Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Journal of Environmental Radioactivity 233 (2021) 106582

Contents lists available at ScienceDirect

Journal of Environmental Radioactivity


journal homepage: http://www.elsevier.com/locate/jenvrad

Estimating population lung cancer risk from radon using a resource


efficient stratified population weighted sample survey protocol – Lessons
and results from Ireland
P. Murphy a, *, A. Dowdall b, S. Long b, B. Curtin a, D. Fenton b
a
School of Mathematics and Statistics, University College Dublin, Belfield, Dublin 4, Ireland
b
Environmental Protection Agency, McCumiskey House, Richview, Clonskeagh, Dublin 14, Ireland

A R T I C L E I N F O A B S T R A C T

Keywords: A 2018 estimate indicates that there were 226,057 radon-attributable lung cancer deaths in 66 countries that had
Radon representative radon surveys. This is a shocking figure, and as it comes from only 66 countries it underestimates
Lung cancer risk the worldwide death toll. Any research that enables countries to conduct representative radon surveys and to
Population weighted radon survey
understand better the risk to citizens from radon is surely welcome. We hope this paper provides a useful
National radon survey
Survey protocol
methodology for estimating population risk.
The estimation of population weighted average indoor radon levels requires statistically valid sampling
methodologies that use a representative sample of occupied homes throughout the country. A literature review
indicates that in many population weighted surveys, the sampling methodology may not have been designed to
do this. This paper describes a simple, resource efficient methodology which produces statistically valid and
reliable estimates based on a small scale sample that is representative of the population distribution. The
resource efficient design of this study enables it to be repeated at frequent intervals providing for a longitudinal
analysis of the population risk from indoor radon.
This survey was conducted in Ireland using 653 measurements and a representative sampling strategy to
provide a baseline population weighted radon exposure for future comparisons. This study estimates the average
population weighted indoor radon concentration in Ireland to be 97.83 Bq m− 3 (95% Confidence Interval 90.69
Bq m− 3 to 105.53 Bq m− 3), and that there are an estimated 350 lung cancer cases and 255 deaths per year due to
radon exposure. The mortality rate of 5.3 per 100,000 due to indoor radon, demonstrates that radon remains one
of the highest preventable causes of death in Ireland.

1. Aims 2. Introduction

This paper has two main aims: Gaskin et al. (2018) quantify the scale of the risk from radon showing
that in 2012 there were 226,057 radon-attributable deaths in 66 coun­
• to describe the design of a multi-stage stratified and population tries that possessed a representative radon survey. This actually un­
weighted survey methodology which may be useful to agencies derestimates the global burden and shows how important it is to be able
interested in measuring the population exposure to indoor radon to accurately estimate population radon risk. Most European countries
levels, (Dubois et al., 2005) and many other countries have conducted radon
• to report on the implementation of this new design in measuring the surveys to establish indoor radon concentrations either at national level
population radon risk in Ireland and to produce associated updated or regional level. While much attention is often given to the measure­
estimates. ment process, no uniform sampling strategy has been employed in these
surveys. A review of survey methodologies demonstrates that some have
used sampling methods that are less than ideal. Convenience sampling

* Corresponding author.
E-mail address: Patrick.Murphy@ucd.ie (P. Murphy).

https://doi.org/10.1016/j.jenvrad.2021.106582
Received 14 March 2019; Received in revised form 3 March 2021; Accepted 7 March 2021
Available online 10 April 2021
0265-931X/© 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
P. Murphy et al. Journal of Environmental Radioactivity 233 (2021) 106582

