Survey ID Number
COVID-19 Rapid Response Phone Survey with Households 2020-2022, Panel
The World Bank in collaboration with the Kenya National Bureau of Statistics and the University of California, Berkeley are conducting the Kenya COVID-19 Rapid Response Phone Survey to track the socioeconomic impacts of the COVID-19 pandemic, the recovery from it as well as other shocks to provide timely data to inform policy. This dataset contains information from eight waves of the COVID-19 RRPS, which is part of a panel survey that targets Kenyan nationals and started in May 2020. The same households were interviewed every two months for five survey rounds, in the first year of data collection and every four months thereafter, with interviews conducted using Computer Assisted Telephone Interviewing (CATI) techniques.
The data set contains information from two samples of Kenyan households. The first sample is a randomly drawn subset of all households that were part of the 2015/16 Kenya Integrated Household Budget Survey (KIHBS) Computer-Assisted Personal Interviewing (CAPI) pilot and provided a phone number. The second was obtained through the Random Digit Dialing method, by which active phone numbers created from the 2020 Numbering Frame produced by the Kenya Communications Authority are randomly selected. The samples cover urban and rural areas and are designed to be representative of the population of Kenya using cell phones. Waves 1-7 of this survey include information on household background, service access, employment, food security, income loss, transfers, health, and COVID-19 knowledge and vaccinations. Wave 8 focused on how households were exposed to shocks, in particular adverse weather shocks and the increase in the price of food and fuel, but also included parts of the previous modules on household background, service access, employment, food security, income loss, and subjective wellbeing.
The data is uploaded in three files. The first is the hh file, which contains household level information. The ‘hhid’, uniquely identifies all household. The second is the adult level file, which contains data at the level of adult household members. Each adult in a household is uniquely identified by the ‘adult_id’. The third file is the child level file, available only for waves 3-7, which contains information for every child in the household. Each child in a household is uniquely identified by the ‘child_id’.
The duration of data collection and sample size for each completed wave was:
Wave 1: May 14 to July 7, 2020; 4,061 Kenyan households
Wave 2: July 16 to September 18, 2020; 4,492 Kenyan households
Wave 3: September 28 to December 2, 2020; 4,979 Kenyan households
Wave 4: January 15 to March 25, 2021; 4,892 Kenyan households
Wave 5: March 29 to June 13, 2021; 5,854 Kenyan households
Wave 6: July 14 to November 3, 2021; 5,765 Kenyan households
Wave 7: November 15, 2021, to March 31, 2022; 5,633 Kenyan households
Wave 8: May 31 to July 8, 2022: 4,550 Kenyan households
The same questionnaire is also administered to refugees in Kenya, with the data available in the UNHCR microdata library: https://microdata.unhcr.org/index.php/catalog/296/
For the KNBS and RDD samples, to make the sample nationally representative of the current population of households with mobile phone access, we create weights in two steps.
Step 1: Construct raw weights combining the two national samples: The current population consists of
(I) households that existed in 2015/16, and did not change phone numbers,
(II) households that existed in 2015/16, but changed phone number,
(III) households that did not exist in 2015/16.
Abstracting from differential attrition, the weights from the 2015/16 KIHBS CAPI pilot make the KIHBS sample representative of type (I) households. For RDD households, we ask whether they existed in 2015/16, when they had acquired their phone number, and where they lived in 2015/16, allowing us to classify them into type (I), (II) and (III) households and assign them to KIHBS strata. We adjust weights of each RDD household to be inversely proportional to the number of mobile phone numbers used by the household, and scale them relative to the average number of mobile phone numbers used in the KIHBS within each stratum. RDD therefore gives us a representative sample of type (II) and (III) households. We then combine RDD and KIHBS type (I) households by ex-post adding RDD households into the 2015/16 sampling frame and adjusting weights accordingly. Last, we combine our representative samples of type (I), type (II) and type (III), using the share of each type within each stratum from RDD (inversely weighted by number of mobile phone numbers). Variable: weight_raw
Step 2: Scale the weights to population proportions in each county and urban/rural stratum: We use post stratification to adjust for differential attrition and response rates across counties and rural/urban strata. We scale the raw weights from step 1 to reflect the population size in each county and rural/urban stratum as recorded in the 2019 Kenya Population and Housing Census conducted by the KNBS (2019 Kenya Population and Housing Census, Volume II: Distribution of Population by Administrative Units, December 2019, Kenya National Bureau of Statistics, https://www.knbs.or.ke/?wpdmpro=2019-kenya-population-and-housing-census-volume-ii-distribution-of-population-by-administrative-units). Variable: weight
In addition to being nationally representative, the data is also weekly representative for all waves except for wave 8. The variable weight_weekly should be used for weekly representative estimates.
