SMC-FCS for Nonlinear Continuous Effects

Why Nonlinear Effects Change Imputation

When a continuous covariate enters the analysis through a square, spline, or another nonlinear function, its relationship with the outcome is not described by a single straight-line slope. That shape matters when the covariate has missing values: plausible imputations should reflect both the covariate distribution and the substantive model’s relationship between the covariate and outcome. An imputation model that ignores the nonlinear effect can produce completed data that do not represent the analysis model adequately.

Substantive-model-compatible fully conditional specification (SMC-FCS) addresses this issue by combining a model for the incomplete covariate with the likelihood from the substantive model. This tutorial focuses on a continuous covariate with a quadratic or other nonlinear effect. It does not cover the details of particular rejection-sampling algorithms or nonlinear effects in logistic substantive models.

Definition: Let $X$ be an incomplete continuous covariate, $Y$ the outcome, and $Z$ other covariates. Let $f_X(x\mid z;\phi)$ be a model for $X$ given $Z=z$, and let $f_Y(y\mid x,z;\psi)$ be the substantive-model likelihood. The SMC-FCS target conditional density for $X$, given $Y=y$ and $Z=z$, is proportional to $$ f_X(x\mid z;\phi) f_Y(y\mid x,z;\psi). $$ The normalized density is obtained by dividing this product by its integral over the possible values of $x$.

For a continuous outcome and a quadratic effect, one possible substantive model is

Y\mid X=x,Z=z \sim N\!\left(\beta_0+\beta_1x+\beta_2x^2+\gamma^\mathsf{T}z,\ \sigma^2\right).

Here $\beta_2$ governs curvature; if $\beta_2=0$, the conditional mean is linear in $x$. The SMC-FCS target density includes the normal outcome likelihood evaluated at the full quadratic mean. It is not enough to treat $x^2$ as an unrelated predictor whose value can vary independently of $x$. Whenever $x$ is imputed, the derived value must be $x^2$.

Conditional Density and Its Justification

The target density follows from the factorization of the joint model for $X$ and $Y$ conditional on $Z$. Under the working factorization $f(x,y\mid z)=f_X(x\mid z;\phi)f_Y(y\mid x,z;\psi)$, conditional-density rules give the density of $X$ given $Y$ and $Z$:

f(x\mid y,z;\phi,\psi) = \frac{f_X(x\mid z;\phi)f_Y(y\mid x,z;\psi)} {\int f_X(u\mid z;\phi)f_Y(y\mid u,z;\psi)\,du}.

The denominator is constant with respect to $x$, so it can be omitted when comparing candidate values or sampling from a distribution proportional to the numerator. This explains the key SMC-FCS adjustment: the covariate model alone supplies a baseline distribution, and the substantive likelihood reweights values according to how well they agree with the observed outcome under the analysis model.

The construction relies on an adequate model for $X\mid Z$, an adequate substantive likelihood for $Y\mid X,Z$, and a valid missingness setup. In particular, standard likelihood-based imputation ordinarily relies on missing at random given the variables in the analysis and an ignorable missingness mechanism. The required variables and predictors of missingness should be represented appropriately. If either working model is seriously misspecified, compatibility with the stated substantive model does not by itself guarantee valid inference.

In practice, SMC-FCS updates the imputation model and substantive-model parameters as part of the iterative procedure and generates draws from the resulting conditional distributions. A proper multiple-imputation analysis also needs to reflect parameter uncertainty, rather than treating fitted parameters as known constants. The precise sampling implementation depends on the model and is a separate computational issue; the target density above is the central statistical requirement for the nonlinear effect.

Worked Example: Comparing Two Candidate Values

Suppose $X\mid Z=z$ is modeled as $N(1,1)$, and the substantive model is $Y\mid X=x\sim N(0.5x-1.505x^2,1)$. For a hypothetical individual with observed outcome $y=0$, compare the unnormalized SMC-FCS target density at $x=0$ and $x=1$. The example checks the arithmetic of combining the two model components.

At $x=0$, the covariate-model density is proportional to $\exp[-(0-1)^2/2]=\exp(-0.5)$. At $x=1$, it is proportional to $\exp[-(1-1)^2/2]=1$. Thus the ratio of covariate-model contributions, value at 0 divided by value at 1, is $\exp(-0.5)$.

