Nonresponse Instrument in Survey Research

• Nonresponse instrument is defined as a family of strategies—including latent proxies, shadow variables, and instrumental variables—to mitigate bias from survey nonresponse.
• It employs models, such as two-parameter logistic and calibration techniques, to estimate response propensities and adjust for missing data.
• These methods are applied for operational improvements in weighting, follow-up design, and robust bias assessment in survey sampling.

Nonresponse instrument denotes a family of methodological devices used to address unit or item nonresponse, especially when missingness is nonignorable. In the literature represented here, the term covers several distinct objects: an internally generated latent proxy for response propensity, a shadow variable or instrumental variable used to identify missing-not-at-random models through exclusion restrictions, and broader operational tools for weighting, calibration, follow-up design, or robustness assessment. The term is therefore not synonymous with a single estimation strategy, and some papers are explicit that what they construct is not an instrumental variable in the econometric sense (Matei et al., 2012, Zhao, 18 Sep 2025).

1. Scope, terminology, and conceptual boundaries

Survey sampling conventionally distinguishes unit nonresponse from item nonresponse. With finite population $U=\{1,\dots,N\}$, sample $s\subset U$, and respondent set $r\subseteq s$, unit response is represented by

$$R_k= \begin{cases} 1,& k\in r,\ 0,& k\in \bar r=s\setminus r. \end{cases}$$

For a study variable $y_j$, item response among respondents is represented by

$r_j=\{k \text{ answers } y_j \mid k\in r\}.$

One paper emphasizes the bridge between the two by quoting the view that unit nonresponse is “just an extreme form of item nonresponse” (Matei et al., 2012).

Across the literature, a nonresponse instrument may mean different things. In one line of work, it is a latent adjustment variable inferred from within-survey item-response indicators and inserted into a response-propensity model. In another, it is a shadow variable $Z$ in a decomposition $X=(U,Z)$, where $Z$ is excluded from the nonresponse propensity but remains informative about $Y$. In a third, it is an instrumental variable that predicts response but is excluded from the outcome model. A separate econometric usage concerns individuals who do not respond to an instrument in treatment selection; that object belongs to the marginal treatment effect literature and is conceptually different from survey nonresponse instruments (Chen et al., 16 Sep 2025, Sun et al., 2023, Martínez-Iriarte et al., 2022).

The common motivation is bias. If the response probability of unit $s\subset U$0 is

$s\subset U$1

nonresponse bias depends on the association between $s\subset U$2 and $s\subset U$3. Under NMAR, that association arises because the outcome itself, or latent factors tied to it, affects participation (Matei et al., 2012).

2. Internal-survey latent proxies for response propensity

A prominent survey-sampling interpretation of nonresponse instrument is an internally constructed latent proxy for willingness to respond. In this formulation, the latent variable $s\subset U$4 is interpreted as the unit’s “will to respond to the survey” or “tendency to respond,” inferred from binary item-response indicators

$s\subset U$5

The corresponding vector is

$s\subset U$6

The paper assumes that $s\subset U$7 are manifestations of a single latent continuous trait $s\subset U$8, estimated by a latent trait model, specifically the two-parameter logistic model

$s\subset U$9

The Rasch model is the special case with common discrimination. The framework relies on three standard assumptions: conditional independence,

$r\subseteq s$0

monotonicity, and unidimensionality (Matei et al., 2012).

Estimation proceeds by marginal maximum likelihood, typically under $r\subseteq s$1, followed by empirical Bayes estimation of $r\subseteq s$2. A practically distinctive step concerns unit nonrespondents, who have no observed item responses. The proposed solution sets

$r\subseteq s$3

adds a single phantom respondent $r\subseteq s$4 with the all-zero response pattern, estimates the latent trait model on $r\subseteq s$5, computes $r\subseteq s$6, and assigns

$r\subseteq s$7

This is the paper’s key construction of an internal nonresponse proxy (Matei et al., 2012).

