Choosing Starting Values for Likelihood Optimization

Why Starting Values Matter

A likelihood optimizer does not begin by knowing which parameter values fit the data. It starts at a supplied point and uses the likelihood and its derivatives to search for a better one. In a longitudinal selection model, this initial point can matter practically: outcome-distribution parameters and response parameters enter the same observed-data likelihood, and the response probability for a dropout can depend on an outcome that was not observed. A starting value that implies near-certain retention for everyone, or nearly impossible retention for everyone, can make numerical calculations unhelpful before optimization has made meaningful progress.

The goal is not to find a clever shortcut that replaces likelihood maximization. It is to provide finite, interpretable initial values that are reasonably compatible with the observed outcome and dropout patterns. A starting value is a computational choice, not an estimate, a test, or evidence that its MNAR assumption is correct.

Model Setup and Assumptions

Consider subjects \(i=1,\ldots,n\) scheduled for \(T\) outcome measurements. Write \(Y_{it}\) for the outcome at visit \(t\), and \(R_{it}=1\) if it is observed and \(R_{it}=0\) otherwise. Assume monotone dropout: once \(R_{it}=0\), all later indicators are zero. For a subject still observed at visit \(t\), let the probability of remaining observed at visit \(t+1\), conditional on the complete outcome vector, be

where \(\operatorname{expit}(u)=1/(1+\exp(-u))\), \(X_i\) denotes covariates, and \(z_{it}\) is a specified vector of covariates and outcome terms. For an MNAR selection model, \(z_{it}\) includes at least one outcome value not observed for some subjects at that transition. The outcome model has density \(f_Y(y_i\mid X_i;\beta,\Sigma)\), with \(\beta\) denoting outcome-model parameters and \(\Sigma\) any covariance parameters.

Under this model, the observed-data likelihood contribution for subject \(i\) is obtained by multiplying the outcome density by the conditional response probabilities and integrating over the unobserved outcomes:

where \(\theta\) collects all parameters, and the integral is over components of \(y_i\) that are missing. The response factors are included only for transitions at which the subject was at risk; for a fully observed final visit there is no subsequent transition to model. This formulation makes an important practical point: a preliminary response fit using observed values or proxies can suggest initial values, but it does not evaluate the full likelihood contribution for a dropout. The likelihood still has to account for the missing outcomes under the specified outcome model.

The strategies below presume that the outcome distribution and response model have been specified, that the observed records follow the stated monotone pattern, and that the parameters being initialized are on a meaningful scale. A simple preliminary fit can be unstable or biased, particularly when complete cases differ from dropouts. Its role here is to give the optimizer a sensible region in which to begin, not to establish the assumptions or resolve weak identification.

Build the Outcome-Model Starting Values First

Start with parameters that can be estimated using the observed outcomes without fitting the selection part. For example, use a simple regression or visit-specific mean model to initialize the outcome means. Initialize variances and correlations using observed data or a simpler fitted outcome model, while checking that the resulting covariance matrix is positive definite. If the eventual model includes covariates, the preliminary model should use the same coding and units where possible. A mean for an outcome recorded in millimeters should not be paired with a response slope interpreted per meter unless the conversion is made explicit.

A useful proxy for a missing outcome is its conditional expectation under the preliminary outcome model. For a bivariate normal pair, if \(Y_1\) and \(Y_2\) have means \(\mu_1,\mu_2\), standard deviations \(\sigma_1,\sigma_2\), and correlation \(\rho\), then

This conditional mean can provide a practical proxy when constructing an initial response fit. It is not a replacement for integrating over the missing outcome: plugging in a single conditional mean generally does not yield the same likelihood as averaging the response probability over the conditional distribution.

Initialize the Response Model

A response-model intercept can often be initialized from the observed retention rate. With a logistic model and no outcome slope, the intercept matching a retention proportion \(\hat p\) is \(\operatorname{logit}(\hat p)=\log\{\hat p/(1-\hat p)\}\), provided \(0<\hat p<1\). With a nonzero outcome slope, however, the intercept must be calibrated jointly with that slope if the goal is to preserve an average retention probability.

One practical workflow is to choose one or more plausible values for the MNAR slope, set preliminary proxies or a reference distribution for the outcome term, and solve for the intercept that matches an observed retention rate over the relevant risk set. This makes the intercept interpretable as a calibration parameter for that particular slope and reference distribution. It does not show that the slope is identified by the observed data or that the chosen slope is scientifically correct.

Compare Starting Strategies

No single initialization is appropriate for every data set. Several strategies can be compared while keeping the outcome-model starts fixed. The following table summarizes their intended role and main caution.

For a monotone longitudinal record, initialize a separate response intercept for each dropout transition when the model allows one. Retention rates can differ across visits, so copying one intercept to every transition may give implausible probabilities. For a logistic slope attached to an outcome measured in its original units, scale matters: a slope per unit may be numerically tiny if the outcome spans thousands of units. Centering and scaling can help, but the parameter interpretation must be mapped correctly. Do not start a logistic probability at exactly zero or one; a finite intercept and slope give probabilities strictly between them.

A concise initialization procedure is:

Diagnostics, Failure Modes, and Limitations

Before optimization, calculate the implied retention probabilities at representative outcome values and covariate combinations. They should be strictly between zero and one and should not be essentially zero or one across nearly the entire risk set unless that pattern is genuinely intended. Check each transition separately, especially when dropout rates change over time. Verify that the outcome covariance start is positive definite and that every parameter is finite. If a small change in an intercept creates a large probability change, inspect the linear predictor and the units of its components.

During fitting, record the initial and final log likelihood, convergence status, and whether parameter values remain within the model's permissible region. A reported optimizer success flag alone is not proof that the result is a valid maximum. A major warning is that estimates or convergence behavior change substantially when reasonable starting strategies are changed. That finding calls for further likelihood and identification diagnostics; it does not justify selecting whichever start produces a preferred scientific conclusion. Detailed analysis of competing maxima is a separate problem.

Starting values from a complete-case outcome fit can be poor when completion is strongly selective. Proxy-based response regressions can also be misleading because the missing outcomes are replaced by predictions rather than integrated over their uncertainty. Likewise, an intercept calibrated to the overall retention proportion may not match retention across covariate strata or transitions. The full likelihood, rather than any preliminary fit, determines the final estimates under the specified model.

A pattern-mixture approach parameterizes outcome distributions by missingness pattern rather than modeling response probabilities conditional on outcomes. It can offer a different way to express assumptions, but it is not merely another initialization for this selection likelihood. Within the selection-model fit, a zero-slope start, calibrated slope starts, and proxy-based starts are alternative computational strategies. None removes the need to state the model assumptions, assess fit and stability, and interpret MNAR parameters cautiously.

Common Mistake

A related error is to use a conditional-mean proxy as if it were the missing outcome itself. If response depends nonlinearly on that outcome, then in general \(E\{\operatorname{expit}(\alpha+\gamma Y)\}\) is not equal to \(\operatorname{expit}\{\alpha+\gamma E(Y)\}\). The proxy may help construct a starting point, but the likelihood calculation must retain the model-based integration over missing outcomes.