Modeling Response Indicators With Observed Predictors

Why Model a Response Indicator?

A missingness rate summarizes how often a value is unobserved, but it can hide important differences between groups. For example, an overall response rate of 70% does not say whether response is similar across age groups or whether one group responds much less often. Modeling a response indicator makes such patterns explicit and gives a common scale for describing associations with observed predictors.

For one variable of interest, define an indicator for each eligible unit. The analysis requires predictors that are observed for the units being modeled; otherwise, the response model itself may exclude units whose predictor values are unavailable. The model describes response patterns in the data. Its fitted associations do not, by themselves, explain the cause of missingness or establish a missingness mechanism.

Definition: For unit \(i\), let \(R_i=1\) if the variable of interest is observed and \(R_i=0\) if it is missing. A response-indicator model specifies how \(P(R_i=1\mid X_i)\) varies with observed predictors \(X_i\). Here, “response” means that the particular value is observed; it need not mean that a person responded to an entire survey.

Logistic Regression for a Binary Indicator

A standard model for a binary response indicator is logistic regression. With one predictor \(X_i\), the model is

R_i\mid X_i \sim \operatorname{Bernoulli}(p_i), \qquad \operatorname{logit}(p_i)=\log\left(\frac{p_i}{1-p_i}\right)=\beta_0+\beta_1X_i, \qquad p_i=P(R_i=1\mid X_i).

The model states that the conditional probability of an observed value follows the logistic curve associated with a linear predictor. The Bernoulli distribution assigns probability \(p_i\) to \(R_i=1\) and \(1-p_i\) to \(R_i=0\). For several observed predictors, the linear predictor can be extended to \(\beta_0+\beta_1X_{i1}+\cdots+\beta_qX_{iq}\). Categorical predictors are represented by indicator variables, with a stated reference category.

Before interpreting a fitted model, check the conditions it relies on. The response indicator must be coded consistently, and the predictors used in the model must be available for every unit included. The log odds should be adequately represented by the chosen predictor terms; a straight-line term for a continuous predictor may be inadequate if the relationship is curved. Standard model-based uncertainty calculations also rely on the assumed independence structure. If units are clustered or repeatedly measured, ordinary independent-observation standard errors may not be appropriate. Finally, the data must contain enough information to estimate the coefficients: empty or extremely sparse predictor groups and perfect prediction can make estimates unstable or prevent finite maximum-likelihood estimates.

Interpretation: In a logistic model, \(\beta_1\) is the change in the log odds of observing the value for a one-unit increase in \(X\), holding other included predictors fixed. The corresponding odds ratio is \(\exp(\beta_1)\). An odds ratio describes a ratio of odds, not a ratio of probabilities.

From Coefficients to Odds and Probabilities

A coefficient is often easiest to interpret after exponentiation. To see why, compare two predictor values \(x\) and \(x+1\) in a one-predictor model. Their fitted log odds differ by

\left[\beta_0+\beta_1(x+1)\right]-\left[\beta_0+\beta_1x\right]=\beta_1.

Exponentiating both sides turns the difference in log odds into a ratio of odds:

\frac{\text{odds}(R=1\mid X=x+1)}{\text{odds}(R=1\mid X=x)} = \exp\!\left(\operatorname{logit}(p_{x+1})-\operatorname{logit}(p_x)\right) = \exp(\beta_1).

To obtain a probability from a fitted linear predictor \(\eta\), use the inverse-logit transformation:

p=\frac{\exp(\eta)}{1+\exp(\eta)} =\frac{1}{1+\exp(-\eta)}.

The probability scale is useful for describing response rates at particular predictor values. It is not a linear rescaling of the coefficient: the same coefficient can correspond to different changes in probability at different starting probabilities.

