Is it possible to transfer some information from your brain to my brain... yes I understand. I read different exchanges between you and other people about the different issues of convergence when using proc glimmix. Thank you again Gwénaël ... View more

Steve, How can I access to your toolbox. Gwénaël ... View more

Thank you very much. I agree that the second question about the convergence is not directly related to the first question but it seem that the coding (class statement or dummy variable) impact the modelling. As suggested by Steve Denham I used 1. method=quad without noptions but the following message occurs NOTE: Convergence criterion (GCONV=1E-8) satisfied. NOTE: At least one element of the gradient is greater than 1e-3. NOTE: Estimated G matrix is not positive definite. and the results of covariance parameter are                                     Covariance Parameter Estimates                        Cov                                         Standard                        Parm       Subject    Group     Estimate       Error                          trt_new    study      meta 1           0           .                          trt_new    study      meta 2           0           . 2. if I add the noptions technique=congra maxiter=1000 gconv=1e-4; the following message occurs ERROR: Floating Point Overflow. ERROR: Termination due to Floating Point Exception 3. while the following program proc glimmix data=temp maxopt=400 pconv=1e-4 class fu study meta; model status = fu study trt1 trt2 / dist=poisson link=log offset=ln_y s ; random trt1 / subject=study type=vc; random trt2 / subject=study type=vc; nloptions technique=congra maxiter=1000 gconv=1e-4; covtest ZeroG; run; gives     Covariance Parameter Estimates                               Cov                            Standard                                Parm    Subject    Estimate       Error                                trt1    study       0.02702     0.03036                                trt2    study       0.01845    0.009520 Thank you for your help. Gwénaël ... View more

For the model 2, I used a order=formatted to obtain the same results because REF did not exist in proc glimmix SAS 9.3. I used dummy variables for treatment variable because it seems that the first following code is faster and converge compared to removing the 2 randoms statement by one (which not converge). proc  glimmix data=temp; class fu study meta; model status = fu study trt1 trt2 / dist=poisson link=log offset=ln_y s; random trt1 / subject=study type=vc; random trt2 / subject=study type=vc; covtest ZeroG; run; proc  glimmix data=temp; class fu study meta trt_new; model status = fu study trt_new/ dist=poisson link=log offset=ln_y s; random trt_new / subject=study type=vc group=meta; covtest ZeroG; run; The objective is to insert a rando effect for treatment effect (A/B) in metaanalysis 1 and a random effect for treatment effect (A/C) in metaanalysis  2. I am trying the following options to resolve the issue of convergent but it does not work in the second program. nloptions technique=congra maxiter=1000 gconv=1e-4; Could you tell me what method used for the estimation in method=PL, QUAD, RSMPL, .... thanks for your advice. Gwénaël ... View more

I am thinking that my confusion come from the 2 following responses  : y= (#event / Person-year) and y=#event. In your last post, you propose the model proc glimmix data=count_data; class study; model LOGtox =  / offset=logPA2 solution;   random intercept   / subject=study ; * to test heterogeneity between study; weight inv_varln; * weight = inverse variance; parms (1) (1) / hold=2; run; But I have an approximation of the variance of tox/person-year or log(tox/person-year). So I have to peform a model where the response is log(tox/person-year) and delete the offset. proc glimmix data=count_data; class study; model LOG(tox/person-years) =  / solution;   random intercept   / subject=study ; * to test heterogeneity between study; weight inv_varln; * weight = inverse variance; parms (1) (1) / hold=2; run; In my last posts I was wondering whether I have to consider the denominator of the response (offset) in the function variance determined by the software or whether the software takes into account only the numerator of incidence rate i.e #events. ... View more

Thank. It's clearer. But based on the paper for log ddd ratios (normal data) and binomial data, is Poisson with internal residual variance determined by software (wihout weight statement) a correct and reasonable approach for incidence rate and gives results similar to the normal distribution with variance of log incidence. ... View more

