Imputing missing data

Peter Levy

18/02/2026

The MetQC app uses multiple gap-fill methods to impute data when the recorded data maybe incorrect or unreliable. This document will outline the methods undertaken, by the app, for each gap-fill method.

QC code value	QC Code	Meaning
0	raw	Original observation: raw data.
1	missing	Missing value: usually only a temporary state, whereafter values are imputed by one of the methods below.
2	time	Time: values are predicted from a GAM fitted to the non-missing data, using the local trend in time plus terms for the diurnal and seasonal cycles. The effect is similar to LOESS filtering but with the addition of the diurnal and seasonal trends, useful when multiple days are missing. Generally this is the best option for most variables.
3	regn	Regression with covariate: values are predicted from a linear regression model fitted to the non-missing data, using some other covariate selected from the data set. Can simply replace missing values with the covariate if fit = FALSE.
4	nightzero	Night-time zero: night-time values are forced to be zero, generally the true for incoming solar radiation variables.
5	noneg	No negative values: any negative values are forced to be zero, true for any quantity which has to be positive e.g. rainfall.
6	zero	All zeros: all values are forced to be zero, mainly used for snow depth in summer.
7	era5	Regression with ERA5 covariate: values are predicted from a linear regression model fitted to using some other covariate selected from the ERA5 data set. ERA5 is a reanalysis data set based on weather forecasting models nudged with observations, run retrospectively so gives a consistent back-casting of weather variables. Generally this is the best option for variables whee a local trend in time is not predictable e.g. rainfall.
8	flagged	Flagged data: manually flagged data, usually only a temporary state, whereafter values are imputed by one of the methods above.

Original observation (raw)

This method simply reverts any previously imputed (either as part of the automated process or by a user) data back to the raw observed data. This can be used to undo any changes.

Missing

This method simply flags data as ‘missing’. No further imputation is undertaken.

Time

Fits a smoothing model (a GAM) of the variable in question against time. New values are predicted from this model.

Regression with covariate

Fits a model between the variable in question and another variable, which the user can pick from a drop down list. New values are predicted from this model.

Night-time zero

Sets values to zero during the night. Values measured during the day are unaffected. Night time is calculated using the position of measurement and the sun position.

No negative values

Sets all selected points that are negative to zero.

All zeros

Sets all selected points to zero.

Regression with ERA5 covariate

Fits a model between the variable in question and the same variable from the ERA5 dataset. New values are predicted from this model.