Once $r\subseteq s$8 is available, unit response propensity is modeled by

$r\subseteq s$9

and item response probabilities $$R_k= \begin{cases} 1,& k\in r,\ 0,& k\in \bar r=s\setminus r. \end{cases}$$0 are estimated from the latent trait model. The resulting estimator for the total of variable $$R_k= \begin{cases} 1,& k\in r,\ 0,& k\in \bar r=s\setminus r. \end{cases}$$1 is

$$R_k= \begin{cases} 1,& k\in r,\ 0,& k\in \bar r=s\setminus r. \end{cases}$$2

The paper is explicit that this is not an instrumental-variable strategy in the exclusion-restriction sense: the latent score is intended to be related to the variable of interest, because that relationship is what makes it useful for reducing nonresponse bias (Matei et al., 2012).

The simulation evidence is substantial. In a binary-outcome setting based on British Social Attitudes abortion items, the naive estimator had around $$R_k= \begin{cases} 1,& k\in r,\ 0,& k\in \bar r=s\setminus r. \end{cases}$$3 relative bias, while the proposed estimator reduced that to around $$R_k= \begin{cases} 1,& k\in r,\ 0,& k\in \bar r=s\setminus r. \end{cases}$$4. In a second simulation with six continuous variables, the naive estimator had roughly $$R_k= \begin{cases} 1,& k\in r,\ 0,& k\in \bar r=s\setminus r. \end{cases}$$5 to $$R_k= \begin{cases} 1,& k\in r,\ 0,& k\in \bar r=s\setminus r. \end{cases}$$6 relative bias, while the proposed estimator reduced this to roughly $$R_k= \begin{cases} 1,& k\in r,\ 0,& k\in \bar r=s\setminus r. \end{cases}$$7 to $$R_k= \begin{cases} 1,& k\in r,\ 0,& k\in \bar r=s\setminus r. \end{cases}$$8. The paper also reports Cronbach’s alpha, two-way and three-way margin residuals $$R_k= \begin{cases} 1,& k\in r,\ 0,& k\in \bar r=s\setminus r. \end{cases}$$9, point-measure correlations, and discussion of PCA of residuals for unidimensionality, because if the selected items do not load on a common response-propensity dimension, $y_j$0 is unlikely to work as an adjustment proxy (Matei et al., 2012).

3. Shadow variables excluded from the nonresponse propensity

A second and influential meaning of nonresponse instrument is the shadow variable. Here the covariates are partitioned as

$y_j$1

and $y_j$2 is a nonresponse instrument if it satisfies

$y_j$3

while

$y_j$4

Thus $y_j$5 is excluded from the nonresponse mechanism once $y_j$6 and $y_j$7 are conditioned on, but remains useful for the outcome model. The literature is explicit that this differs from an ordinary covariate, which may enter both $y_j$8 and $y_j$9 (Chen et al., 16 Sep 2025, Zhao, 18 Sep 2025).

Within this framework, one semiparametric path specifies a parametric data model $r_j=\{k \text{ answers } y_j \mid k\in r\}.$0 and leaves the propensity nonparametric. The key identity is

$r_j=\{k \text{ answers } y_j \mid k\in r\}.$1

which yields a respondent-based pseudo-likelihood. A second path specifies a parametric or semiparametric propensity model and leaves $r_j=\{k \text{ answers } y_j \mid k\in r\}.$2 nonparametric. In the review, examples include

$r_j=\{k \text{ answers } y_j \mid k\in r\}.$3

and the exponential tilting model

$r_j=\{k \text{ answers } y_j \mid k\in r\}.$4

The review also summarizes a doubly robust framework in which consistency for $r_j=\{k \text{ answers } y_j \mid k\in r\}.$5 requires a correct log-odds ratio model and one of two nuisance models (Zhao, 18 Sep 2025).