Dear Lvm, I read a second time the paper you cited above (Meta-analysis with linear and Nonlinear multilevel models using proc mixed and proc nlmixed). The authors used a proc mixed based on log(OR) with weight statement and compare the results using a proc nlmixed with a random effect. In the SAS code of proc nlmixed, I did not seen how they fixed the within-study variance. Very strange ! because the variance function, here, is mu(1-mu)/n for a proportion. Have you some idea about it ! ... View more

For information, please find enclosed a reference Mixed-effects Poisson regression models for meta-analysis of follow-up studies with constant or varying durations. The international journal of Biostatistics. Vol 5, Issue 1, 2009. P.G .Bagos, G.K. Nikolopoulos. Examples are given but using Strata. ... View more

But the results can be more different : I applied these 3 approaches on other data (specific toxicities and not total toxicity). So, no simple solution to this problem. Yes, it was what I am thinking about the comparison of Poisson and Normal distribution. I am wondering whether you will be interest to collaborate on this topic ? ... View more

Please below the results of different approaches                                                                                            Fixed                    Random         -2ResLL Poisson using weight = inverse-variance of incidence rate      -2.6779 (0.1309)       0.3078(0.1058)  31.13 Poisson using weight = inverse-variance of log incidence rate -2.6739(0.1286)       0.2824(0.1030)   30.78 Normal using weight=inverse-variance of log incidence rate*     -2.7431(0.1395)      0.3131(0.1197)   36.05 * because the response is y=log(tox) The -2ResLL is different between the first 2 models and normal model. Poisson model seem better fit than Normal model. Based on these results, could we select the poisson model ? Additionnaly, can you confirm that the random effect as estimated in log scale. THANK ... View more

Thank you for this information but I have some difficulties to choose what variance used to compute the weight. I am only able to calculate the variance of incidence rate or log(incidence rate) but I do not know which used. For a poisson regression, the within-study variance is defined as mean but on which scale : incidence rate or log(incidence rate) ? Could you indicate me on which scale the variance of a poisson model is calculated. The model is log(tox/PA^2) = intercept + random effect. Variance=mean but mean is what ? log(tox/PA^2) or tox/PA^2 or tox or log(tox). Sorry if my questions are stupid. For a normal regression (y=log(tox)), do I calculate the variance of log(tox/PA2) ? After that, I will be able to compare the 2 approches (poisson and normal). Thank you per advance ... View more

To be sure to well understand your proposal and following my last post: In a data set, I add 2 supplementary lines mu=tox; varlnbis=varln/mu;w_varlnbis=w_varln/mu; When I used mu=log(tox) I obtain the following results * weighted approach          -2.6739(0.1286) random=0.2824(0.1030) * fixed the new variance     -2.6778(0.1302) random=0.2980(0.1050)   and with mu=tox the following results * weighted approach          -2.6765(0.1275) random=0.2352(0.099) * fixed the new variance     -2.6778(0.1307) random=0.3047(0.1056)   ... View more

Thank. but I am not sure how to determine the "residual" variance based on the mean. Do I have to take the total number of toxicities (tox) or the incidence rate i.e total number of toxicities divided by person-years. Here A=mu for Poisson model. This question because the weigth is based on the inverse-variance and an estimation of the variance is c/person-years**2 for incidence rate and 1/c for log(IR). In the case of binomial data, I have to use the same approach ? with A=pi(1-pi). Thank par advance ... View more

