Biostatistics (2005), 6, 1, pp. 39–44 doi: 10.1093/biostatistics/kxh016
Quasi-likelihood estimation for relative risk regression models RICKEY E. CARTER, STUART R. LIPSITZ, BARBARA C. TILLEY Department of Biometry and Epidemiology, Medical University of South Carolina, Charleston, SC 29425, USA
[email protected]
S UMMARY For a prospective randomized clinical trial with two groups, the relative risk can be used as a measure of treatment effect and is directly interpretable as the ratio of success probabilities in the new treatment group versus the placebo group. For a prospective study with many covariates and a binary outcome (success or failure), relative risk regression may be of interest. If we model the log of the success probability as a linear function of covariates, the regression coefficients are log-relative risks. However, using such a log–linear model with a Bernoulli likelihood can lead to convergence problems in the Newton–Raphson algorithm. This is likely to occur when the success probabilities are close to one. A constrained likelihood method proposed by Wacholder (1986, American Journal of Epidemiology 123, 174–184), also has convergence problems. We propose a quasi-likelihood method of moments technique in which we naively assume the Bernoulli outcome is Poisson, with the mean (success probability) following a log–linear model. We use the Poisson maximum likelihood equations to estimate the regression coefficients without constraints. Using method of moment ideas, one can show that the estimates using the Poisson likelihood will be consistent and asymptotically normal. We apply these methods to a double-blinded randomized trial in primary biliary cirrhosis of the liver (Markus et al., 1989, New England Journal of Medicine 320, 1709– 1713). Keywords: Bernoulli likelihood; Constrained MLE; Convergence problems; Method-of-moments; Poisson regression.
1. I NTRODUCTION Often logistic regression is used to adjust for covariates when the outcome is binary. The coefficients in the logistic regression model are interpreted as log-odds ratios. As a measure of treatment effect in a randomized clinical trial, the odds ratio (OR) is used less often because it does not measure the relative size of the effect of the new treatment over placebo (Sinclair & Bracken, 1994; DerSimonian & Laird, 1986; Zhang & Yu, 1998). Alternatively, the relative risk (RR) can be used as a measure of effect and is directly interpretable as the ratio of success probability in the new treatment group versus the success probability in the placebo group (Lu & Tilley, 2001; Nam, 1999). In a study where participants are followed prospectively, such as a cohort or clinical trial, the incidence of the event (success) is observable, and the probability of success, the RR and the OR are all estimable. In this article, we consider prospective designs, and instead of a logistic regression model for the success probabilities, we discuss a log–linear regression model for the success probabilities. In particular, if we model the log of the success probability as a linear function of covariates, the regression coefficients c Oxford University Press 2005; all rights reserved. Biostatistics Vol. 6 No. 1
40
R. E. C ARTER ET AL.
are log-relative risks. However, even though the exponential of the regression coefficients are directly interpretable as relative risks, such a log–linear model is not widely used. One reason for this is that a log–linear model with a Bernoulli likelihood can lead to convergence problems in the Newton–Raphson algorithm when the success probabilities are close to one. Wacholder (1986) proposed a constrained maximum likelihood method in order to circumvent these convergence problems, but the constrained maximum likelihood method can still have convergence problems. Zhang & Yu (1998) propose a technique for estimating RR based on the mathematical relationship of OR with RR. To illustrate this relationship, let PA and PB represent the probability of success in groups A and B, respectively. Then, the relative risk can be written as R R = PA /PB = O R/(1 − PB + PB × O R), where the estimate of OR is obtained using logistic regression (Zhang & Yu, 1998). We propose a quasi-likelihood method to obtain consistent estimates of the relative risk parameters even when the outcome is ‘non-rare’. Providing a method for estimating ‘non-rare’ events is important since 0 < | ln(RR)| < | ln(OR)| with the difference between OR and RR increasing in value as the probability of success approaches 1. In addition, unlike the OR, the RR for success does not equal the inverse of the RR for failure. This is problematic in that the definition of a success cannot be altered to improve convergence. For example, when PA = 0.94 and PB = 0.99, the estimated RR is 0.95 and indicates a slight protective trend. If the definition of a ‘success’ were reversed such that PA = 0.06 and PB = 0.01, then the estimated RR would be 6, which is vastly different in interpretation from [0.95]−1 = 1.05. As such, for a fixed risk difference, the ‘relative’ level of improvement diminishes as the probabilities increase, and only RR is capable of producing an estimated level of association consistent with this general understanding.