The substantive mean is $0$ at $x=0$, so the outcome residual is $0-0=0$. At $x=1$, the mean is $0.5-1.505=-1.005$, giving residual $0-(-1.005)=1.005$. With unit variance, the ratio of normal likelihood contributions is $$ \frac{\exp(-0^2/2)}{\exp[-(1.005)^2/2]} = \exp(1.010025/2) = \exp(0.5050125) \approx 1.657. $$

Multiplying the component ratios gives the target-density ratio: $$ \frac{f_X(0\mid z)f_Y(0\mid 0)} {f_X(1\mid z)f_Y(0\mid 1)} = \exp(-0.5+0.5050125) = \exp(0.0050125) \approx 1.005. $$

Thus the unnormalized target density at $x=0$ is about 1.005 times its value at $x=1$. The covariate model favors $x=1$, but the outcome likelihood almost exactly offsets that preference. This ratio does not itself give a probability that $X=0$ or $X=1$: $X$ is continuous, and a density value is not a point probability.

Centering and Scaling a Quadratic Effect

Centering or scaling a continuous covariate can make polynomial models easier to fit and interpret, but it must be done consistently. If $u=(x-c)/s$, where $c$ is a fixed center and $s>0$ is a scale, then $x=c+su$. Substituting this expression into a quadratic mean gives

\beta_0+\beta_1x+\beta_2x^2 = (\beta_0+\beta_1c+\beta_2c^2) +s(\beta_1+2c\beta_2)u+s^2\beta_2u^2.

The model remains quadratic, but its coefficients change. If the substantive analysis is defined in terms of the original $x$, the imputation must represent the same mean function, whether it uses $x$ or $u$. The transformation is deterministic: after drawing or imputing $x$, compute $x^2$ from that same value. Do not impute $x$ and $x^2$ as separate unrelated quantities.

Worked Example: Rewriting a Quadratic After Centering

Consider the hypothetical substantive mean $\eta(x)=2+0.4x-0.05x^2$, where $x$ is measured in years. Set $u=(x-10)/2$, so $x=10+2u$. Express the same mean in terms of $u$, and locate its turning point.

The transformed intercept is $2+0.4(10)-0.05(10^2)=2+4-5=1$. The coefficient of $u$ is $2[0.4+2(10)(-0.05)]=2(0.4-1)=-1.2$. The coefficient of $u^2$ is $2^2(-0.05)=-0.2$. Therefore, $$ \eta(x)=1-1.2u-0.2u^2. $$

For the original scale, the derivative is $d\eta/dx=0.4-0.1x$, which equals zero at $x=4$ years. Equivalently, $u=(4-10)/2=-3$; the derivative of the transformed mean, $-1.2-0.4u$, is also zero at $u=-3$. Because the quadratic coefficient is negative, this point is a maximum. The transformed expression and original expression therefore describe the same curve and turning point.

In an imputation, centering can improve numerical behavior, but the covariate and its quadratic term must still correspond to one another. A value $u=-3$, for example, corresponds to $x=4$ years and $x^2=16$ years squared, not to an independently imputed square.

Nonlinear Shape in the Target Density

A quadratic mean can make the outcome likelihood favor separated regions of the covariate scale or favor values near a turning point. The resulting conditional distribution of $X$ need not be normal, even if the covariate model $X\mid Z$ is normal. This is an important reason not to assume that drawing directly from a simple normal covariate model will preserve a nonlinear substantive relationship.

Worked Example: A Quadratic Likelihood Reweights a Covariate Model

For illustration, take a hypothetical covariate model $X\mid Z\sim N(0,1)$ and substantive model $Y\mid X=x\sim N(x^2,1)$. Suppose the observed outcome is $y=0.5$. Compare the unnormalized target-density weights at the grid points $x=0$ and $x=-1,1$. This finite grid is only a calculation aid; it does not turn the continuous conditional density into point probabilities.

Ignoring constants common to all grid points, the target density is proportional to $$ \exp(-x^2/2)\exp[-(0.5-x^2)^2/2]. $$

At $x=0$, the exponent is $0-0.5^2/2=-0.125$, giving weight $\exp(-0.125)\approx0.8825$. At either $x=1$ or $x=-1$, the exponent is $-0.5-(0.5-1)^2/2=-0.5-0.125=-0.625$, giving weight $\exp(-0.625)\approx0.5353$. The weights are equal at $1$ and $-1$ because both the covariate model and the quadratic outcome mean are symmetric around zero.