and volunteer sampling have both been used in the past and when introduction of building regulations had an average indoor radon level
surveys claim to use random sampling it is not always evident that the of 86 Bq m− 3 while homes built after 1998 (and so subject to the reg­
correct sampling frame has been used. Flaws such as these in the sam­ ulations) had an average of 64 Bq m− 3, indicating the effectiveness of the
pling methodology should be avoided as they have the potential to regulations at reducing indoor radon levels.
introduce bias and to raise questions about the accuracy of the average The GW-NRS was stratified by geography, the distribution of houses
radon concentrations reported. As always in statistics, samples should in Ireland is not homogenous by geography. The GW-NRS was never
ideally be chosen to be representative of the population of interest. intended to measure population risk. Measuring such risk required the
In reporting national average radon concentrations, care should be development of a second survey protocol which would take into account
taken to distinguish between the geographically weighted national the actual distribution of houses in Ireland.
average radon concentration and the population weighted national The primary aim of the 2002 NRS was to determine in detail the
average. These are two different statistics and ideally, they should be geographical distribution of radon levels in dwellings throughout
measured using different surveys. Where they are measured in parallel, Ireland (Fennell et al., 2002). While a population weighted average was
the survey needs to be properly designed to do both (WHO, 2009). reported in the 2002 NRS, that report states clearly that the survey was
However, the literature shows that separate surveys are not used in not a population-based survey but was instead “geographically-based”.
practice and that the population weighted estimates are usually This means that the sample chosen for the 2002 survey was not intended
computed using already-collected data by applying a weighting factor to be representative of the population distribution but was instead
based on census data (Ajrouche et al., 2018; Suzuki et al., 2010; Wrixon chosen to be geographically representative. This lack of representa­
et al., 1988; Health Canada, 2012). tiveness of the population distribution mean in the 2002 NRS means that
We propose a methodology for computing the population weighted the results of that survey and the one described are not directly com­
indoor radon concentration using a sample designed to be representative parable with the 2002 NRS results.
of the national population distribution. In addition to providing a more
representative sample, this method has the advantage of being resource 3. Survey methodology
and time efficient.
This paper acts as an accompaniment to previous work (Dowdall 3.1. Survey design objectives
et al., 2017) which detailed a new Geographic Weighted National Radon
Survey (GW-NRS) used to estimate Ireland’s national average indoor The design objective of this survey was to establish a protocol that
radon concentration in 2015. Prior to that 2015 survey, official esti­ was not resource intensive and could be repeated on a regular basis. The
mates of Ireland’s national radon average and the associated population survey should be capable of accurately estimating the population
risk came from the 2002 National Radon Survey (NRS) (Fennell et al., weighted national average radon concentration. We note that the large
2002). The 2002 NRS was a very time consuming and resource intensive scale 2002 NRS provided risk estimates for each of the 10 km2 grid
project, with 11,319 annual measurements performed during a squares in Ireland, reproducing such grid estimates was not an aim for
seven-year period. The 2002 survey split the country into 10 km grid this project. With both of these objectives in mind, the optimum design
squares using the Irish National Grid system and estimates of the per­ was to use a Multi-Stage Stratified Random Sampling scheme.
centage of houses in each grid that exceeded the national Reference There are two main advantages to a stratified random sampling
Level of 200 Bq m− 3 were produced. These estimates underpin radon scheme over simple random sampling:
preventive measures introduced in Irish Building Regulations since 1998
(Stationery Office, 2008) and workplace requirements introduced in • Firstly, by stratifying the population we are forcing our sample to be
2019 (Stationery Office, 2019). more representative of the population instead of relying on chance to
As detailed in Dowdall et al. (2017) a number of developments have provide that representativeness,
changed the housing environment in Ireland since 2002, including: large • Secondly, it is possible to obtain estimates for each of the established
growth in housing in Ireland between 1996 and 2011, the introduction strata and not just for the entire population.
of the Building regulations in 1998 requiring radon preventive mea­
sures, changes in house types and increased energy efficiency of homes. 3.2. Multi-stage stratified random sampling
These changes required the development of new methods to produce
updated estimates of the national average radon concentration and the As mentioned, the 2002 NRS utilised the Irish National Grid system
population exposure to indoor radon. which splits Ireland into 837 grid squares each with the same 10 km ×
In 2014, a National Radon Control Strategy (NRCS) for Ireland was 10 km dimensions. These grid squares conveniently allow us to create an
adopted by Government (DCCAE, 2014). The NRCS aims to minimise the efficient two stage sampling scheme to select homes for our population
exposure to radon gas for people in Ireland. The Strategy called for the weighted survey.
development of a suite of metrics to help monitor its effectiveness over As the survey was to produce a population weighted estimate we
time. The metrics identified to do this included the geographic weighted would stratify the 837 grids according to their population density. No
and the population weighted national average radon concentrations. information for population density of these grid squares was available so
The scale of the 2002 NRS made it difficult and time consuming to repeat we used housing density instead as a proxy and before conducting our
and the estimates using that survey methodology were not updated. survey we checked and confirmed the validity of this approach, this
Therefore, new survey methodologies were required which would allow validation is described in Section 4.
for easy and time efficient updating of these metrics on a frequent basis We begin by stratifying all of the population of NG = 837 grids into
thereby providing an ongoing series of estimates through time. 10 strata. This is accomplished by ranking all of the grids according to
We divided the task into two separate projects; a geographic housing population density and splitting into 10 deciles, so that each
weighted survey and a population weighted survey. The first which was stratum contains approximately 10% of the housing population.
described in detail in Dowdall et al. (2017) used a Geographic Weighted Ideally each stratum would contain identically 10% of the popula­
Sample Survey to produce an updated national average radon concen­ tion but as we are stratifying by grids and strata contain whole grid
tration and to analyse the results by build date to explore the effect of squares there is a very small deviation from 10%. While the number of
1998 Building regulations aimed at reducing indoor radon levels. The houses Nk, will be approximately the same in each stratum (k) the
results of that project (GW-NRS) showed that the national average radon number of grids in each stratum, NGk, will differ as some grids are
concentration had fallen from 89 Bq m− 3 reported by the 2002 NRS to 77 substantially more populated than others.
Bq m− 3. Analysis by build date showed that homes built before the 1998 Having performed this stratification we now conduct two sampling

