Very interesting thread everybody. I'm always impressed by the ingenuity of the folks in this community
The major problem seems to be the (in)ability to compare measurements made on different rigs, using different drivers, software, etc. I know there are a zillion uncertainties, including how the IEMs are inserted in the coupler (something none of us can probably every exactly replicate). I want to start with a more basic question though. Do you folks know what is on your y axis? I see most FR graphs with the y-axis labeled as "SPL" - and I've committed the same sin, but this is horribly ambiguous. I asked this question on another forum, but so far haven't had any feedback on this issue.
If you don't apply any type of octave-band smoothing, I presume (I would hope) that the Y axis values are actually power spectral density, i.e., dB/Hz. Because if not, the amplitude from your FFT is going to depend on your sample rate and/or sample length. In other words, you could make the exact same measurement again with a different sampling rate and get a different answer - from the exact same IEM/coupler, without even moving anything. I saw mention earlier in this thread of a "decale" smoothing. Can anybody explain what that is? I've never heard of that before. Even if one wants to apply a 'standard' smoothing, like 1/nth octave, technically one still needs to agree on the pass-band edge frequencies first (perhaps less critical if n is large, i.e., 24 or 48). BTW, I've been using REW and have been unable to confirm what its non-smoothed data output corresponds to. If it's described in the REW user manual, I've not been able to find it. However, the ARTA manual does discuss this, starting on page 35, including how it does its octave bin resampling:
I think what's needed - at least as a starting point - is to figure out if we're actually putting the same thing on all our respective y-axes. Do you folks know exactly what output all these various iOS apps are creating?
In my case, AudioTool (Android) does not have the option to change sample rate. I'm stuck @44100hz, which makes my high frequencies look very spiky and inaccurate. (e.g 10000hz has only 4.41 samples, meanwhile 100hz has 441 samples, by logic)
While it helps with accuracy, it doesn't change all that much (accounting for the prediction and calculation of each app)
The smoothing part is pretty much the wild-card here. As far as my understanding goes, smoothing is used to compensate for noise and outside reflection. I always use atleast 1/12th octave smoothing, as anything above gets too inaccurate. So there is an obvious trade-off. We are using very basic measuring systems, hence dealing with such errors is to be expected.
My case with bass-roll is similar to what crinacle posted above. I tried to compensate manually, but I got inconsistent values. SPL = dB, for me.
(My measurement of the LZ A4)
Please correct me, if I am wrong.
Sampling rate and sample length are very, very important - unless we're actually plotting dB/Hz and using the same delta-f frequency bins; then we'll get the same levels on the y axis regardless. We're taking our FFT amplitude output, p, and computing: 10.log(p^2/p_ref^2), where p_ref is your typical 20 microPa. p^2 = power and power goes up the more signal history you have, and the FFT creates more frequency bins as you increase the sample rate for a given fixed time interval. So, as @bartsky says we need to have some multipliers in our FFTs to account for that - one of which should have units of time (1/frequency). So, @FUYU, I believe (at least, sincerely hope) that your units are really dB/Hz. ARTA does this. REW appears to, but I haven't been able to confirm this 100%. I'd be cautious of any piece of software that didn't tell me exactly what it was plotting. 
