Title: Covariates to include in the imputation of missing outcome data in Randomised

Controlled Trials

Background
A randomized controlled trial(RCT) is a type of scientific experiment which aims to reduce bias when
testing a new treatment. The people participating in the trial are randomly allocated to either the group
receiving the treatment under investigation or to a group receiving standard treatment as the
control(Chalmers et al., 1981). RCTs are often considered the gold standard for a clinical trials, and are
often used to test the efficacy or effectiveness of various types of medical intervention and may provide
information about adverse effects, such as drug reactions. Random assignment of intervention is done after
subjects have been assessed for eligibility and recruited, but before the intervention to be studied begins.
After randomization, the two (or more) groups of subjects are followed in exactly the same way and the only
differences between them is the care they receive. For example, in terms of procedures, tests, outpatient
visits, and follow-up calls, should be those intrinsic to the treatments being compared. The most important
advantage of proper randomization is that it minimizes allocation bias, balancing both known and unknown
prognostic factors, in the assignment of treatments(Moher et al., 2010).
Missing data is a common problem within industrial and medical databases (Lakshminarayan et al. 1996),
and is according to (Potthoff et al., 2006) a source of serious problems in statistical analyses of clinical trials
and of other medical studies. Missing data arise when a record has attributes that are missing, or even whole
records might be missing from a dataset. Randomized controlled data is likely of having missing values as
the datasets may be large. The values might be missing because of random errors with equipment or
calculation, attrition due to social or natural processes as for instance death, or respondent refusal, such as a
person not answering certain questions in a survey. Some data might even be intentionally missing
(Humphries, 2013). It is not uncommon to find datasets containing up to 50% missing values (Farhangfar,
Kurgan, and Dy 2008).

Imputation methods are methods that replace missing data with a calculated replacement values. Imputation
methods have been used for improving classifiers' results on datasets containing missing data. Most of the
work experiment with statistical, maximum likelihood or machine learning imputation methods for
comparison and analysis (Tokle, 2017).

Data quality of a dataset depends on the intended use of the data, meaning a dataset may be of good quality
for one use case, but not for another (Han, Pei, and Kamber 2011). A value in an object might be missing, or
even the whole object. Sometimes, whole records are missing from a dataset.

Approaches for dealing with missing values can be either to remove them, calculate substitution values, or
simply ignore them (Tokle, 2017). Common ways to handle missing data is to ignore them, or remove the
cases containing missing values, commonly known as case deletion. However, in some cases, inferring the
values with an imputation method is necessary (Han, Pei, and Kamber, 2011). Finding the right approach to
handle the missing data depends on the application domain and the data itself. Therefore, understanding the
data and its type is important when handling missing data (Humphries, 2013).

Using the wrong methods may change the structure of the data and lead to false information. Another
challenge is the complexity of the imputation methods, as some imputation methods are time consuming.
When using imputation in real time, the best solution might be to ignore the missing values (Han, Pei, and
Kamber 2011), using simple imputation methods, or using data mining methods handling the missing values
internally.

Patient Reported Outcome Measures(PROMs) are tools used to measure patient-reported outcomes (PROs).
PROMs are standardized, validated questionnaires that are completed by patients’ during the perioperative
period to ascertain perceptions of their health status, perceived level of impairment, disability, and health-
related quality of life(Monmouth, 2017). They allow the efficacy of a clinical intervention to be measured
from the patients’ perspective. Questionnaires are given to patients both pre and post operatively to allow
comparison of outcomes pre and post procedure(Black, 2014). In addition to outcomes relating to
interventions, PROMs measure patients’ perceptions of their general health or their health in relation to a
specific disease. PROMs are a means of measuring clinical effectiveness and safety(Black, 2014). PROMs
can be classified as either generic or disease specific. The generic tools measure a variety of aspects of a
broad range of medical conditions, allowing for the overall evaluation of care, quality of life, and cost
effectiveness of interventions (Weldring, 2013). The disease specific PROMs allow individual aspects of a
condition and their impact on outcome to be examined(Monmouth, 2017).

Regular use of patient‐reported outcome measures (PROMs) by health care providers in their routine
practice may help to improve the quality of care, but more evidence is needed before routine use of PROMs
can be recommended(Marshall et al., 2006). PROMs are becoming increasingly used in clinical research.
PROMs typically have several dimensions and use multiple items to measure these dimensions. Missing
items are more likely to occur with PROM data than with clinical data as most PROMs are self-completed
and patients may refuse to answer some or all of the items. Missing PROM values may be imputed.

There are several different methods for imputing missing PROM data: Simple mean, Last Observation
Carried Forward (LOCF), Horizontal mean, regression, Markov chain, hot deck, multiple linear regression
imputation, predictive mean matching, multiple imputation, monotonic multiple imputation and multivariate
Normal imputation. The regulators regard most imputation methods as imperfect. However, they
recommend: specifying imputation methods in advance in the protocol/statistical analysis plan; imputing
missing PROM items according to instrument developer’s guidelines; using several imputation methods
with a sensitivity analysis of the results. It is desirable to report analyses with and without imputation and to
explore different imputation techniques in order that intention to treat analysis can be performed.