2. M ODEL AND NOTATION Suppose Yi is the binary response for subject i, i = 1, . . . , N , where N is the number of independent subjects. Let E[Yi |X i ] = P(Yi = 1|X i ) = Pi be the success probability, where X i is the (K × 1) vector of covariates. We model the success probability using the log link, i.e. ln(Pi ) = X i β, or, equivalently Pi = e X i β , where β is the parameter vector of interest. One can easily show that the elements of β are interpreted as log-relative risks. For the scope of this research, our interest centers on estimating the regression parameter vector β. One can obtain maximum likelihood estimates (MLEs) for relative risk regression by the solving the ˆ = 0, where maximum likelihood score equations S(β)
S(β) =
N ∂ ln L(β) X i (Yi − Pi ) . = ∂β 1 − Pi i=1
(2.1)
However, when the success probability approaches one, the denominator of (2.1) approaches 0, resulting in convergence problems. Wacholder (1986) constrained the likelihood to prevent (1− Pi ) from approaching zero. This approach requires an extra step that determines if e X i β P ∗ , where P ∗ is a predetermined maximum allowed value such as 0.99. This constrained estimation method has not been incorporated into standard statistical software; nonetheless, it is not complicated to program such an algorithm. However, in the simulation summary that accompanies this research, these constrained estimates can have poor properties when the success probability approaches one. The simulation study is available on the journal’s web site. We discuss our approach in the following section.
Quasi-likelihood estimation for relative risk regression models
41
3. Q UASI - LIKELIHOOD USING P OISSON REGRESSION Our approach is to examine the part of the maximum likelihood score estimating equations that causes the convergence difficulties. If one looks at the estimating function in (2.1), one sees that, as Pi approaches 1, the denominator approaches 0, and the ratio approaches ∞. If one replaces (1 − Pi ) by 1 in the denominator of (2.1), then these convergence problems will be avoided, as shown in the simulations = 0, where available on the journal’s web site. The resulting estimating equations are S(β) S(β) =
N i=1
Si (β) =
N
X i (Yi − Pi ).
(3.1)
i=1
from this quasi-likelihood converges ˆ = 0, the estimates β Since E[S(β)] = 0 and we are solving S(β) to β, establishing consistency using method of moments (Gourieroux et al., 1984). Moreover, using a has an asymptotic normal distribution first-order Taylor series expansion and the central limit theorem, β with mean β and covariance matrix N i=1
−1 E[Ii (β)]
N
−1 N
E[Si (β)Si (β) ] T
i=1
−1 E[Ii (β)]
,
(3.2)
i=1
where Ii (β) = −∂ Si (β)/∂β is the information matrix. ˆ which is the robust variance A consistent estimate of (3.2) can be obtained by evaluating (3.2) at β, ˆ estimate proposed by White (1982). Note that, even though β is consistent, the inverse of the information matrix evaluated at βˆ will not provide a consistent estimator of the asymptotic variance since (3.1) is not the Bernoulli score equation. Thus, the robust variance estimator used in generalized estimating equations (GEE) must be used (Zeger & Liang, 1986). The estimating equations in (3.1) are identical to the maximum likelihood estimating equations that would be obtained if one naively (wrongly) assumes that Yi is Poisson with mean (and variance) equal to Pi = e X i β . Therefore, by wrongly specifying the Bernoulli distribution as Poisson, one obtains a set of score equations for consistently estimating RR that can be implemented using any generalized linear model software. However, to obtain the robust variance, one must implement the model in GEE software such as PROC GENMOD in SAS. For example, in SAS PROC GENMOD the log link and the Poisson distribution assignments in the model statement are straightforward, and the robust variance estimate is obtained by the REPEATED statement. If one does not use the REPEATED statement, the inverse of the estimated information matrix will be used to estimate the variance. 4. E XAMPLE The Mayo Clinic conducted a double-blinded randomized clinical trial in primary biliary cirrhosis of the liver (PBC), comparing the drug D-penicillamine (DPCA) with placebo (Markus et al., 1989). The date of randomization and a large number of clinical, biochemical, serologic, and histological parameters were recorded for each of the 312 clinical trial subjects. Detailed discussion of the study can be found in the original manuscript (Markus et al., 1989). Of possible clinical interest is estimating the relative risk of a one-year survival given a group of covariates. Therefore, the outcome Yi equals 1 if subject i survived longer than 1 year, and equals 0 if subject i survived 1 year or less. There is no censoring before 1 year, so Yi is defined for all subjects. Then, Pi = e X i β is the probability of individual i surviving longer than one year, given the covariates X i = [1, stagei1 , stagei2 , stagei3 , genderi , treatmenti , agei ]
