Hello, community

I have a design initially designed to make comparisons both within (before, during, after treatment) subjects and between subjects (disease, control) possible. However, due to difficulties in obtaining high quality samples, the dataset is now more limiting.

Q: Can I perform within-subjects and then between subjects analyses separately, by fitting two different models with two distinct design matrices?

I mean, to get between-subject comparisons:

model.matrix(~Period+libSize+Sex+Class)

where Class = {control, disease} and treating Period = {Before, During, After} as block (individuals repeat across Periods because of repeated measures in 3 occasions).

And for within-subject comparisons:

model.matrix(~Patient+libSize+Sex+Period)

where Patient = {1 ... n} factor specifying individuals.

I managed to do it on DESEq2, limma-voom and edgeR, but I'm not sure if this is allowed, since ideally I'd fit a model to allow both within and between subjects at once by treating Patient as random effect variable (duplicateCorrelation() in limma).

   Patient Period   Class Sex   Batch           run   libSize
   <fct>   <fct>    <fct> <fct> <fct>           <fct> <fct>  
 1 5129    Before   LL    M     08.03.2018am    17    upto17 
 2 5129    During   LL    M     09.03.2018am    13    upto17 
 3 5129    During   LL    M     08.03.2018am    12    upto17 
 4 5129    Finished LL    M     09.03.2018am    17    upto32 
 5 5129    Finished LL    M     08.03.2018pm    15    upto32 
 6 5130    Before   BT    F     08.03.2018am    12    upto17 
 7 5130    During   BT    F     08.03.2018pm    16    upto17 
 8 5130    Finished BT    F     08.03.2018am_pm 15    upto17 
 9 5139    Before   LL    M     08.03.2018am    16    upto32 
10 5139    Finished LL    M     09.03.2018am    12    upto32 
11 5139    Finished LL    M     08.03.2018pm    2     upto17 
12 5140    Before   BT    M     08.03.2018am    18    upto17 
13 5140    During   BT    M     08.03.2018pm    18    upto32 
14 5140    Finished BT    M     09.03.2018am    18    upto32 
15 5141    Before   BT    F     09.03.2018am    18    upto32 
16 5141    During   BT    F     09.03.2018am    17    upto17 
17 5141    Finished BT    F     09.03.2018am    11    upto32 
18 5148    Before   BT    F     08.03.2018am    14    upto32 
19 5148    Finished BT    F     08.03.2018pm    18    upto17 
20 5151    Finished BT    F     08.03.2018pm    14    upto17 
21 5160    Before   LL    M     08.03.2018am    1     upto17 
22 5160    During   LL    M     08.03.2018pm    17    upto17 
23 5160    Finished LL    M     08.03.2018pm    16    upto17 
24 5160    Finished LL    M     09.03.2018am    2     upto32 
25 5162    Before   BT    M     08.03.2018am    4     upto32 
26 5162    Finished BT    M     09.03.2018am    18    upto32 
27 5164    Before   LL    M     08.03.2018am    13    upto17 
28 5164    Finished LL    M     08.03.2018pm    10    upto17 
29 5164    During   LL    M     09.03.2018am    18    upto17 
30 5164    Finished LL    M     08.03.2018pm    17    upto32 
31 5168    Before   LL    M     09.03.2018am    18    upto17 
32 5168    During   LL    M     09.03.2018am    13    upto32 
33 5168    Finished LL    M     09.03.2018am    18    upto17 
34 5170    Before   LL    M     08.03.2018am    12    upto17 
35 5170    During   LL    M     09.03.2018am    15    upto17 
36 5170    Finished LL    M     08.03.2018pm    6     upto32 
37 5170    Finished LL    M     08.03.2018pm    5     upto17 
38 5188    Before   BT    F     08.03.2018pm    16    upto32 
39 5188    During   BT    F     09.03.2018am    17    upto32 
40 5188    Finished BT    F     09.03.2018am    18    upto17 
41 5194    Before   BT    M     08.03.2018pm    7     upto32 
42 5194    During   BT    M     09.03.2018am    9     upto32 
43 5194    Finished BT    M     08.03.2018pm    8     upto17 
44 5195    Before   BT    M     08.03.2018am    17    upto17 
45 5195    During   BT    M     09.03.2018am    18    upto17 
46 5195    Finished BT    M     09.03.2018am    17    upto17 
47 5203    Before   BT    M     08.03.2018am    17    upto32 
48 5203    Finished BT    M     09.03.2018am    18    upto17 
49 5204    Before   BT    F     08.03.2018am    1     upto17 
50 5204    Finished BT    F     08.03.2018am    18    upto17 
51 5208    Before   BT    M     09.03.2018am    18    upto32 
52 5208    During   BT    M     09.03.2018am    15    upto32 
53 5208    Finished BT    M     08.03.2018pm    8     upto17 
54 5213    Before   BT    M     08.03.2018am    17    upto32 
55 5213    During   BT    M     09.03.2018am    15    upto17 
56 5213    Finished BT    M     08.03.2018pm    11    upto32 
57 5215    Before   BT    F     08.03.2018am    17    upto17 
58 5215    During   BT    F     09.03.2018am    18    upto17 
59 5219    Before   BT    F     08.03.2018am    17    upto32 
60 5219    During   BT    F     09.03.2018am    14    upto32 
61 5219    Finished BT    F     08.03.2018pm    18    upto17

The "between subjects" analyses you've done in edgeR and DESeq2 are incorrect because they ignore patient effects. It isn't correct to treat all the samples as biologically independent when in reality you have repeated measures. At best, ignoring patient effects will reduce statistical power to find class and period effects.

The limma package is the only one that can do a between subjects analysis.

You have said yourself that the limma duplicate correlation analysis is "ideal", so why would you not do it that way? I'm not clear what has motivated you to do so many different analyses of the same data.