2
P. Murphy et al. Journal of Environmental Radioactivity 233 (2021) 106582

stages with the aim of choosing a sample of houses such that 10% of the This process involved first creating a copy of the Geodirectory in
sample comes from each of the 10 strata. This provides us with a sample Microsoft Access. Using Structured Query Language (SQL) we created a
that is representative of the population in terms of housing density. series of 837 queries to select buildings from the database that corre­
In Stage 1 sampling, we use simple random sampling (SRS) to select sponded to the grid squares.
10 independent samples of grids, one for each stratum (k). Stage 2 For example to select all buildings in grid square “1607”, we created
sampling involves selecting houses from the samples of grids chosen in a query to retrieve records where addresses were within the Easting
stage 1. Within each stratum we use SRS on each of the grids chosen in coordinates 160,000 and 170,000, and within the Northing coordinates
Stage 1, selecting the same number of houses from each grid in that 70,000 and 80,000, and stating the property use to be residential. A full
stratum. Using this strategy we construct 10 identically sized samples of description of the grid squares is contained in Fennell et al. (2002) and
houses, n1 = n2 = … = n10. To determine the number of grids and houses the procedure for identifying gird square co-ordinates is described in
for use in this sampling scheme we used Monte Carlo simulation as Appendix II of that report.
described in detail in Dowdall et al. (2017). Fig. 1 provides a graphical The records generated by each SQL query were then exported as an
representation of our sampling scheme. excel spreadsheet for permanent availability. Thus we produced a set of
In the following section we describe the sampling frame used in this 837 spreadsheets containing all the residential addresses in the Geo­
project. directory database assigned by grid square. These spreadsheets formed
the sampling frame for our survey. This process was carried out by
Curtin (2016) as part of a Master of Arts degree. It should be noted that
3.3. Sampling frame
this process could easily be automated in visual basic or another pro­
gramming language for future iterations of the survey.
The Geodirectory (2016) is a database containing information on
over 2.2 million addresses in Ireland. It was created in the year 2000
through a collaboration between An Post (Irish Postal Service) and 3.4. Estimation methodology
Ordinance Survey Ireland in order to provide a unique verified standard
address to each property in Ireland. When a property is added to the Having described the methodology for selecting our samples, the
directory it is given its own unique building ID number. Following this, a remaining task of this section is to detail how we will estimate relevant
range of information about each property is added to the record, this statistics from our final data. Because of the stratified nature of our
includes; thoroughfare, county, postal town, whether the property is survey we can produce estimates for each stratum and also an overall
vacant, derelict, a holiday home, and whether the property has a com­ estimate for the total population using weighting techniques for strati­
mercial or residential use. fied surveys (Nguyen and Murphy, 2015). These overall weighted esti­
The database provides exact geographical locations for each property mates are clearly designed to accurately take into account the
in two forms; longitude and latitude and geo-codes. A geo-code is a population distribution in the country and thus differ from the estimates
unique geographical reference for each property in the database. It produced in Dowdall et al., (2017) which were geographically weighted
consists of a generated X and Y coordinate based off a false origin located estimates. We are interested in estimating both the mean indoor radon
off the South-West of the country in the Atlantic Ocean. This is known as levels and the proportion of houses above a given threshold. Table 1
the Irish National Grid. These X and Y coordinates (called Eastings and provides details on how to compute the sample mean and variances at
Northings) express, in terms of kilometres, the distance of the property the stratum level and the overall level. Having estimated the stratified
from the false origin. mean X and variance s2 , we can use these to fit an estimated lognormal
While the Geodirectory contains a large set of variables for each distribution for the population. From this distribution we can compute
building in the database the only variables of interest to us were building the proportion of houses above a given threshold in the population. The
ID, address lines 1–6, county, building use, easting coordinate and background radon level of 6 Bq m− 3 is factored into the computation to
northing coordinate. The database did not record the grid square for ensure a good fit to the lognormal model (Gunning et al., 2014).
each building. We had to create this identification between building and To compute a confidence interval for the mean of the lognormal
grid square for each of the 837 grid squares. distribution we use a method due to Cox (Olsson, 2005). We firstly

Fig. 1. Graphical representation of the multi-stage sampling scheme employed.

3
P. Murphy et al. Journal of Environmental Radioactivity 233 (2021) 106582

Table 1 6000 houses from the sample of 60 grids with 600 houses from each
Sample Statistics under stratified random sampling procedure. stratum. We did not intend to make 6000 measurements but needed this
Stratum Specific Estimate Estimate for Total Population larger sample to account for non-response that would occur when in­
∑nk ∑10 vitations to householders were issued. The response rate to similar in­
Mean i=1 xk,i k=1 Nk xk
10
vitations in the 2002 NRS was typically 21%, ranging from 17 to 36%

xk = X = = Wk xk
n
∑nk k ∑10N k=1
depending on location and in the GW-NRS, an overall response rate of
Variance (xk,i − xk )2 k=1 N k s 2 ∑10
s2k = i=1 s2 = k
Wk s2k
22% ranging from 7 to 50% was observed (Dowdall et al., 2017). The
=
nk − 1 N k=1