42
R. E. C ARTER ET AL.
Table 1. Model-based estimates of log relative risk for individuals diagnosed with primary biliary cirrhosis of the liver One year survival Estimate SEa p-value log(RR) −0.0456 0.2993 0.8790 −0.0245 0.0460 0.5951 0.0390 0.0729 0.5927
Five year survival Estimate SEa p-value log(RR) −1.1477 0.3946 0.0036 −1.1477 0.3390 0.0007 −1.2028 0.3712 0.0012
Parameter
Method
Intercept
Unconstrained Constrainedb Quasi-likelihood
Stage 1
Unconstrained Constrainedb Quasi-likelihood
0.0955 11.2347 0.1439
0.0000 0.0378 0.0379
N/A 0.0000 0.0001
0.9383 0.9384 0.9672
0.2016 0.1976 0.1977
0.0036 <0.0001 <0.0001
Stage 2
Unconstrained Constrainedb Quasi-likelihood
0.0713 0.1535 0.1541
0.0846 0.0395 0.0431
0.3991 0.0001 0.0003
0.8154 0.8154 0.8098
0.1669 0.1669 0.1672
<0.0001 <0.0001 <0.0001
Stage 3
Unconstrained Constrainedb Quasi-likelihood
0.0621 0.1118 0.1416
0.0767 0.0425 0.0429
0.4184 0.0085 0.0010
0.5889 0.5889 0.5858
0.1688 0.1672 0.1673
0.0005 0.0004 0.0005
Gender
Unconstrained Constrainedb Quasi-likelihood
−0.0197 −0.0436 −0.0503
0.1175 0.0174 0.0442
0.8668 0.0123 0.2553
0.1547 0.1547 0.2516
0.2161 0.1681 0.1965
0.4738 0.3573 0.2004
Treatment
Unconstrained Constrainedb Quasi-likelihood
0.0152 0.0172 0.0371
0.0562 0.0212 0.0308
0.7871 0.4179 0.2281
0.0008 0.0008 0.0178
0.1082 0.0974 0.1062
0.9944 0.9938 0.8666
Unconstrained −0.0016 0.0036 0.6620 −0.0034 0.0050 0.4972 Constrainedb −0.0021 0.0010 0.0331 −0.0034 0.0048 0.4773 Quasi-likelihood −0.0037 0.0015 0.0108 −0.0042 0.0052 0.4149 a Robust estimate used for Constrained and Quasi-likelihood. b 1000 iterations used for the one-year survival. Final convergence criterion equaled 0.03 (0.04 reached on 50 iterations). Age
where stageik equals 1 if the disease is histologic stage k, and equals 0 otherwise; genderi equals 1 if female and 0 if male; treatmenti equals 1 if D-penicillamine and 0 if placebo; and agei equals age in years. The crude estimate of the probability of surviving at least one year is approximately 0.9 in the PBC trial data. As a result, it is anticipated that the Bernoulli likelihood will be numerically challenged, so the proposed Poisson likelihood method and Wacholder’s constrained likelihood method with P ∗ = 0.99 are also considered. Table 1 presents the results from each of the three methods. The Bernoulli score equations did not converge to a unique solution in SAS PROC GENMOD (or in generalized linear models programs in Stata or S-Plus), and the information matrix was not positive definite at the solution printed out in SAS PROC GENMOD. As such, the standard error for the Stage 1 indicator variable was set to 0. The constrained likelihood did not fully converge at a convergence criterion of 0.0001 between successive iterations. Increasing the number of iterations from 50 to 1000 failed to improve convergence. In contrast, the Poisson likelihood converged. Looking at Table 1, we see that, compared to the Poisson quasi-likelihood method, the constrained and unconstrained Bernoulli likelihoods
Quasi-likelihood estimation for relative risk regression models
43
give appreciably different estimates for the effect of every covariate. The Poisson quasi-likelihood method produced conclusions on the significance of the covariates that are consistent with earlier analyses (Markus et al., 1989). The second set of results presented in Table 1 are the ones obtained if the five-year survival is considered. In this instance, the crude estimate of the success probability is approximately 0.5 and all three approaches converged. The parameter estimates are similar amongst the three methods and lead to the same statistical conclusions. The following paragraph summarizes the results based on the one-year survival analysis. This example illustrates that the different approaches to estimate the parameters can lead to different, and possibly conflicting, estimates when convergence is questionable. Based on the simulation results and the convergence difficulties with success probabilities close to one, the results obtained from the Poisson quasi-likelihood are preferred for estimating the one-year relative risk of survival. Interpretation of the parameter estimates are in terms of log relative risks. For example, patients with Stage 1 disease are 15% (R R = exp(0.1439) = 1.15) more likely to live at least one year when compared to patients with Stage 4 disease when controlled for the effects of gender, treatment and age. The original analysis did not show a statistically significant effect of D-penicillamine relative to placebo, and the estimated adjusted relative risk value of 1.04 ( p = 0.23) calculated using the proposed method also supports this decision. 