Why Auxiliary Variables Matter
Suppose a survey records income, but some participants do not report it. Response may vary with income: people with different incomes might have different probabilities of reporting. If income is missing, that dependence cannot be addressed simply by grouping participants according to their unreported incomes. But the survey may also record an auxiliary variable, such as age group, for nearly everyone. If age is associated with income and response, conditioning on age can offer an explanation for why response rates differ.
The key question is not just whether the auxiliary variable predicts response. It is whether, after conditioning on the observed information, response still depends on the incomplete value. A useful auxiliary variable can make a MAR description more credible when it captures enough of the observed structure linking response and the incomplete variable. It cannot, by itself, show what response would have been for people whose values are missing.
This is a statement about the missingness mechanism, not a claim that \(Z\) causes response or income. It also does not mean that \(Y\) and \(R\) must be unrelated overall. If the distribution of \(Y\) differs across levels of \(Z\), and response rates differ across those same levels, then \(Y\) and \(R\) can be associated marginally even when the conditional MAR statement holds. The conditioning set matters.
How Conditioning Can Change the Description
The earlier material on MAR and MNAR distinguishes dependence on observed values from dependence that remains on unobserved values. An auxiliary variable adds an observed value to the conditioning information. If it captures relevant differences in both the incomplete variable and response, a response pattern that looks related to \(Y\) overall may be explainable by \(Z\).
The following identity makes the point. For a discrete auxiliary variable, the marginal response probability at a given value of \(Y\) is a weighted average of the response probabilities within the \(Z\) groups:
If MAR holds given \(Z\), substitute \(P(R=1\mid Z=z)\) for the first probability in the sum. The weights \(P(Z=z\mid Y=y)\) can still vary with \(y\), so the marginal response probability can vary with \(y\). Thus, observing a marginal association between \(R\) and \(Y\) does not by itself rule out MAR given an auxiliary variable.
Worked Example: Different Marginal Rates, Equal Rates Within Groups
Consider a hypothetical population of 200 people. Income \(Y\) is classified as lower or higher, and age group \(Z\) is classified as younger or older. The table describes a hypothetical complete-data arrangement, including incomes that would be missing from an actual survey.
| Age group | Income | Total | Income observed | Income missing |
|---|---|---|---|---|
| Younger | Lower | 40 | 20 | 20 |
| Younger | Higher | 60 | 30 | 30 |
| Older | Lower | 60 | 12 | 48 |
| Older | Higher | 40 | 8 | 32 |
Within the younger group, the response rate is \(20/40=0.50\) for lower income and \(30/60=0.50\) for higher income. Within the older group, it is \(12/60=0.20\) and \(8/40=0.20\), respectively. Thus, in this hypothetical arrangement, response depends on age group but not on income within age group: the MAR equality holds given \(Z\).
Now pool across age groups. For lower income, the response rate is \((20+12)/(40+60)=32/100=0.32\). For higher income, it is \((30+8)/(60+40)=38/100=0.38\). The marginal rates differ, even though the conditional rates do not. The different mix of age groups across the two income categories explains this marginal difference.
Because actual missing incomes are not observed, a real survey would not generally provide the full table needed to check these conditional rates directly. The example shows how an auxiliary variable can make MAR a coherent description of a mechanism; it is not a recipe for verifying that description from incomplete data.
What Makes an Auxiliary Variable Useful?
An auxiliary variable is most informative for MAR plausibility when it is associated with the incomplete value and also helps explain differences in response. These associations have different roles. If \(Z\) is associated with \(Y\), it can distinguish groups with different distributions of the incomplete value. If response rates also differ across those groups, conditioning on \(Z\) may account for response differences that would otherwise look like dependence on \(Y\).
Neither association alone is enough to establish the conditional independence in the definition. The next example separates response prediction from accounting for the relationship between response and the incomplete value.
Worked Example: Predicting Response Is Not Enough
In a hypothetical survey, \(Z\) is a recorded region, and \(Y\) is an income category. Each region contains 100 people, with 50 in each income category.
| Region | Income | Total | Income observed | Income missing |
|---|---|---|---|---|
| A | Lower | 50 | 40 | 10 |
| A | Higher | 50 | 20 | 30 |
| B | Lower | 50 | 20 | 30 |
| B | Higher | 50 | 10 | 40 |
Region predicts response: the overall response rate is \((40+20)/100=0.60\) in A and \((20+10)/100=0.30\) in B. But within region A, the response rates are \(40/50=0.80\) for lower income and \(20/50=0.40\) for higher income. Within B, they are \(20/50=0.40\) and \(10/50=0.20\). In both regions, response still differs by income.