Worked Example: Comparing Response Odds Across Two Groups

Hypothetical data: In a group of 60 eligible people, the value of a laboratory variable is observed for 18 of 30 younger people and 12 of 30 older people. Suppose a logistic model uses \(X_i=1\) for younger people and \(X_i=0\) for older people. This two-group model allows a separate response probability for each group, so the fitted group probabilities equal the observed group proportions.

The estimated probabilities are \(\hat{p}_1=18/30=0.600\) for younger people and \(\hat{p}_0=12/30=0.400\) for older people. The model intercept is the older group’s log odds:

\hat{\beta}_0 = \log\left(\frac{0.400}{1-0.400}\right) = \log\left(\frac{0.400}{0.600}\right) = -0.4055.

The coefficient for younger versus older is the difference in the groups’ log odds:

\hat{\beta}_1 = \log\left(\frac{0.600}{0.400}\right) - \log\left(\frac{0.400}{0.600}\right) = 0.4055-(-0.4055) = 0.8109.

Exponentiating gives an estimated odds ratio of \(\exp(0.8109)=2.250\). In this hypothetical sample, the observed-value odds are 2.25 times as high for younger people as for older people. This is a comparison of odds, not a claim that younger people are 2.25 times as likely to have the value observed: the corresponding observed proportions are 0.600 and 0.400.

Worked Example: Converting a Continuous-Predictor Model to Probabilities

Hypothetical fitted model: Suppose the fitted response model for a variable uses age in years and is \(\operatorname{logit}(\hat{p}_i)=-1+0.040\,\text{age}_i\). Treat the coefficients as given fitted estimates for this illustration. The model assumes a linear age association on the log-odds scale over the ages being considered.

For a 40-year-old, the fitted linear predictor is \(-1+(0.040)(40)=0.600\). Therefore,

\hat{p}(40) = \frac{1}{1+\exp(-0.600)} = 0.6457.

For a 60-year-old, it is \(-1+(0.040)(60)=1.400\), giving

\hat{p}(60) = \frac{1}{1+\exp(-1.400)} = 0.8022.

The estimated odds ratio for a 20-year increase is \(\exp((0.040)(20))=\exp(0.800)=2.226\). Under this model, the fitted odds that the value is observed are about 2.23 times as high at age 60 as at age 40. The fitted probabilities differ by about 0.157, or 15.7 percentage points. The odds ratio and probability difference answer different questions and should not be substituted for one another.

Several Predictors and Conditional Associations

A model with multiple predictors can describe how response varies with one observed covariate while accounting for other included covariates. For example, the coefficient for age in a model that also includes education describes the model’s age association at a fixed education value. This is a conditional, model-based association; including more predictors does not automatically make it causal or remove every source of distortion.

Worked Example: Interpreting an Adjusted Group Association

Hypothetical fitted model: Let \(X_i=1\) for rural residence and \(0\) for urban residence, and let \(Z_i=1\) for higher education and \(0\) otherwise. Suppose a fitted model is \(\operatorname{logit}(\hat{p}_i)=-0.200-0.600X_i+0.800Z_i\). The model assumes these additive terms adequately describe the log odds for the combinations being compared.

At lower education, \(Z=0\). The fitted rural probability has linear predictor \(-0.200-0.600=-0.800\), so

\hat{p}(\text{rural},Z=0) = \frac{1}{1+\exp(0.800)} = 0.3100.

The fitted urban probability at the same education level is \(\hat{p}(\text{urban},Z=0)=1/[1+\exp(0.200)]=0.4502\). The model’s rural-versus-urban odds ratio at fixed education is \(\exp(-0.600)=0.5488\). Thus, conditional on education in this fitted model, rural residence is associated with lower odds of observing the value. This coefficient is not a probability difference, and it is not evidence that residence causes a response difference.