Notation in Table 1. number of addresses selected from each grid square differed between
N = Population size. strata since each stratum contained a different number of grid squares. If
10 = Number of strata in the total population. necessary, the sampling procedure can be adjusted by increasing either
Nk = Number of houses in the population that are in stratum k. the number of grids or the number of homes selected per grid to produce
nk = Number of houses to be randomly selected from stratum k. a larger sample.
xk,i = The annual average radon concentration measurement for the ith element Table 2 and Table 3 provide all relevant details on stages 1 and 2 of
of the kth stratum.
the sampling procedure. Fig. 2 shows the actual grids that were selected
Wk = The proportion of the total population belonging to stratum k (also known
and can be contrasted with the grids selected in the GW-NRS. It can
as design weights).
S2k = The sample variance of the kth stratum.
clearly be seen that the grids in the PW-NRS are not uniformly spread
across the country as they were for the GW-NRS but are instead
distributed according to the population distribution in Ireland. In
compute the following intervals for Y = log(X), then we back transform
particular the selection of grids around Dublin city is evident. Approx­
to get confidence intervals for X by exponentiating the end points of this
imately 39% of the Irish population was reported to be living in Dublin
interval.
city and suburbs in the last census (CSO, 2017).
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
SY2 S2Y S4Y
Y + ∓ Zcrit + 4.2. Validation of sample selection
2 n 2(n − 1)

To estimate the population proportion above the reference level of While this survey aims to stratify the country by population, we
200 Bq m− 3 we first fit a lognormal distribution to the data using the mentioned earlier that we did not have accurate population density
mean and variance estimated above. Then computing the relevant information for each grid square that would allow such stratification.
probability from this lognormal distribution we estimate the population Instead we used housing density as a proxy for population density. To
proportion. Using this same method with the end points of the confi­ check whether this substitution was acceptable, we made use of data on
dence interval for the mean we can compute a confidence interval for the average number of persons in private households from the Census of
the population proportion. Population conducted by the CSO (CSO, 2012). Using the CSO data we
were able to compute the average population per household for each of
4. Sample selection and validation the 60 grid squares in our sample. We knew the number of households in
each grid square and so we could compute a weighted average popu­
This section describes in detail the selection of homes to be surveyed lation per household for the total sample of 60 grid squares.
and the validation performed to ensure our sample was representative of The quantity of grid squares for each of the cities or counties
the population. included in the sample was tallied and multiplied by the average pop­
ulation per household for the respective area. Averaging over all these
estimated population per household for the area in the sample, an
4.1. Selecting the sample of houses to be surveyed estimated state average population per household was calculated. Our
computations yielded an average of 2.715 for the 60 grids which com­
In section 3.3 we saw that the sampling frame for this survey consists pares very favourably with the value of 2.7 produced by the CSO for the
of 837 spreadsheets containing residential homes for each 10 km2 grid in entire state. The close agreement between our predictions for the na­
Ireland. These spreadsheets contained a total of 1.783 million dwellings tional average number of persons per household and the official CSO
which is the population for this survey. The stratification procedure census value suggests that our use of housing density instead of popu­
described in section 3.2 was carried out: lation density does not introduce any major source of bias.

• The grid squares were ordered by housing density and then separated
into ten deciles from the most densely populated grid squares down
to the least populated grid squares. Table 2
• Each decile (stratum) contained approximately 178,300 homes. Details of stratification and sample selection procedure in PW-NRS.
Strata No. of grid No. of grid squares No. of homes
The simulation performed in Dowdall et al. (2017) indicated that 60 squares in sampled at stage 1 sampled per
population grid at stage 2
grid squares were required to make up the desired sample. As mentioned
strata
in section 3.2 the number of grids in each stratum varies. The three
strata corresponding to the most densely populated grid squares, con­ 10 (high density sq.) 1 1 600
9 3 3 200
tained only one, three, and six grid squares respectively. In the case of 8 6 6 100
these three most densely populated grids all these grid squares were 7 12 6 100
selected as part of the sample. Using a random number generator, we 6 18 6 100
selected a simple random sample of grid squares from each of the 5 33 6 100
4 66 6 100
remaining seven strata. A greater number of grid squares were sampled
3 108 7 86/85
from the lower density strata, which contained much more grid squares, 2 159 9 67/66
to counteract the low number of grid squares in the more population 1 (low density sq.) 432 10 60
dense strata. This procedure was carried out in order to reach a random
sample of 60 grid squares.
Having completed stage 1 of the sampling we now sampled a total of

4
P. Murphy et al. Journal of Environmental Radioactivity 233 (2021) 106582

Table 3 protocol for the measurement of radon in homes and workplaces (EPA,
Details of survey protocol. 2019). Each measurement was individually seasonally corrected using
Total no. of invitations issued 7593 the start and end date of that measurement.
No. of invitations returned to EPA undelivered 1708 Each participant in the study was also asked to complete a short one
No. of invitations assumed successfully delivered 5885 page questionnaire to gather information on the build date of the home
No. of responses received 1243
and some other building characteristics. Each participant in the survey
% response rate 21
No. of homes selected for measurement 738 received a test report containing their measurement results, if their
No. of valid measurements completed 653 result was found to be above the reference level of 200 Bq m− 3, they
No. of questionnaires completed 598 were provided with advice on radon remediation.
No. of questionnaires completed with valid measurement 574