If these three grid weights are normalized only for a grid-based illustration, their sum is $0.8825+0.5353+0.5353=1.9531$. The normalized grid weights are approximately $0.8825/1.9531=0.4518$ at zero and $0.5353/1.9531=0.2741$ at each of $-1$ and $1$. These are not probabilities for exact values of a continuous variable; they show how the outcome likelihood changes relative density across candidate values.

Assumptions, Diagnostics, and Limitations

For a quadratic SMC-FCS analysis, the central assumptions are substantive and conditional, not merely computational. The substantive model must describe the outcome relationship adequately, including the chosen nonlinear form and its interactions with other covariates if required. The model for $X\mid Z$ must represent the covariate distribution sufficiently well, including relevant predictors and support. The missingness assumption must also be plausible given the information included in the analysis. These conditions are not verified just because the algorithm produces completed data.

Useful diagnostics include checking that completed values of $X$ remain within plausible ranges, inspecting distributions and relationships involving $X$ and $X^2$, and comparing observed and imputed values across relevant covariate or outcome patterns. For a quadratic model, plots of the fitted outcome relationship across the observed covariate range can reveal whether the assumed curvature is plausible. Monitor chains for stability and mixing, and inspect whether imputed values show implausible concentration, extreme tails, or poor coverage in regions where the observed data are sparse.

These checks have limits. A visually stable chain can still target a misspecified model, and agreement in marginal distributions does not establish that the conditional relationship is right. Sparse observations at one end of the covariate range can make the quadratic shape weakly informed; extreme imputations may then be sensitive to model choices. Sensitivity analysis using plausible alternative functional forms or covariate models can reveal this dependence, but cannot eliminate uncertainty caused by inadequate data.

SMC-FCS differs from two common alternatives. Passive imputation constructs a deterministic term such as $X^2$ from an imputed $X$, which preserves the identity but does not by itself ensure that the covariate imputation respects the substantive outcome relationship. Ordinary chained equations can use a conditional model for $X$ that includes $Y$, but a simple chosen form may fail to reproduce the conditional distribution implied by a nonlinear substantive model. SMC-FCS specifically uses the substantive likelihood in the target density. Its additional modeling and computation are not a guarantee against misspecification; they are a way to align the imputation target with the stated analysis model.

Common pitfall: “I imputed $X$, then passively computed $X^2$, so the analysis model’s quadratic effect is automatically respected.” The identity $X^2$ is respected, but the distribution of imputed $X$ may still be incompatible with the outcome likelihood under a quadratic model. SMC-FCS addresses this by reweighting the covariate model with the substantive likelihood; it does not replace the need to check both models and the missingness assumptions.

Practical Workflow

Specify the analysis relationship.
Write the substantive model, including the nonlinear function of the incomplete continuous covariate, before selecting an imputation strategy.

Model the incomplete covariate.
Choose $f_X(x\mid z;\phi)$ to reflect the covariate distribution and relevant predictors, while respecting plausible support.

Form the SMC-FCS target.
Combine $f_X(x\mid z;\phi)$ with $f_Y(y\mid x,z;\psi)$, evaluating the complete nonlinear substantive mean at each candidate $x$.

Generate imputations and check them.
Use an appropriate sampler for the target conditional density, account for parameter uncertainty, and assess convergence, support, and relationships in the completed data.

Check Your Understanding

Answer each question, then compare with the brief solutions.

Write the unnormalized SMC-FCS target density for a missing continuous $X$ under a quadratic normal substantive model.
Why is computing $X^2$ passively after imputing $X$ not, by itself, enough to ensure compatibility with the quadratic analysis model?
If $u=(x-10)/2$ and $x=4$, what are $u$ and $x^2$?
In the third worked example, why are the grid weights at $x=-1$ and $x=1$ equal, and why are they not probabilities that $X$ equals those values?

Solutions. (1) The unnormalized density is $f_X(x\mid z;\phi)f_Y(y\mid x,z;\psi)$, with the quadratic mean substituted into the substantive likelihood. (2) Passive construction preserves the deterministic identity but does not make the distribution of imputed $X$ reflect the substantive likelihood. (3) $u=(4-10)/2=-3$, and $x^2=16$ years squared. (4) Both the standard normal covariate density and the outcome mean $x^2$ are symmetric in $x$, so the weights match; for a continuous variable, the probability of any exact point is zero, and the weights here are density evaluations used for comparison.