{"paper_id":"439204bd-32f4-45a6-9d38-11648edd6cec","body_text":"1 \nAssociation between mobility, non-pharmaceutical interventions, and COVID-19 \ntransmission in Ghana: a modelling study using mobile phone data \n \nHamish Gibbs\n1, Yang Liu1, Sam Abbott1, Isaac Baffoe-Nyarko2, Dennis O. Laryea2, Ernest \nAkyereko2, Patrick Kuma-Aboagye2, Ivy Asante3, Oriol Mitjà4, LSHTM CMMID COVID-19 \nworking group, William Ampofo3, Franklin Asiedu-Bekoe2, Michael Marks5,6,7, Rosalind M. \nEggo1 \n \n \n1Department of Infectious Disease Epidemiology, London School of Hygiene & Tropical \nMedicine, London, United Kingdom. \n \n2Ghana Health Service, Ministry of Health, Accra, Ghana. \n \n3Noguchi Memorial Institute for Medical Research, Accra, Ghana. \n \n4Fight AIDS and Infectious Diseases Foundation, Hospital Universitari Germans Trias i Pujol, \nBadalona, Spain. \n \n5Department of Clinical Research, London School of Hygiene & Tropical Medicine, London, \nUnited Kingdom. \n \n6Hospital for Tropical Diseases, University College London Hospital, London, United \nKingdom. \n \n7Division of Infection and Immunity, University College London, London, United Kingdom. \n \n \nCorrespondence to Hamish Gibbs; Hamish.Gibbs@lshtm.ac.uk\n. \n  \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \nNOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.\n\n2 \nResearch in Context \n \nEvidence before this study \n \nWe searched PubMed and preprint archives for articles published in English that contained \ninformation about the COVID-19 pandemic published up to Nov 1, 2021, using the search \nterms “coronavirus”, “CoV”, “COVID-19”, “mobility”, “movement”, and “flow”.  \nThe data thus far suggests that NPI measures including physical distancing, reduction of \ntravel, and use of personal protective equipment have been demonstrated to reduce COVID-\n19 transmission. Much of the existing research focuses on comparisons of NPI stringency \nwith COVID-19 transmission among different high-income countries, or on high-income \ncountries, leaving critical questions about the applicability of these findings to low- and \nmiddle-income settings. \n \nAdded value of this study\n \n \nWe used a detailed COVID-19 surveillance dataset from Ghana, and unique high resolution \nspatial data on human mobility from Vodafone Ghana as well as Google smartphone GPS \nlocation data. We show how human mobility and NPI stringency were associated with \nchanges in the effective reproduction number (R\nt). We further demonstrate how this \nassociation was strongest in the early COVID-19 outbreak in Ghana, decreasing after the \nrelaxation of national restrictions. \n \nImplications of all the available evidence\n \n \nThe change in association between human mobility, NPI stringency, and Rt may reflect a \n“decoupling” of NPI stringency and human mobility from disease transmission in Ghana as \nthe COVID-19 epidemic progressed. This finding provides public health decision makers with \nimportant insights for the understanding of the utility of mobility data for predicting the spread \nof COVID-19.   \n  \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n3 \nAbstract \n \nBackground: Governments around the world have implemented non-pharmaceutical \ninterventions to limit the transmission of COVID-19. While lockdowns and physical distancing \nhave proven effective for reducing COVID-19 transmission, there is still limited \nunderstanding of how NPI measures are reflected in indicators of human mobility. Further, \nthere is a lack of understanding about how findings from high-income settings correspond to \nlow and middle-income contexts. \n \nMethods: In this study, we assess the relationship between indicators of human mobility, \nNPIs, and estimates of R\nt, a real-time measure of the intensity of COVID-19 transmission. \nWe construct a multilevel generalised linear mixed model, combining local disease \nsurveillance data from subnational districts of Ghana with the timing of NPIs and indicators \nof human mobility from Google and Vodafone Ghana. \n \nFindings: We observe a relationship between reductions in human mobility and decreases \nin R\nt during the early stages of the COVID-19 epidemic in Ghana. We find that the strength \nof this relationship varies through time, decreasing after the most stringent period of \ninterventions in the early epidemic. \n \nInterpretation: Our findings demonstrate how the association of NPI and mobility indicators \nwith COVID-19 transmission may vary through time. Further, we demonstrate the utility of \ncombining local disease surveillance data with large scale human mobility data to augment \nexisting surveillance capacity and monitor the impact of NPI policies. \n \nIntroduction \n \nNations around the world introduced a range of non-pharmaceutical interventions (NPIs) to \nlimit the spread of COVID-19 in the early phases of the epidemic[1]. These NPIs have been \ndiverse, and have included the use of personal protective measures, environmental \nmeasures, physical distancing, restricting movement, and limiting the gathering of people. \nNPIs have been implemented at different times in relation to the progression of local and \nnational disease outbreaks, with some put in place before transmission was established, and \nothers reactive to rises in cases. NPI measures have also overlapped one another in the \ntiming of their application[1,2]. Previous research has attempted to quantify the relative \neffectiveness on COVID-19 transmission of different NPIs[3–6], but modelling the impact of \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n4 \ndifferent intervention strategies includes uncertainty about how different strategies are \nimplemented in practice. Additionally, statistical approaches for estimating the impact of \nindividual interventions can be confounded by the overlapping nature of NPI policies and the \ndifferent mechanisms that interventions use to reduce disease transmission. There remain \nsignificant open questions about methods for reliably isolating and quantifying the individual \neffect of each intervention[7]. \n \nOne approach used by researchers and policymakers to measure the impact of NPIs during \nthe COVID-19 pandemic has been to observe changes in measurements of human \nbehaviour under individual interventions or under a combination of interventions[8–12]. \nPerhaps the most common way to quantify varying patterns of human behaviour is the use \nof human mobility datasets, which measure the locations of individuals using GPS or Call \nDetail Records (CDRs)[13,14]. These mobility datasets have been made available by a \nvariety of network service and mobile application providers[15–17]. Mobility data has been \nused widely during the COVID-19 pandemic to predict the introduction of COVID-19 cases, \nand to monitor and estimate adherence to NPIs including travel restrictions[8,9,18–20], but \nquestions remain about how patterns of mobility and NPI stringency relate to transmission in \nLMIC settings.   \n \nPrevious research has been conducted in Africa on the implications of mobility patterns for \ntransmission of infections other than COVID-19[21,22] and during the COVID-19 epidemic, \nanalysis of movement patterns in Ghana has been conducted to inform policy makers about \nthe volume of reductions coinciding with lockdown interventions in Accra and Kumasi[23]. \nThese indicators may be used as a proxy for social contact[13] and therefore, for potential \nCOVID-19 transmission, although the “link” between movement and disease transmission \nmay decrease due to greater adherence to social distancing or personal protective \nequipment guidelines[24].There remain questions about how mobility indicators can be used \nto estimate COVID-19 transmission and how these indicators reflect behavioural responses \nto NPI measures, particularly in an LMIC context. In this paper, we combine human mobility \nand NPI data to estimate their relationship over time to the progression of the COVID-19 \nepidemic in Ghana.  \n \nHere, we used surveillance data in a sample of 27 districts collected by the Ghana Health \nService on PCR confirmed COVID-19 patients at the district level (administrative level 2, 261 \ntotal districts) in 11 of the 16 regions of Ghana between March and December 2020 to \nproduce individual estimates of R\nt, the real-time reproduction number, for individual districts. \nRt is a time-varying parameter describing the average number of infections derived from a \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n5 \nsingle infection and indicates whether an epidemic is growing (Rt > 1) or decreasing (Rt < 1). \nWe combined estimates of Rt with subnational human movement data from Vodafone \nGhana and Google. These data measure the volume of movement activity (a proxy for social \ncontact), in each district. We then modelled the relationship between human mobility \nindicators and R\nt using a multilevel generalised linear mixed model to assess whether \nchanges in mobility and NPI stringency were associated with changes in Rt in during the \nCOVID-19 epidemic in Ghana. \n \nMethods \n \nStudy Setting  \n \nThe first cases of COVID-19 in Ghana were reported on 12th March 2020[25]. These cases \nwere detected in Accra, the capital city of Ghana and were imported via international \ntravel[25]. Following the announcement of the first COVID-19 cases, the Ghanaian \ngovernment announced the suspension of international travel and the closure of land \nborders to reduce the risk of further introduction[26]. Domestic case numbers grew in March \nand April 2020, leading to the closure of universities and high schools and the \nannouncement of a partial lockdown in the Ashanti and Greater Accra regions, the two most \npopulous regions of Ghana[26].  This lockdown introduced a stay at home order except for \nessential travel including shopping, healthcare, and use of public toilets. Almost all COVID-\n19 NPI restrictions were lifted by July, although restrictions on international travel and \nmandated use of facemasks remained in place until September 2020. \n \nCOVID-19 surveillance data \n \nWe used line list data recording lab-confirmed COVID-19 cases collected by Ghana Health \nServices each day between 12th March and 1st September 2020. These data were collected \nin 11 of the 16 regions in Ghana (administrative level 1: excluding “Ahafo”, “Bono”, “Upper \nWest”, “Volta”, and “Western North” districts due to limited detail of data in these districts). \nPatient-level records were referenced to a standard spatial reference provided by the Ghana \nStatistical Service using patients’ reported district of residence. Using the date of case \nconfirmation and the district of residence, we aggregated individual records into daily case \ncounts of confirmed COVID-19 cases per district (Supplemental Figure 1). Through visual \ninspection, we replaced three outliers in reporting in two districts with a linear interpolation \nbetween the preceding and following records. For these outliers, the number of cases \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n6 \nreported in a district clearly exceeded the overall trend of case reporting (reported cases \ngreater than 5x higher than all previous reports) (Supplemental Table 1). We assumed that \nthese records reflected “late reporting” with samples collected on multiple days reported on \nthe same date. This assumption potentially underestimates the total number of COVID-19 \ncases in these two districts, but it is not possible to approximate when the cases reported in \nthese intervals may have been originally tested.   \n \nDefining stringency indices for NPIs \n \nData on the dates of NPIs implemented in Ghana were provided by Ghana Health Service \ndetailing the starting dates of public health interventions. We used this to define the start \ndates of interventions and augmented it with available news sources and government press \nreleases to create a dataset of the start and end dates of nine intervention measures. Using \nthese intervention data, we constructed a stringency index to measure the stringency of \nCOVID-19 interventions through time, defined daily as the number of active interventions \ndivided by the total number of interventions. This stringency index assigns a uniform level of \nstringency to each intervention measure and records the length of time that the measure \nwas implemented. We also used the OxCGRT stringency index, calculated from a global \ndatabase of NPIs, which is used to construct a stringency index based on a taxonomy of \ngovernment interventions[1] (Supplemental Figure 2). We used the most recent version of \nthe stringency index (as of June 2021), rather