5.2. Study protocol


5. Measurement protocol and data
In this section we provide a brief description and some relevant
5.1. Radon measurement statistics relating to the study protocol. In July 2016, invitations were
sent to the homes selected for our sample. The required number of re­
The measurement of indoor radon levels in this study followed the sponses was achieved in 42 grid squares. However, in 18 grid squares the
same protocol described in detail in Dowdall et al. (2017) and EPA required number of responses was not achieved. In August 2016, a
(2019). Participants received two CR-39 detectors to measure radon second round of invitations was issued to additional homes randomly
levels in the main living area and main bedroom. Detectors were ana­ selected from these 18 grid squares in an effort to achieve the required
lysed by the EPA to ISO 17025 accredited procedures. An internal EPA response rate in these squares.
review of the uncertainties on radon measurements was carried out in
2015 (EPA, 2015). This review found that using a coverage factor of k = 6. Exploratory data analysis
2 (95% confidence interval), the uncertainty was ±20%. This quantifies
the measurement uncertainty, there is also sampling variability from the 6.1. Outlier detection
procedure described in Section 4. The confidence intervals produced
below implicitly include both sources of uncertainty. In this section we report on the exploratory analysis performed to
It is widely reported that radon concentrations in homes can vary identify potential outliers and to establish if the data followed a
considerably with season. Burke et al. (2010) produced a set of seasonal lognormal distribution. Radon data has been empirically shown to
correction factors for indoor radon levels which could be used to closely follow the log-normal distribution in numerous studies across the
calculate a seasonally adjusted annual average radon concentration world over the last 30 years. This empirical result was first observed by
from measurements made at different times of the year. For this survey it Nero et al. (1986). The reason for this close lognormal behaviour is not
was decided that all measurements should be made at the same time of well understood but remains an empirical result, for a discussion see
year. The measurement time chosen was such that the seasonal correc­ Bossew (2010) and Daraktchieva et al. (2014).
tion factors set out in the EPA’s protocol for the measurement of radon in This log-normality would of course yield a large number of outliers if
homes (EPA, 2019) are close to unity. Taking these factors into we plotted the raw radon measurements. To check for outliers in the
consideration, measurements for the survey were scheduled for the three data we produced box plots for each of the ten strata. We subtracted a
month period between September and November inclusive when the background concentration of 6 Bq m− 3 from all of the measurements and
seasonal correction factor for this period is 0.96. The application of then log transformed the data. Fig. 3 shows the box plots for all ten strata
seasonal correction factors is described in Appendix 1 of the EPA and as can be seen we identified only three potential outliers. Similarly,
previous studies in Ireland (Organo and Murphy, 2007) have shown a

Fig. 2. Grid squares selected in GW-NRS (left) and PW-NRS (right).

5
P. Murphy et al. Journal of Environmental Radioactivity 233 (2021) 106582

Fig. 3. Boxplots showing relative absence of outliers in log transformed radon measurements from 10 strata.

number of extreme radon measurements that exceed what would be we also performed a Kolmogorov-Smirnov hypothesis test for each of the
predicted if the data followed a perfectly lognormal model. Most re­ ten strata. This test has the null hypothesis that the data is lognormal and
searchers have found that radon data sets do deviate slightly from small p-values would be a cause for concern as they would indicate a
lognormal in the upper tail of the distribution, usually above 500 Bq m− 3 statistically significant deviation from log-normality. The p-values for
but this does not affect predictions of the proportion above 200 Bq m− 3 these tests are shown in Table 4 and clearly none of these p-values are
(Bossew, 2010; Daraktchieva et al., 2014). below the 0.05 level of significance. As the Kolmogorov Smirnov test is
not the most powerful test we also performed an Andersen Darling test
on this same data. This test is more stringent, however by removing just
6.2. Tests for log-normality 4 outliers from the data set we found the Andersen Darling test also
yielded no significant p-values. We expected that there would be some
To check for log-normality of the data we produced normal quantile deviations at the extreme higher end of the distribution from the
(qq) plots for each of the 10 strata (see Fig. 4). Again we subtracted the 6 lognormal model, this is well discussed in the literature (Bossew 2010
Bq m− 3 background concentration and log-transformed the resulting and Daraktchieva et al., 2014). In our experience with many other data
measurements before producing the plots. While these plots show a very sets from Ireland the lognormal model fits well apart from the very
small deviation from normality the plots are generally well behaved. extreme upper tail above 600 Bq m− 3. This is again true here. In
Because it is not possible to tell if this small deviation is significant

Fig. 4. QQ plots showing approximate log-normality of radon data for the ten strata.

6
P. Murphy et al. Journal of Environmental Radioactivity 233 (2021) 106582

Table 4 Radon Control strategy, future population weighted surveys will be


P-values for Kolmogorov-Smirnov and Andersen Darling tests of normality. compared to this baseline and any reductions or increases observed can
Stratum KS Test p-value on full data AD Test p-value after removing four be statistically test for significance.
No: set outliers

