DNP 851B Module 1 Cleaning the Data and Handling Missing Values Example

Ninety-One of One Hundred Four: Cleaning the Data Set and Handling Missing Values in a Pediatric Asthma Project

Student Name

Doctor of Nursing Practice Program, Aspen University

DNP 851B: Project Data Analysis

Instructor Name

Month Day, Year

Ninety-One of One Hundred Four: Cleaning the Data Set and Handling Missing Values in a Pediatric Asthma Project

Analysis begins before the first statistical test, with the work of making sure that the data set is correct and that every gap in it is understood. This paper describes how the data set for my DNP project was cleaned. Each of the 104 enrolled children, all school-age or adolescent with persistent asthma, received a nurse visit covering a written plan, device coaching, a school link and a later phone check, elements the national asthma guidelines recommend (National Asthma Education and Prevention Program, 2007), then measured asthma control at 12 weeks and counted emergency and urgent care visits. The paper describes the cleaning process, the checks and what they found, the flow of participants from enrollment to analysis, the missing data and the decision about how to handle them.

A Process for Cleaning

One widely used guide describes data cleaning as a repeated cycle of screening for suspect values, diagnosing whether each is an error, a true extreme or a missing value, and editing, correcting only what can be verified and documenting every change (Van den Broeck et al., 2005). It warns against two mistakes: editing values that look odd but are true, and cleaning so late that sources can no longer be checked. The project followed that cycle. Much of the screening had already happened weekly during implementation, so this final pass was a confirmation and a search for anything the weekly checks could not catch, such as scores outside the planned window or inconsistent dates.

Checks and Findings

The table lists the final checks and what they found in the locked data set.

Participant Flow

Of 104 children enrolled, all 104 remain in the analysis of acute care visits and in the process measures, because acute care and delivery data came from three routine sources, not from the questionnaires. For the paired analysis of asthma control, 13 children are missing. Four baseline scores were never captured, all during the week the check-in tablet failed. Six had a 12-week score outside the planned window, most of them completed at a visit in week 16 or 17. Two moved out of the area, and one family withdrew. That leaves 91 children with paired scores from the age-appropriate test within the planned windows.

The same flow will appear as a diagram in chapter four, with the number analyzed shown separately for each outcome so that readers can see why the denominators differ between the control and acute care results.

Did Those Missing Differ?

Missing data bias results when the children who are missing differ in ways related to the outcome. The 13 were compared with the 91 on the variables available for all. Their mean age was similar, 10.1 years against 9.8. A similar share preferred Spanish, 5 of 13 against 38 of 91. They were more often uninsured, 3 of 13 against 6 of 91. For the nine of the 13 who had a baseline score, the mean was 16.0, close to the 91's baseline mean of 16.4. The six whose 12-week score fell outside the window completed it late mostly because of rescheduled visits, not because they were sicker. On this evidence the missing children do not appear to differ much from the others, although the higher share without insurance is noted as a possible source of bias.

Choosing How to Handle Missing Data

Three approaches were considered: complete-case analysis, carrying forward the last value, and multiple imputation. Jakobsen et al. (2017) offer a practical guide for trials: when the share of missing data is small, often below about 5%, complete-case analysis may be enough; when it is large, results may be hypothesis-generating at best; in between, multiple imputation can be considered, particularly when data are plausibly missing at random given observed variables. Here 12.5% of children lack paired scores, which falls in the middle range. Imputation was judged unsuitable for two reasons. First, with only 91 complete cases and few baseline variables, an imputation model would be weak. Second, the project is a quality improvement evaluation, and a transparent analysis the clinic can follow is worth more than a modest gain in precision.

The Decision

The primary analysis will use the 91 complete cases, as the proposal planned. Two sensitivity analyses will test how much the missing data could matter. The first includes the six late scores as if they were on time, since they were completed only a few weeks after the window. The second is a worst-case check for the proportion with well-controlled asthma: all 13 missing children are counted as not controlled at 12 weeks. If the conclusion holds under both, missing data are unlikely to have changed it. Every decision in this paper is recorded in the correction log that accompanies the data set.

Conclusion

The final data set contains acute care data for all 104 enrolled children and paired control scores for 91. Cleaning found few errors, most caught earlier during implementation, and every correction is documented. The 13 missing children appear similar to the others except for insurance, and the planned complete-case analysis will be tested with two sensitivity analyses before any conclusion is drawn.

The cleaned file and its log have been shared with the committee's methods reviewer.

Nothing was deleted from the source files.

References

Jakobsen, J. C., Gluud, C., Wetterslev, J., & Winkel, P. (2017). When and how should multiple imputation be used for handling missing data in randomised clinical trials: A practical guide with flowcharts. BMC Medical Research Methodology, 17, Article 162. https://doi.org/10.1186/s12874-017-0442-1

National Asthma Education and Prevention Program. (2007). Expert panel report 3: Guidelines for the diagnosis and management of asthma (NIH Publication No. 07-4051). National Heart, Lung, and Blood Institute.

Van den Broeck, J., Argeseanu Cunningham, S., Eeckels, R., & Herbst, K. (2005). Data cleaning: Detecting, diagnosing, and editing data abnormalities. PLoS Medicine, 2(10), Article e267. https://doi.org/10.1371/journal.pmed.0020267