than the OxCGRT “legacy stringency index.” \nOxCGRT data also records nine interventions resulting in a change in the stringency index in \nthe study period. We extracted the date of maximal intervention from both indices to \ncompare both stringency indicators, defined as the first date with the highest stringency for \neach index. Although both sources of intervention data reported interventions at a national \nlevel, intervention measures were introduced at different spatial scales in Ghana. School \nclosures (including different educational tracks) and mask mandates, for example, were \nimplemented nationally, while partial lockdown measures were introduced only in Ashanti \nand Greater Accra regions. \n \nR\nt estimation \n \nTo ensure data coverage during the early stages of the COVID-19 outbreak in Ghana, we \nlimited Rt estimation to 27 districts with reported cases before March 30th, 2020 (the \nbeginning of the partial lockdown in Ashanti and Greater Accra), and at least 100 reported \ncases during the entire study period (12th March to 1st September 2020). We chose this \nthreshold to select districts with surveillance resources capable of detecting cases in the \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n7 \nearly epidemic, due to uncertainty about whether case reporting followed the path of the \nepidemic or the availability of testing in districts with irregular case reporting.  \n \nRt estimates were calculated at the district level using the EpiNow2 R package (1.3.2) using \nMCMC, as implemented in Stan[27,28], based on weekly reported cases[29]. Expected daily \ncases were estimated using the renewal equation to weight prior expected cases multiplied \nby the estimated R\nt. Variation in Rt over time was modelled using a mean intercept and an \napproximate Gaussian process with a 3/2 matern kernel on the log scale[30,31]. Unlike in \nAbbott et. al.[30] we modelled R\nt explicitly with the gaussian process and not as a first order \ndifference. This has little impact on retrospective Rt estimates and substantially reduces the \ncomputational overhead. We used a generation time modelled as a gamma distribution with \nmean: 3.6 (standard deviation of mean: 0.7), standard deviation 3.1 (standard deviation of \nstandard deviation: 0.76) and maximum: 15[5,32].  We assumed a negative binomial \nobservation model for reported cases with a day of the week effect modelled as a simplex \nallowing us to model weekly reported cases without manual specification. Rt estimates did \nnot include an estimate of reporting delays as lags were estimated in subsequent analyses. \nTherefore, estimates of Rt on a given date vary as a result of the reported cases on that \ndate. Inference was performed across 4 chains for 2000 samples with a burn-in period of \n250 samples. Convergence was diagnosed using the R hat diagnostic[28]. \n \nDefining a Mobility Indicator from Vodafone Data \n \nWe used Vodafone Ghana Call Detail Records (CDRs) aggregated by the Flowminder \nFoundation prior to data sharing[33]. CDRs record mobile phone connections to the cellular \ntowers routing a call or SMS. CDRs are used by mobile network providers for billing \npurposes. CDRs provide the location of the mobile phone and SIM card based on the \nlocation of the cellular tower routing the signal, most often the nearest one. The precision of \nestimated mobile phone locations depends on the density of cellular towers in an area and \nsignal coverage. It can reach up to 3 km in average coverage conditions and up to 8 km in \ngood coverage conditions. Individual mobility can be estimated by recording a series of \nmobile phone connections to cellular towers over time. We used CDR data aggregated into \nan origin-destination matrix, based on the locations recorded within 24 hours for individual \nmobile phones. The data were censored to remove daily counts of 10 or fewer subscribers \nrecorded for an origin-destination pair, in order to reduce the risk of statistical disclosure of \npersonally identifiable information.  \n \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n8 \nWe generated a normalised index of movement outside of individual districts relative to \nbaseline values from the origin-destination matrix. We used two metrics to calculate this \nnormalised measure of mobility: (1) trips between districts: the daily number of subscribers \ntravelling between pairs of districts, and (2) total subscribers per district: the total number of \nunique mobile phone subscribers recorded in a district on each day. Because of \ninconsistencies between the spatial references used for the mobility data and the case data, \nwe aggregated mobility indicators for Accra, Tema, and Kumasi Metropolitan Areas by \nremoving trips between aggregated districts and calculating the sum of subscribers for these \ndistricts. The mobility data also contained 10 missing dates (4.1%) and we performed linear \ninterpolation for each district for both metrics (trips and subscriber counts) on these dates \n(Supplemental Figure 3). \n \nTo construct the normalised movement index we summed the total number of outgoing trips \nfor individual districts on each day. These values were then normalised by the total number \nof daily subscribers in individual districts to remove bias introduced solely because of varying \nnumbers of subscribers. For each district i and each day t, the normalised number of \noutgoing trips was defined as:  \n/i1/i1/i1/i1/i1 _ /i1/i1/i1 _ /i1/i1/i1/i1 /i1, /i1 /g3404 /i1/i1/i1/i1/i1 _ /i1/i1/i1 /i1, /i1\n/i1/i1/i1/i1/i1 _ /i1/i1/i1/i1/i1/i1/i1/i1/i1/i1/i1 /i1, /i1\n \nThe movement index measures the change in outbound trips from individual districts relative \nto baseline values using September to December 2019 as the baseline period. We chose \nthis baseline as it includes the earliest period for which mobility data was available, but this \nbaseline may not account for seasonal variations in movement patterns during a year.  \n \nBaseline values were then calculated per week day during the baseline period as the median \nof outgoing trips (normalised by the number of subscribers, as above) for each district i and \neach day of the week, j: \n/i1/i1/i1/i1/i1/i1/i1/i1 /i1, /i1/g3404 /i1/i1/i1/i1/i1/i1 /g4666 /i1/i1/i1/i1/i1 _ /i1/i1/i1 _ /i1/i1/i1/i1 /i1, /i1/g4667  \nUsing the baseline values, we calculated the deviation from baseline in the study period as a \npercentage for each district i on each day t given the day of week j of t:  \n/i1/i1/i1/i1/i1/i1/i1 _ /i1/i1/i1/i1/i1/i1 /i1, /i1 /g3404 /g4666 /i1/i1/i1/i1/i1/i1/i1/i1 /i1, /i1 /g3398  /i1/i1/i1/i1/i1/i1/i1/i1 /i1, /i1/g4667\n/i1/i1/i1/i1/i1/i1/i1/i1 /i1, /i1\n/g1499 100  \nThis resulted in a normalised mobility indicator (Supplemental Figure 4). \n \nGoogle Mobility Indicator \n \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n9 \nWe used mobility data from Google as a second measure of human movement[16]. This \ndata records the GPS location of individuals actively using Google services who have \nchosen to share their location data with Google. The data is provided as a measure of \nchanges in activity relative to a baseline in different settings (Residential, Grocery & \nPharmacy, Retail & Recreation, Transit Stations, Workplaces). The dataset documentation \nrecommends consideration of the specifics of mobility in different settings. We chose to use \nonly the mobility indicator from the “Residential” setting because we considered this setting \nto be the most clearly defined setting in the context of Ghana, and because of the relatively \nlower variance of this indicator (Supplemental Figure 5). We calculated the inverse of the \npercent change in residential mobility. This percentage is relative to a baseline period \nbetween 3rd January and 6th February, 2020 which is defined by Google prior to data \nsharing.  \n \nGoogle mobility data is not referenced to known administrative areas but rather to custom \nboundary polygons, which do not closely align with administrative districts in Ghana. To \ncombine this mobility data with the other data sources used in this study, we manually \ndigitized (traced) these features to create a spatial representation of the coverage area of \neach metric. To align Google mobility data with our spatial reference, we assigned Google \nmobility data to those districts with greater than 50% overlap with the administrative areas \ndefined by Google for Accra and Kumasi. This restricted the coverage of Google mobility \ndata to central districts in the Accra and Kumasi Metropolitan Areas (Supplemental Figure \n6).  \n \nStatistical Analysis  \n \nWe assessed the association between NPIs, mobility and median R\nt  while adjusting for \npublic holidays using a two-level multilevel generalised linear mixed model (using a \nGaussian observation model), with random intercepts for individual districts to account for \nlocal variation. We used data from 12th March 2020 to 1st September 2020, during the \nperiod of interventions in Ghana, and before the detection of the Alpha or Beta variants[34]. \nLevel 1 of the multi-level model included the relationship between mobility indicators and the \ninverse NPI stringency with district-level random effects, x to account for correlation between \nmobility and NPI stringency: \n \nLevel 1: Mobility\na ~ xa + NPI stringency \n \nWhere a indicates individual districts.  \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n10 \n \nLevel 2 modelled the relationship between Rt, NPI stringency, residuals from Level 1, and \ndistrict-level random effects: \n \nLevel 2: Rt\na ~ xa + NPI Stringency, i -j + residuals(Level 1a, i -j) + Holidays, i -j  \n \nWhere x is a random effect, a indicates individual districts, j is a lag between 1 and 30 days, \nand i is the original date of data collection. We trained models for each time period and lag \nvalues to determine the optimal lag between R\nt and mobility. We assessed the different \nmodels by comparing the marginal R2, which represents the contribution of fixed effects only, \nfor different lag values and time periods. We also calculated the Median Absolute Error of \neach model: \n \n/i1/i1/i1  /g3404 /i1/i1/i1/i1/i1/i1 /g4666 | /i1/i1/i1/i1/i1/i1/i1/i1/i1 /i1 /g3398  /i1/i1/i1/i1/i1/i1/i1/i1 /i1 | /g4667  \n \nFor i in 1...N values (either predicted or observed) where N is the total number of \nobservations. \n \nCoefficients of the Level 2 model estimate R\nt given inverse NPI stringency, holiday events, \nand the residuals of the Level 1 model (which can be interpreted as “mobility not explained \nby NPI stringency”). The use of inverse NPI stringency means that positive coefficients can \nbe interpreted similarly for NPI stringency and residuals of the Level 1 model. For example, a \npositive coefficient indicates that R\nt will increase as NPI stringency decreases. Independent \nvariables were centred and scaled for all models to allow for comparison between model \ncoefficients.  \n \nHoliday periods included Easter, Eid al-Fitr, Eid al-Adha, and National holidays. We used the \ncustom NPI stringency index and performed a sensitivity analysis using the OxCGRT index \n(Supplemental Section 2). We used the Vodafone mobility index in the main model since it is \navailable in more districts, and performed a sensitivity analysis using the Google mobility \nindex (Supplemental Section 3).  \n \nTo determine if the association between mobility, NPIs, and R\nt at different points of the \nepidemic is time-varying, we repeated model training for varying-length periods from 12th \nMarch to t for t in 19th March to 1st September (7-137 days). To understand the influence of \nvarying time periods on model training, we also conducted a sensitivity analysis training the \nmodel in rolling fixed-length periods of 30, 60, and 90 days (Supplemental Section 4).  \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n \nResults \n \nCOVID-19 epidemic in Ghana \nBroadly, the first wave of the national COVID-19 epidemic in Ghana was characterised by an \nearly increase in cases in March and April 2020, followed by a decline in cases over the \nsummer and a resurgence in June and July 2020 (Figure 1). Using aggregated surveillance \ndata for 27 districts included in the estimation of R\nt, we observed variations in the \nprogression of local epidemics in individual districts (Supplemental Figure 1). Patterns in \neach district varied, with case reports ranging from 1 to 250 cases per day, with districts \nreporting cases on average in 89 of 173 days. Ghana introduced a series of NPIs in \nresponse to the growing number of COVID-19 cases in March 2020 (Figure 1). On 1st April \n2020, a partial lockdown was introduced in the Ashanti and Greater Accra regions requiring \nindividuals to remain at home except for essential errands (shopping, healthcare, use of \npublic toilets). The restrictions also prohibited inter-city travel except for essential services. \nLockdown restrictions remained in place until 28th April 2020. \n \n \n \nFigure 1: Confirmed COVID-19 cases and Non-pharmaceutical interventions. The total number \nof confirmed cases of COVID-19 in districts included in this study. The timeline of different non-\npharmaceutical interventions are indicated with dashed lines. \n \nBoth the custom and OxCGRT stringency indices reflect similar patterns in Ghana: a peak \nstringency coinciding with the introduction of the partial lockdown in Ashanti and Greater \nAccra regions, and a following reduction beginning in July 2020 (Supplemental Figure 2). \nBoth stringency indices also identified similar dates of maximal intervention (OxCGRT: 30th \nMarch 2020, Custom: 1st April 2020). \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n \nChanges in Mobility Indicators \n \nBoth Google and Vodafone mobility indicators show similar patterns in Accra and Kumasi \nmetropolitan assemblies (the two areas for which both indicators are available), showing \napproximately baseline values of movement preceding the identification of the first COVID-\n19 cases. Both datasets show abrupt changes coinciding with the introduction of the partial \nlockdown, followed by a more gradual recovery (Figure 2a). Comparing the week before \n30th March 2020 to the following week, mobility decreased by 24% (Vodafone) and 17% \n(Google) in Accra, and in Kumasi by 23% (Vodafone) and 20% (Google). We compared \nmobility indicators for both districts, finding strong evidence of association between  \nVodafone and Google mobility indicators in Accra (R\n2 = 0.92) and Kumasi (R2 = 0.89), \nmeasured between 12th March and 1st September 2020 (Figure 2b, Supplemental Figure \n7). \n \n \nFigure 2. Mobility indicators in Accra and Kumasi Metropolitan Areas. a) A comparison of the \nVodafone and Google mobility indicators in Accra and Kumasi Metropolitan Areas. b) The correlation \nbetween Vodafone and (inverse) Google mobility indicators. Blue dashed line indicates the best fit \nline. This shows a strong correlation between both mobility indicators across the study period. Note \nthat these data are collected from two different sources (Google: GPS, Vodafone: CDRs) and \ndescribe different aspects of mobility (Google: activity in “residential” areas, Vodafone: travel between \nadministrative districts).\n  \n \n \nDistrict-level estimates of Rt  \n \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\nAfter the first detection of cases on March 12th, reported cases and the number of districts \nreporting cases grew until the announcement of the partial lockdown (Figure 3a-c). We found \nRt above 1 (indicating a growing epidemic) after the first reported COVID-19 cases in \nindividual districts and a subsequent decline coinciding with the period of maximal \ninterventions (Figure 3b-c). This was followed by an increase in Rt during the summer of \n2020. While district-specific epidemics followed a broad trend, transmission in individual \ndistricts was characterised by varying patterns of epidemic progression (Supplemental \nFigure 8). We compared Rt estimates one week before and after the announcement of a \npartial lockdown in Ashanti and Greater Accra regions (the date of maximum intervention), \nfinding that between 25th March and 8th April, Rt decreased in 16 of the 27 districts (7 \nmissing). \n \nFigure 3: Estimates of Rt in individual districts. (a) The number of reported cases in individual \ndistricts. (b) Estimates of Rt for individual districts (median shown). (c) A map of districts included the \nanalysis. Colors indicate individual districts. \n \nAssociation between NPI stringency, mobility, and Rt \n \nWe found an optimal lag of mobility, NPI stringency, and holiday dates of 22 days associated \nwith Rt, measured by the maximum marginal R2 of the multilevel model training across all \nperiods. Because Rt estimates did not include estimated delays from infection to reporting, \nthis lag reflects the delay between mobility and NPI, infection, case detection and reporting. \nSensitivity analysis using OxCGRT stringency index and Vodafone mobility indicator, as well \nas Google mobility indicator and Custom stringency index identified optimal lags of 21 and \n19 days, respectively. \n \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\nWe identified correlation between mobility and inverse NPI stringency using the Level 1 \nmodel (Supplemental Table 2). We found that the Level 2 model explained a greater amount \nof variance in Rt and that the strength of association between Rt and NPI stringency was \nhighest during the early epidemic (Figure 4a, Supplemental Figure 9). Across all training \nperiods (173) and lag values (0-30), the maximum marginal R2 was 0.51 (conditional R2: \n0.64) using data between March 12th and May 12th (Figure 4c). We observed higher \nabsolute error in the beginning of case reporting in specific districts (Supplemental Figure \n10). The model identified strong evidence of a positive association between Rt and both \nNPIs and Residual Mobility (Table 1). Positive coefficients indicate an association between \nNPI stringency and Residual Mobility where Rt increases as NPI stringency decreases and \nmobility increases. Note that a positive coefficient for NPI stringency results from the use of \ninverse stringency in the model. We did not find evidence of association between R\nt and \nholidays (Table 1). For this model, we found 15 district-specific random effects \ndistinguishable from 0 (55.6%) (Supplemental Figure 11). \n  \n \n \nFigure 4. Statistical analysis of Rt. a) The marginal R2 of the multilevel model trained on \nvarying-length periods through time. b) Observed vs Predicted Rt for the model trained \nbetween 12th March and 12th May, 2020. c) The maximum marginal R2 for all periods \ntrained for different lag values from 0 to 30 days. Diagonal where x = y shown as blue \ndashed line. \n \n \nPredictors Estimates CI p \n(Intercept) 1.188 1.136 – 1.240 <0.001 \nNPI 0.241 0.228 – 0.255 <0.001 \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n15 \nMobility Residuals 0.039 0.025 – 0.052 <0.001 \nHolidays -0.015  -0.078 – 0.047 0.632 \nTable 1. Regression coefficients for the multilevel model. Regression coefficients for the \nmultilevel model trained between 12th March and 12th May, 2020. Table shows coefficients, \n95% confidence intervals, and p values for each predictor. \n \nThe performance of the model declined through time from June to September, measured by \ndecreasing marginal R\n2 and increased Median Absolute Error. This reflects a period when \nmobility in most districts was recovering while overall, epidemics decreased. The change in \nmodel performance through time may reflect a “decoupling” of transmission from mobility \nand NPI stringency.  \n \nSensitivity analyses using fixed-length periods of 30, 60, and 90 days identified a similar \npattern of model performance during the early epidemic and an increase in model \nperformance in the later epidemic, relative to models trained on varying-length periods \n(Supplemental Section 4).  \n \nDiscussion \n \nWe found that Rt was associated with human mobility and NPI stringency in the early stages \nof the COVID-19 pandemic in Ghana and that this association decreased through time. We \nalso identified a positive association between residual mobility (mobility not explained by \nNPIs) and Rt. In our sensitivity analyses, we found similar optimal lags for both the Custom \nand OxCGRT stringency indices. These lags were greater than those reported in other \nstudies, (Badr et. al. for example, identified an optimal lag of 14 days between mobility and \nCOVID-19 transmission in the USA). In sensitivity analyses included in Supplemental \nSection 2, using Vodafone and OxCGRT data, we found that the OxCGRT stringency index \nexplained a greater amount of variance in R\nt in the end of the study period, which may be \ndue to higher stringency value of the OxCGRT index in this period. We also performed \nsensitivity analyses using mobility indicators from two providers, Google and Vodafone \nGhana, and detected similar results.  \n \nBetween the introduction of COVID-19 in Ghana and the end of the partial lockdown, we \nobserved a relationship between mobility indicators, NPI stringency and R\nt (at the maximum, \nour model explained approximately 64% of variation in Rt). The strength of this relationship \ndecreased in June and July, and especially after August, when Ghana experienced stability \nin the number of reported cases and approximately constant levels of NPI stringency and \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n16 \nMobility. The declining relationship may indicate a disconnection between mobility and Rt as \nthe effect of mobility was mediated by other behavioural changes. It is also notable that we \ndid not observe a decrease in mobility preceding the end of the second wave of the \npandemic in July. This may indicate that mobility data is most useful in the beginning of the \npandemic, when mobility patterns reflect behavioural changes relevant to disease \ntransmission. \n \nThese findings are in line with those from high-income countries which find associations \nbetween decreases in human movement and a reduction in R\nt[12,20,35,36] and provide \nevidence of the utility of mobility measures for understanding transmission in African \ncountries. In particular, we show how human mobility and NPI stringency related to Rt during \nthe early stages of the Ghanaian COVID-19 epidemic and provide a novel analysis of \nsubnational human mobility indicators and local disease surveillance data in a lower-middle \nincome setting. This analysis improves our understanding of the relationship between NPIs, \nMobility and the progression of COVID-19 in Ghana, and how this relationship varied \nthrough time. Future research should focus further on how human mobility indicators can be \nused as a proxy for social contact (and thereby transmission) and how this link changes \nthrough time. Increasing the spatial extent of case reporting data in Ghana could allow for \nmore detailed research in districts outside of major urban areas.  \n \nWe used an R\nt estimation method that supports uncertain generation times via a Bayesian \nprior with mean 3.6 days (standard deviation of mean 0.7 days) for calculating Rt. The use of \nlonger generation times will lead to greater variance in estimates of Rt. This could translate \ninto larger effect sizes (positive or negative) in the statistical model.  \n \nAnalysing only districts with case counts which were available early in the COVID-19 \nepidemic in Ghana may bias our estimates towards urban populations or populations with \ngreater disease surveillance resources. It is also not possible to determine whether the \ntiming of the first reported cases of COVID-19 in individual districts is related to the \nprogression of local epidemics or to the first availability of PCR testing resources in each \ndistrict. The mobility indicators used in this study rely on the aggregated locations of \nsubscribers to mobile networks (Vodafone) and users of internet services (Google). The \nvolume and reporting of these locations may be influenced by varying patterns of mobile \ndevice usage. The demographics of users of either service may also be different from the \ndemographic of the population of Ghana, particularly for Google data which relies on data \ncollected from internet-connected smartphones[37]. Additionally, we used national, not \ndistrict-specific indices of NPI stringency. Neither index includes intervention measures \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n17 \nwhich may have been implemented in local districts but which are not recorded at a national \nlevel.   \n \nIn this study, we identified evidence of positive associations between mobility, NPI \nstringency, and R\nt and show how the strength of this relationship changed through time. We \nfound that mobility and NPI stringency was able to explain variance in Rt during the early \nepidemic but this pattern declined as the epidemic progressed. This decline may reflect a \ndisconnection between disease transmission and behavioural changes measured by mobility \nand NPI indicators. For policymakers and public health decision makers responding to the \nCOVID-19 pandemic, our findings demonstrate that mobility and NPIs were effective for \nestimating disease transmission during the early epidemic, but that subsequent outbreaks \nmay be more related to factors that are not captured in these data.  \n \nData Availability  \n \nData used in this study included individual Line List data shared with authors by the Ghana \nHealth Service. Use of this data was approved by the LSHTM Research Committee (Ref: \n22477) and the Noguchi Memorial Institute of Medical Research (Ref: 048/20-21). Vodafone \nmobility data used in this study is proprietary data shared by Vodafone Ghana in partnership \nwith the Flowminder Foundation and Ghana Statistical Service. This mobility data is \navailable to researchers by application. Google mobility data used in this study is available in \nthe public domain. Downloads of this data can be found in the references. Code used in this \nstudy is available from: https://github.com/hamishgibbs/ghana_rt_npi_mobility\n. \n \nAcknowledgments \n \nThe following authors were part of the Centre for Mathematical Modelling of Infectious \nDisease 2019-nCoV working group. Each contributed in processing, cleaning and \ninterpretation of data, interpreted findings, contributed to the manuscript, and approved the \nwork for publication: Mark Jit, Rachael Pung, Thibaut Jombart, Billy J Quilty, Anna M Foss, \nCarl A B Pearson, Timothy W Russell, David Simons, Stefan Flasche, Graham Medley, C \nJulian Villabona-Arenas, Emily S Nightingale, Fabienne Krauer, Jiayao Lei, Kerry LM Wong, \nJack Williams, Oliver Brady, Arminder K Deol, Yung-Wai Desmond Chan, Akira Endo, Alicia \nShowering, William Waites, Ciara V McCarthy, Nikos I Bosse, Kiesha Prem, Naomi R \nWaterlow, Yalda Jafari, Rachel Lowe, Paul Mee, Megan Auzenbergs, Kevin van Zandvoort, \nJoel Hellewell, Adam J Kucharski, Samuel Clifford, Mihaly Koltai, Christopher I Jarvis, James \nW Rudge, Fiona Yueqian Sun, W John Edmunds, Quentin J Leclerc, Simon R Procter, \nMatthew Quaife, Stéphane Hué, Gwenan M Knight, Nicholas G. Davies, David Hodgson, \nGeorgia R Gore-Langton, Petra Klepac, Emilie Finch, Jon C Emery, Katherine E. Atkins, \nKatharine Sherratt, Alicia Rosello, Sophie R Meakin, Rein M G J Houben, James D Munday, \nSebastian Funk, Lloyd A C Chapman, Frank G Sandmann, Rosanna C Barnard, Charlie \nDiamond, Damien C Tully, Kaja Abbas, Amy Gimma, Kathleen O'Reilly. \n \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n18 \nThe following funding sources are acknowledged as providing funding for the named \nauthors. This research was partly funded by the Bill & Melinda Gates Foundation (INV-\n003174: YL). EDCTP2 (RIA2020EF-2983-CSIGN: HPG, OM, RME, MM). This project has \nreceived funding from the European Union's Horizon 2020 research and innovation \nprogramme - project EpiPose (101003688: YL). HDR UK (MR/S003975/1: RME). This \nresearch was partly funded by the National Institute for Health Research (NIHR) using UK \naid from the UK Government to support global health research. The views expressed in this \npublication are those of the author(s) and not necessarily those of the NIHR or the UK \nDepartment of Health and Social Care (16/137/109: YL; NIHR200908: RME). UK DHSC/UK \nAid/NIHR (PR-OD-1017-20001: HPG). UK MRC (MC_PC_19065: RME, YL). Wellcome \nTrust (210758/Z/18/Z: SA). \n \nFunding \n \nThe following funding sources are acknowledged as providing funding for the working group \nauthors. BBSRC LIDP (BB/M009513/1: DS). This research was partly funded by the Bill & \nMelinda Gates Foundation (INV-001754: MQ; INV-003174: JYL, KP, MJ; INV-016832: SRP; \nNTD Modelling Consortium OPP1184344: CABP, GFM; OPP1139859: BJQ; OPP1183986: \nESN; OPP1191821: KO'R, MA). BMGF (INV-016832; OPP1157270: KA). CADDE \nMR/S0195/1 & FAPESP 18/14389-0 (PM). DTRA (HDTRA1-18-1-0051: JWR). Elrha \nR2HC/UK FCDO/Wellcome Trust/This research was partly funded by the National Institute \nfor Health Research (NIHR) using UK aid from the UK Government to support global health \nresearch. The views expressed in this publication are those of the author(s) and not \nnecessarily those of the NIHR or the UK Department of Health and Social Care (KvZ). ERC \nStarting Grant (#757699: JCE, MQ, RMGJH). ERC (SG 757688: CJVA, KEA). This project \nhas received funding from the European Union's Horizon 2020 research and innovation \nprogramme - project EpiPose (101003688: AG, KLM, KP, MJ, PK, RCB, WJE). \nFCDO/Wellcome Trust (Epidemic Preparedness Coronavirus research programme \n221303/Z/20/Z: CABP, KvZ). This research was partly funded by the Global Challenges \nResearch Fund (GCRF) project 'RECAP' managed through RCUK and ESRC \n(ES/P010873/1: CIJ, TJ). HPRU (NIHR200908: NIB). Innovation Fund (01VSF18015: FK). \nMRC (MR/N013638/1: EF, NRW; MR/V027956/1: WW). Nakajima Foundation (AE). NIHR \n(16/136/46: BJQ; 16/137/109: BJQ, CD, FYS, MJ; 1R01AI141534-01A1: DH; Health \nProtection Research Unit for Modelling Methodology HPRU-2012-10096: TJ; NIHR200908: \nAJK, LACC; NIHR200929: CVM, FGS, MJ, NGD; PR-OD-1017-20002: AR, WJE). Royal \nSociety (Dorothy Hodgkin Fellowship: RL; RP\\EA\\180004: PK). Singapore Ministry of Health \n(RP). UK MRC (LID DTP MR/N013638/1: GRGL, QJL; MC_PC_19065: NGD, SC, TJ, WJE; \nMR/P014658/1: GMK). Authors of this research receive funding from UK Public Health Rapid \nSupport Team funded by the United Kingdom Department of Health and Social Care (TJ). \nUKRI (MR/V028456/1: YJ). Wellcome Trust (206250/Z/17/Z: AJK, TWR; 206471/Z/17/Z: \nOJB; 208812/Z/17/Z: SC, SFlasche; 210758/Z/18/Z: JDM, JH, KS, SFunk, SRM; \n221303/Z/20/Z: MK; UNS110424: FK). No funding (AKD, AMF, AS, DCT, JW, SH, YWDC). \n \n \nReferences \n \n1.  Hale T, Angrist N, Goldszmidt R, et al. A global panel database of pandemic policies \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n19 \n(Oxford COVID-19 Government Response Tracker). Nat Hum Behav. 2021 \nApr;5(4):529–38. \n2.  PHSM [Internet]. [cited 2021 Jul 3]. Available from: \nhttps://www.who.int/emergencies/diseases/novel-coronavirus-2019/phsm \n3.  Liu Y, Morgenstern C, Kelly J, et al. T he impact of non-pharmaceutical interventions on \nSARS-CoV-2 transmission across 130 countries and territories. BMC Med. 2021 Feb \n5;19(1):40. \n4.  Sharma M, Mindermann S, Rogers-Smith C,  et al. Understanding the effectiveness of \ngovernment interventions in Europe’s second wave of COVID-19. medRxiv. 2021 Mar \n26;2021.03.25.21254330. \n5.  