opinion so I’ve been what I call a tiebreaker before so the editor
contacted me and said hey the reviewer and the author’s can’t seem to agree on
this I don’t know what to do with it can you please take a look and give me your
opinion so don’t worry if there’s this back and forth process as long as you
feel like you’re you’re grounded in research so some final thoughts just to
wrap up wrap up the talk ideally you should begin working with a statistician
before you even start your study so before you collect any data come talk to
one of us have your statistical collaborator review the manuscript – for
submissions their co-author on the paper you should be getting their approval any
way to submit it include your statistician in the revision process
don’t go at it alone and then in terms of space so I often I’m working with
clinicians have heard well we don’t have space to put all those details how do
you handle that right specifically related to the methods include an
appendix with all of the details and you could even include code in your appendix
which I recommend doing if possible so that someone can reproduce your results
reproducibility going forward is crucial so be as open as you can provide what
you can’t fit in the manuscript in an appendix good luck with your manuscripts
and thanks for listening and we have a few minutes for questions I think [waiting for questions in the audience] I think it depends on the statistician
we’re doing the paper I know it some statisticians for me I’m a little bit
more pragmatic about it I think it’s okay as long as you’re looking about
what is the effect is what is it for myself whether that trend is really
meaningful or not but some statisticians feel very strongly that you shouldn’t
even say that at all

You May Also Like

About the Author: Oren Garnes

Leave a Reply

Your email address will not be published. Required fields are marked *