Steve, I ran your 2 proposed models and compare with a Negative Binomial random model using weight statement(=inverse variance) and a NB random model with type=simple. Note: I updated the number of fixed variance parameters. 1 - Model with weight (inverse variance) proc glimmix data=count_data /*empirical*/;                         class study;                         model tox =  / dist=nb link=log offset=logPA2 solution;                         random intercept   / subject=study ;                         weight inv_varln;                         where groupe=1; run;                                                                      -2 Res Log Pseudo-Likelihood       29.74                                Generalized Chi-Square             16.87                                Gener. Chi-Square / DF              0.99                                     Covariance Parameter Estimates                                                                  Standard                              Cov Parm     Subject    Estimate       Error                              Intercept    study        0.2245      0.1309                              Scale                     0.8371      1.8751                                      Solutions for Fixed Effects                                            Standard                   Effect       Estimate       Error       DF    t Value    Pr > |t|                   Intercept     -2.6600      0.1244       17     -21.38      <.0001 2 - random Model fixing variance covariance  (R) proc glimmix data=count_data;                         class study;                         model tox =  / dist=nb link=log offset=logPA2 solution;                         random intercept   / subject=study;                         random _residual_  / subject=study group=stud y type=simple;                   parms  (0)            (0.125) (0.333) (0.031) (0.2) (0.0625)                                                   (0.0169) (0.005) (0.007) (0.055) (0.0476)                                                   (0.083) (0.033) (0.1)  (0.0625)  (0.0212)                                                   (0.025) (0.0625) (0.0169) (0.0060)             / hold=(2 to 19);                         where groupe=1; run;                                -2 Res Log Pseudo-Likelihood       29.74                                Generalized Chi-Square             16.87                                Gener. Chi-Square / DF              0.99                                     Covariance Parameter Estimates                                                                          Standard                      Cov Parm         Subject    Group       Estimate       Error                      Intercept        study                    0.2244      0.1308                      Residual (VC)    study      study 1       0.1250           .                      Residual (VC)    study      study 2       0.3330           .                      Residual (VC)    study      study 3      0.03100           .                      Residual (VC)    study      study 4       0.2000           .                      Residual (VC)    study      study 5      0.06250           .                      Residual (VC)    study      study 6      0.01690           .                      Residual (VC)    study      study 7     0.005000           .                      Residual (VC)    study      study 8     0.007000           .                      Residual (VC)    study      study 9      0.05500           .                      Residual (VC)    study      study 10     0.04760           .                      Residual (VC)    study      study 11     0.08300           .                      Residual (VC)    study      study 12     0.03300           .                      Residual (VC)    study      study 13      0.1000           .                      Residual (VC)    study      study 14     0.06250           .                      Residual (VC)    study      study 15     0.02120           .                      Residual (VC)    study      study 16     0.02500           .                      Residual (VC)    study      study 17     0.06250           .                      Residual (VC)    study      study 21     0.01690           .                      Scale                                     0.8399      1.8777                                      Solutions for Fixed Effects                                            Standard                   Effect       Estimate       Error       DF    t Value    Pr > |t|                   Intercept     -2.6600      0.1244       17     -21.38      <.0001 3 - Model with no random effect and type=chol(1) for matrix R proc glimmix data=count_data;                         class study;                         model tox =  / dist=nb link=log offset=logPA2 solution;                         random intercept/subject=study group=study type=chol(1);                         parms             (0.125) (0.333) (0.031) (0.2) (0.0625)                               (0.0169) (0.005) (0.007) (0.055) (0.0476)            (0.083) (0.033) (0.1)  (0.0625)  (0.0212)                                                   (0.025) (0.0625) (0.0169) (0.0060)            /  hold=(1 to 19);                          where groupe=1; run;                                -2 Res Log Pseudo-Likelihood       99.51                                Generalized Chi-Square            113.25                                Gener. Chi-Square / DF              6.66                                     Covariance Parameter Estimates                                                                        Standard                        Cov Parm     Subject    Group       Estimate       Error                        CHOL(1,1)    study      study 1       0.1250           .                        