Ferguson N, Laydon D, Nedjati Gilani G,  et al. Report 9: Impact of non-pharmaceutical \ninterventions (NPIs) to reduce COVID19 mortality and healthcare demand [Internet]. \nImperial College London; 2020 Mar [cited 2020 Sep 17]. Available from: \nhttp://spiral.imperial.ac.uk/handle/10044/1/77482 \n6.  Flaxman S, Mishra S, Gandy A, et al. Estimating the effects of non-pharmaceutical \ninterventions on COVID-19 in Europe. Nature. 2020 Aug;584(7820):257–61. \n7.  Soltesz K, Gustafsson F, Timpka T, et al. The effect of interventions on COVID-19. \nNature. 2020 Dec;588(7839):E26–8. \n8.  Gibbs H, Liu Y, Pearson CA, et al. Changing travel patterns in China during the early \nstages of the COVID-19 pandemic [Internet]. Public and Global Health; 2020 May [cited \n2020 Aug 27]. Available from: \nhttp://medrxiv.org/lookup/doi/10.1101/2020.05.14.20101824 \n9.  Gibbs H, Nightingale E, Liu Y, et al. Detecting behavioural changes in human \nmovement to inform the spatial scale of interventions against COVID-19. medRxiv. \n2021 Feb 15;2020.10.26.20219550. \n10.  Buckee CO, Balsari S, Chan J, et al. Aggregated mobility data could help fight COVID-\n19. Science. 2020 Apr 10;368(6487):145–6. \n11.  Jacobsen GD, Jacobsen KH. Statewide COVID-19 Stay-at-Home Orders and \nPopulation Mobility in the United States. World Med Health Policy. 2020;12(4):347–56. \n12.  Association between mobility patterns and COVID-19 transmission in the USA: a \nmathematical modelling study - The Lancet Infectious Diseases [Internet]. [cited 2021 \nMay 13]. Available from: https://www.thelancet.com/journals/laninf/article/PIIS1473-\n3099(20)30553-3/fulltext \n13.  Tizzoni M, Bajardi P, Decuyper A, et al. On the Use of Human Mobility Proxies for \nModeling Epidemics. PLOS Comput Biol. 2014 Jul 10;10(7):e1003716. \n14.  Grantz KH, Meredith HR, Cummings DAT, et al. The use of mobile phone data to \ninform analysis of COVID-19 pandemic epidemiology. Nat Commun. 2020 Sep \n30;11(1):4961. \n15.  Disease Prevention Maps [Internet]. Facebook Data for Good. [cited 2020 Sep 17]. \nAvailable from: https://dataforgood.fb.com/tools/disease-prevention-maps/ \n16.  Overview - Community Mobility Reports Help [Internet]. [cited 2021 May 11]. Available \nfrom: https://support.google.com/covid19-mobility/answer/9824897?hl=en-GB \n17.  COVID\n/i119 - Mobility Trends Reports [Internet]. Apple. [cited 2021 May 11]. Available \nfrom: https://www.apple.com/covid19/mobility \n18.  Quilty BJ, Diamond C, Liu Y, et al. The effect of inter-city travel restrictions on \ngeographical spread of COVID-19: Evidence from Wuhan, China. medRxiv. 2020 Apr \n21;2020.04.16.20067504. \n19.  Chen Z-L, Zhang Q, Lu Y, et al. Distribution of the COVID-19 epidemic and correlation \nwith population emigration from Wuhan, China. Chin Med J (Engl) [Internet]. 2020 Feb \n28 [cited 2021 Apr 15]; Available from: \nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC7147281/ \n20.  Kissler SM, Kishore N, Prabhu M, et al. Reductions in commuting mobility correlate \nwith geographic differences in SARS-CoV-2 prevalence in New York City. Nat \nCommun. 2020 Sep 16;11(1):4674. \n21.  Marshall JM, Wu SL, Sanchez C. HM, et al. Mathematical models of human mobility of \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint \n\n20 \nrelevance to malaria transmission in Africa. Sci Rep. 2018 May 16;8(1):7713. \n22.  Wesolowski A, Eagle N, Tatem AJ, et al. Quantifying the Impact of Human Mobility on \nMalaria. Science. 2012 Oct 12;338(6104):267–70. \n23.  Monitoring human mobility during the COVID-19 lockdown in Ghana using mobile \nphone data from Vodafone [Internet]. Migration data portal. [cited 2021 Aug 10]. \nAvailable from: http://migrationdataportal.org/data-innovation/human-mobility-COVID-\n19-Ghana-mobile-phone-data \n24.  Nouvellet P, Bhatia S, Cori A, et al. Reduction in mobility and COVID-19 transmission. \nNat Commun. 2021 Feb 17;12(1):1090. \n25.  Ghana Health Service. Ghana Confirms 2 Cases of COVID-19 [Internet]. [cited 2021 \nJul 3]. Available from: \nhttps://www.ghanahealthservice.org/covid19/downloads/covid_19_first_confirmed_GH.\npdf \n26.  Ofori-Adjei D, Lartey M, Koram KA. Ghana and the COVID-19 pandemic. Ghana Med \nJ. 2020 Dec;54(4 Suppl):1–2. \n27.  Guo J, Gabry J, Goodrich B, et al. rstan: R Interface to Stan [Internet]. 2020 [cited 2021 \nAug 25]. Available from: https://CRAN.R-project.org/package=rstan \n28.  Stan [Internet]. stan-dev.github.io. [cited 2021 Aug 25]. Available from: //mc-stan.org/ \n29.  Estimate Real-Time Case Counts and Time-Varying Epidemiological Parameters \n[Internet]. [cited 2021 Aug 5]. Available from: https://epiforecasts.io/EpiNow2/ \n30.  Abbott S, Hellewell J, Thompson RN, et al. Estimating the time-varying reproduction \nnumber of SARS-CoV-2 using national and subnational case counts. Wellcome Open \nRes. 2020 Jun 1;5:112. \n31.  Riutort-Mayol G, Burkner P-C, Andersen MR , Solin A, Vehtari A. Practical Hilbert space \napproximate Bayesian Gaussian processes for probabilistic programming. 2020; \n32.  Mee P, Alexander N, Mayaud P, et al. Tracking the emergence of disparities in the \nsubnational spread of COVID-19 in Brazil using an online application for real-time data \nvisualisation. medRxiv. 2021 May 4;2021.04.30.21256386. \n33.  Vodafone Ghana, Flowminder. Ghana Human Mobility CDR Aggregates. 2021 Jun 5; \n34.  Covid-19: Africa’s game-changing variants [Internet]. The Africa Report.com. 2021 \n[cited 2021 Oct 15]. Available from: https://www.theafricareport.com/107175/covid-19-\nafricas-game-changing-variants/ \n35.  Bergman NK, Fishman R. Correlations of Mobility and Covid-19 Transmission in Global \nData. :56. \n36.  European Commission. Joint Research Centre. How human mobility explains the initial \nspread of COVID-19: a European regional analysis. [Internet]. LU: Publications Office; \n2020 [cited 2021 Aug 16]. Available from: https://data.europa.eu/doi/10.2760/61847 \n37.  Milusheva S, Bjorkegren D, Viotti L. Assessing Bias in Smartphone Mobility Estimates \nin Low Income Countries. In: ACM SIGCAS Conference on Computing and Sustainable \nSocieties (COMPASS) [Internet]. Virtual Event Australia: ACM; 2021 [cited 2021 Oct \n20]. p. 364–78. Available from: https://dl.acm.org/doi/10.1145/3460112.3471968 \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted November 17, 2021. ; https://doi.org/10.1101/2021.11.01.21265660doi: medRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}