Income has the same distribution in regions A and B, so region does not capture differences in income composition here. More importantly, the conditional response rates are unequal across income categories. Therefore, this table does not support MAR given region, even though region is a strong predictor of response. Response prediction is useful evidence about response patterns, but it is not the MAR condition.
Checking the Limits of a Proxy
An auxiliary variable may be related to both response and the incomplete variable yet still leave a residual relationship between response and the incomplete value within its groups. This can happen when the auxiliary measure is coarse, noisy, or omits relevant distinctions. Grouping by broad age bands, for example, may not capture the response differences associated with age within each band.
Worked Example: A Related but Incomplete Proxy
Suppose a hypothetical survey records a broad age group \(Z\) and asks for income \(Y\). The table again describes a complete-data arrangement to make the mechanism transparent.
| Age group | Income | Total | Income observed | Income missing |
|---|---|---|---|---|
| Younger | Lower | 80 | 48 | 32 |
| Younger | Higher | 20 | 8 | 12 |
| Older | Lower | 20 | 6 | 14 |
| Older | Higher | 80 | 16 | 64 |
Age group is strongly related to income: the higher-income share is \(20/100=0.20\) among younger people and \(80/100=0.80\) among older people. It is also related to response: the response rate is \((48+8)/100=0.56\) among younger people and \((6+16)/100=0.22\) among older people.
However, response still varies by income within each age group. Among younger people, the rates are \(48/80=0.60\) for lower income and \(8/20=0.40\) for higher income. Among older people, they are \(6/20=0.30\) and \(16/80=0.20\). Since these rates differ within both groups, the exact MAR equality does not hold given this \(Z\). The proxy is informative, but it does not fully account for the response pattern in this hypothetical mechanism.
Assumptions, Diagnostics, and Comparisons
To use \(Z\) as an observed conditioning variable, its values must be available for the people whose \(Y\) values are missing. The MAR statement is also specific to the conditioning set: a variable can make MAR more plausible relative to one set of observed information without guaranteeing that all relevant information has been included. In applications, researchers usually have to judge this plausibility using how the data were collected, subject-matter knowledge, and observed associations.
Adequate overlap matters for learning about groups: if some combinations of \(Z\) and observed \(Y\) are rare or absent, comparisons within those groups will be unstable or unavailable. A sparse table does not prove that MAR fails, but it limits what the data can reveal about response patterns. A useful diagnostic is to compare response rates across observed \(Z\) groups and examine whether the auxiliary variable also distinguishes values or distributions of \(Y\) among respondents. These checks can identify weak proxies or limited overlap; they cannot directly reveal \(Y\) for nonrespondents.
The alternatives differ in what they assume about response. MCAR requires response to be independent of both \(Y\) and observed variables such as \(Z\). MAR allows response to vary with observed \(Z\), while requiring no residual dependence on \(Y\) after conditioning on \(Z\). MNAR describes dependence on the incomplete value that remains after conditioning on the observed information. Adding a useful auxiliary variable may make MAR more plausible than MCAR, but it does not rule out MNAR.
The practical value of a well-chosen auxiliary variable is therefore not that it certifies a mechanism. It provides observed information that can explain response variation and may support a defensible MAR description. The strength of that description depends on whether the variable is measured for nonrespondents, captures relevant differences, and leaves little reason to expect residual dependence on the missing value.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Write the MAR condition for an incomplete variable \(Y\), response indicator \(R\), and fully observed auxiliary variable \(Z\).
- In a hypothetical study, response rates are 0.70 and 0.30 in two \(Z\) groups. Does this alone establish MAR? Why or why not?
- Suppose the response rates for lower and higher \(Y\) are 0.45 and 0.45 in one \(Z\) group, and 0.25 and 0.25 in another. What does this pattern indicate in the hypothetical complete data?
- Why can marginal response rates differ by \(Y\) even when the MAR condition holds given \(Z\)?
Solutions. (1) The condition is \(P(R=1\mid Y,Z)=P(R=1\mid Z)\). (2) No. These rates show that \(Z\) predicts response, but MAR additionally requires that, within each \(Z\) group, response probability not depend on \(Y\). (3) The equal rates within each group are consistent with MAR given \(Z\) in this hypothetical complete data, though they do not prove that a real incomplete-data mechanism is MAR. (4) The distribution of \(Z\) can differ across values of \(Y\). Averaging group-specific response rates with different \(Z\) weights can therefore produce different marginal rates.