[remlapmot]

• New function lsmm() fits the double logistic structural mean model (LSMM) of Vansteelandt and Goetghebeur (2003), which estimates a causal odds ratio, by two-step GMM with multiple instruments as described by Clarke et al. (2015). It has print() and summary() methods, the summary includes the Hansen J-test of the overidentifying restrictions, and covariates are handled as in msmm(). It is the same estimator as the ivlsmm command in the Stata package ivonesamplemr, and its estimates agree with ivlsmm.

• tsps() and tsri() with link = "logmult" now converge to the solution of the estimating equations. Previously the GMM optimizer often stopped at its iteration limit, giving slightly inaccurate estimates and a non-convergence warning. The starting values are now obtained by solving the second stage estimating equations exactly by Newton-Raphson, rather than from a gamma regression with zero outcomes replaced by 0.001. The estimates now agree with the Stata commands ivtsps and ivtsri in the ivonesamplemr package.

• tsps() and tsri() no longer give a spurious non-convergence warning when the optimizer has reached the minimum of the GMM objective function (which is 0 because these models are exactly identified).

• msmm() with estmethod = "gmm" or "gmmalt" now uses a tighter convergence tolerance for the Nelder-Mead optimizer (reltol = .Machine$double.eps, maxit = 1E4), so that the estimates converge to the minimum of the GMM objective function. Previously the default tolerance could stop the optimizer early, so estimates depended on the starting values and could differ in the second or third decimal place from the solution of the estimating equations. Estimates from exactly identified models now agree with ivtools::ivglm() G-estimation and with Stata's ivpoisson gmm, multiplicative.

• The parameters in the output of tsps() and tsri() are now labelled by stage and variable name. First stage parameters are prefixed stage1:, the second stage exposure and covariate parameters take their variable names (previously, e.g., xhatxhat and xhatcovariates for tsps()), and the first stage residual in tsri() is labelled stage1:residuals (previously resres). The same labels are used when the user specifies t0, and tsps() and tsri() now give a clear error message if t0 has the wrong length. The documentation of t0 has been corrected.

• The print and summary methods for tsps() and tsri() fits now only exponentiate the exposure parameter in the causal risk ratio or causal odds ratio table. Previously all the second stage parameters were exponentiated and presented as causal ratios, including the intercept, any covariates, and, for tsri(), the coefficient of the first stage residuals.

• msmm() now distinguishes covariates (variables included in both the exposure and instrument parts of the formula) from exposures. The causal risk ratio table (crrci) now only contains the exposures; previously it also reported the exponentiated covariate coefficients as causal risk ratios. The covariate adjusted model itself is unchanged, and is the same model as fitted by Stata's ivpoisson gmm, multiplicative with exogenous covariates. The returned object has a new covariates element, the print methods list the covariates adjusted for and note that E[Y(0)] is then estimated at covariate values of 0, and covariate adjustment is documented with an example in ?msmm. Covariates with the "tsls" and "tslsalt" methods now give a clear error message.