Stratum 1 0.53 0.10 7.3. Proportion of homes exceeding the reference level
Stratum 2 0.87 0.53
Stratum 3 0.77 0.17 Using the procedure from section 3.4 we have also estimated the
Stratum 4 0.49 0.20
proportion of homes exceeding the 200 Bq m− 3 reference level at 7.50%
Stratum 5 0.92 0.47
Stratum 6 0.49 0.07 with a 95% confidence interval of 6.46%–8.66%. This equates to a
Stratum 7 0.19 0.06 rounded estimate of 134,000 homes exceeding the reference level in
Stratum 8 0.95 0.54 Ireland, with a rounded 95% confidence interval of 115,000 to 154,000.
Stratum 9 0.73 0.07 Nationally the number of people exposed to radon above the reference
Stratum 10 0.81 0.77
level is estimated as 361,000, rounded 95% confidence interval
(310,000, 417,000).
particular the lognormal model allows accurate predictions for each of
the ten strata of the proportion of homes above the reference level of 200 7.4. Lung cancer risk
Bq m− 3. Predicting the proportion of homes above a much higher
threshold such as 1000 Bq m− 3 would not be reliable using this model. According to Darby et al. (2005), the risk of lung cancer increases by
an estimated 16% (95% confidence interval 5.0%–31%) per 100 Bq m− 3
7. Results of usual indoor radon assuming a linear no threshold dose response. The
2017 Population Weighted NRS (PW-NRS) analysis of radon levels
7.1. Radon measurement results yielded an estimated 97.83 Bq m− 3 (95% confidence interval (90.69,
105.53)).
While there was not perfect uniformity in the measurements made We can quite easily combine this information to produce an estimate
across each stratum, approximately equal numbers of measurements of the additional lung cancer risk in Ireland. Darby et al. determined that
were obtained from each decile. We did not require perfect uniformity as the risk is linear in nature and that the increased risk due to an increase
the estimation procedure outlined in section 3.4 includes a weighting of 1 Bq m− 3 radon is constant regardless of the level of radon. This
mechanism to ensure estimates are unbiased. Table 5 provides a concise linearity in the risk allows us to determine that there is an increase in
summary of the measurement data from this study. risk of 0.16% per 1 Bq m− 3.
A total of 64 homes surveyed had radon concentrations exceeding The linearity also means that if we wish to compute an estimate for
the 200 Bq m− 3 national reference level. The maximum concentration the risk for the entire population we can apply the risk to the 97.83 Bq
measured was 1393 Bq m− 3 in a home in Co. Sligo. The home is a two m− 3 from the PW-NRS. Thus we arrive at an estimate of 0.16 × 97.83 =
storey house built in a high radon area; that is an area where more than 15.65% for the additional risk due to radon. Thus the total number of
10% of homes are estimated to have a radon level above the national lung Cancer deaths/cases is 115.65% of what it would be if there were
reference level. As the house was built between 1998 and 2005, it is no radon in Ireland.
assumed to comply with the building regulations requiring radon pre­ According to National Cancer Registry of Ireland data (NCRI, 2020)
ventive measures in homes built since July 1998. As indicated earlier all the mean annual number of lung cancer deaths in Ireland (2015–2017)
householders with measurements exceeding the reference level were is 1883 deaths per year. So the mean annual number of deaths in the
provided with remediation advice. absence of any radon would have been 1628 = 1883/1.1565. Thus we
estimate the number of excess deaths due to radon at 255 per year (95%
confidence interval (81, 464)). Based on the CSO figure of 4.76 million
7.2. Population weighted national average radon level for the Irish population in the 2016 census (CSO, 2016), this equates to a
mortality rate of 5.3 per 100,000 people per year that can be attributed
Utilising the stratification methodology from section 3.4 we esti­ to exposure to indoor radon.
mated the national average population weighted indoor radon level as The NCRI reported that the total average number of lung cancer cases
97.83 Bq m− 3. The 95% confidence interval for this mean is 90.69 Bq per year is 2588 (NCRI, 2020). Applying the methodology just described
m− 3 to 105.53 Bq m− 3. This confidence interval provides us with our to these lung cancer cases we estimate that the number of excess lung
first properly constructed baseline for all future studies of the population cancer cases due to radon is 350 per year (95% confidence interval (111,
weighted radon exposure in Ireland. As this is one of a suite of metrics to 638)). This is equivalent to 7.3 excess lung cancer cases per 100,000
be tracked over time to assess the effectiveness of Ireland’s National people per year.

Table 5
Summary of radon measurement data.
Decile No. of grid squares in No. of grid squares selected per No. of measurements per Radon concentration range (Bq No. of measurements >200 (Bq
decile decile decile m− 3) m− 3)

10 1 1 55 14–148 0
9 3 3 63 16–274 5
8 6 6 64 14–233 1
7 11 6 65 15–1083 14
6 18 6 64 17–1393 10
5 33 6 63 20–239 3
4 66 6 61 15–1251 7
3 108 7 71 20–385 7
2 159 9 76 14–635 12
1 432 10 71 17–504 5
Entire 14–1393 64
sample