5. D ISCUSSION The simulation study on the journal’s web site examines the statistical properties of the proposed approach relative to the unconstrained Bernoulli likelihood and the constrained Bernoulli likelihood (Wacholder’s method). Samples sizes of 100, 300, 500 and 700 were considered for outcomes that ranged from rare (Pi ≈ 0) to non-rare (Pi → 1). The simulations show that our proposed approach has good statistical properties under many varying conditions. Perhaps the strongest attributes of the Poisson quasi-likelihood are its observed 100% convergence rate and negligible bias. In all of the simulations, no simulation replicate failed to converge when the Poisson quasi-likelihood was used even when the Bernoulli likelihood and constrained likelihood converged infrequently. One limitation of the model is that it is undefined for retrospective (or case-control) studies. However, in this setting, the outcome of interest is typically rare and RR estimated by the OR would be a reasonable data summary. A general strategy for fitting a RR model could be to fit the correct (Bernoulli) likelihood to see if the model converges and fit the Poisson quasi-likelihood for comparison. If convergence for the Bernoulli likelihood is questionable or completely fails, then the results from the Poisson quasi-likelihood should be used. If the success probability is much less than 1, then the constrained likelihood may produce estimates that are more precise; however, the parameter estimates from the Poisson quasi-likelihood are consistent, and existed for all simulation configurations, regardless of the sample size or the crude estimate of the success probability. In conclusion, the Poisson quasi-likelihood is a valid and robust tool for estimating a RR regression model. ACKNOWLEDGEMENTS We are grateful for the support provided by grants 1 U10 DA013727, HL69800, AHRQ 10871, and CA 70101 from the National Institutes of Health in the United States. R EFERENCES D ER S IMONIAN , R. AND L AIRD , N. (1986). Meta-analysis in clinical trials. Controlled Clinical Trials 7, 177–188. G OURIEROUX , C., M ONFORT, A. Econometrica 52, 681–700.
AND
T ROGNON , A. (1984). Pseudo maximum likelihood methods: Theory.
44
R. E. C ARTER ET AL.
L U , M. AND T ILLEY , B. C. (2001). Use of odds ratio or relative risk to measure a treatment effect in clinical trials with multiple correlated binary outcomes: data from the NINDS t-PA stroke trial. Statistics in Medicine 20, 1891–1901. M ARKUS , B. H., D ICKSON , E. R., G RAMBSCH , P. M., F LEMING , T. R., M AZZAFERRO , V., K LINTMALM , G. B., W IESNER , R. H., VAN T HIEL , D. H. AND S TARZL , T. E. (1989). Efficiency of liver transplantation in patients with primary biliary cirrhosis. New England Journal of Medicine 320, 1709–1713. NAM , J.-M. (1999). Power and sample size for testing homogeneity of relative risks in prospective studies. Biometrics 55, 289–293. S INCLAIR , J. C. AND B RACKEN , M. B. (1994). Clinically useful measures of effect in binary analyses of randomized trials. Journal of Clinical Epidemiology 47, 881–889. WACHOLDER , S. (1986). Binomial regression in GLIM: estimating risk ratios and risk differences. American Journal of Epidemiology 123, 174–184. W HITE , H. (1982). Maximum likelihood estimation of misspecified models. Econometrica 50, 1–26. Z EGER , S. L. AND L IANG , K.-Y. (1986). Longitudinal data analysis for discrete and continuous outcomes. Biometrics 42, 121–130. Z HANG , J. AND Y U , K. F. (1998). What’s the relative risk? A method of correcting the odds ratio in cohort studies of common outcomes. Journal of the American Medical Association 280, 1690–1691. [Received March 17, 2004; revised May 13, 2004; accepted for publication May 17, 2004]