Brms - Variable selection ordinal model

Modeling
1

I wrote to you last May but I have a big problem with interpretation any output from brms package. Can you help to me? My output is this:

Ordinal response, Cumulative family, 46 variables, 54 observations, horseshoe prior for all predictors, all predictors are quantitative numeric variable [0-1] values.

Only 1 variable has not include 0 between values of CI. I have some problem to determinate the discriminant variables with baesian approach. Thank you so much.

summary()

  Horseshoe prior Population-Level 
                          Estimate Est.Error l-95% CI u 95% CI 
Intercept[1]           -19.64       14.33     -53.65     4.16          
Intercept[2]           -13.38       13.77     -45.24    10.03              
CDKN2B_seq_              4.40        5.40      -2.66    17.55              
DDIT3_P1313_R            4.47        4.92      -2.42    16.39               
ERN1_P809_R             -3.46        4.45     -13.86     3.23            
GML_E144_F              -5.67        3.01     -12.67    -0.46          
HDAC9_P137_R            -2.27        3.4       -9.70     3.53          
HLA.DPA1_P205_R         -3.52        3.42     -11.19     1.83        
HOXB2_P488_R             1.98        3.92      -4.78    11.17         
IL16_P226_F             -6.03        5.92     -21.10     2.01           
IL16_P93_R              -3.34        3.95     -12.63     2.80            
IL8_P83_F               -1.67        2.08      -6.23     1.93              
MPO_E302_R              -2.72        4.77     -14.44     4.43          
MPO_P883_R              -2.17        3.67     -10.58     3.89         
PADI4_P1158_R            2.97        3.37      -2.18    10.76          
SOX17_P287_R             4.53        4.06      -1.50    13.72          
TJP2_P518_F              3.67        4.99      -3.46    16.05             
WRN_E57_F               -2.03        4.06     -11.85     4.78            
CRIP1_P874_R            -0.66        2.91      -7.50     4.76           
SLC22A3_P634_F           1.05        3.71      -5.82     9.03         
CCNA1_P216_F             2.85        4.22      -3.76    12.64      
SEPT9_P374_F             0.03        3.61      -7.44     7.59        
ITGA2_E120_F             0.82        3.96      -6.47     9.73 
ITGA6_P718_R            -0.12        4.30      -8.88     8.50 
HGF_P1293_R             -1.08        4.43     -11.16     7.22 
DLG3_E340_F             -1.59        2.73      -7.93     3.04 
APP_E8_F                -0.21        4.22      -9.14     8.29 
SFTPB_P689_R            -1.47        4.21     -11.16     6.26 
PENK_P447_R              1.29        3.38      -5.13     8.71
COMT_E401_F             -0.27        3.10      -6.81     6.16 
NOTCH1_E452_R           -0.12        4.25      -9.01     8.53      
EPHA8_P456_R            -0.68        4.45     -10.59     7.96 
WT1_P853_F               0.20        3.53      -7.31     7.42 
KLK10_P268_R            -2.09        3.57     -10.09     4.11 
PCDH1_P264_F             0.06        2.67      -5.68     5.60 
TDGF1_P428_R             0.48        2.75      -5.04     6.65 
EFNB3_P442_R             0.48        3.94      -7.55     9.33 
MMP19_P306_F            -0.67        4.16      -9.73     7.50    
FGFR2_P460_R             1.17        3.09      -4.32     8.20     
RAF1_P330_F              0.15        3.90      -8.13     8.32      
BMPR2_E435_F             0.02        4.31      -9.32     9.40    
GRB10_P496_R             1.16        2.74      -3.99     7.09     
CTSH_P238_F              0.34        4.21      -8.29     8.94   
SLC6A8_seq_28            1.53        2.70      -2.92     8.04     
PLXDC1_P236_F            0.90        3.96      -6.80     9.76      
TFE3_P421_F             -0.48        3.01      -7.30     5.83       
TSG101_P139_R           -0.19        4.13      -8.95     8.46

Thank you so much.

2

Hi, sorry for not getting to you earlier.

I fear that with this model and dataset you just cannot learn anything useful. 54 observations are not even enough to inform an intercept-only model with high precision. If there is a big difference, you migth learn something about one predictor, maybe even two predictors, but I don’t see how how you could learn anything about more than 40 predictors.

I would expect the single variable that excludes zero from the 95% CI to be most likely a fluke (the CI is still quite close to zero and very wide) and would not interpret it in any way.

Hope that helps at least a bit.