7
P. Murphy et al. Journal of Environmental Radioactivity 233 (2021) 106582

8. Discussion advantage, it allows the PW-NRS to be repeated on a frequent basis and


thus enables longitudinal comparisons of population risk to be pro­
8.1. Efficiency of the new survey design duced. The results presented here provide a baseline for those future
studies. This methodology can easily be adjusted by increasing the
This new survey PW-NRS protocol took 11 months to complete, we sample size to decrease the width of the confidence intervals on any
began by issuing invitations in July 2016 and all radon measurements future estimates and thus increase the statistical power of future studies.
were completed by February 2017. Analysis of the data was completed The results we have reported here, in particular the mortality rate of
by May 2017. In the new survey protocol, the time required for the 5.3 per 100,000 due to indoor radon, demonstrate that radon remains
radon measurement process was just 6 months. Because of the smaller one of the highest preventable causes of death in Ireland. Our estimate of
sample size we were able to greatly reduce the number of invitations 255 deaths per year from radon exposure along with the 141 fatalities
issued and radon detectors required resulting in an associated significant reported last year from road traffic accidents provides a very clear and
reduction in staff resources. The PW-NRS allowed us to estimate the easily understood comparison of the dangers posed by radon. This
population weighted national average radon level with a level of accu­ striking comparison greatly contrasts with the lack of public awareness
racy that will enable comparisons with future population weighted of the risk caused by radon and the very high profile given to road traffic
surveys, the width of the 95% confidence interval was approximately 7 deaths.
Bq m− 3.

8.2. Discussion of population weighted average Declaration of competing interest

This new PW-NRS has produced an estimate 97.83 Bq m− 3 for the The authors declare that they have no known competing financial
population weighted national average indoor radon concentration and interests or personal relationships that could have appeared to influence
was specifically designed to provide a sample that was representative of the work reported in this paper.
the population distribution. The modelling used here and the estimates
produced were all based on computations performed at the decile level References
and not at the grid square level. This means that all of the estimates in
Ajrouche, R., Roudier, C., Cléro, E., et al., 2018. Quantitative Health Impact of Indoor
this paper were based on sufficiently large sample sizes to provide es­
Radon in France Radiation and Environmental Biophysics. https://doi.org/10.1007/
timates that will allow comparisons with future population weighted s00411-018-0741-x, 2018.
studies. Bossew, P., 2010. Radon: exploring the log normal mystery. J. Environ. Radioact. 101
(10), 826–834. October 2010.
Burke, O., Murphy, P., Organo, C., Fenton, D., Colgan, P.A., 2010. Estimation of seasonal
8.3. Comments regarding number of homes and people exceeding the correction factors through Fourier decomposition analysis- a new model for indoor
reference level radon levels in Irish homes. J. Radiol. Prot. 30, 433–443.
CSO, 2012. http://www.cso.ie/en/media/csoie/census/documents/census2011profile
4/Profile4_Statistical_Tables_and_Appendices_and_Form.pdf. (Accessed 30 November
Our 2017 survey provides a central estimate of 134,000 homes 2016).
exceeding the reference level. The most useful statistic that should be CSO, 2016. https://www.cso.ie/en/media/csoie/releasespublications/documents/po
used for comparison with future studies concerns the proportion of pulation/2017/Chapter_1_Population_change_and_historical_perspective.pdf.
(Accessed 16 September 2020).
homes with levels exceeding the threshold. We estimated this at 7.50% CSO, 2017. http://www.cso.ie/en/media/csoie/releasespublications/documents/popula
with 95% Confidence Interval of (6.46%, 8.66%). tion/2017/Chapter_2_Geographical_distribution.pdf. (Accessed 3 March 2018).
Curtin, B., 2016. Population Weighting in the National Radon Survey. MINOR THESIS for
M.A. STATISTICS. University College Dublin.
8.4. Significance of lung cancer estimates Darby, S., Hill, D., Auvinen, A., Barros-Dios, J.M., Baysson, H., Bochicchio, F., Deo, H.,
Falk, R., Forastiere, F., Hakama, M., Heid, I., Kreienbrock, L., Kreuzer, M.,
We estimate the number of excess deaths due to radon is 255 per year Lagarde, F., Mäkeläinen, I., Muirhead, C., Oberaigner, W., Pershagen, G., Ruano-
Ravina, A., Ruosteenoja, E., Rosario, A.S., Tirmarche, M., Tomásek, L., Whitley, E.,
and number of excess lung cancer cases due to radon is 350 per year.
Wichmann, H.E., Doll, R., 2005. Radon in homes and risk of lung cancer:
That is, as many as 13.5% of all lung cancer cases and 13.5% of all lung collaborative analysis of individual data from 13 European case studies. Br. Med. J.
cancer deaths may be due to radon. In 2005 the Radiological Protection 330, 223–228.
Daraktchieva, Z., Miles, J.C.H., McColl, N., 2014. Radon, the lognormal distribution and
Institute of Ireland and National Cancer Registry of Ireland estimated
deviation from it. J. Radiol. Prot. 34 (1).
that 13% of all lung cancer cases were due to radon. (RPII and HSE, DCCAE, 2014. National radon Control strategy. Wexford: Department of the
2005). As the levels of smoking in the population decreases the pro­ Communications, Climate action and Environment. https://www.dccae.gov.ie/doc
portion of lung cancers due to radon will inevitably increase. uments/National%20Radon%20Control%20Strategy.pdf. (Accessed 11 June 2018).
Dowdall, A., Murphy, P., Pollard, D., Fenton, D., 2017. Update of Ireland’s national
According to the Road Safety Authority of Ireland (RSA, 2019) 141 average indoor radon concentration – application of a new survey protocol.
people died in road traffic collisions in Ireland in 2019. The NCRI (NCRI, J. Environ. Radioact. 169–170, 1–8.
2020) reports that the annual average (2015–2018) number of deaths Dubois, G., et al., 2005. An Overview of Radon Surveys in Europe. European Commission
– Joint Research Centre. EUR 21892 EN. https://rem.jrc.ec.europa.eu/RemW
due to cervical cancer were 84 deaths and 154 from skin cancer. When eb/publications/EUR_RADON.pdf. (Accessed 30 March 2018).
we compare these figures with the estimate of 255 deaths due to radon EPA, 2015. Report on Radon Uncertainties 2015. Internal Report.
exposure we can see that radon is a very significant element of the EPA, 2019. EPA Protocol for the Measurement of Radon in Homes & Workplaces. May
2019. http://www.epa.ie/pubs/advice/radiation/Measurement%20Protocol%20H
preventable deaths in Ireland each year. omes%20and%20Workplaces_May%202019.pdf. (Accessed 16 September 2020).
Fennell, S.G., Mackin, G.M., Madden, J.S., McGarry, A.T., Duffy, J.T., O’Colmáin, M.,
9. Conclusion Colgan, P.A., Pollard, D., 2002. Radon in Dwellings, the Irish National Radon Survey.
RPII-02/1. Radiological Protection Institute of Ireland, Dublin. www.epa.ie/pubs/re
ports/radiation/radonindwellingstheirishnationalradonsurvey.html. (Accessed 31
This survey protocol enabled us to issue invitations, complete the May 2016).
measurement process and produce estimates of the population weighted Gaskin, J., Coyle, D., Whyte, J., Krewksi, D., 2018. Global estimate of lung cancer
mortality attributable to residential radon. Environ. Health Perspect. 126 (5).
national radon average, the proportion of homes exceeding the refer­
Geodirectory, 2016. https://www.geodirectory.ie. (Accessed 15 February 2016).
ence level and estimates of the number of lung cancer cases and deaths Gunning, Pollard, Finch, 2014. An outdoor radon survey and minimizing the
from radon in a period of less than one year. This short time frame can be uncertainties in low level measurements using CR-39 detectors. J. Radiol. Prot. 34,
compared with the almost decade long process required in the larger 457–467, 2014.
Health Canada, 2012. Cross-Canada Survey of Radon Concentrations in Homes.
scale 2002 NRS. https://www.canada.ca/content/dam/hc-sc/migration/hc-sc/ewh-semt/alt_for
The short time frame required for this protocol has another mats/pdf/radiation/radon/survey-sondage-eng.pdf. (Accessed 15 April 2018).

