Hi, I simulated a dataset with 500 profiles of a biomarker measured every 3 weeks during 2 years. My structural model is physiological and so a little complex. It can be compared to a biexponentiel function with a time of rupture D. I have 4 fixed effects, 4 random effects (3 are log-transformed and 1 logit-transformed) and a constant residual error on log(Y+1) to estimate. My code is the following: proc nlmixed data=dataset; parms mu_A=0.05 mu_B=80 mu_C=0.3 mu_D=140 omega_A=0.1 omega_B=0.6 omega_C=1.5 omega_D=0.6 sigma=0.36; A=mu_A*exp(eta_A); B=mu_B*exp(eta_B); C=mu_C/(mu_C+(1-mu_C)*exp(-eta_C)); D=mu_D*exp(eta_D); YD=0.23*B/(A*(1-C)-0.046+0.23)*exp((A*(1-C)-0.046)*D)+(B-0.23*B/(A*(1-C)-0.046+0.23))*exp(-0.23*D); if (TIME<D) then Y= 0.23*B/(A*(1-C)-0.046+0.23)*exp((A*(1-C)-0.046)*TIME)+(B-0.23*B/(A*(1-C)-0.046+0.23))*exp(-0.23*TIME) ; else Y=0.23*B*exp((A*(1-C)-0.046)*D) /(A-0.046+0.23)*exp((A-0.046)*(TIME-D))+(YD-0.23*B*exp((A*(1-C)-0.046)*D)/(A-0.046+0.23))*exp(-0.23*(TIME-D)) ; lY=log(Y+1); model logY ~ normal(lY, sigma**2); random eta_A eta_B eta_C eta_D ~ normal([0,0,0,0], [omega_A**2, 0, omega_B**2, 0, 0, omega_C**2, 0, 0, 0, omega_D**2]) subject=ID; run; I would like to use the adaptative Gauss quadrature or Laplace approximation. This code works for the methods FIRO or for non-adaptative Gauss quadrature (but NOAD needs a lot of QPOINTS and takes a long time). But for the adaptative Gauss quadrature, I have the error message: No valid parameter points were found and the initial negative log-likelihood is 1.15792E77. I don’t understand this error since my dataset is simulated given this model and I initialize parameters with the good values. I tried simpler models, notably with only one or two random effects or with smaller omega. Some work, if omegas are enough small, or if only A and B have random effects. Is my model too complex for SAS? Is there a problem with the rupture time D? Have I too much variability? My goal is then to write my own likelihood function and to fit real data, so I could not control their variability. Thanks in advance
... View more