In this project, the imputation methods and their effect on statistical analysis and some computer simulation
and analysis to compare the different methods of imputation (with and without the randomised treatment
group as a covariate) with the view to developing guidance on whether or not to include the treatment group
as a covariate in the imputation model conclusions will be compared using the data from several RCTs with
PROMs. The project would then involve simulation and analysis to compare the different methods of
imputation (with and without the randomised treatment group as a covariate) with the view to developing
guidance on whether or not to include the treatment group as a covariate in the imputation model.

Objectives of the project

The main objective of the project is to compare different imputation methods and their effect on statistical
analysis using the data from several RCTs with PROMs data.
Specifically, the project intends to:
1. Analyse PROMs data with and without imputation so as to explore different imputation techniques.
2. Compare different imputation methods and their effect on statistical analysis.
3. Applying computer simulation and analysis to compare the different methods of imputation (with
and without the randomised treatment group as a covariate) with the view to developing guidance on
whether or not to include the treatment group as a covariate in the imputation model.

Methodology
The project aims to explore and compare the efficacy of several different methods for imputing missing
PROM data with the view to develop guidance on whether or not to include the treatment group as a
covariate in the imputation model. The following imputation methods will be considered to achieve the
objectives: Simple mean, Last Observation Carried Forward (LOCF), Horizontal mean, regression, Markov
chain, hot deck, multiple linear regression imputation, predictive mean matching, multiple imputation,
monotonic multiple imputation and multivariate Normal imputation.
Participants and sample
An approximate number of patients(sample) required in the study is 385. This number is obtained from
formula given by Cochran(1963) using a confidence level of 95% , a precision rate of ±5%, and variability
of 50% (maximum variability in the population). The projects will not consider the gender of a patient. That
is both males and female patients have equal chance of being chosen as sample units. Also, those 385 will be
chosen randomly from a target population by using simple random sampling.
Intervention and research design
The research project will be longitudinal study where the 385 patients shall be followed for time between 5
to 7 months, and during this period of time the PROMs data will be collected. The patients shall be divided
randomly into two(treatment and control) groups.
Data collection
The dataset will be collected by using PROMs whereby the records of each patient will be taken monthly
for successive period of 5 to 7 months.
Data analysis
The data will be analysed by using several imputation techniques as mention in methodology above. All
analyses will be done at 5% level of significance.
References
Melissa Humphries. Missing Data & How to Deal: An overview of missing data_.
In: Population Research Center. University of Texas. Recuperado de: http://www.
google. com/url (2013) (cit. on pp. 39, 41, 46, 51_55).
J. Han, J. Pei, and M. Kamber. Data Mining: Concepts and Techniques. The Morgan
Kaufmann Series in Data Management Systems. Elsevier Science, 2011. isbn:
9780123814807. https://books.google.no/books?id=pQws07tdpjoC (cit. on
pp. 39_41, 43, 44, 69).
Alireza Farhangfar, Lukasz Kurgan, and Jennifer Dy. _Impact of imputation of
missing values on classi_cation error for discrete data_. In: Pattern Recognition
41.12 (2008), pp. 3692_3705. issn: 0031-3203. doi: http://doi.org/10.1016/j.
patcog.2008.05.019. http://www.sciencedirect.com/science/article/pii/
S003132030800201X (cit. on pp. 3, 39_41, 74).
Kamakshi Lakshminarayan et al. Imputation of Missing Data Using Machine
Learning Techniques_. In: Proceedings of the Second International Conference on
Knowledge Discovery and Data Mining. KDD'96. Portland, Oregon: AAAI Press,
1996, pp. 140_145. http://dl.acm.org/citation.cfm?id=3001460.3001489
(cit. on pp. 3, 39, 74).
Marit Iren Rognli Tokle. Comparison of Missing Data Imputation Methods for Improving Detection of
Obstructive Sleep Apnea, 2017. http://www.duo.uio.no/

Chalmers TC, Smith H Jr, Blackburn B, Silverman B, Schroeder B, Reitman D, Ambroz A (1981). "A
method for assessing the quality of a randomized control trial". Controlled Clinical Trials. 2 (1): 31–
49. doi:10.1016/0197-2456(81)90056-8. PMID 7261638.
Moher D, Hopewell S, Schulz KF, Montori V, Gøtzsche PC, Devereaux PJ, Elbourne D, Egger M, Altman
DG (2010). "CONSORT 2010 explanation and elaboration: updated guidelines for reporting parallel group
randomised trials". Br Med J. 340: c869. doi:10.1136/bmj.c869. PMC 2844943 . PMID 20332511

Susan Marshall , Kirstie Haywood , Ray Fitzpatrick (2006). Impact of patient‐reported outcome measures on
routine practice: a structured review. Available from:

https://doi.org/10.1111/j.1365-2753.2006.00650.x (accessed 25 July, 2018)

Monmouth Partners. A Guide to Patient Reported Measures - Theory

, Landscape and Uses. Available from
http://www.monmouthpartners.com/assets/pdf/A%20Guide%20to%20Patient%20Reported%20Measures.pd
f (accessed 25 July, 2018).

Black N, Varaganum M, Hutchings A. Relationship between patient reported experience (PREMs) and
patient reported outcomes (PROMs) in elective surgery. BMJ Qual Saf 2014;23: 534–42

Weldring T, Smith SM. Patient-Reported Outcomes (PROs) and Patient-Reported Outcome Measures
(PROMs). Health Serv Insights, 2013;6:61