8
P. Murphy et al. Journal of Environmental Radioactivity 233 (2021) 106582

NCRI, 2020. National Cancer Registry of Ireland, 2020. Cancer in Ireland 1994 – 2018 RSA, 2019. RSA Annual Report 2019 (Accessed 6 April 2021).
with Estimates for 2018 – 2020. Annual Report of the National Cancer Registry. Stationery Office, 2008. Building Regulations 1997 Technical Guidance Document C
(Accessed 6 April 2021). (2004 Edition) - Site Preparation and Resistance to Moisture (Radon Prevention).
Nero, A.V., Schwehr, M.B., Nazaroff, W.W., Revzan, K.L., 1986. Distribution of airborne Department of the Environment, Heritage and Local Government, Dublin.
radon-222 concentrations in US homes. Science 234, 992–997. Stationery Office, 2019. S.I. 30 of 2019. Radiological Protection Act 1991 (Ionising
Nguyen, Murphy, 2015. To weight or not to Weight? A statistical analysis of how weights Radiation) Regulations 2019.
affect the reliability of the quarterly national household survey for immigration Suzuki, G., et al., 2010. A nation-wide survey on indoor radon from 2007 to 2010 in
research in Ireland. Econ. Soc. Rev. 46 (No. 4), 567–603. Winter 2015. Japan. J. Radiat. Res. 51, 683–689, 2010.
Olsson, U., 2005. Confidence intervals for the mean of a log-normal distribution. J. Stat. Who, 2009. WHO Handbook on Indoor Radon: a Public Health Perspective. http://www.
Educ. 13 (1), 2005. who.int/ionizing_radiation/env/9789241547673/en/. (Accessed 11 December
Organo, C., Murphy, P., 2007. The Castleisland Radon Survey - follow-up to the 2018).
discovery of a house with extremely high radon concentrations in County Kerry (SW Wrixon, A.D., Green, B.M.R., Lomas, P.R., Miles, J.C.H., Cliff, K.D., Francis, E.A.,
Ireland). J. Radiol. Prot. 27, 275–285. Driscoll, C.M.H., James, A.C., O’Riordan, M.C., 1988. Natural Radiation Exposure in
Rpii, Hse, 2005. Health Risks Due to Exposure to Radon in Homes in Ireland. RPII. https UK Dwellings. NRPB R-190.
://www.epa.ie/pubs/reports/radiation/RPII_NCRI_Radon_Health_Risks_2005.pdf.
(Accessed 15 April 2018).

You might also like