It is a common error that you may find the reason why by googling it. Condition and subjects variables are nested. You cannot perform such linear model with your study design. If you show your study design, maybe we could help you to choose the best statistical model to apply.

Dear Gordon Smith, thank you first of all for your reply. I have designed models for paired data before but here the design is a bit more complicated due to the variables involved and the large number of samples. For the latter reason I was already planning to rewrite the code for limma-voom because both DESeq2 and edgeR, even with a simpler design, require extremely long computational times. But still the design of the model poses some problems for me. However, I also understand that the presented experimental scheme is only a simplification of the actual dataset and therefore I will try to present it again as far as possible in a more understandable (I mean biologically understandable) way below by adding the requested clarifications. Now, "specimen" represents each individual sequenced sample; "patient" (corresponding to sample_ID in the previous example) indicates the pairing (same patient evaluated at baseline and at follow up); "time" is the period when the specimen was collected, i.e. at baseline or at follow up; and "disease_grade" is the different grades of a disease from better to worse (0 to 4). Some patients may have the same disease grade at both baseline and follow-up, while others may improve or worsen over time. So, I want to test the gene expression differences in patients who improve, worsen and do not change their disease grade. Obviously, I am not interested in all possible combinations of changes in disease grade at follow-up compared to baseline, which is effectively 25 combinations, but more realistically I need to test 8 different combinations, which corresponds to my real data. For example, I might have 18 patients (i.e. 36 paired samples) who change their disease grade from grade_1 (baseline) to grade_2 (follow_up), 17 patients (i.e. 34 paired samples) who change their disease grade from grade_2 (baseline) to grade_3 (follow_up), and 19 patients (38 paired samples) who improve their disease grade by having grade_3 at baseline and grade_2 at follow_up... and so on. I hope I have not bored you too much with this experimental problem, but I would really appreciate any feedback on this issue. Thanks.

      specimen    patient      time disease_grade
1   specimen_1 patient_01  baseline       grade_0
2   specimen_2 patient_01 follow_up       grade_0
3   specimen_3 patient_02  baseline       grade_1
4   specimen_4 patient_02 follow_up       grade_2
5   specimen_5 patient_03  baseline       grade_1
6   specimen_6 patient_03 follow_up       grade_1
7   specimen_7 patient_04  baseline       grade_2
8   specimen_8 patient_04 follow_up       grade_2
9   specimen_9 patient_05  baseline       grade_2
10 specimen_10 patient_05 follow_up       grade_2
11 specimen_11 patient_06  baseline       grade_3
12 specimen_12 patient_06 follow_up       grade_3
13 specimen_13 patient_07  baseline       grade_0
14 specimen_14 patient_07 follow_up       grade_0
15 specimen_15 patient_08  baseline       grade_1
16 specimen_16 patient_08 follow_up       grade_2
17 specimen_17 patient_09  baseline       grade_3
18 specimen_18 patient_09 follow_up       grade_2
19 specimen_19 patient_10  baseline       grade_4
20 specimen_20 patient_10 follow_up       grade_3
21 specimen_21 patient_11  baseline       grade_4
22 specimen_22 patient_11 follow_up       grade_3
...