CHOL(1,1)    study      study 2       0.3330           .                        CHOL(1,1)    study      study 3      0.03100           .                        CHOL(1,1)    study      study 4       0.2000           .                        CHOL(1,1)    study      study 5      0.06250           .                        CHOL(1,1)    study      study 6      0.01690           .                        CHOL(1,1)    study      study 7     0.005000           .                       CHOL(1,1)    study      study 8     0.007000           .                        CHOL(1,1)    study      study 9      0.05500           .                        CHOL(1,1)    study      study 10     0.04760           .                        CHOL(1,1)    study      study 11     0.08300           .                        CHOL(1,1)    study      study 12     0.03300           .                        CHOL(1,1)    study      study 13      0.1000           .                        CHOL(1,1)    study      study 14     0.06250           .                        CHOL(1,1)    study      study 15     0.02120           .                        CHOL(1,1)    study      study 16     0.02500           .                        CHOL(1,1)    study      study 17     0.06250           .                        CHOL(1,1)    study      study 21     0.01690           .                        Scale                               0.006000           .                                      Solutions for Fixed Effects                                            Standard                   Effect       Estimate       Error       DF    t Value    Pr > |t|                   Intercept     -2.7003     0.04557       17     -59.26      <.0001 4 - random Model and type=chol(1) for R matrix proc glimmix data=count_data;                         class study;                         model tox =  / dist=nb link=log offset=logPA2 solution;                         random intercept/subject=study;                         random intercept/subject=study group=study type=chol(1);                         parms  (0)            (0.125) (0.333) (0.031) (0.2) (0.0625)                               (0.0169) (0.005) (0.007) (0.055) (0.0476)            (0.083) (0.033) (0.1)  (0.0625)  (0.0212)                                                   (0.025) (0.0625) (0.0169) (0.0060)            /  hold=(2 to 19);                          where groupe=1; run;                                -2 Res Log Pseudo-Likelihood       29.73                                Generalized Chi-Square             16.89                                Gener. Chi-Square / DF              0.99                                     Covariance Parameter Estimates                                                                        Standard                        Cov Parm     Subject    Group       Estimate       Error                        Intercept    study                    0.1043           .                        CHOL(1,1)    study      study 1       0.1250           .                        CHOL(1,1)    study      study 2       0.3330           .                        CHOL(1,1)    study      study 3      0.03100           .                        CHOL(1,1)    study      study 4       0.2000           .                        CHOL(1,1)    study      study 5      0.06250           .                        CHOL(1,1)    study      study 6      0.01690           .                        CHOL(1,1)    study      study 7     0.005000           .                        CHOL(1,1)    study      study 8     0.007000           .                        CHOL(1,1)    study      study 9      0.05500           .                        CHOL(1,1)    study      study 10     0.04760           .                        CHOL(1,1)    study      study 11     0.08300           .                        CHOL(1,1)    study      study 12     0.03300           .                        CHOL(1,1)    study      study 13      0.1000           .                        CHOL(1,1)    study      study 14     0.06250           .                        CHOL(1,1)    study      study 15     0.02120           .                        CHOL(1,1)    study      study 16     0.02500           .                        CHOL(1,1)    study      study 17     0.06250           .                        CHOL(1,1)    study      study 21     0.01690           .                        Scale                                 0.1103     0.09373                                      Solutions for Fixed Effects                                            Standard                   Effect       Estimate       Error       DF    t Value    Pr > |t|                   Intercept     -2.6504      0.1239        0     -21.39       . The random model with weight statement (model 1) and the random model with type=simple (model 2) give similar results. Note: When I use a poisson model the scale parameter is not defined. I need to insert a random _residual_  while scale parameter estimation is by default with Negative binomial. The last 3 models give a Pearson/df close to 1 (control of overdispersion). The model 3 indicates the need to adjust for overdispersion chiSquare/df=6.66 (>>1).But  I realize I  make a mistake because I wrongly fixed the parm hold=(1 to 19)  leading to fix the scale parameter. The model  4 (random effect type=chol(1)) give no SE estimation for random effect  and DF=0 for fixed effect. How explain  it ?  Very strange. Thank per advance for you help. Gwénaël ... View more