Handling Missing Values in Longitudinal Panel Data With Multiple Imputation

• Author: Rebekah Young,David R. Johnson
• Date: 01 February 2015
• DOI: http://doi.org/10.1111/jomf.12144
• Published date: 01 February 2015

R Y University of Washington

D R. J The Pennsylvania State University∗

Handling Missing Values in Longitudinal Panel Data

With Multiple Imputation

This article offers an applied reviewof key issues

and methods for the analysis of longitudinal

panel data in the presence of missing values. The

authors consider the unique challenges associ-

ated with attrition (survey dropout), incomplete

repeated measures, and unknown observations

of time. Using simulated data based on 4 waves

of the Marital Instability Over the Life Course

Study (n=2,034), they applied a xed effect

regression model and an event-history analysis

with time-varying covariates. They then com-

pared resultsfor analyses with nonimputed miss-

ing data and with imputed data both in long

and in wide structures. Imputation produced

improved estimates in the event-history analysis

but only modest improvements in the estimates

and standard errors of the xed effects analysis.

Factors responsible for differences in the value

of imputation are examined, and recommenda-

tions for handling missing values in panel data

are presented.

The use of longitudinal panel (prospective)

survey data is common in the area of family

Department of Biostatistics, Collaborative Health Studies

Coordinating Center, University of Washington,Box

354922, Seattle, WA98195.

∗Department of Sociology, The Pennsylvania State

University,211 Oswald Tower, University Park, PA16802

(drj10@psu.edu).

Key Words: event history analysis, xed effects, longitudinal

data, missing data, multiple imputation, panel data.

research. From 2010 to 2014, approximately

287 quantitative and qualitative research arti-

cles (excluding theory development, research

reviews, comments, rejoinders, and method-

ological innovation articles) were published in

the Journal of Marriage and Family (JMF). Of

these, 176 (61%) analyzed longitudinal data.

Data on the same individuals or families at

multiple points in time provide for stronger

inferences about change processes and allow

for more control of unmeasured differences

between individuals that can bias study nd-

ings (Johnson, 1995, 2005). What tempers

these advantages is the large amount of miss-

ing data found in many longitudinal studies.

Nearly all of the JMF articles explicitly men-

tioned the presence of missing values and study

dropout—suggestive of the widespread concern

with missing data in panel studies.

Few guidelines for the analysis of longitudi-

nal panel data in the presence of missing values

are accessible to family researchers. Moreover,

no clear appraisals of the consequences of

different ways of handling missing data are

readily offered. Existing guidelines tend to be

directed toward statisticians or focus on types of

longitudinal data rarely found in the family lit-

erature, such as randomized clinical trials (e.g.,

Daniels & Hogan, 2008; Enders, 2011; Hedeker

& Gibbons, 2006; National Research Council,

2010) or data sets with few cases but many

waves, such as cross-national time-series stud-

ies (e.g., Honaker & King, 2010). Methods for

handling missing values have been addressed

in the family literature (e.g., Acock, 2005;

Journal of Marriage and Family 77 (February 2015): 277–294 277

DOI:10.1111/jomf.12144

278 Journal of Marriage and Family

Johnson & Young, 2011; Young & Johnson,

2013), but these resources focus primarily on

cross-sectional data. Although much of what we

know about the approaches to handling missing

values in cross-sectional situations applies to

longitudinal panel data, panel data have char-

acteristics that complicate the application of

techniques such as multiple imputation (MI).

Such complications, along with a lack of acces-

sible guides to help address these issues, may

be contributing to the limited use of modern

methods like maximum likelihood (ML) or MI

among the many studies in the area of family

that use longitudinal data (Jelicic, Phelps, &

Lerner, 2009).

In this article, we review standard approaches

to handling missing data in longitudinal panel

studies, apply several techniques to a simu-

lations study based on an empirical family

research problem using a multiwave panel data

set, and assess how different strategies have

consequences for the research ndings. Our

focus is on missing values in panel data sets

with large numbers of respondents but small

numbers of survey waves administered at xed

intervals—typical conditions for data sets found

in much family research. Missing data MI

strategies with xed effect, pooled time-series

models and event-history (Cox proportional

hazard) models are examined. Our review of the

methods used in 176 JMF articles suggests that

the most common models for analyzing lon-

gitudinal data were event history (19%), xed

effects (18%, or 19% including change scores),

and mixed effect or multilevel (17%, or 22%

including growth curve), followed by linear

regression (16%), logistic regression (15%), and

structural equation models (10%, or 15% includ-

ing growth curve and latent class analysis). Less

common methods for analyzing longitudinal

data included multinomial regression (5%) and

qualitative analysis (2%). (Note that percentages

sum to more than 100% because many articles

used more than one method.)

B

Longitudinal panel studies have several fea-

tures that complicate the techniques commonly

applied when handling missing data. Unlike

cross-sectional data sets, longitudinal data

sets have both within-wave and whole-wave

missingness. Longitudinal data analysis meth-

ods require a particular data structure (long

vs. wide) that creates issues when handling

missing data (Lloyd, Obradovic, Carpiano, &

Motti-Stefanidi, 2013). Other complications

of missing values in longitudinal data include

repeated measures; time-to-event models; non-

random study dropout; and statistical procedures

that routinely handle some, but not all, sources

of missing data.

Two Sources of Missing Values

Missing values in panel data can occur in vari-

ables within a wave and when a full waveof data

is missing for a respondent. Within-wave miss-

ing values result from typical item nonresponse

that is found in any cross-sectional study. These

missing values occur when a valid response is

not recorded for a survey question either because

the participant chose not to answer the question

or an interviewer failed to record the answer.

Item nonresponse occurs most frequently for

sensitive questions (e.g., regarding income or

sexual behavior) and questions that are difcult

to answer (e.g., recalling a date; De Leeuw, Hox,

& Huisman, 2003). Within-wavemissing data in

panel studies can also occur when questions are

included in only some study waves.

Whole-wave missingness occurs when

respondents do not participate in all data col-

lection time points. In a four-wave panel study,

for example, a respondent may participate only

in the rst two waves before dropping out of

the study. This produces missing data on all

variables in the two subsequent waves. The

result is a substantial amount of missing data for

the time period covered by the wave, although

time-invariant characteristics (e.g., date of birth)

may be carried over from an earlier wave. When

respondents are missing entire waves of data,

too little information is available in the wave

to inform the data analysis, and information

on time-varying change is lost because of the

missing waves.

Attrition in longitudinal panel studies (or

study dropout) has received much attention

in the literature, and several strategies for

statistically evaluating and adjusting for the

consequences of attrition have been developed

(Groves, Dillman, Elting, & Little, 2002; Little,

1995). The attrition literature focuses heavily

on the potential for biased statistical estimates

that could result from overlooking attrition. In

medical clinical trials, for instance, the dropouts

from the trial may have been persons for whom