To construct panel weights, we follow the approach outlined in Himelein (2014): “Weight Calculations for Panel Surveys with Subsampling and Split-off Tracking”. In each household we follow one target respondent. Wherever households split, only the current household of the target respondent was interviewed. The weights for the wave 1 and 2 balanced panel are constructed by applying the following steps to the full sample of Kenyan nationals:
0. Wave 1 cross-sectional weights after post-stratification adjustment are used as a base. W_1 = W_wave1
1. Attrition adjustment through propensity score-based method: The predicted probability that a sample household was successfully re-interviewed in the second survey wave is estimated through a propensity score estimation. The propensity score (PS) is modeled with a linear logistic model at the level of the household. The dependent variable is a dummy indicating whether a household that has completed the survey in wave 1 has also done so in wave. The following covariates were used in the linear logistic model: Urban/rural dummy, County dummies, Household head gender, Household head age, Household size, Dependency ratio, Dummy: Is anyone in the household working, Asset ownership: Radio, Asset ownership: Mattress, Asset ownership: Charcoal Jiko, Asset ownership: Fridge, Wall material: 3 dummies, Floor materials: 3 dummies, Connection to electricity grid, Number of mobile phones numbers household uses, Number of phone numbers recorded for follow-up, Sample dummy for estimation with national samples
2. Rank households by PS and split into 10 equal groups
3. Calculate attrition adjustment factor: ac (attrition correction) = the reciprocal of the mean empirical response rate for the propensity score decile
4. Adjust base weights for attrition: W_2 = W_1 * ac
5. Trim top 1 percent of the weights distribution (), by replacing the weights among the top 1 percent of the distribution with the highest value of a weight below the cutoff. W_3 = trim(W_2)
6. Apply post-stratification in the same way as for cross-sectional weights (step 2) Variable: weight_panel_w1_2
The balanced panel weights including waves 3, 4, 5, 6, 7 and 8 were constructed using the same procedure. Variables: weight_panel_w1_2_3, weight_panel_w1_2_3_4, weight_panel_w1_2_3_4_5, weight_panel_w1_2_3_4_5_6, weight_panel_w1_2_3_4_5_6_7 and weight_panel_w1_2_3_4_5_6_7_8.
Variable names were kept constant across survey waves. For questions that remained exactly the same across survey waves, data points for all waves can be found under one variable name. For questions where the phrasing changed (even in a minimal way) across waves, variable names were also changed to reflect the change in phrasing.
Extended missing values are used to indicate why a value is missing for all variables. The following extended missing values are used in the dataset:
.a for ‘Don’t know’
.b for ‘Refused to respond’
.c for ‘Outliers set to missing’
.d for ‘Inconsistency set to missing’ (used for employment data as explained below)
.e for ‘Field Skipped’ (where an error in the survey tool caused the question to be missed)
.z for ‘Not administered’ (as the variable was not relevant to the observation)
To address potential inconsistencies in the employment data, some data points had to be dropped for waves 2, 3, and 4. Despite the random allocation of households to enumerators, high variability was observed in reported employment across enumerators. To reduce inconsistencies, data on employment collected by some enumerators were set to missing. For each enumerator, the mean proportion of households without any employment was calculated. For waves 2 and 3, the 95 percent confidence interval of this mean proportion was established across all enumerators. Enumerators who displayed a proportion of households with no employment above the upper bound of the confidence interval were dropped. For wave 4, those enumerators with a mean proportion of households without any employment 1 standard deviation above the mean proportion across all enumerators were dropped. This resulted in dropping the data on employment for 596 households in wave 2, 1,109 households in wave 3, and 380 households in wave 4. To account for the dropped observations in the survey weights, the variable ‘weight_labor’ should be used for weighing when analyzing data on employment. It is constructed in the same way as the cross-sectional ‘weight’ variable, but only considering observations for which the data on employment is kept.
More detailed data on children was collected from between 3 and 7, compared to waves 1, 2 and 8. In waves 1 and 2, data on children, e.g. on their learning activities, was collected for all children in a household with one question. Therefore, variables related to children are part of the ‘hh’ data for waves 1 and 2. Between waves 3 and 7, questions on children in the household were asked for specific children. Some questions covered all children, while others were only administered to one randomly selected child in the household. This approach allows to disaggregate data at the level of the child household members, and the data can be found in the ‘child’ data set. The household level weights can be used for analysis of the children’s data. In wave 8, detailed information on children was dropped, as the questionnaire focused on other topics.
The education status of household members, except for the respondent, was imputed for rounds 1 and 2. For rounds 1 and 2, only the education status of the respondent was elicited, while for later rounds the education status for each household member was asked. In order to evaluate outcomes by the household member’s education status, information on education was imputed for waves 1 and 2, using the information provided for all household members in waves 3, 4, and 5. This resulted in additional information on the education status for household members in round 1 and 2, which was not yet available for earlier versions of this data.
Some questions are not asked repeatedly across waves such that their values were imputed. For some questions, answers are not possible or unlikely to change within two months between survey waves such that households were not asked about them in all waves. The questions on assets owned before March 2020 were only asked to households when they are interviewed for the first time. The questions on the dwelling’s wall and floor material as well as the household’s connection to the power grid was not asked for all households in wave 2 and 3, where only new households and those who moved were covered by these questions. Questions on the main source of electricity in the households and types of assets owned were not asked in wave 8. The missing values those variables have when they were not asked, are imputed from the answers given in earlier waves.
Improved quality insurance algorithms lead to minor revisions to wave 1 to 5 data. Based on additional data checks, the team has made minor refinements to wave 1 to 5 data. The identification of the household members that were the respondent or the household head was refined in the rare cases where it was not possible to interview the same respondent as in previous waves for a given household such that another adult was interviewed. For this reason, for about 2 percent of observations the household head status was assigned to an incorrect household member, which was corrected. For <1 percent of households the respondent did not appear in adult level dataset. For about 1 percent of observations in wave 5 the respondent appeared twice in the adult level dataset.
Data from questions on COVID-19 vaccinations from wave 7 have been dropped from the dataset. Self-reported vaccination rates were significantly higher compared to official administrative records (most likely due to social desirability bias), therefore reported data on vaccinations was deemed unreliable.. Consequently, data on vaccination status and questions using the vaccination data as a validation criterion were dropped from the datasets.