We try to collect data from a number of reports on the same disaster, each specifying whether its figures refer to individuals or households, the reporting terms and sources used, the publisher, the title of the source document and the date of publication. When possible we triangulate the figures using competing reports. Sometimes, however, our estimates are derived from a single report. In others, they are the aggregation of a number of reports that together cover the wide geographical area affected. This dataset allows us to better interpret the context of the figure in each report. In determining our estimates, it is vital that the data selected represents the most comprehensive figure from the most reliable source available for that event at the time when data was collected. Reporting bias We are aware that our methodology and data may be subject to different types of reporting bias, some of which are detailed below. Unequal availability of data: Global reporting tends to emphasise large events in a small number of countries where international agencies, funding partners and media have a substantial presence, or where there is a strong national commitment and capacity to manage disaster risk and collect information. Under-reporting of small-scale events: These are far more common, but less reported on. Disasters that occur in isolated, insecure or marginalised areas also tend to be under-reported because access and communications are limited. “Invisible” IDPs: There tends to be significantly more information available on IDPs who take refuge at official or collective sites than on those living with host communities and in other dispersed settings. Given that in many cases the vast majority fall into the second category, figures based on data from collective sites are likely to be substantial underestimates. Real-time reporting is less reliable, but later assessments may underestimate: Reporting tends to be more frequent but less reliable during the most acute and highly dynamic phases of a disaster, when peak levels of displacement are likely to be reached. It becomes more accurate once there has been time to make more considered assessments. Estimates based on later evaluations of severely damaged or destroyed housing will be more reliable, but they are also likely to understate the peak level of displacement, given that they will not include people whose homes did not suffer severe damage but who fled for other reasons. Our estimates for some disasters are calculated by extrapolating from the number of severely damaged or destroyed homes or the number of families in evacuation centres. In both cases we multiply the housing and family data by the average number of people per household. Estimating average household size Primary sources often report the number of homes rendered uninhabitable or the number of families displaced, which we convert into a figure for IDPs by multiplying the numbers by the average household size (AHHS). There is, however, no universal dataset with updated and standardised AHHS data for all countries. Given the potentially significant influence of AHHS on our estimates, we have continued to update the data and methodology we use to calculate it. This year we used a linear extrapolation obtained with improved methodology developed for the GRID 2016.3 The AHHS and therefore our estimates are subject to a margin of error, which means that by applying a particular value we may underestimate or overestimate real figures. If possible we review and update the AHHS every year and, as a general rule, when data is expressed in household or family units, we estimate the number of displaced people according to the AHHS for the year when the data is captured. This applies particularly to figures obtained from historical or retrospective research, notably in protracted or prolonged displacement cases where using a contemporary household size without accounting for demographic changes would have lead to an underestimate for an event that occurred in 2008 (see table A.3). METHODOLOGICAL ANNEX 101