From:Leonid GibianskySubject: [NMusers] model diagnostics Date: Thu, 01 May 2003 13:07:17 -0400 Dear All, It was unusually quiet recently in the list, let me suggest an idea for discussion: There were recent messages that mentioned model diagnostics other than usual estimates and standard errors of estimates. It would be good to have an expect advise on what to do with COVARIANCE MATRIX OF ESTIMATE CORRELATION MATRIX OF ESTIMATE INVERSE COVARIANCE MATRIX OF ESTIMATE What they offer on top of what we recover from estimates and standard errors of estimates ? Do you always check them ? Any criteria what is good, what is bad ? Practical advices what to do if something looks bad ? Thanks, Leonid _______________________________________________________ From:"Bachman, William" Subject: [NMusers] from Ken Kowalski re: model diagnostics Date:Fri, 9 May 2003 08:32:40 -0400 Leonid, Attached please find a 4-page discussion on the $COV step that I excerpted from a best practices document that my company is preparing. Hopefully you'll find this useful. Ken COV Step Estimation The $COV statement is used to estimate the asymptotic variance-covariance matrix of the estimates in q, W, and S. The default matrix used in NONMEM is the R-1SR-1 matrix where R denotes the hessian matrix (i.e., matrix of second derivatives with respect to the parameters evaluated at the final estimates) and S denotes the cross-product gradient matrix (i.e., matrix obtained from the cross product of the gradient vector and its transpose where the gradient is a vector of first derivatives evaluated at the final estimates). When the random effects (h's and e's) are normally distributed the inverses of both the R and S matrices are consistent estimators of the covariance matrix of the parameter estimates. However, in the presence of non-normal random effects the R-1SR-1 matrix is a more robust estimator of the covariance matrix (1). For categorical and other non-continuous data (i.e., using the LIKELIHOOD option with the $EST statement), consistency (asymptotically unbiased) and other optimal properties of the estimates are highly dependent on the assumption of normality of the random effects. For these types of data, the MATRIX=R option should be used with the $COV statement. It is good practice to only report out models for which the COV step runs successfully. COV step failures should not be ignored as they usually imply ill conditioning of some aspect of the model. Two common warning messages that NONMEM reports out when the COV step fails are: 1) the R or S matrix is singular, and 2) the R matrix is non-positive semi-definite. The singularity condition implies an infinite set of solutions (estimates) of the parameters can result in the same OFV suggesting that the likelihood surface is very flat near the final estimates. When the likelihood surface is extremely flat, rounding errors may occur and the estimation may not converge to a final solution. The non-positive semi-definite condition suggests that the stationary point is not a global minimum but rather a saddle point. In which case, in some direction of the parameter space the OFV can go to **. Both of these conditions usually imply that the model is over-parameterized in some aspect of the model related to q, W, or S. Regardless of which warning message is reported the solution to the ill conditioning or over-parameterization is the same. Simplification of the structural and/or statistical aspects of the model is usually necessary. Although the reported OFV may not be a global minimum when the COV step fails, in practice it is reasonable to identify a more parsimonious (fewer parameters) model with a successful COV step that has an OFV close to that reported for the over-parameterized model. In some cases, the more parsimonious model may lead to a lower OFV than the over-parameterized model even when the two models are hierarchical. This can happen when the over-parameterized model converges to a local optimum rather than a global minimum. When the COV step runs successfully, it is also good practice to inspect the correlation matrix of the estimates of q, W, and S reported in the NONMEM output to ensure that the fitted model is stable. When one or more correlations are near *1 the model may still be over-parameterized and unstable. A model may be unstable if the COV step fails for one set of 'reasonable' starting values but runs successfully with another set of starting values. In this setting one or more correlations may be near *1 (a near singular condition) but numerically nonsingular so that the COV step runs successfully. When this occurs, standard errors for one or more of the parameters are usually quite large relative to their parameter estimates. For these reasons, simply changing starting values to obtain a successful COV step usually does not resolve the ill conditioning of the model. Bates and Watts (2) suggest that when one or more correlations exceed 0.99 in absolute value the model is over-parameterized. Given the large number of parameters often involved in fitting population models this guidance suggests that further evaluation and possible simplification of the model to guard against over-parameterization should be considered when one or more correlations exceed 0.95 in absolute value. It should also be noted that ill conditioning could exist without any one correlation exceeding 0.95. An alternative approach for assessing ill conditioning is to inspect the eigenvalues of the covariance matrix. This can be performed by using the PRINT=E option on the $COV statement. Specifically, the ratio of the largest eigenvalue to the smallest eigenvalue, referred to as the condition number, is a measure of ill conditioning. A condition number exceeding 1000 is indicative of severe ill conditioning (3). This document recommends routine use of the PRINT=E option and calculation of the condition number. When the condition number is high (>1000) often there is a cluster of correlations that are relatively high even if no one correlation exceeds 0.95. The condition number provides a simple statistic for assessing the degree of ill conditioning, while inspection of the correlation matrix can provide insight into the source of the ill conditioning. When over-parameterization exists, it may be that the over-parameterized model is scientifically plausible but the data do not support estimating all of the parameters. For example, an Emax model may be postulated but the concentration-response relationship may not exhibit sufficient curvature (plateauing) to accurately estimate the Emax and EC50. In this situation the correlation between the fixed effects estimates for Emax and EC50 may be highly correlated. A simplification of the model assuming a linear concentration-response may adequately describe the data leading to a more stable model. Another example involving ill conditioning due to specification of random effects is the situation where an Emax dose-response model is fit to data arising from a parallel group dose-ranging trial. In this setting it may be unrealistic to expect stable estimation of interindividual random effects on both Emax and ED50, as there is little information on dose-response within an individual (i.e., each individual only receives one dose level). To guard against over-parameterization, inspection of the correlation matrix of the estimates (q, W, and S) from the NONMEM output should be performed at all stages of model development. Specifically, the stability of the base model should be investigated prior to covariate model building. If the COV step runs successfully but one or more pairwise correlations are near one, say for example, between Emax and EC50, then introduction of covariates on Emax and EC50 may exacerbate the instability of the model perhaps leading to convergence problems (rounding errors) or COV step failures. One may falsely conclude that the instability is due to inclusion of covariate effects when in fact the instability is associated with the structural model and imprecision in the estimates of the structural parameters (i.e., base model q's). In general, model development should not be guided solely based on the models that converged and had successful COV step estimations without attempting to understand the underlying cause of the over-parameterization associated with the models that had convergence and/or COV step failures. Understanding the nature of the over-parameterization when it occurs can be helpful in postulating a more parsimonious model. References 1 Beal, S.L., and Sheiner, L.B. NONMEM Users Guide - Part II: Users Supplemental Guide. NONMEM Project Group: University of California, San Francisco, 1988, p 21. 2 Bates, D.M., and Watts, D.G. Nonlinear Regression Analysis and its Applications. Wiley, NY, 1988, pp. 90-91. 3 Montgomery, D.C., and Peck, E.A. Introduction to Linear Regression Analysis. Wiley, NY, 1982, pp. 301-302. _______________________________________________________ From: "Kowalski, Ken" Subject:RE: [NMusers] from Ken Kowalski re: model diagnostics Date: Fri, 9 May 2003 09:12:41 -0400 Bill, Thanks for converting my word document to email text so that it could be distributed to Nmusers. All, In the write-up below regarding COV step estimation, certain symbols were converted and may be a bit confusing. Here is the translation: q = theta W = omega S = sigma h = eta e = epsilon ** = +/- infinity *1 = +/-1 Sorry for the confusion. Ken _______________________________________________________