Hello Adane Teshome Kefale. For the particular example you mention (adherence), I would not dichotomize the variable at all. Carving it into categories squanders information, and causes a whole host of other problems. See this list of problems:

https://discourse.datamethods.org/t/categorizing-continuous-variables/3402

Regarding sensitivity analysis, you said that many articles you have seen reported it, but were not clear about their methodology. What did those articles report? E.g., did they use graphs to show how the value of some measure of effect size depended on some assumption or decision? That is what I would expect to see. Thanks for clarifying.

Please check the attached screenshot search for full details. Best wishes, David Booth

Thanks David Eugene Booth and Rani P Ramachandran

Thanks David Eugene Booth and Rani P Ramachandran

Hi Bruce Weaver

Thanks for your insight. Regarding other studies, they simply report that sensitivity was done, followed by the result (either similar with the primary analysis or different). They didn't report what type of statistical test they performed.

It's change in what YOU DID not some other test method. They want to know how your answer changes if you change important parameters a little bit in what YOU DID ALREADY. This is not some other statistical test. David Booth. Here's a dumb example:

Suppose you have the equation 2x=6. X=3 is the solution. They wish to know how sensitive the solution is if you Change 6 to say 6.05 calculate the new x and compare it to x=3. That's sensitivity. GOOD LUCK DB

Hi Adane Teshome Kefale. The opening paragraph of the Greenhouse & Iyenger (2009) chapter that David Eugene Booth pointed to earlier does a great job, IMO, of explaining what sensitivity analysis is. The context for that chapter is meta-analysis, but the explanation works well in any context. here is that first paragraph (with boldface added for emphasis).

--- Start of excerpt ---

At every step in a research synthesis, decisions are made that can affect the conclusions and inferences drawn from that analysis. Sometimes a decision is easy to defend, such as one to omit a poor quality study from the meta-analysis based on prespecified exclusion criteria, or to use a weighted average instead of an unweighted one to estimate an effect size. At other times, the decision is less convincing, such as that to use only the published literature, to omit a study with an unusual effect size estimate, or to use a fixed effects instead of a random effects model. When the basis for a decision is tenuous, it is important to check whether reasonable alternatives appreciably affect the conclusions. In other words, it is important to check how sensitive the conclusions are to the method of analysis or to changes in the data. Sensitivity analysis is a systematic approach to address the question, “What happens if some aspect of the data or the analysis is changed?”

--- End of excerpt ---

When I first heard about sensitivity analysis (many years ago), I was confused, because the software I was using at the time (SPSS) did not have any options or commands for a procedure called "sensitivity analysis". I eventually understood that it simply meant repeating the analysis in question (perhaps several times) while changing some aspect of the analysis. David gave one simple example above. I gave a somewhat more complicated example in my Aug 12 post in this thread:

• https://www.researchgate.net/post/Is-it-possible-to-do-a-dependent-t-test-if-pre-post-were-anonymous-Are-there-any-alternatives

As you'll see if you read that thread, the context is that someone knows the means and SDs for paired pre-test and post-test scores, but does not have the individual scores and does not know r12, the correlation between pre and post. I suggested that if one could estimate a range of plausible values for r12, one could compute a corresponding range of plausible paired t-test values. I've just modified the code I used in that thread to plot the 95% CI for the paired mean difference as a function of the estimated correlation, r12. Suppose the plausible range of values for r12 was .35 to .65. The attached plot shows that the 95% CI for the paired mean difference excludes 0 only near the top end of that plausible range--see the attached figure. This is the sort of thing I had in mind earlier when I suggested using graphs to show how the results of your primary analysis (e.g., the paired t-test in this example) vary depending on how you change some part of the input (the value of r12 in this example). I hope this helps somewhat.

PS- Here is the Stata code that produced the figure.

// Part 2

// Plot 95% CI for the mean diff as a function of r12

clear

set obs 99

generate byte n = 12

generate m1 = 21

generate m2 = 22.75

generate sd1 = 2.730301

generate sd2 = 3.250874

* Let r12 range from .1 to .9 in steps of .1

generate r12 = _n/100

* Use the observed value of r12 on the last row

generate SE = sqrt(sd1^2/n + sd2^2/n - 2*sd1*sd2*r12/n)

generate diff = abs(m1-m2)

generate t = diff/SE

generate p = Ftail(1,n-1,t^2)

generate tcrit = invt(n-1,.975)

generate lower = diff-tcrit*SE

generate upper = diff+tcrit*SE

*list, clean noobs

quietly summarize diff

local MeanDiff = r(mean)

twoway scatter lower r12, ms(oh) msize(small) || ///

scatter upper r12, yline(0 `MeanDiff') ///

xline(.35 .65) ms(oh) msize(small) ///

xlab(0(.1)1, grid) ylab(-1(1)4, grid) ///

title(95% CI for paired mean difference as a function of r12)

Hi

Email me to send you many articles on sensitivity analysis and its global manifestations.

Thanks