Diagnostics and Alternative Summaries

A fitted response model is useful only to the extent that its structure is reasonable and its data support the estimates. Begin by comparing observed response proportions with fitted probabilities across meaningful groups or ranges of a continuous predictor. Large systematic discrepancies can point to a missing nonlinear term, an omitted interaction, or another inadequacy in the chosen mean model. For a continuous predictor, plotting grouped observed proportions against fitted probabilities can reveal curvature that a single linear log-odds term misses. Calibration checks are descriptive; a close match does not establish that the model captures every relevant pattern.

Also inspect the data for sparse combinations, extreme fitted probabilities, and influential observations. If a predictor or combination of predictors perfectly separates observed from missing values, ordinary logistic maximum likelihood may produce a coefficient that grows without bound rather than a stable finite estimate. A warning, very large standard error, or fitted probabilities extremely close to zero or one should prompt investigation rather than confident interpretation. Dependence from repeated observations or clusters needs methods that reflect that structure. If a predictor is itself missing for some eligible units, a model restricted to units with complete predictors may not describe response in the full eligible group; report the restriction and consider how it affects the target description.

Simple response proportions or a two-way table are often the clearest starting point for a categorical predictor. They show the group denominators and make the observed response-rate difference transparent. A logistic model is a compact extension that can combine predictors and express associations as odds ratios, but its form and assumptions need checking. A linear probability model instead regresses the zero-one indicator directly on predictors; its coefficients are probability differences, but fitted values can fall below zero or above one, and ordinary variance assumptions are often unsuitable for a binary outcome. The choice depends on the descriptive aim and modeling assumptions, not on a rule that one method is always preferable.

Common Mistakes and Limits

Reversing the indicator without changing the interpretation. If \(R=1\) means observed, a positive coefficient describes higher odds of observation. If instead \(R=1\) means missing, the interpretation reverses. State the coding before fitting or interpreting the model.
Calling an odds ratio a probability ratio. An odds ratio of 2 does not mean the response probability doubles. Calculate fitted probabilities when probability-scale comparisons are needed.
Reading an association as a cause. A fitted coefficient describes an association under the specified model. It does not show that changing a predictor would change response.
Treating a fitted response pattern as a complete explanation of missingness. The model uses observed predictors and the response indicator. Its associations alone do not establish why values are missing or settle which missingness mechanism applies.
Ignoring the model form and data support. A reported coefficient can be misleading if the log-odds form is inadequate, important groups are sparse, predictors are unavailable for many units, or observations are dependent in a way the analysis ignores.

The practical role of this analysis is therefore bounded but valuable: it quantifies how observed values are distributed across observed covariates and helps identify response patterns that a single overall rate conceals. Report the indicator coding, the analysis sample, the predictors and model form, and interpretable fitted comparisons. Treat the results as descriptions of response associations, not as a substitute for assumptions about unobserved values.

Check Your Understanding

Answer each question, then compare with the brief solutions.

If a model codes \(R=1\) as observed, what does a positive coefficient for a binary predictor indicate about the fitted response odds?
A fitted coefficient is \(0.300\) per additional year of age. Calculate the odds ratio for a five-year increase and interpret it.
For a fitted linear predictor \(\eta=-0.400\), calculate the fitted probability of observation.
Why is it incorrect to interpret an odds ratio of 2 as meaning that a response probability is twice as large?
Name two warning signs that an ordinary logistic response model may be unreliable.

Solutions. (1) At predictor value 1, the fitted odds of observation are higher than at value 0, with other included predictors held fixed; the odds ratio is the exponential of the coefficient. (2) \(\exp((0.300)(5))=\exp(1.500)=4.482\). The fitted odds of observation are about 4.48 times as high over a five-year increase, under the model; this is not a probability ratio. (3) \(\hat{p}=1/[1+\exp(0.400)]=0.4013\), or about 40.1%. (4) Odds and probabilities are different quantities; the probability depends on the starting odds, so an odds ratio does not specify a probability ratio. (5) Examples include perfect prediction, sparse predictor groups, very large standard errors, extreme fitted probabilities, and systematic mismatch between observed group rates and fitted probabilities.