Because the shadow variable is often not known in advance, one paper studies instrument, variable, and model selection under nonignorable nonresponse. For a candidate $r_j=\{k \text{ answers } y_j \mid k\in r\}.$6, it compares two estimators of

$r_j=\{k \text{ answers } y_j \mid k\in r\}.$7

through the validation criterion

$r_j=\{k \text{ answers } y_j \mid k\in r\}.$8

The selected set is then refined by nonparametric variable selection using

$r_j=\{k \text{ answers } y_j \mid k\in r\}.$9

Under regularity conditions, the paper states

$Z$0

and the final pseudo-likelihood estimator of $Z$1 is consistent and asymptotically normal (Chen et al., 16 Sep 2025).

Categorical shadow variables create a separate identification problem because completeness can fail even in simple models. One paper therefore replaces completeness with a verifiable sufficient condition based on the respondents’ outcome model. Under

$Z$2

and the requirement that for each $Z$3 there exist $Z$4 such that

$Z$5

the parameter $Z$6 is identifiable. In the fully categorical case, the paper shows that completeness is necessary and sufficient for identifiability (Beppu et al., 2023).

4. Response-predicting instruments excluded from the outcome model

A distinct IV tradition defines the nonresponse instrument in the opposite way: $Z$7 affects response behavior but is excluded from the outcome model. In this line, the full data are

$Z$8

the observed data are

$Z$9

and the target is the population mean

$X=(U,Z)$0

The central IV conditions are

$X=(U,Z)$1

To encode nonignorability, the response mechanism is parameterized through an extended propensity score

$X=(U,Z)$2

where

$X=(U,Z)$3

Under the preferred factorization,

$X=(U,Z)$4

the nuisance components $X=(U,Z)$5 and $X=(U,Z)$6 are variation independent. The resulting estimator is doubly robust in the paper’s specific sense: consistency holds if $X=(U,Z)$7 is correct and either $X=(U,Z)$8 or $X=(U,Z)$9 is correct (Sun et al., 2023).

The empirical illustration uses HIV testing refusal in Mochudi, Botswana, with interviewer experience as the instrumental variable. The complete-case estimate of HIV prevalence is $Z$0 $Z$1, the MAR/IPW estimate is $Z$2 $Z$3, and the proposed estimator is $Z$4 $Z$5. The estimated selection-bias parameter is

$Z$6

suggesting that HIV-positive individuals may have been less likely to participate in testing, though the estimate is not statistically significant (Sun et al., 2023).

A more recent formulation introduces a multiplicative instrumental variable model for MNAR outcomes. The observed data are

$Z$7

with target functional $Z$8 defined through

$Z$9

The missing-case quantity is

$Y$0

The assumptions are

$Y$1

and the multiplicative selection model

$Y$2

with $Y$3 for $Y$4. For binary $Y$5, the identification formula is a single-arm Wald ratio: $Y$6 where

$Y$7

The paper states that under the assumptions, any regular statistical functional of the missing outcome is nonparametrically identified, and it develops semiparametric multiply robust IV estimators (Zhang et al., 26 Sep 2025).

5. Model-assisted, calibration, and benchmark-based instruments

In a broader survey-methodological sense, nonresponse instrument can denote a practical inferential device that combines response weighting with prediction or calibration rather than an exclusion-restriction variable. One paper studies model-assisted estimators under MAR by treating nonresponse as a second phase of sampling. With a working model

$Y$8

the practical estimator is

$Y$9

which reduces to the standard nonresponse-adjusted Horvitz–Thompson estimator

$s\subset U$00

when the prediction term is dropped. Response probabilities are mainly estimated by calibration through

$s\subset U$01

In the GREG case, calibration forces the auxiliary-total discrepancy term to vanish, so the troublesome remainder disappears asymptotically (Eustache et al., 2022).

The simulation hierarchy is explicit. When both the response model and the working model fit well, the proposed estimator and $s\subset U$02 have bias near zero, but the proposed estimator can attain very small standard deviations; in scenario 1 with GREG, $s\subset U$03 and $s\subset U$04, versus $s\subset U$05 for HT and $s\subset U$06 for imputation. When the response model is wrong but the working model is good, the proposed estimator remains best among feasible estimators; with GREG in scenario 2 it still has $s\subset U$07 and $s\subset U$08, while NWA has $s\subset U$09 and $s\subset U$10 (Eustache et al., 2022).

