Table 2.
Missing data features reported by the studies
| Features reported | Number of papers (N = 82) (%) |
|---|---|
|
Was the amount of missing data reported? | |
| Yes |
66 (80%) |
| Missing data reported for each follow-up wave used in the analysis |
35 |
| A general statement was made about the amount of missing data or the amount of completed follow-up (how many participants attended at least one wave or only the final follow-up wave) |
22 |
| Indicated number that completed all waves of follow-up (i.e. number included in final sample) |
6 |
| Indicated amount missing for certain (key) variables |
3 |
| No |
16 (20%) |
|
Assessed differences between individuals with complete data and those with incomplete data? | |
| Yes |
26 (32%) |
| Provided a table comparing distributions of key exposures and outcome variables for those with missing and non-missing information |
6 |
| Table not provided but some summary statistics included in text |
4 |
| General comment provided (did not include a table or summary statistics or included p-values only) |
16 |
| No |
56 (69%) |
|
Reasons were given for the missing data |
13 (16%) |
|
Statistical method for handling missing data† |
|
| Method not stated |
14 (16%) |
| Complete-case analysis assumed |
9 (11%) |
| Complete-case analysis |
54 (66%) |
| Weighted |
1 |
| Unweighted |
53 |
| Exclude participants with missing data at any repeated waves of exposure |
38 |
| Exclude participant data record for waves of data collection with missing exposure data†† |
15 |
| Missing Indicator Method |
1 (1%) |
| Mean value substitution |
3 (4%) |
| Last Observation Carried Forward |
7 (9%) |
| Multiple Imputation |
5 (6%) |
| Details provided for the multiple imputation: | |
| Indicated how many imputations were performed |
4 |
| Indicated which variables were included in the imputation model |
2 |
| Compared results from multiple imputation with complete case analysis |
3 |
| Performed a sensitivity analysis under different assumptions for missing data |
4 |
| Fully Bayesian Model | 1 (1%) |
† Three papers used more than one method to handle their missing exposure data.
†† These studies assessed both exposure and outcome measures repeatedly over the waves of data collection. The data were analysed using either Generalised Estimating Equations (GEEs) or mixed-effects models to account for the correlated outcome data within an individual and excluded participant data records for waves where the exposure data were missing.