External benchmarks can play a closely related role. In multiple imputation for voter turnout, known voter turnout totals and demographic margins are incorporated directly into the missing-data model. The paper defines a hybrid MD-AM model comprising a pattern-mixture model for unit nonresponse and selection models for item nonresponse. In the North Carolina CPS application, the known turnout rate is about $s\subset U$11, and the intercept matching algorithm draws

$s\subset U$12

and adjusts the unit-nonresponse effect on voting so that the imputed completed-data total matches the benchmark. The paper is explicit that these margins are not instruments in the classical IV sense; they are calibration targets and identification restrictions embedded in the imputation mechanism. The final estimated overall turnout is about $s\subset U$13, close to the auxiliary target by construction (Tang et al., 2022).

6. Sensitivity analysis, follow-up design, and non-instrument alternatives

When no credible nonresponse instrument is available, several papers replace point identification with sensitivity analysis or decision-theoretic tools. One approach concerns whether to stop or continue data collection under possibly nonignorable unit nonresponse for multivariate continuous variables. The method fits a finite mixture of multivariate normal distributions to respondents,

$s\subset U$14

and then generates nonrespondent imputations by replacing $s\subset U$15 with scenario-specific $s\subset U$16. Follow-up sample sizes are compared using utility measures

$s\subset U$17

and a cost function

$s\subset U$18

In the Census of Manufactures application, moving from no follow-up to following up on $s\subset U$19 or $s\subset U$20 of nonrespondents sharply reduces the error measures, while gains beyond about $s\subset U$21 appear limited relative to added cost (Paiva et al., 2015).

Another line develops worst-case resistance testing as a nonresponse-bias diagnostic. WCRT asks how many nonrespondents, with what effect size, would be needed to reverse a study’s conclusion. For correlations, it constructs “n-curves” plotting the required nonresponse count against the assumed nonresponse effect size. In the empirical example, the observed correlation between shopping experience and satisfaction is about $s\subset U$22; at $s\subset U$23, it would take $s\subset U$24 nonrespondents if $s\subset U$25, $s\subset U$26 if $s\subset U$27, and $s\subset U$28 if $s\subset U$29 to negate significance. For the weaker correlation between intentions and enjoyment, about $s\subset U$30, the corresponding numbers are $s\subset U$31, $s\subset U$32, and $s\subset U$33. The paper is explicit that WCRT is not a questionnaire or psychometric scale, but a statistical diagnostic / sensitivity-analysis tool (France et al., 2023).

A more general alternative dispenses with instruments altogether and treats nonresponse as a partial-identification problem. In panel-data stochastic-dominance testing, the observed data are $s\subset U$34 with monotone baseline nonresponse, and inference is based on sharp upper and lower bounds for dominance functionals. The test uses pseudo-empirical likelihood and a design-effect-adjusted likelihood-ratio statistic compared to a $s\subset U$35 critical value. The paper is explicit that it does not offer a classical instrumental variable for nonresponse; instead, it provides an assumption-indexed bounding and testing device for settings in which no credible exclusion variable exists (Tabri et al., 2024).

The cumulative literature therefore supports a precise terminological conclusion. In one usage, a nonresponse instrument is a latent internal proxy for willingness to respond; in another, it is a shadow variable excluded from the nonresponse mechanism but informative for the outcome; in another, it is a response-predicting instrumental variable excluded from the outcome model; and in broader survey practice it may refer to a methodological instrument for weighting, calibration, follow-up design, or robustness assessment. Treating these objects as interchangeable obscures the identifying assumptions that each requires (Matei et al., 2012, Zhao, 18 Sep 2025, Sun et al., 2023, Paiva et al., 2015).